The United States began building the Panama Canal in 1904 and opened it in 1914. When Panama opened its larger locks in 2016, the first vessel through was *COSCO Shipping Panama*, chartered by a Chinese shipping company and built by Hyundai Samho in South Korea. American yards ought to be competing for more of that commercial work, using allied expertise to expand production here while we rebuild the fleet that carries American goods.

The long way around

The Panama Canal Authority restored the maximum draft for its larger locks to 49 feet on September 28 after rainfall improved. Daily reservation slots will increase to 33 on October 15. Ships still depend on freshwater to rise through the locks, drawing from reservoirs that also supply nearby communities. Panama plans to build the Río Indio reservoir between 2027 and 2031, leaving exporters to navigate several more dry seasons before the additional water arrives.

American Pacific ports and transcontinental railroads already carry freight around the canal, and Pacific Northwest grain terminals give farmers direct access to Asian markets. Expanding those facilities would let more cargo use them when Panama restricts passage. America also needs pipelines between producing regions and markets that depend on fuel delivered by sea. Magellan and the route around Cape Horn through the northern Drake Passage remain costly alternatives, with longer voyages and severe weather.

Before the war with Iran began in February, Hormuz carried roughly one-fifth of global petroleum consumption and more than one-fifth of global liquefied natural gas trade. Saudi Arabia’s East-West Pipeline carries crude across the country to the Red Sea, while the UAE’s pipeline to Fujairah reaches an export terminal outside Hormuz. Those land routes have helped keep oil moving, although attacks have threatened the Saudi line as well. Both have finite capacity, some already in use, and neither can carry Qatar’s LNG.

On October 12, Saudi Arabia’s energy minister discussed increasing the kingdom’s pipeline export capacity to around 10 million barrels a day over two to three years. The proposal includes improving the East-West system and potentially converting a domestic gas pipeline temporarily to carry oil.

Tanker operators have also used shuttle vessels to carry oil through Hormuz, transferring cargo to other ships outside the strait. Renewed attacks and congestion have made that arrangement more difficult, with some vessels proceeding directly to India. The crews doing those voyages remain exposed each time they return for another cargo.

Maersk and Hapag-Lloyd announced in September that four more joint services would return to Suez, joining two already using the route, subject to regional stability. Houthi attacks around Bab el-Mandeb continue to threaten the southern entrance to the Red Sea. Diversions around the Cape of Good Hope keep ships at sea longer, burning additional fuel and delaying their availability for the next cargo. American and allied navies help protect the shorter route, while shipowners decide whether to use it.

Singapore handled 44.66 million twenty-foot equivalent units of container traffic in 2025, connecting cargo across an enormous transshipment network. The adjoining Singapore and Malacca straits carry energy and goods to China, Japan and South Korea. Sunda and Lombok offer longer alternatives through Indonesia, subject to different vessel-draft limits. Beijing depends on continued access to these waters even as it develops the forces to pressure its neighbors.

A northern opening

*Dubai Tower* left Qingdao, its final Chinese port, on August 19 and reached Teesport in England on September 9 after crossing the Russian Arctic. Sea Legend advertised a season of eight scheduled departures from China to Europe. Ice conditions forced adjustments during the crossing. On October 5, tracking placed *Dubai Tower* in Suez during its return to Asia, on the southern route Sea Legend had already described.

Russia supplies icebreaker support and administers permits along its northern coast, giving Moscow substantial influence over the cost and conditions of passage. Chinese operators are accumulating experience in that system as the seasonal opportunity develops. Alaska faces Russia across the Bering Strait, at the Pacific entrance to the route, while Greenland lies near its Atlantic approaches. We already have an Arctic coast to develop, and cooperation with Denmark and Greenland should strengthen American access and security at the other end.

The Coast Guard has awarded contracts for up to 11 Arctic Security Cutters. Finnish yards will build part of the initial fleet, with first deliveries expected in 2028 and the first American-built cutter expected in 2029.

Nome’s planned port expansion would give larger vessels a place to operate along Alaska’s western coast. Construction there should proceed alongside the cutters, so the ships arrive with shore facilities able to support their crews and sustain operations.

Who builds the fleet

Chinese yards produced 56.1% of the world’s new ship capacity by deadweight tonnage in 2025 and took 69% of new orders on the same basis, according to China’s industry ministry. State-owned groups build merchant vessels and warships, sustaining industrial skills and suppliers across both businesses. Foreign commercial orders help support that system, while COSCO’s shipping services and overseas terminals extend China’s participation from construction into the movement and handling of cargo. American and allied navies continue protecting trade carried aboard ships that a strategic competitor increasingly manufactures. Beijing has rehearsed blockade operations around Taiwan and applied coast guard pressure against the island. A blockade would threaten semiconductor deliveries and give the Chinese Communist Party leverage over manufacturers well beyond Taiwan. The United States needs operational ships to deter that coercion and maintain deployments if deterrence fails.

In February, the White House’s Maritime Action Plan identified only eight American yards capable of building ships longer than 400 feet. The Government Accountability Office found that public yards completed only about 11% of attack-submarine depot maintenance on time over fiscal years 2016 through 2025. Submarines waiting for maintenance leave commanders with fewer boats available for patrols. Congress and the Navy should expand staffed dry docks and repair facilities alongside new construction, preserving more of the fleet’s availability as additional ships enter service.

Work on the waterfront

On October 10, Matson christened *Makua* at Hanwha Philly Shipyard, the first of three new Jones Act-compliant containerships in a roughly $1 billion construction program. The ship is scheduled for delivery in the first quarter of 2027 for Matson’s Hawaii and China-Long Beach Express services. Matson is reusing a name from its earlier fleet: the original *Makua* was a freighter commissioned by the U.S. Shipping Board and acquired by the company in the 1920s.

Hanwha completed its acquisition of Philadelphia’s shipyard in 2024 and subsequently announced a $5 billion expansion plan. Orders for ten medium-range tankers give the yard a pipeline of commercial work, with the first delivery expected in early 2029. Philadelphia’s development authority advanced a lease proposal for roughly 47 additional acres in September. The investment would add production at an operating American yard, drawing on a Korean company with extensive experience building ships for international customers.

At Ingalls in Mississippi, HD Hyundai and Huntington Ingalls Industries are testing mechanized welding equipment in work that largely relies on manual welding. We should make it attractive for Korean and Japanese builders to establish more production here, with repeated orders large enough to justify the equipment and workforce they would need.

Anduril’s October 6 announcement proposed $3.7 billion in private investment at Sparrows Point for a facility manufacturing Virginia-class submarine components and modules. Its Navy contract is worth up to $2.9 billion, with payments tied to demonstrated production. The site retains waterfront and transportation connections from its steelmaking and shipbuilding years, giving a new manufacturer access to industrial infrastructure that would be difficult to assemble elsewhere. The yard had built *Ancon*, the steamship that made the Panama Canal’s inaugural transit in 1914. Maryland and Baltimore County should organize construction approvals around the planned factory, and employers should work with local schools to prepare people for the jobs it will create.

Orders worth building for

American yards built 2,710 Liberty ships during World War II, using standardized designs and production methods that brought thousands of workers into shipbuilding. Today’s yards need years of dependable work to invest in their own equipment and train the next generation. The administration’s August shipbuilding memorandum proposes using allied designs for selected naval and support ships while requiring participating suppliers to develop American production. Federal buyers should place successive orders for mature designs and resist changing them after construction begins. Commercial tankers and cargo ships should also help sustain the yards and suppliers between military orders.

The Maritime Administration counted just 190 American-flagged, self-propelled oceangoing merchant ships of at least 1,000 gross tons in August, including those serving domestic routes. Ships in international commerce sustain mariners whose qualifications also prepare them to carry military cargo during a crisis. Cargo preference requirements and the Maritime Security Program help preserve that workforce. A larger American merchant fleet should carry more of our exports and provide more careers at sea, with federal ship financing helping operators afford the vessels American yards build.

The Jones Act generally reserves cargo movements between American ports for American-built ships with qualifying American ownership and crews. It creates demand for American construction, although the price of those vessels can discourage coastal shipping. Expanding financing for domestic fleet renewal would give operators more reason to order ships and cargo owners more opportunity to move goods along our coasts. The investment belongs alongside expanded port facilities and rail connections, making waterborne transport a stronger part of the country’s internal freight network.

Senators Mark Kelly and Todd Young, with Representatives John Garamendi and Trent Kelly, have assembled a bipartisan coalition behind SHIPS for America legislation. Congress should use that support to expand ship financing and establish long-term demand for American construction. China-related maritime fees remain suspended during a one-year period that began last November. Companies considering a new yard need contracts they can finance over many years, beyond the next round of trade negotiations.

Federal and state agencies should review dredging and waterfront construction together under deadlines that let companies schedule the work. Apprenticeships should connect to employers with signed orders, giving trainees a path into paid work as the facilities open.

Philadelphia’s ten tanker orders should lead to further commercial contracts, keeping the yard in production and expanding employment through the 2030s. The Navy should likewise use its sustained demand to expand the number of qualified manufacturers, building on projects such as Sparrows Point and the American Arctic cutters. Over the next decade, those investments should leave the country with an expanding commercial fleet and enough construction and repair capacity to keep more of its Navy at sea. American businesses would gain more ways to reach customers abroad, backed by an industry here capable of sustaining that commerce.

The War Production Board approved the first section of the Big Inch in June 1942. Washington financed construction, and a private company built the government-owned system from Texas to the Northeast. The Big Inch carried crude oil, while the Little Big Inch carried refined products. Completed in 1943 and 1944, the lines moved fuel overland and freed tankers for the shorter voyage across the Atlantic.

After the war, coal producers and railroads fought conversion of the lines to natural gas because they feared the competition. Texas Eastern acquired them in 1947 and began moving gas from the Southwest to eastern customers. Builders subsequently installed roughly half the mainline gas transmission network still operating today during the 1950s and 1960s. Generations of American businesses have benefited from that willingness to connect distant parts of the country, with new employers arriving long after the original construction crews had left.

Westbound

Phillips 66 stopped making fuel at its Los Angeles refinery late last year, and Valero finished idling the processing units at its Benicia refinery in April. California still has operating refineries, but these losses have made an already isolated fuel market more dependent on replacement supply. Valero told investors it would meet its Northern California obligations through imports or other supply arrangements after Benicia stopped refining. California's major refineries have also supplied most of Nevada's transportation fuel and nearly half of Arizona's, extending the consequences well beyond the state.

California has no operating pipeline bringing refined fuel west from the major refining centers in Texas and the Midwest. Natural gas pipelines already serve the state, and fuel pipelines carry California products east into Arizona and Nevada. Gasoline, diesel and jet fuel needed to replace lost California production generally have to arrive by ship. The California Energy Commission lists India, South Korea and the United Kingdom among the state's overseas fuel suppliers, alongside deliveries from other American coasts.

The United States produced a record 13.6 million barrels of crude oil a day in 2025. American refineries on the Gulf Coast produce more fuel than their region consumes, yet California remains outside the pipeline network that could bring that supply west. A country with our resources should have dependable domestic routes between its producing regions and its major markets. We should be building additional pipelines across the continental United States, including the east-west connections that would let more Americans benefit from the energy produced and refined here.

Phillips 66, Kinder Morgan and HF Sinclair approved their Western Gateway investment in August to bring fuel from Gulf Coast and Midwestern refineries into Arizona and California. A new line from Borger, Texas, to Phoenix would connect with an existing line whose flow would be reversed toward California. Designed to carry 230,000 barrels a day, the system is targeting completion in 2029, subject to permits. Western Gateway should receive the approvals needed to meet that date, opening the western market to more American fuel and competition among domestic suppliers.

Refine it here

U.S. refining capacity fell by more than 250,000 barrels a day during 2025, to 18.2 million at the start of this year, before Benicia's shutdown. Small expansions at other plants offset some of the losses, but the national total still moved backward. We need more American refining capacity to replace lost domestic production and supply markets that have become more dependent on finished fuel from abroad. New connections to those markets would also give refiners more customers to justify expansion.

ExxonMobil added 250,000 barrels a day at its Beaumont, Texas, refinery in 2023, roughly the capacity of a midsized refinery. The $2 billion expansion connected additional processing capacity with crude arriving from the Permian Basin in West Texas and southeastern New Mexico. Construction began in 2019 and continued through the pandemic. The company brought it into service on schedule and on budget, adding production that the country could use as demand recovered.

Other expansions should follow where existing refineries have room to grow, and new refineries should remain possible where demand and access to crude support them. Modernization should include equipment for processing different grades of oil, along with the pipelines bringing crude in and carrying finished fuel out.

U.S. imports of Venezuelan crude rose from 200,000 barrels a day in January to 701,000 in July, the latest month in the Energy Information Administration's published series. Much of Venezuela's production is heavy crude, well suited to the complex refineries American companies built along the Gulf Coast. Those plants have equipment for breaking down heavier oil and removing sulfur, allowing them to turn a difficult raw material into valuable transportation fuels. Valero said in July that it had shifted toward more Venezuelan crude following disruptions to Canadian supply in June.

American shale production supplies a great deal of lighter crude, so importing heavier barrels can complement domestic production. The refinery's equipment and the prices of available crude determine the blend it processes. Buying a suitable raw material abroad while doing the processing in the United States retains American industrial work. Replacing a closed refinery's output with imported gasoline transfers that work to the supplier, even though Americans continue buying the finished product.

American refineries operate under enforceable air and water requirements, and modern equipment can improve their efficiency while reducing pollution. New capacity should incorporate strong emissions controls and replace less efficient processing where practical. Building and modernizing here keeps that production within reach of American inspectors, gives us control over the standards it must meet and preserves the skilled workforce needed to improve it.

American capacity

China received 69% of Venezuela's crude exports in 2023, according to EIA's country assessment. American refiners are now buying substantially more Venezuelan oil, and their ability to handle those heavy barrels gives them a basis for competing for the processing work. We should welcome that business into American plants and build the connections needed to sell their output throughout the country.

China imported 11.1 million barrels of crude a day in 2024, with 92% arriving by sea. The United States has both enormous domestic production and refineries capable of drawing on additional supplies from abroad. Expanding our plants and connecting them to more markets would strengthen an industry that serves Americans while supplying other countries. American refiners should compete for more of our allies' fuel purchases, using those sales to support further investment here and take business from rival suppliers.

Pennsylvania, Ohio and West Virginia have developed shale gas fields much closer to northeastern customers than the older southwestern sources. New pipelines should carry more of that supply into the Northeast and south to growing industrial markets. Gas remains a fuel for factory furnaces and a raw material for chemical production, as well as a source of electricity for computing and manufacturing. The industrial investment America wants to attract requires physical connections to that supply, including storage and dependable delivery to the plants that will use it.

Approve and build

Dominion Energy and Duke Energy canceled the Atlantic Coast Pipeline in 2020 after its estimated cost rose from 4.5–5 billion to $8 billion. The project was intended to serve Virginia and North Carolina, and the companies said customers had recently renewed commitments for about 90% of its capacity. Even after winning a Supreme Court case over its route beneath the Appalachian Trail, they cited continuing permit uncertainty and delays in abandoning it.

California has encouraged a transition away from petroleum while still relying on refineries to supply most of its transportation fuel. Companies considering investments that take decades to recover have to judge how long the state will welcome their operations. Federal and state policy should support continued investment in the capacity that demand still requires, including upgrades that make existing plants cleaner and more productive. The cost of replacing lost production falls on consumers well before a government's desired transition is complete.

Congress should establish a clear national interest in moving energy across state lines and give major interstate projects a dependable path to construction. Agencies should review applications together under an enforceable schedule, and courts should resolve challenges promptly. Once a project has met the requirements, its approval should remain usable through changes in administrations. Developers need a decision they can rely on long enough to finance and finish the work.

Landowners deserve compensation, and neighboring communities deserve safe construction and enforcement of pollution limits. Those obligations should be settled through a process capable of authorizing a project and letting work proceed. Environmental group Food & Water Watch has pledged to fight the Northeast Supply Enhancement pipeline, intended to bring more gas to New York City and Long Island, until it is defeated. Elected officials who support construction have to defend the decision to build even when opponents would prefer the project never existed.

Republicans Shelley Moore Capito of West Virginia and Mike Lee of Utah joined Democrats Sheldon Whitehouse of Rhode Island and Martin Heinrich of New Mexico on a permitting bill introduced September 30. Their Bipartisan American Affordability and Jobs Act would protect approved projects against certain later reversals and change parts of the review process. It gives Congress a current opportunity to make construction more dependable. The coalition should carry that work through both chambers, with a framework broad enough to advance pipelines and refining alongside new electricity infrastructure.

Western Gateway's planned 2029 opening should be followed by other completed connections, with additional American refining capacity feeding them. The next decade ought to leave the country better able to move its energy east and west, and to supply regions whose industrial demand outgrows the network they inherited. Building that system would keep more processing and manufacturing in the United States while giving future businesses infrastructure they won't have to spend years trying to secure. We have the resources to support a much larger American industrial economy, and we should approve and build the facilities that let us use them.

Arsenal-2 will supply components and large assemblies for Virginia-class attack submarines, with General Dynamics Electric Boat and HII’s Newport News Shipbuilding retaining final assembly. Anduril says its Navy contract is worth up to $2.9 billion, with payments contingent on demonstrated production outcomes. Initial operations are expected in 2030, followed by a gradual increase in production as the workforce and facilities develop.

Anduril would enter a program with established designs, funded orders and two experienced final-assembly yards. The Navy should make that kind of entry attractive to more American companies, using its purchases to bring new capital and new manufacturing methods into submarine construction. We need an American maritime industry that keeps expanding and improving as China builds the naval power it uses to pressure its neighbors.

Before the submarines

The steamship that made the Panama Canal’s inaugural transit on August 15, 1914, came from Sparrows Point. Maryland Steel had launched it as the *Shawmut* in 1901, years before it became the *Ancon*.

Bethlehem acquired the shipyard and adjoining steelworks in 1916. The yard built merchant vessels through the First World War, then produced 101 ships during the Second World War with as many as 8,000 workers. In 1953, it led the world in new ship construction by deadweight tonnage.

Bethlehem owned both the yard and its principal source of steel, with the mill next door supplying material for ship construction. Beginning with 28,000-deadweight-ton tankers in 1949, the company developed ten standard tanker designs as demand grew. Bethlehem invested in facilities for those vessels, including the graving dock Anduril now proposes to use.

Bethlehem launched its last newly constructed ship at Sparrows Point in 1990 and sold the yard in 1997. Its successor, Baltimore Marine Industries, entered bankruptcy in 2003 after a prospective cruise-ship customer failed to secure financing and the yard was left with little work. Steelmaking on the larger property ended in 2012.

New ownership began redeveloping the 3,300-acre property in 2014 into what became Tradepoint Atlantic. Port operations and logistics businesses made use of the waterfront, highway access and rail connections to CSX and Norfolk Southern. That redevelopment kept the property in commercial use and gave Anduril an operating industrial site around which to build its factory. As America expands naval production, it should make more use of industrial land whose waterfront and transport connections would be difficult to assemble again from scratch.

Building for the Pacific

At China’s Changxing Island shipbuilding complex in Shanghai, commercial vessels and naval surface combatants appear together under construction in satellite imagery analyzed by the Center for Strategic and International Studies. Researchers documented a similar arrangement at Longxue Island in Guangzhou in their March 2025 study. Both sites belong to the state-owned China State Shipbuilding Corporation, which builds merchant ships for foreign customers alongside vessels for the Chinese navy. Overseas orders help sustain an industrial system serving Beijing’s military ambitions.

The Pentagon’s 2025 China military report describes the Joint Sword exercises of 2024 as rehearsals for maritime and air superiority, a blockade of Taiwan’s key ports and precision strikes on the island. Beijing has been developing the forces to coerce a self-governing democracy and obstruct American intervention. The United States and its allies need the means to prevent the Chinese Communist Party from deciding who can trade, travel or govern themselves in the region.

Admiral Samuel Paparo identified American undersea dominance as an advantage in deterring aggression against Taiwan during his April 2025 Senate testimony, while warning that the margin was shrinking. Preserving it requires submarines available to commanders, supported by an industry that can build their successors. American engineers have spent decades developing these boats, and the country ought to be able to manufacture enough of them to sustain the advantage their work created.

In April 2026, the Congressional Budget Office reported that attack-submarine production had averaged 1.1 boats annually over the preceding two years, although the Navy had generally been purchasing two a year. CBO also put Virginia-class construction an average of four years behind the delivery dates specified in the original contracts. Those delays deprive the fleet of submarines Washington has already decided it needs, prolonging a shortage while China’s naval industry continues to develop.

Electric Boat and Newport News must also build twelve Columbia-class ballistic missile submarines to replace the Ohio class. The Navy’s baseline production goal is one Columbia and two Virginias annually, drawing on many of the same trades and suppliers. Columbia supports the nuclear deterrent, so Washington has to expand attack-submarine output alongside a program it cannot casually postpone.

Under AUKUS, the security partnership with Britain and Australia, the United States intends to sell Virginia-class submarines to Australia. Replacing those boats raises the American production requirement to roughly 2.33 annually. We should be able to equip an ally strengthening the defense of the Pacific while maintaining our own fleet. A larger American submarine industry would give Washington more ability to sustain both commitments as Beijing expands its military power.

Room for another builder

Austal already builds and outfits submarine sections in Alabama, then ships them by barge to a prime yard for integration. CBO counts more than twenty companies participating in outsourced construction for Electric Boat and Newport News. Anduril would join that network with a new factory and its own capital, adding manufacturing in Maryland to a program whose final-assembly yards are in Connecticut and Virginia.

CBO estimated that approximately 70% of suppliers of critical submarine components had no competitors. The Navy should use its unfilled demand to attract American firms willing to establish qualified production, including companies that developed their manufacturing methods in other industries.

Anduril says it intends to recruit from commercial aerospace and automotive companies, as well as high schools and vocational schools. Maryland has proposed a workforce training center at or near Tradepoint Atlantic to help prepare people for the new manufacturing jobs. CBO described a generation of longtime shipyard workers retiring, leaving crews at many yards less experienced. A factory at Sparrows Point should give more Marylanders the opportunity to learn those trades and develop careers making the equipment their country needs.

Building the yard

The Maritime Administration’s 1972 annual report described Sparrows Point’s new plate-cutting and blasting equipment, along with a panel shop capable of handling ship subassemblies weighing 200 tons. Bethlehem equipped the yard to fabricate larger sections and build the tankers its customers were ordering. Anduril now proposes to arrange new fabrication areas beside high-bay assembly halls, shortening the distances between fabrication, assembly, testing and integration.

Anduril calls its proposal a software-defined shipyard, with reconfigurable production lines and ArsenalOS connecting fabrication and outfitting to material movement, inspection and documentation. Its plans include digital work instructions, sensing and embedded quality controls. The Navy’s SUBSAFE program requires traceable evidence of the material and work performed on covered components, so the software will have to support the records required to qualify production.

Anduril plans to manufacture Virginia-class components, including torpedo tubes and test assets, at a separate 160,000-square-foot facility in Orange County, California, two years before the Maryland yard opens. A company spokesperson has confirmed that the California facility is SUBSAFE certified, and Anduril has already started building torpedo-tube prototypes there. The company says the earlier operation will give it time to develop manufacturing methods and train the workforce before Arsenal-2 opens.

Anduril intends to take on more complicated modules and eventually hopes to manufacture entire hull sections. Its engineers would develop production methods against existing submarine designs, learning to build more of the boat as the company proves itself to the Navy and its prime contractors. Additional manufacturers with that experience would give the Navy more firms to draw on for future weapons and submarine designs.

Orders worth investing in

Anduril says it will bear most of the execution risk under a contract that links Navy payments to production outcomes. President Trump announced a 40% government stake in the shipyard, and Anduril has since clarified that the stake would be in a new subsidiary created to operate Arsenal-2. The government should publish the rights accompanying that stake and explain how the ownership arrangement relates to the production contract and the company’s private investment.

The Navy plans to keep buying Virginia and Columbia submarines after Arsenal-2’s target opening, while replacing the boats it sells to Australia. Contracts covering successive production runs would help manufacturers finance expansion and retain trained crews, with further awards dependent on cost, quality and delivery.

Anduril said in its October 6 announcement that permitting and site preparation would begin immediately, with construction following the necessary approvals. Maryland and Baltimore County should organize the reviews they control around the proposed operating date, while Anduril supplies complete plans and resolves outstanding requirements.

The basin is still there

The graving dock entered service in 1971. At 1,200 feet long and 200 feet wide, it accommodated Sparrows Point’s 265,000-deadweight-ton tankers, with an intermediate gate allowing crews to divide it for smaller vessels. In 2020, the floating dry dock *Titan* arrived for repairs. Bethlehem had built a dry dock large enough to dry-dock another dry dock.

Anduril identifies the basin as an asset for moving large components, although it hasn’t published a detailed operating plan for it. Its proposed manufacturing halls would stand beside infrastructure built for an earlier generation of ships, within a property that Tradepoint has kept in commercial use. The Navy’s orders now give a new American manufacturer a reason to invest there, and Maryland has the opportunity to help thousands of people enter the work.

Anduril’s engineers and crews should be able to build on that experience as the company earns further orders, refining the factory and learning to manufacture more complicated assemblies. Other qualified American companies ought to have opportunities to enter the business and grow alongside it. As Beijing expands the fleet it uses to pressure its neighbors, the Navy’s submarine purchases should leave America with more qualified manufacturers and more people trained to construct its next generation of submarines.

Isaias is the Atlantic’s ninth named storm of 2026 and would become its first hurricane. Through October 6, the preceding eight storms had generated 10.8 units of Accumulated Cyclone Energy, the index that combines storm intensity and duration, roughly 89% below normal for that point in the season. Across the Americas, the eastern Pacific had recorded seventeen named storms, seven hurricanes and five major hurricanes by early October 7, with ACE running 96% above normal. Polo reached Category 5 strength on September 22, with sustained winds of 180 mph.

NOAA’s historical database last records an entire Atlantic season without a hurricane in 1914, the year the Panama Canal opened to commercial traffic. The *SS Ancon* made the canal’s inaugural official transit on August 15, shortly after the outbreak of World War I, while the Atlantic produced just one recorded tropical storm that season. The database also lists no hurricanes in 1907; however, ships and coastal observers could miss storms that modern satellites would detect. Those early records establish how unusual a hurricane-free season would be, with considerably less certainty about what happened far offshore.

Isaias is developing over water near 30–31°C, with abundant moisture and relatively modest wind shear. The hurricane center expects stronger shear as the storm approaches the coast; however, its Wednesday forecast still anticipates a hurricane reaching the northern Gulf.

A Current Named for Christmas

Peruvian fishermen noticed warm coastal water and declining catches long before scientists could measure sea-surface temperatures across the Pacific. They called the phenomenon *El Niño*, a reference to the Christ child, because they associated its arrival with Christmas. Researchers eventually connected those coastal changes to a much larger rearrangement of winds, ocean temperatures and tropical rainfall.

NOAA’s September assessment found a strengthening El Niño and assigned a greater than 90% probability to a very strong event during the coming fall and winter. Warm water and shifting tropical convection alter atmospheric circulation far beyond the equatorial Pacific. Across much of the tropical Atlantic, stronger upper-level winds increase vertical wind shear, displacing thunderstorms from the circulation beneath them and making sustained organization more difficult. In the eastern and central Pacific, El Niño often supplies warmer water and more favorable atmospheric conditions for hurricane development.

NOAA’s August Atlantic outlook gave a 75% probability of a below-normal season and projected two to six hurricanes. It also described a tendency for Atlantic activity during El Niño years to shift toward the Gulf and subtropics as conditions become less favorable across the main tropical development region. A storm finding warm water, moisture and manageable shear in the Gulf fits within that broader pattern. Its eventual track will determine which communities and facilities face the strongest winds, heaviest rain and highest water.

The Treasure Fleet

A fleet of twelve ships assembled in Havana in 1715 with silver, coins and other cargo from Spain’s American colonies. The War of the Spanish Succession had disrupted shipments to Europe, and the Crown needed the money. Delays kept the convoy in port until late July, leaving heavily loaded sailing ships to negotiate the narrow waters between Florida and the Bahamas during hurricane season.

On July 31, a hurricane wrecked or sank eleven ships and killed roughly 1,000 of the 2,500 sailors and passengers aboard the Spanish vessels. The French ship *Grifon* had sailed ahead and escaped. Survivors established salvage camps along the Florida coast, where English privateer Henry Jennings later raided the recovered treasure. Subsequent discoveries of coins and wreckage helped give the region its name, the Treasure Coast.

Manzanillo handled more than 2.45 million twenty-foot-equivalent container units during the first seven months of 2026, accounting for 43.4% of the container traffic through Mexican ports. Manufacturers using the harbor gain access to Asian suppliers and customers through a port that concentrates a substantial share of the country’s maritime trade. A Pacific storm near that coast can delay cargo even if its strongest winds remain offshore: heavy seas can interrupt vessel movements, wind can halt crane operations, and flooding can sever road and rail connections inland.

Shipping companies can reroute vessels or postpone arrivals; however, they still need an available berth, a functioning terminal and an inland connection at the destination. Moving cargo to another port can add trucking distance, handling charges and days to a delivery schedule. Those costs reach manufacturers and customers who may never experience the storm itself.

A Few Degrees of Longitude

Refineries along the Texas and Louisiana coasts operate beside offshore production fields, petrochemical plants, export terminals and pipelines serving customers far inland. EIA’s January 2026 figures put operable refining capacity across its Gulf Coast district at approximately 9.88 million barrels a day. Grain shipments also move down the Mississippi toward Gulf export terminals, while LNG facilities connect American gas production to overseas buyers.

Katrina and Rita crossed this region within weeks of each other in 2005. During Rita, operators shut in 100% of Gulf oil production and roughly 81% of natural-gas production. At the peak of refinery outages that September, 4.9 million barrels a day of capacity stood idle, nearly 29% of the national total. The storms damaged 457 underwater pipelines, and cumulative lost offshore crude production exceeded 162 million barrels by June 1, 2006.

Refinery outages reduced the supply of finished fuels, while damaged pipelines and closed terminals complicated deliveries from facilities that could still operate. Repairs prolonged the disruption after the storms passed. For Isaias, the relevant questions now concern the facilities within reach of its wind, surge and rainfall, the precautions operators take before landfall, and how quickly they can restore service afterward. The October 7 forecast doesn’t establish a particular refinery shutdown or production loss.

In Puerto Rico, Hurricane Maria damaged the electrical and transportation systems supporting pharmaceutical and medical-device factories in 2017. FDA subsequently monitored about forty critical drug products manufactured primarily or exclusively on the island, including fourteen sole-source products for which no alternative drug products were available. Maria also aggravated existing shortages of intravenous solutions, prompting FDA to arrange additional imports. Hospitals across the mainland depended on production lines, power supplies and shipping routes that ran through the island.

The Price of a Coordinate

Mexico arranged $175 million of Pacific hurricane protection through a World Bank catastrophe bond issued in 2024, alongside $125 million for Atlantic named-storm risk in a separate issuance that year. The Pacific coverage increased from $125 million after Hurricane Otis struck Acapulco in 2023 and triggered a $60 million payout under the previous bond. Both 2024 issuances had scheduled maturities in April 2028.

Investors receive interest for accepting the risk that a qualifying disaster will reduce the principal they recover. Mexico’s hurricane coverage uses parametric triggers tied to an event’s location and severity, allowing a payout without waiting for adjusters to total every damaged building. A storm’s path can therefore determine whether the government receives funds and whether investors lose capital, as well as which communities require assistance.

The Pacific bond attracted investors from North America, Europe, Bermuda and Asia. Its financing spreads part of Mexico’s disaster exposure among institutions far from the threatened coast; however, a payout still leaves the government and local businesses with the work of repairing roads, restoring electricity and reopening facilities.

When Quiet Means Dry

Farmers in Honduras and Nicaragua entered October after prolonged dry spells and high temperatures had hurt the first maize harvest. GEOGLAM’s October Crop Monitor reported below-average yields across Central America and the Caribbean and continued dry conditions at the beginning of the next planting season. Its regional outlook also identified below-normal rainfall risks for October through December in association with the strengthening El Niño.

Tropical disturbances supply rain to farms and reservoirs even when they never develop hurricane-force winds. A season with fewer hurricanes can still bring damaging floods from weaker systems, while extended gaps between rain events can leave crops short of water. Water authorities need rainfall over the catchments that feed their reservoirs, and farmers need it during the stages when their crops can use it. A hurricane developing in the Gulf offers little information about whether those needs will be met elsewhere in the basin.

A farmer’s reduced harvest, a ship’s delayed arrival and an insurer’s hurricane claim arise from different combinations of weather and exposure. A basin-wide storm count captures only part of what those businesses have experienced.

Halsey’s Forecast

Admiral William Halsey’s Third Fleet was operating east of the Philippines in December 1944 when Typhoon Cobra caught it during attempts to refuel. Three destroyers, *Hull*, *Monaghan* and *Spence*, capsized, killing 775 men aboard those ships. The fleet lost roughly 790 sailors in total, along with 146 aircraft, and twenty-seven ships suffered damage. Gerald Ford, then an officer aboard the carrier *Monterey*, nearly went overboard while responding to a fire after aircraft broke loose in the hangar.

Fuel shortages, difficulty refueling in heavy seas and errors in locating the typhoon constrained commanders’ decisions as conditions worsened. The losses removed sailors, aircraft and ships from the war without an enemy attack.

The Central Pacific Hurricane Center’s October 7 summary put accumulated cyclone energy at more than five times normal for that point in the year. Hawaii’s naval facilities, airfields, ports and electrical systems support both a civilian population and American operations across that ocean. Commanders can move ships and aircraft ahead of dangerous weather when time and conditions permit; however, those assets still depend on fuel, maintenance facilities, communications and functioning ports when they return. Damage to the supporting infrastructure can restrict operations after the immediate danger has passed.

The hurricane center expects Isaias’s wind field to expand as it interacts with an approaching trough, with rain and wind extending well east of its center. Along the northern Gulf, operators and emergency managers face forecasts that can change which ports suspend traffic, when offshore crews leave and which roads remain available for evacuation.

McKinsey published a chart this week that turns one of the grander promises of artificial intelligence into something considerably more useful: a calendar. Across a group of publicly disclosed AI-enabled drug programs, companies have reached preclinical candidates and first-in-human trials much faster than the timelines associated with conventional large-pharma development. The comparison deserves some caution from the start. These are selected programs that survived long enough and worked well enough for companies to talk about them, while the conventional benchmark also captures some of the organizational drag inside large pharmaceutical companies: portfolio reviews, sequential handoffs, competing priorities and the usual machinery of a very large organization.

Even after allowing for that, the compression is difficult to dismiss. McKinsey puts the traditional trip from an initial drug hit to a preclinical candidate at roughly 32 to 50 months, followed by another 16 to 20 months before the drug reaches its first human trial. Schrödinger nominated SGR-1505 as a development candidate about ten months after beginning its MALT1 program. Recursion and Exscientia reached a candidate with EXS21546 in nine. Generate:Biomedicines moved one program from hit to first-in-human testing in 17 months; Iambic and Absci have disclosed timelines around two years.

Those numbers shouldn't be treated as though every clock started from the same place. Schrödinger was working against MALT1, an established oncology target, and its achievement is primarily a story about computational chemistry. Recursion's REC-1245 program began further upstream with a biological relationship uncovered through its maps of cellular data. Insilico Medicine went further still with rentosertib, using its platform to identify TNIK as a potential target for idiopathic pulmonary fibrosis and then designing a drug against it. A ten-month hit-to-candidate program on known biology and an eighteen-month target-to-candidate program on comparatively novel biology are both interesting; they aren't the same accomplishment.

In The Age of Therapeutic Abundance, we looked at a broader change already appearing across medicine: more powerful therapies are beginning to collide with the capacity of the healthcare system to manufacture, finance and deliver them. AI drug discovery sits considerably further upstream. Faster candidate generation doesn't create therapeutic abundance by itself; it increases the number of experiments that can reach the places where biology, clinical development and eventually the healthcare system decide which ones survive.

Rentosertib has now made it farther through that sequence than most. Insilico moved from its AI-identified target to a preclinical candidate in about 18 months, and doctors in China dosed the first patients in a Phase III trial this month. The study will enroll roughly 320 people with idiopathic pulmonary fibrosis and follow them for 52 weeks, measuring whether the drug can slow the decline in lung function. The attractive part of the discovery curve has therefore met a rather less compressible calendar: researchers now have to give people the drug for a year and see what happens.

Eight Billion Possibilities

Schrödinger's MALT1 program offers the cleanest view of what has already changed. Researchers were looking for a small molecule that could inhibit MALT1, a protein involved in signaling pathways associated with several B-cell cancers, while satisfying the collection of other requirements that routinely kill otherwise promising molecules. A compound can bind tightly to the desired protein and still fail because it doesn't dissolve properly, can't cross the right membrane, disappears too quickly inside the body or binds somewhere else with unfortunate consequences.

Medicinal chemistry has always dealt with those tradeoffs through iteration. Researchers design compounds, make the most promising ones, test them, inspect the results and decide what to make next; the chemistry itself may be clever, but a considerable amount of calendar time disappears while ideas wait to become molecules and molecules wait to become data. Schrödinger pushed far more of that search into computation before asking the laboratory to make anything.

Its platform evaluated roughly 8.2 billion compounds while looking for promising chemical series. Researchers later assessed more than 5,000 proposed designs against the combination of potency, solubility, permeability and other properties they wanted, narrowing the field to 43 that satisfied their criteria closely enough to deserve further attention. Across the program, they synthesized 129 compounds, including 78 in the lead series, before selecting SGR-1505 as the development candidate about ten months after the work began.

Eight billion possibilities ultimately produced 129 physical compounds, which gets closer to the industrial significance of AI drug discovery than the familiar language about machines “inventing” medicines. The computer didn't eliminate medicinal chemistry, and it certainly didn't produce a perfect molecule on the first try. It allowed an enormous number of bad or mediocre ideas to die as calculations rather than compounds, preserving the laboratory for the much smaller number worth making.

The evidence for AI in drug discovery is strongest, at least for now, as an economics-of-search story. Schrödinger can consider billions of compounds without making billions of compounds; weak ideas can die earlier and promising ones can reach the laboratory faster. What hasn't been demonstrated is the more consequential claim that candidates produced by these systems are more likely to survive Phase II, Phase III and regulatory review. Getting to the answer faster has real value even when the answer is no. Getting to yes more often would be something else entirely.

The Search Moves Upstream

Recursion's REC-1245 program already stretches the search beyond chemistry. Its platform identified a previously unknown regulatory relationship involving RBM39 and CDK12, giving researchers a possible way to attack cancers with particular genomic vulnerabilities; the company then developed a molecule intended to degrade RBM39 and moved the program into IND-enabling work in less than 18 months after synthesizing roughly 200 compounds. The first patient entered a Phase I/II trial in December 2024.

On September 23, Anthropic reported an experiment that pushed the same idea much further into basic biology. Rather than asking Claude to optimize a molecule against a target researchers had already chosen, Anthropic's scientists gave it a high-level instruction to search a large genomic database for interesting examples of reverse transcriptases, enzymes that copy RNA into DNA. Roughly 950 Claude agents then spent 21 hours working through the data, using about 210 million tokens while gathering more than 200,000 reverse transcriptases and identifying roughly 3,500 candidate systems.

The agents narrowed those candidates to 20 for deeper analysis. While examining one unusual family, Claude noticed a repeating pattern of DNA beside the gene for a reverse transcriptase, compared its structure with known biological systems and searched the literature for evidence that anyone had described the combination before. Anthropic's scientists subsequently investigated the candidate in the laboratory and now describe it as an array-associated reverse transcriptase, or ART, a previously uncharacterized enzyme system found in bacteriophages with a repeat structure reminiscent of CRISPR.

The underlying reverse transcriptase itself had appeared in earlier research; Anthropic says Claude appears to be the first to recognize the neighboring DNA repeats and accessory protein as the defining features of a larger system. The work is also a preprint, Anthropic still doesn't know what ART actually does, and every wet-lab experiment was performed by human scientists. This is an interesting biological discovery, not a new gene-editing therapy discovered autonomously by a chatbot.

Still, it is a different kind of result from searching billions of candidate molecules. The useful object being searched was biology itself: hundreds of thousands of genomic sequences containing systems whose functions were poorly understood or completely unknown. Anthropic says an expert scientist might spend weeks or months carrying out the sort of analysis its agents compressed into roughly a day, and the agents weren't merely retrieving a previously documented answer. They were looking for anomalies, developing hypotheses about them and deciding which candidates merited scarce human attention.

The asymmetry that appeared in Schrödinger's chemistry program shows up again, just one level earlier. Claude could generate and discard thousands of biological leads before scientists chose which ones deserved experiments, much as computational chemistry can discard enormous numbers of molecules before researchers synthesize them. The scarce resource shifts from producing another hypothesis toward knowing which hypothesis deserves a piece of the physical world.

Biology Has a Different Clock

Anthropic's experiment complicates any clean line between fast computation and slow biology, because some of the work normally described as biological discovery can plainly be accelerated too. Reading literature, searching genomic neighborhoods, comparing protein families and noticing unusual patterns are intellectual search problems; agents can run more of those searches in parallel than a laboratory could ever staff with scientists. If that approach generalizes, the supply of plausible biological hypotheses could increase very quickly.

The answer to those hypotheses still comes from somewhere else. Anthropic's agents could identify ART in 21 hours, but they couldn't determine its actual function simply by thinking harder about the sequence. Scientists had to express proteins, run biochemical experiments and examine what the system did; months of additional work may follow before anyone knows whether ART is merely an interesting curiosity or the beginning of another useful biological tool.

Drug development encounters the same distinction with much higher stakes. A gene can appear repeatedly among people with a disease without causing the disease, while a protein that looks essential in a cell line can prove surprisingly optional inside a person. Even sophisticated experimental models simplify something. Mice aren't small humans, organoids aren't patients and cancer, fibrosis or neurodegeneration doesn't reorganize itself into tidy training data because researchers have acquired more GPUs.

McKinsey consequently finds AI considerably more mature in molecule design than in target identification, preclinical research or translation into humans. Anthropic's result suggests that target discovery and hypothesis generation may begin moving faster as well, although generating a promising biological explanation and demonstrating that it actually causes disease remain different jobs. The feedback arrives more slowly, contains more noise and often leaves several explanations alive at once; cells also take however long cells take to grow, while a chronic disease may need months before researchers can reliably see whether anything changed.

A beautifully optimized molecule aimed at the wrong mechanism remains a beautifully optimized failure, just as a clever genomic hypothesis can disappear after its first serious experiment.

Producing either one faster still has substantial economic value because researchers reach the bad news earlier and redirect time and capital toward another idea. The larger prize comes if better computational search also improves which hypotheses scientists choose to test in the first place.

The structure of pharmaceutical AI deals suggests that the industry understands the difference. McKinsey puts the headline value of major AI drug-discovery partnerships at $29.5 billion in 2025, but only about $600 million of that came as upfront payments. Most of the nominal value sits behind clinical and commercial milestones; among 38 deals signed during 2023 and 2024, less than one percent of the maximum milestone value had been paid by June 2026. Some of that simply reflects the years required for drug development, although the contracts also place most of the money beyond the point where a computationally attractive idea has to survive biology.

More Candidates, Same Patients

A discovery organization that can produce several credible candidates in the time once required for one eventually runs into a problem that another billion calculations cannot solve. Clinical trials need investigators, sites and patients; complicated therapies need specialized manufacturing; every additional program also needs the capital required to carry it through years of development. Those resources don't automatically multiply because the front end becomes more productive.

That is where the connection to The Age of Therapeutic Abundance becomes more direct. The report looked downstream, where better therapies increasingly collide with manufacturing capacity, specialist infrastructure, insurance rules and the ability to pay for treatment. Faster discovery can create a similar problem earlier in the process. Abundance at the front of a pipeline can produce congestion in the middle long before any of those molecules become abundant medicines.

Clinical trials make the mismatch easiest to see because one of their scarce resources is often a person. Two companies developing treatments for the same relatively narrow disease population may need many of the same investigators, hospitals and patients, while oncology programs can compete for people whose tumors carry very specific mutations. Faster candidate generation doesn't create another supply of eligible patients; it places more protocols in front of the same population.

Manufacturing presents another version as molecules become more complicated. A small molecule generally poses a different industrial problem from an antibody, cell therapy or antibody-drug conjugate; specialized facilities and expertise cannot always be conjured at the same pace as another computationally promising candidate. Development capital has its own limit as well. A company may be able to discover ten plausible drugs without possessing the money to carry ten drugs through clinical trials.

Anthropic's biology experiment contains a miniature version of the same problem. Nine hundred and fifty agents generated thousands of candidate systems in less than a day, but the laboratory couldn't run thousands of serious follow-up experiments at the same speed. Anthropic narrowed the field to 20 for deeper analysis and ultimately put physical resources behind a much smaller number. Once hypothesis generation becomes abundant, experimental attention becomes more valuable.

None of those constraints makes faster discovery less valuable. They change what the next valuable unit of capacity looks like. Once chemical and biological search become cheap enough, another billion virtual molecules or another thousand genomic hypotheses matter less than another reliable disease model, another high-quality clinical site, or a better way to decide which candidate deserves the scarce experiment.

The Test

Rentosertib is interesting because it has crossed from the part of the process where AI has already demonstrated speed into the part where the larger claim gets tested. Insilico didn't merely optimize a drug against a familiar target; its platform helped identify TNIK as a disease target before designing the molecule that eventually became rentosertib. A successful Phase III study would therefore say something more consequential than another ten-month medicinal-chemistry program.

It still wouldn't settle the question by itself. One successful drug can be an important proof point without establishing that AI-originated candidates are systematically more likely to work. Over the next several years, the more revealing number will be attrition: whether cohorts of AI-enabled drugs survive Phase II and Phase III at higher rates than comparable conventional programs, after accounting for disease, modality and the maturity of the underlying target.

Anthropic points toward another result worth watching. If agents can repeatedly surface genuinely novel biological systems or disease mechanisms that subsequently survive prospective laboratory testing, the improvement will have moved from searching chemical space into choosing what biology deserves to be acted upon. Cellular foundation models, perturbation datasets, genomic agents and other systems are all aimed somewhere in that direction, although it is precisely where the information is messiest and experimental validation remains indispensable.

For now, several clocks are moving at once. Schrödinger considered billions of compounds and reached a candidate in about ten months; Recursion moved new biological insight into IND-enabling work in less than 18; Anthropic's agents searched hundreds of thousands of proteins and surfaced a previously uncharacterized biological system in 21 hours; Insilico turned an AI-identified target into a drug that has now reached Phase III. None of those results establishes that AI-originated medicines are more likely to work, but the searchable part of biology is expanding well beyond molecules.

Rentosertib will spend the next 52 weeks helping answer the harder question. Claude needed 21 hours to find something unusual in a genome; unfortunately, at least for now, human lungs still need a year to tell researchers whether a drug works.

On September 25, 2015, Xi Jinping stood beside Barack Obama in the Rose Garden after an arrival ceremony on the South Lawn. Both governments wanted to finish a bilateral investment treaty, and Washington reiterated its commitment to encourage exports of American commercial high technology for civilian use in China. Agriculture, aviation, health, climate, infrastructure and finance all produced agreements or working groups, the usual machinery for keeping a state visit alive after the motorcades had gone home.

Obama also pressed Xi on cyber theft and Beijing’s construction campaign in the South China Sea, where dredgers were turning reefs into islands large enough for runways and military facilities. The governments announced that neither would conduct or knowingly support cyber theft of trade secrets for commercial advantage. Xi said China did not intend to militarize its South China Sea outposts, an assurance that became harder to defend as missile systems, radars and military aircraft followed. None of that prevented the same visit from producing plans for more investment and more technology trade; Washington still expected the difficult parts of the relationship to fit inside a much larger economic one.

Trump hosted Xi at Mar-a-Lago in 2017, and Joe Biden met him near San Francisco in 2023. Washington gives us the cleaner comparison. On Thursday, Trump will host Xi at the White House, four months after visiting him in Beijing and less than a year after the two men met in Busan. Treasury Secretary Scott Bessent and Vice Premier He Lifeng met in New York on Sunday, with trade, critical minerals and artificial intelligence still on the table. The state visit will begin with plenty of business unfinished.

What Grew

The United States imported $483.2 billion in goods directly from China in 2015 and ran a $367.3 billion goods deficit. By 2025, imports had fallen to $308.7 billion and the deficit to $202.7 billion. That decline is real, though anyone tempted to read it as a clean retreat from Chinese manufacturing should spend a little time with the rest of the Asian trade tables.

Vietnam’s exports to the United States have tripled since the 2018–19 tariff increases. Much of that growth reflects genuine production in Vietnam, including by Chinese-owned firms. A Federal Reserve analysis found that those firms accounted for about a quarter of Vietnam’s goods exports to the United States in 2020–23, up from 11% immediately after the tariffs, while many still bought inputs from China. The customs label records where a product underwent its legally significant transformation; it does not identify every supplier or investor behind it.

Four months before Xi’s 2015 state visit, the State Council released Made in China 2025. Beijing’s diagnosis was unusually blunt: China had become a manufacturing giant, “large but not strong,” and still depended too heavily on foreign technology higher up the value chain. The plan targeted semiconductors, robotics, electrical equipment, aerospace, ships, machine tools and new-energy vehicles, among other industries. A U.S.-China Economic and Security Review Commission assessment at the end of the plan’s first decade found that Beijing missed some targets and met or exceeded others; across most of the sectors it singled out, Chinese firms gained ground.

China had about 1.5 terawatts of installed generation capacity at the end of 2015; by the end of 2025, it had 3.89. Its solar capacity went from about 42 gigawatts to 1,200, and wind rose from roughly 130 to 640. The solar capacity added in those ten years alone equaled more than three-quarters of China’s entire power system when Xi stood beside Obama in the Rose Garden. Installed capacity is not the same as electricity delivered, particularly for solar and wind, but it tells us how much equipment and construction the country brought into service.

China installed 295,000 industrial robots in 2024, 54% of the worldwide total, pushing its operating stock past two million. Japanese and European manufacturers once supplied most of the robots automating Chinese plants; Chinese makers now supply a majority of their own domestic market. The production equipment has become another Chinese industry, rather than merely something imported to make Chinese factories more efficient.

Chinese yards already delivered 36% of the world’s ships by tonnage in 2015; by 2024, their share had reached 53%, according to the U.S.-China Commission. Foreign owners took delivery of more than 70% of the tonnage those yards produced between 2019 and 2024, so global commerce keeps much of the system busy whether Beijing places a naval order or not. Roughly 40% of China’s annual ship production comes from yards with strong military links; civilian orders keep welders, suppliers and facilities working at sites that also build for the navy and other maritime forces. The Commission counts 78 additions to China’s navy between 2015 and 2023, against 20 for the United States.

Missile production leaves fewer convenient numbers behind. Bloomberg traced listed Chinese companies connected to the country’s two principal state-owned missile conglomerates and found 81 that disclosed producing important missile components in 2025, up from 32 in 2013. The firms reported 189 billion yuan in combined sales, much of it from businesses other than missiles. That figure cannot tell us how many weapons rolled out the factory door, though the expanding supplier network deserves attention in its own right.

China still imports some of the machinery needed to make its most sophisticated products, especially high-end computer-controlled machine tools and leading-edge semiconductor equipment. Nor can a commercial shipyard casually switch from containerships to modern destroyers, or an electric-vehicle plant become a missile factory on command. The useful overlap lies farther down: electronics, batteries, motors, specialized materials, machine tools and workers accustomed to making complicated things in large quantities.

China accounts for more than 90% of global refining for gallium, graphite, manganese and rare earths, according to the International Energy Agency. It also supplied more than 90% of the growth in global copper-smelting capacity since 2005, lifting its share of capacity from about 15% to half. A mine in Australia, Africa or the United States solves only part of the problem if its output still has to pass through Chinese processing before a manufacturer can use it.

The property sector has meanwhile suffered years of losses, household demand remains weak relative to production, and China’s population continues to shrink and age. Beijing subsidized inefficient firms and built excess capacity along with useful capacity; industrial policy never acquired a magical exemption from bad investment. What stands out is how much physical capability China added while carrying those problems at the same time.

Busan

When Trump and Xi met in Busan on October 30, 2025, their agenda had moved a long way from the investment treaty of 2015. China had announced new rare-earth export controls; Washington was restricting advanced technology; Beijing had retaliated against American semiconductor companies. Agricultural purchases and fentanyl precursors sat in the same negotiation as minerals used in motors, electronics and weapons.

The Busan arrangement suspended or deferred some Chinese mineral controls, addressed retaliation against American semiconductor companies, restored agricultural purchases and brought commitments on fentanyl precursors. Washington adjusted portions of its tariffs in return. Both governments retained the industrial leverage that had brought those issues to the table.

Trump’s May visit to Beijing created a Board of Trade and a Board of Investment and brought agreements on rare-earth supplies, agriculture and Boeing aircraft. The White House described its aim as a “constructive relationship of strategic stability.” That language leaves plenty of room for competition, while giving negotiators somewhere to take problems before the presidents have to settle them personally.

Some problems have made it back onto the presidents’ desks anyway. Bessent, He Lifeng and U.S. Trade Representative Jamieson Greer met in New York ahead of this week’s summit, and the tariff truce faces a November deadline. The American side proposed a channel for notifying each other about serious AI incidents, while Washington continued pressing Beijing for more dependable critical-mineral supplies. The two governments still trade hundreds of billions of dollars in goods each year, yet increasingly negotiate over which technologies, materials and equipment should cross the Pacific at all.

The Machine Behind the Model

American private AI investment reached $285.9 billion in 2025, more than 23 times the Chinese figure measured by Stanford. Government guidance funds make that an incomplete comparison of the two countries’ total spending, but American companies remain central to advanced accelerators, cloud computing and frontier-model development. They can also raise capital for absurdly expensive computing projects at a speed few countries can match.

Chinese systems have nevertheless stayed close to the frontier. American and Chinese models traded places near the top of major rankings repeatedly beginning in early 2025; Stanford put the leading American model only 2.7% ahead of the leading Chinese model on its measure in March 2026.

Training and serving large AI systems takes chips and data centers; the data centers need transformers, cooling equipment, transmission and generation. Semiconductor fabs need electricity, water, chemicals, specialized machinery and skilled workers of their own. The United States holds formidable advantages in chips, capital and research, but thousands of megawatts of new demand are trying to connect to electrical systems where permitting, interconnection and equipment can hold a large project up for years.

China’s leading-edge chip industry still depends on foreign equipment, which American export controls deliberately make harder to obtain. At the same time, Beijing spent the last decade adding generation, electrical equipment, batteries, power electronics, mature-node chip capacity and industrial machinery on a scale that gives its technology companies a very different physical environment in which to work. Compute has to run somewhere, even when the models and chips receive most of the attention.

Modern weapons draw from parts of that same economy. Autonomous systems need sensors, batteries, communications equipment, motors and software; missiles need guidance components, chemicals, metals and machine tools; ships require engines, steel, cable, pumps and thousands of ordinary industrial parts before anybody installs the combat system. Beijing’s military-civil fusion policies seek to pull researchers and suppliers beyond the traditional defense firms into the PLA’s orbit. A submarine component or specialized missile seeker still needs its own narrow supplier base, but the depth of the surrounding industrial system matters as soon as demand rises beyond what peacetime defense lines were built to handle.

The Window

The PLA’s centenary goals make 2027 a useful marker for measuring Chinese capability around Taiwan, though Xi did not set an invasion appointment. Geography will matter in the opening fight: Taiwan sits close to Chinese airfields and missile forces, while much of the American military would have to cross the Pacific. Ships, aircraft, submarines, launchers and ready magazines would determine what either side could do first.

If the fighting continued, those inventories would start to disappear. In CSIS Taiwan wargames, American stocks of some long-range precision munitions ran out in less than a week; aircraft would need parts and maintenance, damaged ships would need yards, and cratered runways would need repair crews. RAND has examined how an intense U.S.-China war could also become a prolonged one. China would face serious disruption if attacks damaged its ports, factories, transport links or power supply; neither side would get to mobilize its industry under laboratory conditions.

An unclassified Navy briefing estimated that China has roughly 230 times the shipbuilding capacity of the United States measured in gross tons produced over time. That measure largely captures commercial construction. It says nothing about a 230-to-one ability to build nuclear submarines or destroyers with advanced combat systems. The Congressional Research Service still treats the underlying gap as a serious concern, especially if both navies take heavy losses and then have to repair or rebuild what remains.

Public information cannot tell us how efficiently a Chinese yard could repair a battle-damaged destroyer while an adversary was also trying to destroy the yard. A large commercial dry dock helps, but it cannot produce a specialized radar or propulsion system on demand. Missile production presents the same difficulty: the supplier network has grown, while finished production rates, stockpiles and the resilience of individual plants remain largely opaque. China enters that uncertainty with a much larger maritime and manufacturing system than it had when Xi last visited Washington.

The United States has put money into missile plants, submarine suppliers, shipyard workforces and Pacific infrastructure. Guam is receiving new defenses; Japan is fielding longer-range weapons; and the Philippines has opened additional locations to American forces. GAO reported in August that maintenance delays and idle time cost the Navy more than 15,000 attack-submarine operational days over the previous decade. As of June 2025, Virginia-class yards were building at roughly one submarine per year, half the Navy’s target; the two boats delivered in 2025 both arrived more than three years late.

Congress can fund a program much faster than a yard can train welders, qualify suppliers, install equipment or finish a submarine already years behind schedule. That lag matters before a conflict. After both sides begin using what they have already bought, it may matter even more.

What Xi Sees

Xi returns after eleven years that included the trade war, Huawei, COVID, semiconductor controls, Hong Kong, Xinjiang, repeated Taiwan crises, rare-earth restrictions and a revival of American industrial policy. Much of the commercial relationship survived. American consumers still buy Chinese goods; Chinese manufacturers still need some American and allied technology; American industry still relies on Chinese processing and components. The two presidents have reasons to preserve room for business and to prevent the competition from turning into a military crisis.

The immediate agenda is crowded enough: minerals, tariffs, semiconductor controls, AI and a truce with a November deadline. An agreement on any of them could matter to the companies waiting on a shipment or a license. China’s factories and shipyards will keep operating after the ceremony, however, and American submarine yards will keep working through the backlog they already have. New semiconductor fabs must qualify equipment and ramp production; mineral projects outside China must build competitive processing capacity; utilities must connect new loads to generation and transmission that often have yet to arrive.

None of the peacetime output figures tells us how China’s industrial system would perform under wartime attack, and the American equivalents look less comfortable once existing stocks and spare capacity run down. China has spent the eleven years between state visits building many of the systems that would sustain both its technological ambitions and its military power. The American response has begun, but a large share of it remains somewhere between appropriation and output. Those schedules will still be there when the motorcade leaves.

The Charter gives the organization work to do: solve international problems, improve standards of life, coordinate governments that often have little reason to cooperate. UN80 has spent this year confronting how much machinery accumulated around that work. Silicon Valley’s better companies developed under a less forgiving discipline; projects have owners, products reach users early, evidence arrives quickly, and a bad result eventually has to change where the money and people go. Peacekeeping and humanitarian relief obviously aren’t software products, but an institution can still know what it is trying to accomplish, measure whether the work is getting there, close things that have outlived their purpose and move resources toward whatever is producing better results.

Forty Thousand Instructions

More than 85% of active Secretariat mandates contained no instructions for review or termination when the UN examined them. Renewal consequently became the ordinary course for a large share of the work, a phenomenon the mandate review describes, with some restraint, as “inertia.” Secretariat entities cited roughly 4,000 mandate sources while preparing the original 2026 program budget, 36% of them cited by more than one entity; around 50 citations referred to instructions that appeared outdated or inactive. Recurring resolutions often changed little from one session to the next, separate reports covered similar subjects, and individual entities sometimes planned work without a complete view of what the rest of the system was already doing.

Member states created those obligations over decades through resolutions and decisions, after which the Secretariat converted them into programs, meetings, reports, staffing and budget requests. Wars produced peacekeeping missions, famines produced relief programs, new diseases produced health initiatives and successive development agendas generated their own institutions and reporting requirements. A program could make perfect sense when adopted and remain difficult to reconsider years later, after the emergency changed, another organization entered the field or the original political bargain lost its relevance. Governments usually have constituencies urging them to create something new; considerably fewer organize around reopening an old diplomatic compromise merely to close an office.

UN80 is reducing staff positions by 21% in the Secretariat’s revised 2026 operating model, relocating posts from expensive duty stations and combining administrative systems that developed separately across the organization. The new mandate rules require clearer objectives for future work and give member states procedures to adapt, merge or retire what already exists; the 2027 program plans already cite substantially fewer mandate sources per Secretariat entity after an initial review for relevance and comparative advantage.

Four Years

OpenAI was founded in 2015 and released ChatGPT publicly in November 2022. Less than four years later, companies have reorganized work around large language models, governments have rewritten policy around them and the company’s chief executive is appearing before the Security Council because the technology has acquired national-security implications. A bad product can survive hype, internal politics and an impressive amount of venture capital, but users leave, competitors ship something better, benchmark scores move and eventually somebody has to explain why the expensive thing still does not work.

DeepMind’s AlphaFold moved from research project to ordinary scientific infrastructure on an even shorter timetable. The AlphaFold Protein Structure Database opened in 2021 and expanded the following year to more than 200 million predicted structures, covering nearly every protein then catalogued by science. More than three million researchers in over 190 countries now use AlphaFold, including more than one million users in low- and middle-income countries; the work behind the system received part of the 2024 Nobel Prize in Chemistry three years after the database became publicly available.

Researchers still have to determine whether a predicted structure leads anywhere useful, design compounds, test them, manufacture whatever survives the trials and get the resulting medicine to patients. AlphaFold removed a large amount of work from one part of that sequence without removing the rest. When a structure that might previously have consumed months of laboratory effort is already available in a database, scientists reach the next bottleneck sooner; laboratory capacity, clinical trials and manufacturing become more valuable accordingly.

The World Bank’s 2026 World Development Report argues that developing economies could use AI to accomplish in a decade what might otherwise take a century, particularly where progress has been limited by shortages of skilled people. Steam technology took about 80 years to diffuse to lower-income countries, electricity around 40 and the internet roughly 20; middle-income countries accounted for half of ChatGPT’s worldwide traffic within six months of its release. The report documents AI already being used for medical screening, agricultural and weather advice, education and government services, including smaller systems capable of working through basic phones and voice interfaces rather than the enormous computing infrastructure required to train a frontier model.

Expertise Without Waiting

Poor countries have spent generations with too few doctors, engineers, teachers, agronomists and experienced civil servants. Training more remains necessary and takes years; importing them is expensive and often temporary. Software that can reproduce parts of their expertise cheaply won’t eliminate those professions, although it can increase how much useful knowledge reaches places that previously had very little of it. A nurse with better diagnostic assistance, a farmer receiving a more accurate local forecast or a teacher generating material in a language poorly served by existing textbooks gains capability without waiting for an entire professional workforce to arrive.

Electricity and connectivity still determine how much of that capability leaves the screen. The World Bank identifies power, internet access, skills, data and institutional quality among the principal constraints on AI adoption in developing economies. Generation and transmission have to expand; broadband has to reach more communities; clinics and schools need equipment that works reliably; ministries need procurement systems capable of buying useful technology before it becomes obsolete. Better models increase the value of those investments because the connection increasingly provides access to expertise as well as information.

The development ecosystem that grew through the 1990s and 2000s learned to work through multi-year grants, implementing partners, consultations, monitoring frameworks and reports because the underlying scarcities usually moved slowly. Doctors still took years to train, power plants took years to build, laboratories were expensive and administrative capacity accumulated over generations. Those structures could coordinate difficult work across dozens of countries, although they also made activity considerably easier to document than completion. A program could hire staff, disburse its budget, hold its consultations, submit its reports and still leave the underlying constraint largely where it found it.

AI can distribute pieces of expertise more cheaply; biotechnology can compress parts of scientific discovery; better forecasting can reduce uncertainty; automation can increase what a small organization or government can administer. Cheap diagnostic assistance does little for a clinic without medicine, better agricultural forecasts still need to reach the farmer, and faster drug discovery raises the value of factories capable of manufacturing what scientists find. Electricity, broadband, laboratories, factories and competent procurement offices remain stubbornly relevant, although new technology gives development institutions more opportunities to remove bottlenecks instead of organizing indefinitely around them.

The Old Operating Model

The General Assembly responded to AI by establishing a Global Dialogue on AI Governance and a 40-member scientific panel. The first dialogue met in Geneva in July, another is scheduled for New York next May, and the panel will produce annual assessments and thematic briefs; a separate AI Governance for Humanity Lab in Valencia supports work across the system. None of those bodies can regulate AI directly.

An issue becomes a mandate; the mandate produces a panel, office or program; the new structure needs staff, meetings, reports and a budget; the governments that created it rarely include an expiration date. No single step has to be unreasonable for the accumulated result to become difficult to manage. UN80 is now sorting through tens of thousands of earlier instructions created by the same process, identifying work that overlaps, has already been completed or no longer needs to continue in its existing form.

Cash, customers and competitors interrupt that sequence more often in a startup. Teams put unfinished products in front of users and change them when the assumptions prove wrong; capital moves among approaches, employees get reassigned and companies disappear when the evidence remains bad for long enough. Large technology companies become bureaucratic too, obviously. Younger firms generally encounter the consequences sooner.

Peacekeeping missions, humanitarian operations and diplomatic negotiations will never produce feedback as clean as a product dashboard, nor should they. Their objectives can still be defined with more precision, duplicated functions can be consolidated, costs can be compared with results, and a program whose original purpose disappeared twenty years ago can eventually disappear with it. UN80 already accepts most of those propositions; otherwise there would be little reason to review the mandates in the first place.

A consultation can identify an obstacle, a report can establish what governments do not know and an NGO can provide expertise a state does not possess. Did more electricity reach the community afterward? Did the clinic treat more patients? Did farmers receive better information? Did the ministry become capable of performing the work itself? Did the new technology reach people who could benefit from it? The UN’s mission concerns those outcomes rather than the institutional forms used to produce them.

What Survives Review

The 2027 program plans already cite about 30% fewer mandate sources per Secretariat entity after the initial review. Payroll, travel, technology services and other administrative functions are being consolidated; posts are moving from high-cost locations; member states now have procedures for adapting, combining and retiring existing work. An organization with global ambitions still has to know who is responsible for something, what it costs, whether another office is already doing it and when the work is finished.

Sequencing the first human genome consumed years and billions of dollars through the Human Genome Project; protein structures that occupied laboratories for months can now be retrieved from a database; translation, coding and scientific analysis that were expensive enough to ration are becoming inexpensive enough to distribute much more widely. Faster discovery increases the need for manufacturing, cheap intelligence increases the value of electricity and broadband, and better tools become more useful when governments and institutions can deploy them quickly.

Technology companies revisit constraints because removing one can create an enormous economic reward. Some turn out to be stubborn physical limits, some merely move somewhere else, and some disappear after enough engineering effort and capital arrive. Development institutions have a different incentive structure, but the people they serve would benefit from the same willingness to reconsider problems that have gradually become accepted as permanent.

Altman will spend Wednesday answering questions from diplomats about technologies that barely existed when the Security Council first put AI on its agenda three years ago. Nothing in that encounter requires a larger UN role in governing his industry; UN80 has already supplied a sufficiently difficult problem inside the building. The organization is trying to consolidate overlapping work, shorten an enormous list of mandates, reduce administrative duplication and develop a mechanism for deciding when an activity has accomplished enough to end.

Under the new mandate rules, member states will spend the coming year deciding which existing instructions should continue, which can be combined and which have run their course. The Secretariat still has thousands of mandates and decades of accumulated machinery to work through.

Brookhaven National Laboratory leads the $11.5 million GridFM 2.0 project, which aims to increase planning throughput by more than 10,000 times and run some calculations more than 1,000 times faster than conventional methods. A utility trying to connect a new power plant, accommodate a data center or decide where another transmission line would do the most good could spend the morning testing an enormous number of possible grids and, at least in theory, know by afternoon which one works best. Then it might discover that the transformer required to build that grid won't arrive until 2029 or 2030.

DOE says large transformers for substations and generators now carry lead times of three to four years in some cases. Distribution transformers, the smaller units hanging from utility poles or sitting on concrete pads outside buildings, typically took three to six months to procure in 2019; utilities now wait one or two years, sometimes longer, while demand has risen 41%.

GridFM could therefore compress weeks of planning into hours while leaving another clock almost untouched. Somebody still has to turn the answer into copper, steel and insulation, and the factories that do that work already have years of orders ahead of them.

The Part That Weighs Hundreds of Tons

Transformers change voltage; electricity travels long distances more efficiently at very high voltages, then utilities step it down repeatedly before it reaches a factory, a house or a row of server racks at a usable level. More than 90% of the electricity Americans consume passes through a large power transformer somewhere along the way.

Utilities order the largest transformers to fit particular substations and operating requirements, and some weigh hundreds of tons. Moving one from the factory to its eventual home can require specialized railcars or trailers, route studies, bridge reviews and heavy-lift equipment; unlike a server or a generator small enough to move on an ordinary truck, a large transformer can force the transportation system to organize itself around the machine. Utilities can't simply keep a few universal spares in a warehouse and dispatch one when something breaks.

America also doesn't make enough of them. Roughly 82% of the large power transformers that U.S. utilities installed in 2019 came from abroad, according to DOE, while American manufacturers still rely on foreign suppliers for much of the grain-oriented electrical steel that forms transformer cores.

That dependence follows the transformer through almost every stage of production. An American factory may assemble the final machine, but materials may arrive from overseas; the order then enters a factory backlog measured in years before the finished transformer begins a journey planned around its extraordinary weight and dimensions. There aren't many places in that chain where a planner can simply delete six months.

Forty Thousand Ways to Order a Transformer

Distribution transformers present a different version of the same problem because American utilities have spent decades ordering them in an astonishing number of variations.

Utilities specify different bushings, mounts, voltage ratings, connections, accessories and other features. Some differences reflect genuine engineering requirements; others survive because a utility chose a design years ago, wrote it into its standards and kept buying the same thing. Specifications behave a little like old contract language in that respect: once a clause enters the form, it can remain there long after everyone has stopped asking why somebody inserted it in the first place.

DOE says utilities now use almost 40,000 distribution-transformer configurations nationwide. Take only ten common design attributes and the number of possible combinations climbs above 80,000, which turns the familiar gray box hanging from a utility pole into a surprisingly difficult manufacturing problem.

Factories have to accommodate more designs and shorter production runs; utilities hold inventories that don't necessarily fit one another's systems, so a perfectly good transformer sitting in one company's yard may do little for another utility trying to restore power twenty miles away. Variety has value when engineering demands it. When history rather than engineering produces the variety, however, everybody pays to preserve the distinction.

DOE has therefore worked with utilities and manufacturers to narrow specifications, identify common designs and eliminate legacy variations that no longer serve much purpose. Electric systems really do differ, so nobody can standardize the entire country into a handful of boxes; still, 40,000 configurations contain a fair amount of bureaucratic sediment alongside the engineering, and removing even some of it could help manufacturers run longer production batches while allowing utilities to stock and share more interchangeable equipment.

The Factory Has a Lead Time Too

Standardizing a transformer doesn't manufacture one. Producers still need copper, electrical steel, insulation, winding equipment, core assembly, testing facilities and people who know how to put all of it together; when manufacturers decide to increase output, they also need more buildings and specialized machinery, some of which can take a year or two to procure.

In April, the White House invoked the Defense Production Act for transformers, substations, high-voltage circuit breakers, electrical core steel and other grid equipment. The presidential determination described America's capacity to design, manufacture and deploy major grid infrastructure as “dangerously limited,” citing long procurement times, inadequate domestic production and dependence on foreign suppliers.

Manufacturers have started spending accordingly. Hitachi Energy broke ground in June on a $457 million expansion in South Boston, Virginia, which the company says will become the country's largest large-power-transformer factory; Siemens Energy is directing part of a $1 billion U.S. manufacturing investment toward additional transformer production in North Carolina and a new high-voltage switchgear plant in Mississippi. Prolec, meanwhile, is spending $140 million in Goldsboro, North Carolina, to increase annual medium-power-transformer production from roughly 220 units to 420, and the company expects the expansion to take about two and a half years.

Hitachi added another project on September 15, announcing a $528 million transformer factory in Gallman, Mississippi. The company plans to build a plant more than twice the size of its existing Crystal Springs operation and more than double its production capacity there; with the new project, Hitachi puts its broader American manufacturing expansion at roughly $1.5 billion. Construction should begin late this year, while Hitachi expects the factory to start producing transformers in 2029.

That date matters because industrial capacity has its own supply chain. Utilities can't solve a three-year transformer backlog merely by demanding more transformers; manufacturers have to order machinery, expand plants, recruit and train workers, qualify production lines and test the equipment before another unit reaches the market. The obvious answer to a shortage is more production, but the machinery that creates more production belongs to the same stubborn physical economy as the product itself.

McKinsey put some scale around the problem three days after Hitachi's announcement. Its analysis estimated that Europe and North America had 914 gigavolt-amperes of transformer demand against 566 GVA of domestic supply in 2025, a gap of roughly 38%. Manufacturers could add about 186 GVA of announced capacity by 2030 under McKinsey's model; demand keeps growing in the meantime, however, leaving an estimated annual shortfall of roughly 312 GVA even after those expansions come online.

The analysis becomes more interesting when McKinsey strips away some of the demand that usually dominates the discussion. In one downside scenario, it removed data-center demand entirely and sharply reduced renewable construction; the model still produced a transformer shortfall of roughly 180 GVA in 2030. Data centers plainly make the problem harder, particularly because a handful of very large projects can alter a utility's load forecast much faster than anyone can expand a factory, but the manufacturing constraint runs deeper than the AI buildout.

Utilities also have aging equipment to replace, new generation to connect, industrial loads to serve and ordinary grid investment that they deferred for years.

So manufacturers are spending billions to relieve a shortage whose solution arrives on roughly the same schedule as the shortage itself. Hitachi's newest factory will begin producing equipment in 2029; a utility placing an order for certain large transformers today may also expect delivery around 2029 or 2030. Capacity is coming, certainly, but nobody gets to teleport through the years required to build it.

172 Tons at a Time

Ukraine offers a harsher demonstration of what those physical constraints mean once equipment becomes urgently necessary. Russia has repeatedly attacked the country's power system, destroying generating capacity and grid equipment while Ukraine and its allies race to replace it; through its Civil Protection Mechanism, the European Union has sent more than 7,200 transformers since Russia launched its full-scale invasion.

One of the more remarkable efforts began with an entire thermal power plant in Lithuania. Crews dismantled the plant and spent eleven months moving it to Ukraine in 149 separate shipments containing 2,399 tons of equipment; forty loads exceeded ordinary dimensions, and some of the transformers and stators weighed about 172 tons each.

At that scale, infrastructure stops behaving like the diagrams used to describe it. The replacement equipment first has to exist somewhere; crews then have to disconnect it, prepare it for transport and load it onto a vehicle capable of carrying it. Engineers have to find roads and bridges that can bear the weight, officials have to clear border crossings and permits, and cranes and specialized crews have to meet the cargo at the other end before anyone can install it and connect it to a live electrical system.

Software can help with nearly every decision surrounding that process. An engineer can identify the missing section of a network, calculate how to route power around it and determine what equipment should replace it; better models can perform those calculations much faster, while AI may eventually test thousands or millions of alternatives during the same period. None of those gains, however, change the logistical fact that somebody still has to move a machine weighing 172 tons across actual roads.

Military logistics has always imposed the same discipline. Sophisticated systems attract attention because they embody technological advantage, while sustained operations depend on ammunition, spare parts, maintenance, transportation and the industrial capacity to replace whatever breaks. Electrical infrastructure obeys much the same rule, although its supply chain usually attracts considerably less attention until the lights go out.

Every Server Rack Reaches the Grid

Commentators usually describe the AI race in terms of chips, GPUs, models, compute and talent, all of which matter enormously. Eventually, however, every server rack reaches an electrical system assembled from much less glamorous machinery.

A data center needs electrical distribution equipment inside and around the building; it also needs a connection to a substation full of transformers, breakers and switchgear. That substation connects to transmission lines and generation, which bring their own layers of equipment, while semiconductor fabs and advanced manufacturing plants require much of the same infrastructure. Even the factories that manufacturers are building to produce more transformers, breakers and switchgear need additional electrical equipment of their own.

Demand therefore begins to compound in slightly strange ways. The industries driving new electricity consumption need more grid equipment; the factories expanding grid-equipment production consume power and require grid connections themselves, while the mines, mills and component suppliers feeding those factories add another layer. The feedback loop doesn't continue forever, obviously, but it helps explain why adding industrial capacity involves more than announcing a factory and pouring concrete.

GridFM 2.0 could still matter enormously precisely because it may help utilities avoid some of those physical requirements. If a billion simulations reveal a safe way to reroute power around a constraint, show that existing equipment can carry more load or identify a different transmission upgrade that eliminates the need for another transformer, the avoided order has real value; removing one major piece of long-lead equipment can move a project's completion date by months or even years.

Faster planning also gives utilities something the manufacturing system badly needs: warning. A company that can see several years ahead to the point where its network runs out of capacity can reserve manufacturing slots, order equipment and begin building substations before the load actually arrives. GridFM can't shorten every factory backlog, but it may allow a utility to encounter that backlog earlier, when time remains to do something about it.

Eventually, though, the simulation produces an answer that has to leave the computer.

Then Procurement Begins

Suppose GridFM works exactly as DOE hopes. A utility feeds it the network, demand forecasts, generation plans and proposed projects; the system tests a billion possibilities and identifies the best answer in a day. At that point the intellectual problem may have collapsed from weeks into hours, while the industrial problem has barely begun.

Steel mills have to roll electrical steel and manufacturers have to wind copper, assemble transformers and test them. Hitachi has to finish factories in Virginia and Mississippi; Prolec has to install more machinery, while manufacturers everywhere have to find space in order books already crowded with other utilities' projects. Carriers need permits and route plans for oversized loads, and substations need land, concrete, electricians and lineworkers before anybody energizes the equipment.

That doesn't diminish the value of faster planning. Knowing what to build in a day beats spending weeks or months finding the same answer, especially when a poor planning decision can tie up scarce equipment and capital for years; GridFM may also help utilities avoid some orders entirely, or at least place the unavoidable ones before everybody else discovers that they need the same equipment.

Rather, the improvement changes where the delay lives. Once software strips time out of planning, procurement, manufacturing, transportation and construction consume more of the schedule, and the constraints that once hid behind engineering studies become much harder to ignore. DOE puts some large-transformer lead times at three to four years; McKinsey says portions of the market can stretch toward five.

GridFM is trying to let utilities reason about an enormous physical system at something approaching machine speed; the factories, mills, railcars, trailers, cranes and crews that turn those answers into an electrical system still operate according to the stubborn tempos of materials, labor and construction. If a utility places an order today, 2029 isn't merely a date on a planning chart. It may be the year the transformer arrives, and, in Hitachi's case, the year one of the new factories meant to relieve the shortage finally begins making them.

The case asks what Huawei did years ago when it wanted American technology. The hardware arriving while the jury sits tells a different part of the story: what Huawei did after Washington decided it should no longer have access to it.

Eight Years

Federal prosecutors first charged Huawei in 2018 and publicly unsealed the case the following January. A broader racketeering indictment followed in 2020, alleging that the company spent years obtaining technology from American firms through fraud and deception, including T-Mobile's Tappy phone-testing robot, source code, antenna technology and other trade secrets. Prosecutors also accuse Huawei of concealing its relationship with an Iranian affiliate so that it could move money through the American financial system despite U.S. sanctions. Huawei denies the charges; it argues that prosecutors have recast legitimate business activity and isolated employee misconduct as a criminal enterprise.

The trial may run for roughly three months, although getting to a jury took eight years. Huawei didn't spend those years standing still. It lost access to critical American technology, watched its smartphone business collapse, built its own operating system, redesigned processors around increasingly restrictive export controls, found substitute suppliers and eventually worked its way back toward its old scale.

Peter Thiel described an earlier version of this problem when he talked about PayPal's first years as a “race between technology and politics.” PayPal was moving into a part of the financial system where the rules were still catching up, and Thiel's instinct was to grow quickly enough that regulators would eventually encounter an established network rather than a proposal they could simply stop. He later described the result as a fait accompli.

Obviously, PayPal's fights with financial regulators bear little resemblance to a federal racketeering prosecution involving allegations of trade-secret theft and sanctions evasion. The useful part of Thiel's observation is the clock. Companies ship products, engineers redesign systems, suppliers disappear and new suppliers emerge while institutions work through conduct that may already be several product generations old.

The Justice Department understood that problem when it sued Microsoft in 1998. Janet Reno sought preliminary relief because waiting for the full case threatened to give Microsoft enough time to capture the browser market; she warned that the company's conduct could inflict “irreversible harm to competition” while the litigation crawled forward. That wasn't bureaucratic impatience. DOJ recognized that a remedy arriving after the market had reorganized itself might be legally satisfying and commercially irrelevant.

A generation later, the department sued Google over search in October 2020, before ChatGPT existed as a public product. Judge Amit Mehta found in 2024 that Google had illegally maintained monopolies in general search and search advertising; by the time he reached remedies in 2025, however, generative AI had altered the market enough that he wrote it had “changed the course of this case.” The court had finally caught the market it set out to regulate, only to discover that the market had kept moving.

Huawei makes the timing problem more complicated because Washington isn't dealing solely with a company responding to incentives inside a market. It is dealing with a Chinese national champion operating inside an industrial policy that increasingly treats every foreign dependency as a vulnerability worth eliminating.

Before the Verdict

Washington didn't wait for a Brooklyn jury before acting against Huawei. In May 2019, only months after prosecutors publicly unveiled the first case, the Commerce Department placed Huawei and dozens of its affiliates on the Entity List and restricted their access to American technology. Commerce tightened those controls the following year by reaching certain semiconductors manufactured overseas with American software, technology or semiconductor equipment.

That second move mattered because the supply chain extended far beyond whatever country appeared on the finished chip. Huawei could design a processor in China and contract with a foundry somewhere else; the foundry could still depend on American design software, fabrication equipment or other tools. Washington therefore reached deep into factories it didn't own and jurisdictions it didn't control because American technology occupied enough indispensable positions in the semiconductor stack.

Huawei felt the pressure quickly. After losing access to advanced chips and Google's mobile ecosystem, the company saw annual revenue fall about 29 percent in 2021, while its premium smartphone business suffered even more visibly. Engineers had to redesign products around worse options; managers had to locate substitute suppliers; the company began spending heavily to recreate technologies it had previously been able to purchase.

Those years bought Washington something valuable: time. They also handed Huawei a remarkably detailed shopping list.

There is a recurring problem in economic coercion here that goes well beyond semiconductors. A country often has the greatest leverage over something precisely because the target hasn't yet bothered to reproduce it. Once the supplier weaponizes the dependency, the economics of duplication change. A substitute that looked wasteful at normal commercial prices can look perfectly rational when the alternative is strategic exposure. Nobody builds a second refinery, payments network or machine-tool ecosystem because redundancy is elegant; they build one because someone has just demonstrated why relying on the first was dangerous.

Huawei received that lesson rather emphatically.

Building Around the Controls

HarmonyOS began as an answer to the possibility that Huawei could lose access to Google's Android ecosystem; the emergency passed, but the operating system remained and grew into something larger. Huawei also pushed deeper into semiconductor design, enterprise software, automotive technology and domestic supply chains while Beijing and Chinese industry directed enormous amounts of money and attention toward many of the same dependencies.

The recovery now appears in Huawei's accounts. Revenue reached 880.9 billion yuan in 2025, close to its 2020 peak, while a premium smartphone business that looked badly wounded several years ago is again releasing expensive flagship devices built around Huawei's own Kirin processors. The Mate XT2 makes the point rather conspicuously: it runs without the Google ecosystem Huawei once depended upon and uses another Kirin processor produced under semiconductor restrictions designed, among other things, to make precisely this technological climb harder.

The restrictions did make it harder. That's important, because treating the reappearance of an advanced Huawei phone as proof that export controls failed throws away most of the useful information. Controls can matter by raising costs, reducing performance, constraining production or forcing a company to spend six years rebuilding something it once ordered from a supplier. Time has economic and strategic value even when it doesn't produce permanent technological paralysis.

Huawei's newest AI roadmap shows where the adaptation is going next. At Huawei Connect in September, the company said development of its next Ascend generation was running ahead of its previous schedule, with the Ascend 960DT now planned for the first quarter of 2027 and additional generations following annually through the end of the decade. A faster product cadence matters, although the more interesting announcement concerned everything Huawei intends to build around those processors.

Its new Peerium architecture and UnifiedBus interconnect aim to connect thousands of processors into much larger computing systems. Huawei says an Atlas 960E SuperPoD will contain 4,096 NPUs; combinations of those systems could eventually reach hundreds of thousands of processors, while the company has described larger configurations approaching one million NPUs.

Huawei is describing engineering targets, of course, rather than demonstrating that it can economically operate a million-chip system with acceptable yields, power consumption, reliability and utilization. Still, the direction deserves attention because Huawei isn't limiting the contest to whether one Ascend accelerator can match one Nvidia accelerator. It is trying to make that comparison less important.

If a Chinese accelerator remains less capable, Huawei can compensate with more accelerators, faster interconnects, larger clusters, better scheduling software and tighter integration across the stack. That approach doesn't make the disadvantages disappear; in some respects it makes them more expensive. More chips require more memory, packaging, networking, electricity and cooling, while enormous clusters introduce reliability and software problems of their own. Rather, Huawei is shifting the problem upward. Washington constrains the processor; Huawei pushes toward fabrication. Washington pressures fabrication; Huawei needs better memory and packaging. Individual chips lag; Huawei starts treating the cluster as the product.

The history of computing is full of this sort of substitution. Engineers rarely respond to a constraint by politely accepting it; they move the burden somewhere else in the system. Mainframes yielded work to distributed machines, slow storage encouraged elaborate caching, expensive bandwidth produced compression, and weaker processors have often survived by multiplying. There are limits to every workaround, obviously, but there is no reason to expect Chinese engineers to approach export controls with less imagination than engineers approach every other unpleasant design constraint.

Huawei itself supplied a useful caution against getting carried away with the roadmap. Eric Xu acknowledged this month that the company still can't produce enough AI computing equipment to satisfy Chinese demand. That admission tells us considerably more about the current effectiveness of the controls than a photograph of an Ascend processor on a conference stage.

Where the Leverage Still Lives

High-bandwidth memory, or HBM, has become one of the more important pressure points. HBM sits close to an AI processor and moves data into it quickly enough to keep the chip working near full capacity; advanced accelerators require large quantities of it, while U.S. restrictions have made the best HBM substantially harder for Chinese firms to obtain.

Huawei's systems strategy makes that constraint more interesting, not less. If the answer to a weaker chip is more chips, then the company also needs more memory, more advanced packaging, more networking equipment, larger power systems and enough manufacturing capacity to produce all of those components with tolerable yields. Architecture can compensate for some shortcomings in the underlying silicon; it can also magnify demand for almost everything surrounding it.

Chinese firms are working their way through those dependencies as well. CXMT said this month that its fifth-generation DRAM manufacturing platform had entered mass production, another step in China's attempt to reduce its reliance on foreign memory suppliers. Conventional DRAM isn't leading-edge HBM, and mass-producing one shouldn't be treated as evidence that China can suddenly manufacture the other at the volumes and performance levels frontier AI systems require. It does, however, show memory moving along a familiar path: foreign dependence creates political urgency; political urgency attracts capital; capital produces substitutes, initially worse and more expensive, which improve over time.

China has made extraordinary progress in semiconductor manufacturing, but leading-edge production still depends on categories of equipment and technology controlled by the United States and allied countries. Lithography receives most of the attention because enormous machines firing droplets of molten tin with lasers are, admittedly, more interesting than metrology. Yet deposition, etching, inspection, electronic-design software, advanced memory, packaging and dozens of less glamorous parts of the process can constrain output just as effectively.

Huawei once depended heavily on foreign chips it could simply buy. Restrictions shifted more of the problem toward fabrication; advances in fabrication increase the importance of equipment, memory, yields and production scale. Larger clusters raise the importance of interconnects and electricity. Better domestic memory will eventually put more pressure on packaging or manufacturing tools. Semiconductor independence isn't one technological breakthrough waiting to happen; it is a long exercise in discovering the next annoying thing that somebody else still makes better.

That is why arguments over whether export controls have “worked” tend to age badly. Huawei's 2021 collapse was real; its recovery is real too. Chinese firms now produce advanced chips that many observers several years ago doubted they could build under the controls, while manufacturing and memory constraints continue to make those chips harder and more expensive to produce at scale. Both facts matter because the controls aren't a binary machine that either prevents technological progress or fails. They alter the cost and speed of that progress, while the target spends money and engineering effort trying to make each restriction less relevant.

The dependency keeps moving; so does the leverage.

Outside Brooklyn

The jury still has a consequential legal question to answer. Prosecutors must prove the criminal charges beyond a reasonable doubt; Huawei continues to deny them, while testimony now includes evidence concerning the company's alleged relationship with Skycom in Iran and its acquisition of American technology. Whatever verdict the jury eventually returns will resolve questions about liability for conduct that stretches back years.

Meanwhile, the industrial contest that grew around the case continues on an entirely different clock. Suppliers change while witnesses testify; Chinese engineers redesign chips and memory systems while lawyers argue about conduct from the 2010s; Huawei is planning processors it expects to ship in 2029 before the Brooklyn jury has decided what happened a decade earlier.

Washington has also begun competing farther downstream. This month, the Export-Import Bank approved nearly $100 million in financing for Africell to purchase American and allied telecommunications equipment as the United States tries to challenge Huawei's position in African networks. That effort matters because the contest has moved beyond preventing Chinese firms from acquiring American technology. Washington increasingly has to compete against Chinese technology after Huawei has already become somebody else's incumbent supplier.

Incumbency creates its own kind of leverage. Telecom operators don't replace networks because a geopolitical argument became more persuasive; they replace them when financing, equipment, maintenance, compatibility and switching costs make the alternative worthwhile. Huawei spent years benefiting from that arithmetic abroad, sometimes with substantial support from Chinese policy banks. If Washington wants countries to choose something else, export controls alone can't make the sale.

A technological chokepoint remains most valuable while the other side still depends on it. Pressure on Huawei helped turn semiconductor independence into a strategic priority in Beijing; the investment unleashed by that pressure now targets precisely the technologies that gave Washington leverage in the first place. That doesn't mean Washington made a mistake by using the leverage. It means leverage depreciates.

For now, the United States and its allies still possess a considerable amount of it. Huawei acknowledges that it can't make enough AI equipment to satisfy domestic demand; advanced memory remains constrained; leading-edge manufacturing equipment remains difficult to replace; Chinese fabrication still carries cost, yield and performance disadvantages that matter enormously when a company has to produce chips by the tens or hundreds of thousands rather than display one functioning device.

Those are the measurements worth following now: how many Ascend accelerators Huawei can actually ship, what they cost, where the memory comes from, how quickly fabrication yields improve, how much performance Huawei can recover through system architecture, which imported tools remain hardest to replace and, ultimately, how much time those dependencies still buy.

The jury in Brooklyn is deciding what Huawei did when it wanted American technology. Washington has a different question to answer now: how much American technology does Huawei still need?

Much of the law concerns people, factories, transportation, inventories and the records connecting them. Provincial authorities have to maintain mobilization plans and current information on available manpower and resources. Companies assigned military research, production, conversion, expansion or repair work must retain the equipment, materials, technology, data, software and technical personnel needed to perform it, while central authorities are directed to examine the industrial and supply chains supporting military production.

A fair amount of mobilization work can therefore happen while China still looks like it is conducting ordinary administration. A reserve roster can be checked against current employment records. Officials can find out whether trucks listed in an old transportation plan still exist and who controls them. A factory carrying a military assignment can discover that one of its inputs now comes from a single supplier hundreds of miles away.

The work is mundane because large organizations usually fail in mundane ways. Plans age, people move, suppliers disappear, equipment gets reassigned and two offices discover that they have both counted the same resource. A country that finds those problems during an exercise has time to do something about them.

October 1 falls 182 days before April 1, 2027, placing six months of implementation inside the first spring period of the Vulnerability Window. China will enter that spring in the PLA’s centenary year. The Pentagon has described 2027 as a short-term military modernization marker within a program extending through 2035 and 2049; whatever China builds over those six months, the administrative systems behind the force can change considerably faster.

What Changes on October 1

The 2010 law already allowed mobilization when China’s sovereignty, national unification, territorial integrity or security came under threat. The revised statute keeps those grounds and adds China’s “development interests.” Because Beijing has long treated Taiwan as an issue of national unification, the revision expands the statute without creating the underlying Taiwan rationale from scratch.

The more consequential changes sit farther down in the machinery.

China’s mobilization system covers reserve personnel and military production along with the civilian resources required to support them. The new law formalizes national and provincial mobilization plans, implementation plans and statistical surveys of mobilization potential; it also requires a system for dynamically updating that information.

Consider a factory assigned to repair military equipment. The assignment may have been written when a particular machine was still on the floor and several experienced technicians worked there. Years later, the equipment may have changed and the technicians may be gone. A component once available from three suppliers may now come from one. The revised law requires authorities to assess military industrial and supply-chain security, while assigned companies must maintain the equipment, materials, supporting products, technology, data, software and technical teams needed for their missions and regularly exercise those plans.

Transportation creates the same problem across a much larger network. A major contingency would draw on civilian ports, commercial shipping, rail, trucks, warehouses and distribution systems alongside dedicated military logistics. Strategic-material provisions already contemplate civilian organizations handling storage, transportation and distribution, while the wider mobilization framework gives officials reason to test the assumptions underneath those plans before they become urgent.

Officials can discover whether the vessel named in a file is suitable for the intended cargo, whether a port can handle the traffic assigned to it or whether several agencies have all assumed access to the same rail capacity. China has conducted mobilization exercises and incorporated civilian transportation into military planning for years. October 1 gives that existing system a revised statutory framework and starts another round of plans, surveys, exercises and administrative work that can be compared with what came before.

Before Anyone Mobilizes

A large call-up eventually becomes difficult to hide. Reservists stop reporting to civilian jobs, factories alter schedules, commercial vessels disappear from normal service and ports or rail networks begin moving cargo in unusual patterns.

The preceding months can look much more ordinary.

A province updates personnel records. Officials inspect stocks held by a company. A factory runs through an assigned production task. Transportation planners discover during an exercise that the route in their plan cannot handle the expected load. None requires China to change its visible military posture, and each leaves one fewer problem to discover during a crisis.

For an operation across the Taiwan Strait, those details would continue to matter after the first forces moved. Units would consume ammunition and fuel. Damaged equipment would move toward repair facilities. Replacements would travel toward the coast and casualties would enter a medical system still serving the civilian population.

The civilian infrastructure underneath that effort would face its own pressures. Commercial ships can be lost. Ports can stop operating. A factory prepared for emergency production can lose power or access to an upstream supplier. Rail capacity on a spreadsheet disappears quickly if a bridge or terminal stops functioning.

The statute cannot make those constraints disappear. It gives Chinese authorities more opportunities to find some of them in advance.

A replacement part with no second supplier can be identified before production stops. An obsolete reserve roster can be corrected. Officials can discover that two agencies planned around the same trucks. None of those corrections changes the number of missiles or ships China possesses, though each can affect the amount of time lost when existing forces begin drawing on the wider economy.

That creates a harder problem for outside observers. Another commercial-shipping survey in Fujian may simply be another survey. A factory exercise may reflect routine compliance with the revised law. An updated reserve database may never become visible outside the bureaucracy at all.

The useful information accumulates over time: surveys larger than previous versions, exercises repeated more frequently, unfamiliar civilian participants appearing in military support roles, or industrial and transportation preparations beginning to move together. The pattern carries more information than any single event.

The 182 Days

Six months passes quickly in a shipyard. A submarine already under construction may move farther down the line, but China cannot create a new fleet over the winter. Aircraft production remains bounded by existing factories, new crews still require training and major facilities follow construction schedules measured in years.

Government files move faster.

Between October and April, provinces can review reserve assignments and rerun exercises after something fails. Companies can check old military-production plans against current machinery and suppliers. Authorities responsible for defense industry can trace dependencies through a supply chain, find single points of failure and decide which ones warrant another supplier, another stockpile or another plan.

The process will also produce bad information. Companies have incentives to offer optimistic estimates of emergency capacity. Local officials can satisfy requirements on paper. An exercise can validate an unrealistic plan because nobody involved wants to tell a superior that it failed. Implementation will reveal considerably more than the text of the statute by itself.

The same winter passes for Taiwan and its partners. Weapons already under contract continue through production and delivery, bases change, stockpiles move and logistics plans get exercised. Some weaknesses that exist in October will remain in April; others will look different six months later.

That is the basic problem behind the Vulnerability Window. China can keep adding military capability while changes elsewhere alter the relative balance it would face. Mobilization readiness is one of the harder parts of that balance to measure because ships and aircraft leave visible inventories while repair throughput, replacement capacity, spare rail capacity and industrial resilience usually do not.

The revised law directs Chinese officials toward exactly those less visible systems.

Another calendar reaches a regular point later in 2027. The Communist Party held its 20th National Congress in October 2022, and the Party constitution calls for a national congress every five years under ordinary circumstances, placing the next regular congress in 2027. The calendar overlap matters as context rather than proof of a common timetable; the spring operating period, PLA centenary benchmark, mobilization-law implementation and Party leadership cycle will all occupy the same year.

By then, the revised mobilization system will no longer be a newly published statute. It will have a record.

What to Watch

Provincial implementation should begin filling in details that the national law leaves broad. Local regulations can show which agencies acquire responsibilities, while exercises give some indication of what officials are actually testing.

Civilian transportation should leave a relatively visible trail. Commercial-shipping surveys, heavy-transport drills, port exercises and rail activity can be compared with previous years. Industrial-conversion and repair exercises may identify companies expected to support military production, while procurement, subsidies or new capacity can reveal where authorities found weaknesses in particular supply chains.

Other pieces will remain harder to see. Strategic-material planning may appear through storage construction or local regulations. Supply-chain reviews can surface through support for particular producers. Personnel databases may leave little public evidence at all.

Fujian will attract the most immediate attention because it faces Taiwan, though the system required to sustain a cross-Strait operation extends far beyond the coast. Components can be manufactured far inland. Equipment can be repaired there as well, with road and rail carrying material in both directions. A mobilization system built for a large war is a national network.

October 1 starts a six-month record. By April, Chinese authorities will have spent roughly half a year working under the revised law, enough time for provincial rules, exercises and administrative activity to begin showing what changed after the statute took effect.

That comparison will matter more than the law on the day it begins.

Iran still has a 19-foot American autonomous vehicle it can take apart.

Dive-LD can operate for as long as ten days and descend 6,000 meters. More than a cubic meter of its interior is available for payloads, with software, mechanical and electrical interfaces that let operators change sensors and mission equipment without redesigning the vehicle around them. The same hull can support intelligence collection, mine warfare, seabed work and other missions.

Iranian engineers can now inspect how that hull was actually built. They can pull apart the propulsion and power systems, look at cooling and pressure management, identify components and trace the physical connections between them. They can see how Anduril divided the internal volume and how it designed the payload interfaces. Many of the individual parts may already exist on the commercial market. The useful information lies in how engineers chose them, arranged them and made the whole thing work underwater for days at a time.

If the Pentagon builds the autonomous force it wants, this won't be the last American machine an adversary gets to examine.

Built to Be Risked

The Pentagon calls systems like Dive-LD “attritable,” a word close enough to “disposable” that the distinction can seem academic until someone has to decide whether to send one into danger. A cruise missile disappears by design. An attritable vehicle should come home and go out again, but losing one has to remain an acceptable outcome when the mission warrants it.

Price matters. Sending a $2.5 million unmanned vehicle toward a minefield or hostile coastline creates a different problem from exposing a submarine, aircraft or ship along with the crew aboard it. Removing the people, though, only solves part of the problem. A drone that costs nearly as much as the crewed platform it supplements and exists in similarly small numbers will still spend much of its life being protected as a scarce asset.

The larger ambition is numerical. Hundreds or thousands of aircraft, surface vessels and undersea vehicles can spread sensors and weapons across more territory, enter places commanders would hesitate to send people and keep operating as some of the machines disappear. The Pentagon has used the phrase “precise mass” for the idea, an attempt to recover quantity after decades in which the United States steadily exchanged numbers for increasingly capable individual platforms.

For a long time, that exchange worked. Better submarines, aircraft, satellites, sensors and precision weapons allowed American forces to do more with fewer ships and airplanes. China presents a different problem because its industrial base can support enormous quantities of ships, missiles and other military hardware.

Kathleen Hicks put the imbalance plainly when, as deputy defense secretary, she announced Replicator in August 2023. China had more ships, more missiles and more people. Replicator would add large numbers of autonomous systems around America's most capable platforms, giving commanders more sensors, more weapons and more pieces of hardware they could afford to lose.

The initial goal called for “multiple thousands” of autonomous systems within 18 to 24 months. Replicator concentrated first on All-Domain Attritable Autonomous systems, or ADA2. Hicks had a better description: “small, smart, cheap, and many.”

The Pentagon struggled with the last word. By the original August 2025 deadline it had fielded hundreds of systems rather than thousands, although thousands more were under contract and moving through production. Producing autonomous weapons by the thousand requires a procurement system and an industrial base built for that pace.

The Acquisition Clock

A major weapons program can spend more than a decade getting to its first operational capability. That timetable can make sense for a nuclear submarine or bomber expected to serve for decades. Autonomous systems carry processors, sensors and software that can move through several generations while a traditional program is still defining requirements and selecting contractors.

Anduril built much of its business around avoiding that sequence. The company identifies military problems, spends private capital developing hardware and brings products to the government after much of the engineering work has already happened. The Pentagon still has to test the machines, integrate them with existing forces and decide whether to buy them. Anduril and its investors have simply taken on more of the risk before the government writes the first large check.

This doesn't offer a universal model for defense procurement. Nobody is going to bootstrap a nuclear submarine program in a warehouse with venture capital. Submarines, stealth bombers and intercontinental missiles require specialized infrastructure, enormous engineering organizations and years of systems integration. Autonomous vehicles occupy friendlier ground for a company willing to build first and sell later.

Dive-LD followed that path. The Navy's Advanced Undersea Systems program worked with the Defense Innovation Unit to find commercial approaches to large-displacement unmanned underwater vehicles, eventually awarding prototype agreements to Anduril, Kongsberg Discovery and Oceaneering.

That puts real machines in the water before the government settles on a design for the next decade. Sailors can use them, discover what breaks and find shortcomings that don't appear in a presentation. Designs that work can move toward production while their electronics still belong to the current generation.

For autonomous systems, procurement time therefore ends up inside the performance calculation. A drone doesn't become useless because its processor is three years old, but the advantage of buying commercial technology erodes quickly if government purchasing takes long enough for the commercial market to move several generations ahead.

Getting to “Many”

Fast contracting doesn't create an attritable force if the factory can only build prototypes and small production runs.

Anduril announced a Rhode Island facility in 2024 designed to produce as many as 200 autonomous underwater vehicles each year. Arsenal-1, its much larger project in Ohio, calls for more than $900 million in planned investment and five million square feet of manufacturing space. The company says almost 90 percent of its products can be built primarily from commercially available components and materials.

Those numbers matter more to the attritable concept than the terminology does. Losing a $2.5 million vehicle looks manageable if another can come off the line behind it. The calculation changes when a specialized component has a two-year lead time or the entire fleet consists of a few dozen machines that the factory can't quickly replace.

For much of the post-Cold War period, the United States could lean heavily on technological superiority. A relatively small number of exceptionally capable weapons could accomplish an extraordinary amount, and the ability to manufacture replacements quickly didn't always sit at the center of force planning.

A war against another industrial power would put the production line back in the fight. A missile magazine that empties faster than factories can refill it has a hard limit. So does a drone program built around components that can't be supplied in volume. An autonomous vehicle that works beautifully in demonstrations but remains stuck between prototype and production never becomes mass.

Anduril has also tried to make the hardware less dependent on a single family of vehicles. Its Lattice software connects Anduril equipment and third-party systems, combines sensor information and gives operators a way to supervise autonomous systems across air, land and sea. If the software can coordinate a changing mix of platforms, commanders don't have to preserve one specific machine simply because the rest of the architecture depends on it.

Replicator took a similar approach at the department level. Rather than create one enormous autonomy program, spend years writing requirements and then choose a contractor to build it, the Pentagon looked for systems already emerging from the military services and private companies and accelerated the ones mature enough to use.

The first publicly disclosed tranche included AeroVironment's Switchblade 600 loitering munition along with maritime and aerial systems. Later tranches added more platforms and software for coordinating autonomous vehicles under contested conditions. The Pentagon considered hundreds of commercial companies; dozens received contracts, and 75 percent qualified as nontraditional defense contractors.

In June, the Air Force selected Anduril and General Atomics to produce its first Collaborative Combat Aircraft, autonomous aircraft intended to fly alongside crewed fighters. Anduril is no longer supplying only experimental systems at the edge of the force. It has started competing inside major aircraft programs.

A lost Dive-LD is easier to replace in that world because there are more factories, more suppliers and more machines. It also means the military will put many more machines into situations where replacement becomes necessary.

What Iran Gets

The logic works well enough from the commander's side. A $2.5 million unmanned vehicle can perform useful work in waters where sending sailors carries a much larger consequence. If the fleet owns enough of them, losing one doesn't derail the mission or gut the inventory.

Capture adds a cost that doesn't show up in the unit price.

Software can be protected. Encryption, secure boot, compartmentalized mission software, protected cryptographic keys and automatic zeroization can make a captured computer far less useful. Particularly sensitive sensors can stay ashore when the mission doesn't require them, and modular payloads limit what any single captured vehicle reveals about the rest of the fleet.

Iran doesn't need to defeat the encryption to learn from the machine in front of it. Engineers can measure the hull, cut it open and trace the propulsion system. They can examine batteries, cooling, seals, pressure management and internal support structures. They can see what Anduril considered worth spending weight on, where it accepted commercial components and how much volume each function consumed.

The value of that information is easy to underrate because none of it sounds exotic. Integration is often the hard part. Buying a battery, motor, processor or sensor doesn't tell an engineer how to package those components into a vehicle that can dive thousands of meters, operate for days and leave enough internal space for useful military payloads. An intact vehicle answers some of those questions directly.

Iran has tried this before. After recovering an American RQ-170 Sentinel in 2011, Tehran spent years displaying aircraft it said drew from the captured design. Analysts still dispute how much later Iranian hardware actually incorporated American technology from the RQ-170. Iran's effort to study and reproduce elements of the aircraft isn't seriously in doubt.

Autonomous warfare creates many more opportunities for that kind of exploitation because the machines have to operate where they can be lost. The same force design that makes commanders more comfortable risking them also produces more crashes, breakdowns, navigation errors and recoveries by whoever happens to reach the wreck first.

Designing for Capture

Engineers can make the next captured vehicle less useful than the last one. They can protect keys and storage, isolate mission software, add anti-tamper systems and move sensitive capabilities into interchangeable payloads. Mission planners can leave a particular sensor behind when the intelligence risk exceeds the value it would add.

Every protective measure has to fit inside the same hull. Anti-tamper equipment uses weight and space. Additional security costs money and engineering time. Better sensors make the vehicle more useful to the United States and more interesting to whoever recovers it. More onboard computing helps the vehicle navigate and make decisions after communications fail or someone starts jamming them, which also puts more capable computing hardware inside a machine that may someday sit in a foreign laboratory.

This is where the economics of attrition start imposing their own discipline. Add enough specialized protection and exquisite hardware and the cheap autonomous vehicle starts moving toward the cost and complexity that the concept was meant to escape.

A force with a few extraordinarily expensive systems can spend extraordinary amounts protecting each one. A force with thousands of autonomous vehicles has to accept more mundane failures. Motors break. Communications disappear. Vehicles run aground, drift, surface in the wrong place or never return.

According to the Pentagon, this Dive-LD drifted for more than a day before Iran recovered it. As the autonomous fleet grows, the Navy and the rest of the military will have to assume that some lost vehicles won't remain lost for long. The practical design question is how much an adversary can learn when one of them ends up on a workbench.

Otis came ashore a few hours later with estimated winds of 165 mph. Its wind speed had increased by roughly 110 mph in 24 hours, one of the fastest periods of hurricane intensification ever observed in the Western Hemisphere. Forecast models had badly underestimated what was about to happen, leaving more than a million people in and around Acapulco with very little time to prepare.

WeatherNext 3, the global AI weather model released by Google DeepMind and Google Research on September 3, would not necessarily have predicted Otis. Rapidly intensifying tropical cyclones remain difficult, and Google makes no such claim. The model does attack one of the problems exposed by storms that change quickly: a forecast can already be hours behind the atmosphere by the time it starts running.

WeatherNext 3 can incorporate fresh satellite observations and launch a new short-range forecast every hour. Google tested those hourly runs against the same model operating on the more traditional six-hour schedule; the fresher forecasts produced roughly two to three additional hours of useful precipitation lead time.

A hurricane warning has plenty to do in three hours. Forecast information has to reach emergency managers, become an official warning, move through television stations and phones, get families onto roads, bring fishing boats back toward port and give hospitals and utilities time to move people and equipment. Storms do not wait for any of it.

Six Hours

The starting point for a modern weather forecast is an enormous reconstruction project.

Satellites measure clouds, temperature and water vapor from orbit. Weather balloons rise through the atmosphere carrying instruments. Commercial aircraft collect temperature and wind data along their routes. Ships and buoys cover parts of the oceans; radar watches precipitation move across land; thousands of surface stations record conditions at particular points on the ground.

The European Centre for Medium-Range Weather Forecasts, or ECMWF, pulls observations like these together into an analysis, its best estimate of the state of the atmosphere at a particular time. Traditional numerical models then divide that atmosphere into a three-dimensional grid and repeatedly solve equations describing the movement of air, heat, moisture and pressure.

The analysis is enormously valuable, but assembling it takes time. Major global forecast cycles have traditionally run every six hours, and some of the observations feeding them are already several hours old by the time the model begins. A thunderstorm can organize, a front can move or a tropical system can strengthen considerably during the gap.

Lewis Fry Richardson ran into the timing problem more than a century ago. The British mathematician attempted one of the first numerical weather forecasts by working through the equations by hand; calculating six hours of weather took several weeks. Richardson later imagined solving the problem with a giant “forecast factory” staffed by tens of thousands of human calculators working on different pieces of the atmosphere at the same time.

Computers eventually built the factory. AI has made it much faster.

Models such as WeatherNext learn from decades of atmospheric data how one weather state tends to evolve into another. They can generate a global forecast in minutes without repeatedly solving the full set of physical equations used by a traditional model. WeatherNext 3 still starts with ECMWF analysis data, so the satellites, numerical models, observations and meteorological infrastructure underneath conventional forecasting remain very much in the picture.

The new model adds another source: hourly mosaics from geostationary weather satellites.

Every Hour

Geostationary satellites orbit at the same rate Earth rotates, leaving them fixed over roughly the same portion of the planet. That makes them particularly useful for watching weather develop. Clouds build and dissipate underneath them; water vapor shifts; temperatures change; tropical convection that was scattered an hour ago can begin organizing around a circulation.

WeatherNext 3 uses those observations to update the atmospheric state between the main forecast cycles instead of simply waiting for another complete analysis. The newest satellite information can be less than an hour old.

Its four primary daily runs extend 15 days into the future. The hourly runs between them extend 48 hours, where the value of fresher observations is greatest. Each ensemble generates 64 plausible futures rather than a single deterministic forecast, giving forecasters a range of possible outcomes as small differences in the atmosphere compound.

Surface forecasting has changed too. WeatherNext 3 produces many variables at roughly 10-kilometer resolution, while temperature and humidity forecasts trained directly against weather stations operate at about 5 kilometers and account for local geography and elevation. Its precipitation models learn from satellite and radar-derived measurements. The system also forecasts tropical-cyclone characteristics, cloud cover, solar radiation and winds around turbine height, the sort of weather that determines how much electricity a wind or solar fleet may produce the following day.

Google found the largest advantage from the hourly satellite updates in precipitation. Compared with forcing WeatherNext 3 to run only on the six-hour schedule, the hourly system pushed useful precipitation skill roughly two to three hours farther forward.

Three hours will not look especially dramatic on a global weather map. It looks different on a radar screen when a line of thunderstorms is building over Dallas or a tropical cyclone is beginning to reorganize offshore.

Five Percent

Across full-year 2024 testing, Google reports that WeatherNext 3 improved roughly 5% over WeatherNext 2 on many upper-atmosphere variables. In forecast time, the company says that amounts to around six additional hours before WeatherNext 3 degrades to the same level of skill.

The surface results are larger. Against weather stations withheld from training, Google reports short-range temperature improvements reaching 30% over WeatherNext 2 and 40% over ECMWF's physics-based ensemble. Its specialized precipitation forecasts reduced error by as much as 60% against NASA's IMERG satellite dataset, 30% against high-quality U.S. radar observations and 10% against rain gauges at early lead times.

Those results come from Google. Brightband's Operational WeatherBench provides a useful second set of eyes by running live forecasts from WeatherNext 3 against ECMWF's traditional IFS and AI-based AIFS systems, NOAA's GFS and GEFS, Microsoft's Aurora, Nvidia's Atlas and other models. WeatherNext 3 is performing near the front across many of the conventional atmospheric variables Brightband tracks.

The specialized precipitation forecasts are harder to compare independently because they target different observational products and resolutions from the models on the existing precipitation leaderboard. The large radar and satellite improvements remain primarily Google-reported results.

Google also compared WeatherNext 3 with ECMWF's AIFS ensemble during July and August and reported an average advantage of roughly 10% across upper-atmosphere variables during the first forecast week. The comparison covered only six weeks, and the researchers caution against putting too much weight on small differences over such a short period.

ECMWF and NOAA are moving in the same direction. ECMWF already operates its own AI forecasting system. NOAA has put an AI global model, an AI ensemble and a hybrid AI-plus-physics ensemble into operations; NOAA says the AI global model can use as much as 99.7% less computing than its traditional counterpart, while early testing of the AI ensemble showed roughly 18 to 24 additional hours of useful forecast skill over the existing Global Ensemble Forecast System.

The weather agencies that spent decades building numerical forecasting are now deploying AI alongside it.

The Forecast Factory

AI does not make the expensive parts of meteorology disappear. Satellites still have to be built and launched. Radar networks have to operate. Aircraft, balloons, ships, buoys and surface stations have to keep collecting observations. NOAA, ECMWF and other meteorological agencies maintain the archives, models and scientific infrastructure that WeatherNext itself uses.

Another forecast, however, has become much cheaper to produce.

NOAA's reported 99.7% reduction in computing requirements gives some sense of the scale. Traditional global numerical forecasting has historically required the sort of supercomputing capacity available to major governments and large research institutions. Once an AI model is trained, it can generate additional global forecasts at a fraction of that computational cost.

Richardson's imaginary forecast factory no longer needs 64,000 calculators, and increasingly it does not need one of the world's largest supercomputers either.

That changes how far sophisticated forecasting can travel. A country without ECMWF-scale computing can still receive output from an advanced global model. Google is already feeding WeatherNext forecasts into Search, Maps and Gemini, while businesses and governments can access the underlying data through Google Cloud and Earth Engine.

The observations still have to come from somewhere. A satellite gap over Africa or poor surface coverage in a developing country does not disappear because the forecast model uses AI. But the expensive calculation that comes after those observations no longer has to be rebuilt separately in every country that wants the result.

Between 1970 and 2021, weather, climate and water-related disasters caused more than 2 million deaths and $4.3 trillion in reported losses worldwide. Disaster mortality is nearly six times higher in countries with limited early-warning coverage than in countries with strong systems. The World Meteorological Organization estimates that 24 hours of warning before a dangerous event can reduce the resulting damage by as much as 30%.

Most WeatherNext forecasts will have considerably less dramatic consequences. Farmers decide when to plant or harvest. Airlines move aircraft around thunderstorms. Construction crews decide whether concrete can be poured. Utilities estimate the next day's electricity demand and renewable generation. Ships choose departure times and routes. A 2024 national study estimated the annual value of weather forecast information to the American public at roughly $102 billion.

The better forecast is useful long before anybody evacuates.

After the Forecast

Acapulco also shows where the model stops helping.

By 10 p.m. on October 24, 2023, the National Hurricane Center knew Otis had become a catastrophic hurricane and told the public exactly that. The storm was already about 55 miles offshore.

From there, meteorology had to hand the problem to everyone else.

Warnings had to reach people who were asleep or away from television. Local officials had to make decisions. Hotels and hospitals had to protect guests and patients. Boats still offshore had to get back. Families without cars needed another way to move. Roads had to remain passable long enough for any of it to work.

Google explicitly warns that WeatherNext 3 is not an official warning system. A better forecast does not fix communications, transportation, shelter capacity or local government, and three extra hours are worthless if nobody acts on them.

Otis strengthened by roughly 110 mph in a day. WeatherNext 3 can now take another look at the atmosphere every hour.

Sometimes that extra look buys two or three hours.

Then everything on the ground has to work.

Key Findings

• GLP-1 drugs are changing metabolic medicine.: Semaglutide has reduced major cardiovascular events, while investigational retatrutide has produced Phase III weight loss approaching 30%. Researchers are already testing whether muscle can be preserved as weight falls.
• CRISPR gene therapy is already medicine.: Gene-edited stem cells are approved for sickle cell disease, shifting the challenge toward manufacturing, financing and delivery.
• Cancer treatment is becoming individualized.: Personalized mRNA cancer treatment, genomics and biomarkers are increasingly linking diagnosis, patient selection and therapy.
• CAR-T is moving beyond cancer.: Early human research is testing whether severe autoimmune diseases can sometimes be pushed into durable remission rather than managed indefinitely with immunosuppressive drugs.
• Healthcare access is becoming the bottleneck.: Prior authorization, insurance design, drug pricing, manufacturing capacity and specialist treatment infrastructure increasingly determine whether effective medicine actually reaches an eligible patient.

An investigational obesity drug has now produced average weight loss approaching 30% in Phase III. CRISPR-edited stem cells are approved medicine. A cancer treatment manufactured around the mutational fingerprint of an individual patient has produced a positive Phase III result. Researchers are testing whether autoimmune diseases can be pushed into durable drug-free remission, whether muscle can be preserved while fat disappears, and whether artificial intelligence can find useful drug targets and molecules faster than conventional discovery.

These advances come from different technologies and operate against different diseases, but they are arriving at roughly the same time. Medicine is gaining more ways to intervene upstream in disease, target the biology of an individual patient, or alter the biological process itself.

GLP-1 Drugs and the New Metabolic Medicine

Obesity medicine makes the scale easiest to see. Semaglutide produced 14.9% mean weight loss after 68 weeks in the 1,961-patient STEP 1 trial. Tirzepatide moved mean reductions beyond 20% in SURMOUNT-1. Lilly reported in May 2026 that participants receiving 12 milligrams of the investigational drug retatrutide in Phase III lost an average of 70.3 pounds, or 28.3% of body weight, over 80 weeks; 45.3% lost at least 30% of their starting weight. Among participants with baseline BMI of at least 35 who continued into a 104-week extension, average loss reached 30.3%.

Obesity sits upstream from an enormous amount of chronic disease. Measured national data show obesity in 40.3% of American adults and overweight in another 32.1%. CDC estimates that 40.1 million Americans had diabetes in 2023 and 115.2 million adults had prediabetes. SELECT has already shown that semaglutide can reduce the combined risk of cardiovascular death, nonfatal myocardial infarction or nonfatal stroke by 20% in adults with overweight or obesity and established cardiovascular disease but without diabetes.

Researchers are now trying to manipulate what kind of tissue disappears during aggressive weight loss. In a randomized Phase II trial published in Nature Medicine in June 2026, the anti-myostatin antibody apitegromab preserved 54.9% of the lean mass that otherwise would have been lost during tirzepatide treatment. Regeneron is pursuing a similar strategy with semaglutide and its own muscle-preservation combinations. Weight loss is beginning to become a problem of body composition rather than a single number on a scale.

Scale is developing alongside efficacy. FDA approved orforglipron, marketed as Foundayo, in April 2026 as a once-daily oral small-molecule GLP-1 receptor agonist for chronic weight management. Its pivotal Phase III trial produced 11.1% mean weight reduction at 72 weeks at the highest studied dose. That is less than the best injectable therapies, but a pill that is easier to manufacture, distribute and take can serve a very different market. Tens of millions of patients are unlikely to be treated with one drug at one dose through one delivery system.

Personalized mRNA Cancer Treatment and Precision Medicine

Cancer medicine is moving in the opposite direction, toward increasingly individualized treatment. On August 19, 2026, Merck and Moderna announced that intismeran autogene, an mRNA therapy manufactured around the mutational fingerprint of each patient's tumor, met both its recurrence-free-survival and distant-metastasis-free-survival endpoints in a randomized Phase III melanoma trial involving 1,137 patients. Each product can encode as many as 34 neoantigens selected from that patient's cancer.

Nine days later, BioNTech terminated a Phase II trial of another individualized mRNA treatment in resected colorectal cancer after an independent monitoring board concluded that continued treatment was unlikely to change the efficacy outcome and identified a numerical imbalance in overall survival.

Those two results belong together. Personalized cancer medicine has advanced far enough that different products, combinations and tumor types are succeeding and failing in meaningful human trials. Patient-specific pharmaceutical manufacturing has moved beyond a laboratory demonstration.

CRISPR Gene Therapy Is Already Medicine

Cell and gene therapy have crossed an even more basic threshold. Casgevy became the first FDA-approved CRISPR-based treatment in 2023, and in July 2026 its indication was expanded to patients age two and older with sickle cell disease with recurrent vaso-occlusive crises or transfusion-dependent beta thalassemia. The treatment edits a patient's own blood-forming stem cells to increase fetal hemoglobin production.

The science is extraordinary. The delivery is difficult. Patients require cell collection, specialized manufacturing, myeloablative conditioning, hospitalization and long-term follow-up. Casgevy entered the market at roughly $2.2 million per patient, while Lyfgenia, a different gene therapy for sickle cell disease, was listed at roughly $3.1 million.

The financing problem was foreseeable enough that CMS had already selected a Medicaid-focused Cell and Gene Therapy Access Model for testing in February 2023, before either therapy reached the U.S. market. The model now negotiates outcomes-based arrangements on behalf of participating Medicaid programs. Thirty-two states, the District of Columbia and Puerto Rico are participating, representing approximately 84% of Medicaid beneficiaries with sickle cell disease. The medicine is forcing insurance accounting to operate on a different clock.

CAR-T and Autoimmune Disease

The same pressure is beginning to appear elsewhere. A 2024 New England Journal of Medicine case series followed 15 patients with severe refractory autoimmune disease after CD19 CAR-T treatment. All eight patients with lupus met the study's remission criterion, while patients with inflammatory myositis and systemic sclerosis achieved major clinical improvements; immunosuppressive therapy was discontinued. Fifteen highly selected patients cannot establish a standard of care, but they are enough to ask a different question: whether selected immune diseases can sometimes be reset deeply enough to produce durable remission rather than managed indefinitely with recurring immunosuppression.

AI Drug Discovery Reaches Phase III

Artificial intelligence belongs in this story more narrowly than the hype sometimes suggests. FDA's Center for Drug Evaluation and Research had experience with more than 500 submissions containing AI components from 2016 through 2023. Rentosertib, whose TNIK target and small-molecule design were developed through an AI-directed platform, produced a preliminary signal in a 71-patient randomized Phase IIa study in idiopathic pulmonary fibrosis and has since advanced into Phase III; the first patient in the GENESIS-IPF-3 trial was dosed on September 10, 2026. Computational target discovery and molecule design have moved from software demonstrations into late-stage human drug development while remaining subject to ordinary pharmaceutical uncertainty.

Peptides, BPC-157 and the Gray Market

The peptide gray market shows another side of therapeutic abundance: people are already experimenting before formal medicine has settled the evidence. BPC-157, TB-500, MOTS-C and other compounds are being marketed for recovery, metabolism, body composition and longevity despite limited human data and uneven quality control.

Demand is not evidence that these compounds work. It is evidence that people are willing to seek pharmacological tools outside the conventional system, creating both a safety problem and a research signal.

Prior Authorization and Healthcare Access

The scientific advance is arriving inside an access system already struggling with ordinary medicine. A KFF analysis of CMS reporting data found that Medicare Advantage insurers made approximately 52.8 million prior-authorization determinations in 2024 and denied roughly 4.1 million requests in whole or in part. Only 11.5% of denials were appealed, but 80.7% of appealed denials were overturned. The 2025 AMA Prior Authorization Physician Survey found an average of 40 prior authorizations per physician each week, consuming roughly 13 hours of physician and staff time.

CMS is trying to modernize that process. Its 2024 rule imposed new electronic and timing standards for many medical services; in April 2026, the agency proposed separate pharmaceutical requirements that would impose electronic standards and, in several federal programs, maximum decision periods of 24 hours for urgent requests and 72 hours for standard requests beginning in 2027.

American laboratories can sequence a tumor and manufacture an individualized therapy around its mutations while the country is still building uniform electronic rules for how an insurer approves the prescription.

The financing problem runs deeper. Beginning July 1, 2026, Medicare's GLP-1 Bridge began providing qualifying Part D beneficiaries certain obesity medicines for a $50 monthly copayment through a national demonstration operating outside ordinary Part D payment. Private employers are moving in the other direction in some cases. Forty-nine percent of large employers covered GLP-1 drugs for weight loss in 2025, but 6% dropped coverage in 2026, another 5% planned or were considering doing so for 2027, and 27% tightened or planned to tighten utilization controls.

A therapy can work and still be difficult to finance. The payer buys the drug today; the avoided heart attack, dialysis episode, disability claim or cancer recurrence may come years later, perhaps after the patient has moved to another insurer.

For much of modern medicine, finding something that worked was the dominant constraint. That constraint is beginning to loosen in several important fields. The pressure is moving downstream into manufacturing, diagnostics, financing, insurance rules, treatment capacity and the time required to get an eligible patient through all of them.

The United States does not need to approve every drug, pay every price or eliminate scrutiny. Failed drugs should fail. Weak evidence should remain weak evidence. Clinically interchangeable products should compete aggressively on price.

Once strong evidence exists and a patient clearly qualifies, however, the system has to be capable of acting on that information.

Therapeutic abundance will matter only if medicine can actually reach the people it was invented for.

View or download full report (PDF)

In less than a week, the Houthis have swept down Yemen’s Red Sea coast to Bab el-Mandeb. They took Mocha and Dhubab, directly opposite Perim Island; they also captured the Hanish Islands farther north. Houthi terrorists now occupy the Yemeni coast overlooking one of the world’s busiest and most important shipping lanes.

Saudi Arabia is taking the advance seriously. Crown Prince Mohammed bin Salman reportedly asked Washington for military help against the Houthis as the front moved south. Shipowners, charterers, insurers and governments are watching the same coastline.

Despite all of this, Bab el-Mandeb remains open. Tankers and cargo ships are crossing it, Suez is operating, and Saudi crude continues to load on the Red Sea coast. The immediate problem is a sharp increase in the threat to shipping in a strait that, because of the situation in Hormuz, the world cannot afford to lose.

The Other Way Out

The reason Bab el-Mandeb matters so much more today starts about 1,200 miles to the east.

In the fourth quarter of 2025, 21.6 million barrels of crude oil and petroleum liquids crossed the Strait of Hormuz every day. By the second quarter of this year, throughput had fallen to 4.9 million barrels per day. The effective closure of the vital sea lane (we’re not going to argue over whether it was ever technically “closed”) forced the Gulf Arab oil producers to rapidly change how they got their product to market.

Saudi Arabia surged crude into its East-West Pipeline, or Petroline, which carries oil from the Kingdom’s eastern production system across the country to Yanbu, where it can load onto tankers in the Red Sea.

Aramco, the state-owned Saudi oil juggernaut, pushed Petroline to its 7 million barrel-per-day maximum and says it changed roughly eighty years of normal export patterns in eight days.

That workaround is now under pressure too. On September 10, drones struck Petroline several times in the Riyadh and Medina regions, injuring people and forcing Saudi Arabia to shut the line while crews assessed the damage. Riyadh later said the drones came from Iraq.

The timing could hardly be worse. Reuters estimates roughly 4 million barrels per day had been moving through Petroline to Yanbu, and the port may have only five to seven days of export stocks available if the pipeline stays offline. The route Saudi Arabia built to get around Hormuz is suddenly depending on stored oil at one end and a threatened strait at the other.

Saudi crude could now reach the world market through the Red Sea in much greater volumes. Oil and petroleum liquids moving through Bab el-Mandeb jumped from 5.4 million barrels per day in the fourth quarter of 2025 to 8.1 million in the second quarter of 2026. That helps explain why events along a narrow strip of Yemeni coastline suddenly matter well beyond Yemen.

Bab el-Mandeb connects the Red Sea with the Gulf of Aden and Arabian Sea. At its narrowest it is roughly eighteen miles across. Southbound ships from Suez pass through it toward the Indian Ocean; northbound traffic uses it to reach Suez. Ships that decide the transit is too dangerous have another option, but it is a costly one: turn south and go around Africa.

Eighteen Miles of Water

The Houthis, an Iranian-backed terrorist proxy and part of Tehran’s so-called “Axis of Resistance,” have been shooting at ships in the Red Sea for years, so their ability to threaten Bab el-Mandeb is hardly new. What changed this week is where they can do it from.

Their advance through Mocha and Dhubab puts Houthi forces directly along the Yemeni approaches to the strait. They have also reached Perim Island, which sits inside Bab el-Mandeb itself. Reuters reported the move based on four Yemeni government sources, and AP separately reported that the Houthis seized the island. That gives them an even better perch over traffic entering and leaving the Red Sea.

They have already demonstrated the threat from the coast. On August 12, the TIHAMAH was hit off Mocha and several mariners were killed. MARAD continues to warn ships throughout the southern Red Sea and Bab el-Mandeb about anti-ship missiles, UAVs, unmanned surface vessels, small boats and boarding attempts. Moving those capabilities closer to the strait gives the Houthis shorter lines to the water and more coastline from which to operate.

For now, despite the alarm, ships are still going through. Preliminary tracking counted 28 commodity-vessel transits on September 10, right around the recent average; 25 crossed the day before. AIS data miss some ships operating dark, but we are nowhere close to seeing owners abandon the route wholesale.

Many owners did reroute after the first round of Houthi attacks in 2023 and 2024. Sending an Arabian Sea-to-Europe tanker around the Cape of Good Hope adds roughly fifteen days; depending on the voyage, the deviation can add several thousand miles and two or three weeks at sea.

That extra distance gets expensive quickly because the ship stays employed longer, burns more bunkers and returns to the available tonnage pool later. Across enough voyages, there are simply fewer ships available when charterers need them.

Freight

The Baltic Exchange, the London institution that publishes benchmark shipping rates used by owners and charterers around the world, tracks the cost of hiring tankers on major routes. One of its newer benchmarks follows Very Large Crude Carriers, or VLCCs, moving oil from the Gulf of Oman to China. A VLCC carries roughly 2 million barrels of crude.

Since March, the quoted freight rate on that route has more than doubled, rising about 130%. By September 4, a VLCC making the voyage could earn more than $260,000 per day; a week later, Reuters put the freight charge at roughly $11.50 for every barrel aboard the ship.

For a fully loaded VLCC, that works out to roughly $23 million just to move the cargo.

Tankers are spending longer on voyages, getting pulled onto different routes and becoming harder to find on short notice. A ship that spends another two weeks going around Africa is unavailable for its next cargo during those two weeks. And when you divert enough ships, the pool of available ships starts thinning quickly.

Owners willing to transit the Red Sea will price in war-risk insurance coverage and the danger to the ship and crew; others will take the Cape or decline the cargo. The strait can, therefore, become very expensive long before traffic stops moving through it.

Bunkers

Going around Africa also means burning a lot more fuel. In shipping, that fuel is called bunker fuel, or simply “bunkers.”

When ships diverted away from the Red Sea in 2024, Singapore, the world’s largest bunkering port and a major refueling stop on Asia-Europe routes, sold a record 54.9 million tonnes of bunker fuel, up 6% from the year before. Singapore’s maritime authority attributed part of that increase to the longer voyages around the Cape of Good Hope.

Those extra miles add up quickly. A tanker that spends another two weeks at sea keeps burning fuel every day, and hundreds of ships making the same decision can push bunker demand materially higher even if the amount of cargo moving around the world barely changes.

Fuel is already getting more expensive in the region. Fujairah, one of the world’s largest bunkering ports and located just outside the Strait of Hormuz, saw the price of very-low-sulfur fuel oil rise from about $852 per metric ton on September 4 to $959 a week later, an increase of roughly 13% in seven days. Compared with late August, the price is up about 21%. Supplies across the major marine-fuel grades are also extremely tight.

More ships diverting around the Cape would add demand on top of that. They also have to refuel somewhere, which shifts business toward ports positioned along the longer route. Singapore saw exactly that during the last major Red Sea disruption.

For shipowners, the math is straightforward: take the shorter Red Sea route and accept the security and insurance costs, or sail thousands of miles farther and buy a lot more fuel.

Running Bab el-Mandeb

MARAD, the U.S. Maritime Administration, now advises U.S.-flagged vessels in the region to turn off their Automatic Identification System, or AIS, when safe navigation allows it, vary course and speed to make targeting harder, and refuse Houthi demands for voyage information or AIS activation. Ships also maintain contact with U.S. Naval Forces Central Command while operating in the area.

AIS is the system ships use to broadcast their position, course and speed to other vessels and authorities. Normally, that makes navigation safer, but off Yemen, it can also tell an attacker exactly where a ship is and where it is headed.

That puts masters in an uncomfortable position. Do you keep your AIS on and make yourself easier to track, or turn it off and lose one of the basic tools used to keep commercial traffic separated in a crowded sea lane?

Insurance, for obvious reasons, has moved with the threat. War-risk insurers have expanded the areas they consider exposed to conflict to include Bab el-Mandeb, the Yemeni and Saudi Red Sea coasts, and a much larger section of the Gulf of Aden and Indian Ocean. Owners can still insure the voyage, but that risk now gets priced directly into the cost of sending a ship through. And that’s a price that’s going to be pumped to consumers along with the oil shipments.

Merchant ships move through these waters around the clock, and the threat can come from shore-based missiles, drones, unmanned surface vessels and small boats. No navy, despite the best efforts of a fairly large coalition of nations, can put a warship next to every tanker or contain every launch site along the Yemeni coast.

That leaves the final decision with the owner and master of each ship. They have to decide whether the shorter Red Sea route is worth the security risk, insurance cost and operating restrictions, or whether another two weeks around Africa makes more sense. And they have to make that decision with the Houthis perched like vultures over Bab el-Mandeb.

Researchers led by the Howard Hughes Medical Institute’s Janelia Research Campus, working with Google Research and other collaborators, recently published the finished connectome of an entire male fruit fly central nervous system. The map contains 166,691 neurons, more than 125 million synaptic connections and 11,691 identified cell types, stretching from the brain and optic lobes through the ventral nerve cord. For the first time in a male fly, researchers can trace an essentially complete structural path from sensory inputs toward the neurons that ultimately control movement.

The wiring diagram is already useful for more than anatomy. In 2024, another team used measured connectivity from the fruit fly visual system to construct a model of more than 45,000 neurons, then used deep learning to infer properties of neurons and synapses that hadn’t been measured directly. When researchers tested the model against decades of neuroscience experiments, it reproduced neural responses measured across 26 previous studies.

A separate team working with Google DeepMind has built an anatomically detailed virtual fruit fly body inside a physics simulator. The model, based on an adult female fly, can walk and fly with realistic movement; its nervous system is controlled by an artificial neural network trained on real fly behavior rather than a biological reconstruction of the fly brain.

The new male connectome hasn’t been integrated with that anatomically detailed body model into a validated whole-animal simulation, but researchers are already trying.

Biologists, like nature itself, have, until now, always been constrained by biology, itself. Change a gene, breed the animal, administer the compound, grow the cells, wait for the response and measure what happened. Every additional test consumes animals, reagents, equipment, technicians and time; researchers therefore spend enormous effort deciding which experiments are worth running before anything reaches the bench.

A laboratory that can physically test a limited number of ideas can now inspect far more of them, first, eliminate weak candidates and concentrate its equipment and people on experiments more likely to produce useful information. Even better, the model doesn’t have to reproduce an organism perfectly to save considerable time and money; it only has to become good enough at identifying dead ends that researchers stop testing all of them individually in the “real world.”

The Arc Institute’s Virtual Cell Challenge is pushing directly at that problem. Its 2026 competition asks models to predict how cells respond to genetic changes in biological contexts they haven’t previously seen perturbed. That is much closer to the problem confronting an actual laboratory, where the possible experiments vastly outnumber the experiments anyone has the time or money to perform.

One living fly can still provide the observation that settles a question, except researchers may reach that experiment after a computer has already discarded thousands of less promising ones. The physical experiment becomes more targeted, and the data generated by it can feed the next generation of models.

That makes high-quality experimental data more valuable as well. The new connectome required years of electron microscopy, reconstruction and validation; virtual-cell models depend on similarly expensive biological datasets. An experiment that once answered a specific research question can also produce data used across thousands of later computational experiments, allowing the value of the original laboratory work to extend well beyond the study that paid for it.

Eventually, this runs into another constraint: laboratories still have to test what the models produce. If computation can generate useful hypotheses faster than scientists can validate them, automated laboratories, robotics, better assays, organoids and faster experimental systems become more important. The amount of physical biology may not fall in every field; researchers may instead run better-selected experiments at much greater scale.

The FDA is expanding its use of New Approach Methodologies, including computational models and human-derived systems, when they can provide reliable evidence without automatically requiring an animal study. In March, the agency released draft guidance on validating and incorporating these methods into drug development; NIH committed more than $150 million this year to human-based research methods and created a dedicated office in June to develop and scale computational tools, 3D human tissue systems and other alternatives to animal models.

Whole-organism physiology is still hard to reproduce on a computer, so some questions will continue to require living animals. What can change much faster is how many experiments have to reach an animal in the first place. A drug company choosing among 10,000 compounds has a very different animal-testing requirement if computational screening can eliminate 9,500 of them before the remaining candidates ever leave the computer.

Animal experimentation has long served two purposes at once: testing serious hypotheses and searching for the hypotheses worth taking seriously. Computation is beginning to take over more of the search.

There are obvious economic reasons to accelerate that process. Animals cost money, experiments take time and failed candidates consume laboratory capacity that could have been used somewhere else. There are also obvious ethical benefits if fewer animals are needed to produce the same or better scientific results.

We’re unlikely to reach the Lab Rat Utopia anytime soon. But a growing share of the work once performed by breeding an animal, administering something to it and waiting to see what happens can move upstream into computation, reserving living systems for the questions that survive the screening process.

The fruit fly has spent more than a century helping scientists search for answers one generation and one experiment at a time. We’re getting better at searching before the experiment begins; for a billion future fruit flies, that could make all the difference.

Conflict expands that gray market because armed groups can seize mines, tax roads and control who moves material; government officials lose access to production areas, customs routes become harder to police and traceability programs withdraw. The trade keeps moving through whoever is willing to buy under those conditions, while companies that need clean title, documented origin, sanctions clearance and lenders willing to approve the transaction can be pushed out of it entirely.

That difference matters in eastern Congo because the minerals don't stop coming out of the ground when the state loses control. Production continues, and the companies and traders able to operate through uncertain ownership, weak enforcement and questionable paperwork gain access to material that more heavily regulated competitors may be unable to touch.

Rubaya shows how this works. The mining district in North Kivu supplies roughly 15% of the world's coltan, the ore used to produce tantalum for electronics, aerospace systems and other advanced manufacturing. M23 seized Rubaya in April 2024 and built a parallel administration around the mines, traders, transport routes and taxes.

Reuters reported that minerals were being trucked toward Rwanda, sometimes at night, while United Nations investigators found that Congolese coltan had been mixed with Rwandan production before export. Once ore crosses the border and is combined with material from other sources, determining its origin depends heavily on records created and enforced along the route; those records become considerably less useful when armed groups and smugglers control parts of the system.

Rubaya had already been part of illicit trading networks before M23 took control. The U.S. Treasury sanctioned the Congolese mining company CDMC after finding that it sold minerals sourced and smuggled from areas then controlled by the PARECO-FF armed group to East Rise and Star Dragon, two Hong Kong companies. Treasury has said minerals from eastern Congo are frequently smuggled through Rwanda before reaching major refining and processing countries such as China, and its sanctions have identified companies in Congo and China participating in the international trade in conflict-linked material.

Gold has moved through a similar system. In June 2026, Treasury sanctioned Gasabo Gold and said Rwandan forces and M23 personnel had earlier in the year moved at least 60 kilograms of gold from occupied areas of South Kivu through Rwanda and into the Kigali refinery. Moving production through those channels makes its original source progressively harder to establish as it passes through traders, borders and processors.

Western companies face practical limits on what they can buy under those conditions. A concession with disputed title, an armed group collecting money along the transport route, questionable customs documents or an uncertain chain of custody can fail lender review, create insurance problems and expose a company to sanctions or anti-corruption liability. Reuters reported this year that American and other Western companies pursuing projects in Congo were being slowed by anti-bribery reviews, title verification, community-impact requirements and other compliance obligations.

Chinese firms enter the same market with a much larger existing position. They have spent years building mines, processing relationships, infrastructure and commercial networks across Congo; Reuters estimates that Chinese firms control more than 70% of the country's copper-cobalt and other rare-mineral assets. Treasury's findings concerning conflict-linked trade involve particular companies and transactions, while China's larger mining position includes enormous conventional industrial operations, but the structure of the market still matters: when uncertainty removes potential competitors from a transaction, the companies and networks that remain have fewer rivals for the material.

Traceability programs are meant to preserve provenance as ore moves from mine to export. ITSCI uses mine-level tagging and supply-chain records for tin, tantalum and tungsten, but suspended operations in areas under M23 control. ITSCI has acknowledged that tags can be misused and origin can be misrepresented while disputing broader claims that its system has failed; UN investigators have nevertheless documented Rubaya minerals crossing into Rwanda and being mixed with Rwandan production. Once material from several sources enters the same stream, the paperwork has to establish distinctions that the physical shipment no longer can.

This is one reason Washington has connected the security settlement in eastern Congo to its minerals strategy. The United States brokered the June 2025 agreement between Congo and Rwanda, followed by the Washington Accords signed by Presidents Félix Tshisekedi and Paul Kagame in December. M23 remained outside the state-to-state agreement and has negotiated separately with Kinshasa through the Qatar-backed Doha process; ceasefire monitors made their first field deployment under that mechanism on August 24, although M23 continues to control substantial territory in eastern Congo.

Treasury has said the Washington framework is intended to produce a “fully licit and transparent regional minerals sector,” and the economic agreements surrounding it tie investment to better control of mineral production and trade. Its August 2025 sanctions action was explicit about the American interest involved: conflict minerals were financing armed groups while preventing law-abiding companies from investing and restricting U.S. and allied access to materials needed for national defense.

The U.S.-Congo minerals agreement attacks the commercial problem from the other direction. Congo agreed to provide U.S. investors rights of first offer on designated strategic assets and certain mineral offtake while the governments pursue formalization of artisanal mining, traceability programs and technical assistance. A mine with recognized ownership, reliable records and a secure route to market can attract financing and buyers that would have rejected the same deposit under armed control.

Bisie has already shown how closely security and commercial access can be linked. The major tin mine restarted operations after U.S. diplomatic pressure helped reduce fighting around it, according to Reuters; the threat of renewed clashes remains, but a mine that can keep workers on site and material moving has a much better chance of maintaining contracts, financing and transportation than one repeatedly evacuating because the front line has moved.

Stabilizing eastern Congo won't unwind the Chinese mining positions already established elsewhere in the country. CMOC's Tenke Fungurume and Kisanfu copper-cobalt operations are enormous industrial assets, while Zijin has a major position in Kamoa-Kakula and began exporting lithium concentrate from the disputed Manono project to China this summer. Those holdings were assembled through years of investment and aren't dependent on control of Rubaya.

The more immediate opening lies in assets that remain underdeveloped, contested or commercially difficult to reach. If improved security and traceability bring more of that production into documented supply chains, Western companies don't have to displace an existing Chinese operator to gain access; they can compete for material and projects that previously could not clear their own requirements.

Washington is also trying to build the commercial routes needed to move that material. DFC has provided $553 million for rehabilitation of the Lobito Atlantic Railway, giving copper and cobalt from Congo's southern mining belt a more efficient route through Angola to the Atlantic; DFC describes the project as part of its effort to counter Chinese dominance in critical-mineral supply chains. The Gécamines-Mercuria partnership is creating another channel, and DFC says it has already sold and started shipping approximately 100,000 tons of copper committed to the United States.

The volumes entering the American market are starting to become significant. U.S. imports of Congolese copper reached a record 53,290 metric tons in July, up from fewer than 32,000 tons during all of 2024; Congo accounted for 23.9% of U.S. copper imports that month. The surge came from the established copper industry in southern Congo rather than the conflict-affected mines of the east, but it demonstrates that American buyers can absorb substantial Congolese production when the material can move through commercial channels they are willing to use.

Eastern Congo presents the harder version of the same problem. Mine formalization, functioning traceability, secure transport routes and enforceable ownership give more buyers the ability to finance and purchase production; stronger border enforcement and sanctions also make it harder for armed groups and intermediaries to turn control of a mine or road into revenue.

Smuggling will continue wherever the money justifies the effort, and tags, customs documents and government officials can all be compromised. Those realities don't require the conclusion that formalization is futile; they explain why armed control and broken traceability have commercial value in the first place. A market in which provenance can be obscured favors participants willing or able to operate inside that opacity.

As recognized authorities regain control of mines and transport routes, that advantage begins to narrow. More production can meet the requirements imposed by lenders, insurers and downstream buyers, Kinshasa has a better chance to collect taxes and royalties, and armed groups have fewer opportunities to take revenue from the movement of the mineral.

China's position in Congo will remain formidable because much of it rests on mines, infrastructure and processing relationships built over decades. Eastern Congo presents a different contest, however, because a portion of the advantage there has come from the disorder itself; reducing that disorder allows companies previously excluded by the gray market to compete for what comes out of it.

Rubaya remains worth watching because M23 still controls the district, the Doha process remains unfinished and normal traceability has not returned. The coltan is still coming out of the ground, and somebody is still buying it.

He has some experience with biological problems yielding faster than expected. Hassabis and DeepMind researcher John Jumper won the 2024 Nobel Prize in Chemistry for AlphaFold, an AI system that predicts the three-dimensional shapes of proteins, the molecular machinery involved in almost everything a living cell does.

Protein structure helps determine how a protein behaves, what may be going wrong in disease and where a drug might interact with it. Scientists once had to determine many of those structures through painstaking experiments that could take months or years. AlphaFold has now predicted more than 200 million of them, covering nearly every protein catalogued by science, and DeepMind says more than three million researchers in 190 countries use its database.

DeepMind is beginning to do something similar with another enormous biological search space. Its new AlphaGenome Atlas contains predictions for the molecular effects of roughly nine billion possible single-letter changes in human DNA, including both protein-coding genes and the much larger portion of the genome involved in regulating how those genes behave. Researchers cannot experimentally test nine billion mutations one by one; the point is to use computation to narrow the field before deciding which ones deserve scarce laboratory time.

The pattern is becoming familiar. AlphaFold dramatically expanded the amount of protein structure scientists could inspect without first determining it experimentally. AlphaGenome is trying to push more of genetic variation through the same computational filter. Problems that once began in the laboratory increasingly begin on a computer.

DeepMind is also pushing farther into the work scientists can do before an experiment even begins. Its Co-Scientist system, built on Google’s Gemini models, can read scientific literature, propose explanations, argue against them and rank the ideas it thinks deserve further investigation.

Microbiologist José Penadés and his team at Imperial College London gave it an unresolved question involving the spread of antibiotic resistance after spending most of a decade working out the mechanism themselves. Co-Scientist returned five possible explanations in two days and put essentially the same answer the researchers had reached at the top of its list. Penadés was surprised enough to ask Google whether the system had somehow seen his unpublished work.

More interesting are the experiments where humans didn’t already know the answer. Co-Scientist proposed existing medicines that could be repurposed against acute myeloid leukemia, an aggressive blood cancer, and follow-up work found activity against leukemia cells among several of its choices. In a separate liver-fibrosis project, researchers tested AI-generated treatment ideas in human organoids, small laboratory-grown models of human tissue, and found promising anti-fibrotic effects among several of the proposed targets. The results were published in Nature in May.

A promising hypothesis is still a long way from an injectable cure, but this is where Hassabis’s ambition starts to make more sense. Parts of drug discovery that now take years could eventually take months or weeks because researchers can search a much larger space of possibilities before committing scarce laboratory time to one of them.

Sam Altman has made essentially the same argument from the other end of the pipeline. While calling for a much larger buildout of AI infrastructure, the OpenAI CEO offered a deliberately provocative example: 10 gigawatts of compute might help AI figure out how to cure cancer.

Anthropic CEO Dario Amodei, who trained as a biophysicist, has gone further, estimating that sufficiently capable AI could compress 50 to 100 years of biological progress into five to ten years.

The interesting part of Amodei’s argument is where he thinks the acceleration stops. Biology remains stubbornly physical. Cells have to grow. Experiments have to run. Equipment has to be built. Drugs eventually have to be manufactured, administered to people and watched long enough to find out whether they actually work.

DeepMind calls this the “validation bottleneck,” and it may become one of the defining infrastructure problems of AI-driven science.

Scientists already compete for time in specialized facilities. Now imagine research agents producing hundreds or thousands of plausible hypotheses faster than laboratories can investigate them. The model can move immediately from paper to paper, target to target and molecule to molecule; the next step may still require sequencing, microscopy, animal work or a technician running the same assay for several days.

AI can make the queue much longer before it makes the laboratory any faster.

Drug companies have noticed. Alnylam Pharmaceuticals agreed in June to a collaboration with Inceptive valued at as much as $2 billion, including $30 million up front, pairing more than two decades of Alnylam’s RNA-interference data with Inceptive’s biological foundation models. Boehringer Ingelheim has reached an agreement with Owkin to use AI and patient data in cancer and immunology research, joining AstraZeneca and Sanofi among the large pharmaceutical companies working with the firm.

OpenAI entered the market directly in April with GPT-Rosalind, a model built specifically for biology, drug discovery and translational medicine. Researchers can use it across genomics, protein analysis, medicinal chemistry, scientific literature and experiment planning; OpenAI says Amgen, Moderna and the Allen Institute are among the organizations working with the system.

The models are only one piece of the system. Biology also needs usable genomic and clinical data, sequencing machines, high-end imaging, robotics and laboratories capable of processing a much larger volume of experiments.

DeepMind has specifically called for investment in automated labs, where robotic equipment can carry out experiments, collect the results and feed them back into an AI system that helps decide what to try next. Instead of a model handing a scientist a list of ideas to work through over the following month, parts of the discovery loop can begin to run continuously.

That changes the infrastructure question considerably. If AI can propose experiments around the clock, eventually somebody has to build laboratories capable of keeping up with it.

The federal government already owns pieces of that system. President Trump’s Genesis Mission, launched last November and expanded in July with more than $5 billion in announced federal commitments, is linking Department of Energy computing, government datasets, national laboratories and scientific instruments through a common AI-for-science platform. Its health programs include using molecular, genomic and clinical data to search for new uses for existing drugs and applying federal supercomputing to pediatric cancer and chronic disease.

Four federal agencies are also working on autonomous laboratories that combine robotics, AI and real-time experimental analysis. The Department of Energy’s national laboratories bring supercomputers and specialized scientific facilities that would be prohibitively expensive for most universities or companies to reproduce. Under Genesis, those assets are increasingly being connected into a broader AI-driven scientific system.

The industrial chain stretches much farther than the data center. Power plants and compute sit at one end; automated laboratories, scientific instruments, clinical-trial networks and pharmaceutical manufacturing sit farther down the line. If discovery accelerates without comparable investment in those later stages, the bottleneck simply moves.

China is assembling many of the same pieces. Beijing’s plans for AI-enabled science call for systems that can identify biological targets, design potential drugs and predict their efficacy and safety, while China’s National Medical Products Administration has separately laid out plans to use AI throughout drug review, approvals, inspections and surveillance. Its 2030 target includes dedicated computing infrastructure, specialized models and higher-quality datasets rather than simply attaching AI tools to the existing regulatory system.

The resulting competition won’t be determined by who has the cleverest model in isolation. It will also depend on the datasets feeding those models, the laboratories available to test their ideas, the patients available for trials, the regulators reviewing the results and the factories capable of producing whatever works.

A few months saved at several points in that chain starts to matter very quickly when multiple companies or countries are chasing the same biological target.

The economics could change as well. Rare diseases often attract less research because there are relatively few patients over whom to spread the cost of years of development, while difficult biological problems can consume enormous resources before producing a credible lead. If AI makes the early search cheaper, researchers can examine more existing medicines, pathways and genetic evidence before deciding which experiments deserve real money.

None of this requires Hassabis to literally cure every disease within a decade.

His timetable may prove wildly optimistic. Altman’s 10 gigawatts may never produce the cancer breakthrough he imagines, and Amodei’s compressed century depends on AI systems more capable than the ones available today. But those forecasts are increasingly sitting on top of something tangible.

AlphaFold already turned a decades-old scientific challenge into a resource used around the world. AlphaGenome is extending the same approach into the immense search space of human genetic variation. Co-Scientist has generated biomedical hypotheses that survived early experimental testing. Pharmaceutical companies are putting serious money behind AI-driven discovery, while governments are beginning to connect supercomputers, scientific instruments and automated laboratories into the same system.

The question is becoming less theoretical: what happens if the cost of generating a good scientific idea falls much faster than the cost of testing one?

We may find ourselves with more promising biology than the existing scientific system can physically process. That would be a very good problem to have, but it would still be a problem.

The cure factory requires a factory.

Now, in ordinary times, China can replace those barrels without much trouble, but in a future crisis, Beijing now starts with one less piece on the board.

At the end of 2025, China was still taking a large share of Venezuela’s crude. Vortexa estimated shipments at roughly 470,000 barrels per day during the year, about 4.5% of China’s seaborne crude imports, while internal PDVSA shipping data put exports to China closer to 642,000 barrels per day. The difference partly reflects years of Venezuelan crude moving through middlemen and appearing under other origins in Chinese customs data. The dark fleet is gonna dark fleet, after all. Either way, hundreds of thousands of barrels of discounted Venezuelan heavy crude were moving east, much of it to independent Chinese refiners.

Obviously, that traffic changed quickly in January 2026. A raid by U.S. special forces has a tendency to be disruptive, and this proved no different. Two China-flagged supertankers sailing toward Venezuela to collect crude associated with debt repayment turned around and headed back toward Asia, and by August 28 S&P Global reported that China had taken no Venezuelan barrels since December 2025.

Venezuela kept exporting. In August, shipments averaged about 1.17 million barrels per day, including 553,000 to the United States, 297,000 to India and 260,000 to Europe. Chevron exported about 286,000 barrels per day from its Venezuelan operations, while Vitol and Trafigura together handled roughly 597,000. The barrels did not disappear. New buyers and traders simply took over a trade that had pointed heavily toward China.

The Treasury has since put that shift into the rules governing the market. General License 46D, issued August 27, allows an “established U.S. entity” to lift, buy, market, transport, sell and resell Venezuelan oil under specified conditions. The license excludes transactions with Venezuelan or U.S. entities owned or controlled by Chinese persons, along with entities operating joint ventures with them, although OFAC allows an established U.S. company to resell Venezuelan crude to a Chinese buyer.

China can therefore still buy Venezuelan oil through the authorized trade, but Treasury now sets the terms of that channel and Chinese firms face explicit barriers on the Venezuelan side. Companies selling the oil outside the United States also have to report the parties, quantities, values and ultimate destinations to the U.S. government. Washington now gets a fairly detailed look at a trade Beijing once accessed much more directly, and with far less scrutiny.

China had much more invested in Venezuela than the cargoes. Sinopec and CNPC held some of the largest foreign oil positions in the country; Reuters identified roughly 2.8 billion and 1.6 billion barrels of entitlement reserves respectively, and CNPC remained a producer through the Sinovensa venture after it stopped lifting crude directly in 2019. China Concord Resources was still expanding in Venezuela in 2025, bringing a floating production facility to Lake Maracaibo for a project worth more than $1 billion that aimed to raise output at two fields from roughly 12,000 to 60,000 barrels per day and send the crude to China.

North American Blue Energy Partners has now taken over some of those positions. NABEP received 100-year rights covering 17 Venezuelan fields with roughly 65 billion barrels of reserves, and Reuters reported that Chinese and Russian firms, including China Concord, Sinopec and CNPC, had operated several of them. The U.S. government holds a 35% stake in NABEP through warrants structured to protect that position from dilution, along with access to 20% of production at cost and preferential rights over additional barrels.

President Trump made the strategic purpose unusually explicit when he announced the agreement on August 28. The White House subsequently called it the “biggest oil deal in world history,” said it gives the United States majority control over more than 65 billion barrels of Venezuelan reserves, and emphasized that most of the additional fields moving under NABEP had previously been controlled or operated by Chinese and Russian firms. The administration says NABEP could invest as much as $100 billion in Venezuelan oil infrastructure, with millions of barrels of additional production eventually processed through U.S. refineries. The precise legal meaning of “majority control” is more complicated than the White House formulation suggests, but the administration is leaving little ambiguity about what it wants the agreement to accomplish: move a major Venezuelan reserve base out of adversaries’ hands and into a U.S.-led commercial system.

Beijing noticed. Chinese Foreign Ministry spokesman Guo Jiakun responded on September 1 that China-Venezuela cooperation enjoys protection under international law and the laws of both countries and that China’s legal rights and interests in Venezuela “must be guaranteed.” Washington appears to have reached a rather different conclusion about who gets to decide what those interests are worth.

China also still has billions of dollars tied up in Venezuela. Chinese creditors committed roughly $106 billion between 2000 and 2018, much of it through arrangements linked to oil repayment, before Beijing stopped making large new commitments and began reducing its exposure. Estimates place Venezuela’s remaining Chinese debt at roughly $10 billion to $15 billion. PDVSA records reviewed by Reuters show that only a small share of the roughly 642,000 barrels per day sent toward China in 2025 directly serviced debt; most of the trade had long since become ordinary oil commerce.

China depends heavily on imported crude, with EIA putting the seaborne share of those imports at 92%. Venezuela offered Chinese companies a politically friendly reserve base large enough to support decades of investment and much more production than the country delivers today. It spread Chinese petroleum interests beyond the Persian Gulf and western Pacific, even if a trans-Atlantic supply line would remain exposed to U.S. pressure during a conflict.

The United States now occupies more of that position. U.S.-approved companies handle a growing share of exports, American producers are expanding, and the NABEP agreement puts the U.S. government directly into fields that include former Chinese positions. Chevron announced on September 2 that it plans to invest $7 billion and double its Venezuelan production to about 600,000 barrels per day over five years. Eni is also expanding, including a new 25-year production-sharing agreement for the Junin 5 heavy-oil project.

None of this turns Venezuela into an instant energy superpower, but it sets it back on the firm trajectory of the 90s and early 2000s of a true Western Hemisphere Petrostate. The country still produces only about 1.1 million to 1.2 million barrels per day, and years of underinvestment have left terminals, upgraders and upstream infrastructure needing extensive work. Raising output by millions of barrels will take years of construction and capital. Lawyers and energy analysts have also questioned the NABEP agreement’s legal basis, contract structure and lack of competitive bidding. Its 65 billion barrels remain underground, and Venezuelan politics will continue to matter to contracts stretching out for a century.

A year ago, several hundred thousand Venezuelan barrels were still moving toward Chinese buyers, Chinese companies held producing and prospective field positions, and new Chinese capital was funding projects designed to send more crude east. By September 2026, U.S.-approved sellers controlled the authorized path for Chinese purchases, Chinese tankers had withdrawn from the old trade, Chinese operators had lost positions in fields now tied to a company partly owned by the U.S. government, and Western capital was moving into the same production base.

It remains to be seen how quickly Venezuela can recover from decades of socialism and neglect. The damage is enormous, and rebuilding an oil industry that was systematically hollowed out will take years. But Venezuela still has the resource base, Western capital is returning, and for the first time in a generation the country is moving toward a very different economic relationship with the United States. That should offer some genuine optimism for Venezuela’s future. For China, the picture is simpler: one of the world’s great oil reserves is moving back toward the Western economic orbit, and Beijing now has one less piece on the board.

Building the Infrastructure for American Power Beyond Earth

For most of the space age, the rocket was the problem. This report argues that American success in reusable launch is moving the constraint into the systems around it: manufacturing, secure payload processing, roads and ranges, communications, spectrum, skilled workers, and the logistics required to sustain activity in orbit.

Launch counts alone cannot describe usable national capacity. A faster range can expose a processing shortage; new processing bays can shift pressure toward factories, integration teams, or spacecraft commissioning. The national task is to keep removing those constraints as demand grows, while reducing dependence on shared infrastructure whose failure could affect several providers at once.

Resilience Requires Reconstitution

Proliferated military satellite networks can withstand the loss of individual spacecraft, but sustained attrition creates a different test. The report proposes measuring how many militarily useful spacecraft the United States can manufacture, integrate, process, launch, commission, and connect to operational networks within 30, 90, and 180 days. Prepared spares and new production should be measured separately, with the limiting constraint identified at each interval.

Cape Canaveral and Vandenberg should be treated as strategic ports, with secure processing, utilities, transportation, communications, range systems, and specialized workers planned as an interconnected system. Distributed launch options and practical common interfaces can preserve usable pathways when a provider or location becomes unavailable.

A National Space Infrastructure Strategy

The report calls for an interagency space-infrastructure scorecard, investment in common-user assets and secure payload processing, and competitive government demand for orbital transfer, servicing, and refueling. It distinguishes shared public and strategic infrastructure from speculative commercial ventures that must prove their own economics.

Space Force should develop the mobility, logistics, commercial integration, and infrastructure resilience needed to preserve military freedom of action under disruption, while civilian agencies retain their safety, spectrum, science, and exploration responsibilities. The analysis extends this infrastructure approach to sustained lunar activity and argues that the United States should use its current lead to build durable capacity as China works to expand its own reusable-launch and satellite-production systems.

View or download full report (PDF)

The additional sailings are arriving during a measurable change in Arctic conditions. The Arctic recorded its warmest October-to-September year since observations began in 1900 during 2024-25, and since 1980 its annual air temperature has risen nearly three times as fast as the global mean. Every one of the 19 lowest September sea-ice minimums in the satellite record has occurred during the past 19 years. Ice more than four years old, which is generally thicker and harder for ships to penetrate, covered about 1.72 million square kilometers on average in 1985-2004; only about 95,000 square kilometers remained at the September 2025 minimum, a decline of roughly 95 percent. End-of-summer ice cover in 2025 was 28 percent smaller than it had been in 2005.

The calendar is shifting along with the amount and age of the ice. NOAA calculates that Arctic melt onset has moved earlier by about 4.1 days per decade; in 2025 it began roughly two weeks earlier than at the start of the satellite melt-onset record in 1979. September sea-ice extent has been declining at roughly 12 percent per decade, while the remaining pack is younger and less dominated by thick multiyear ice. That doesn't produce a straight line toward easy navigation every summer, since winds can push ice into particular seas and choke a route even during a low-ice year, but it does change how many weeks a commercial operator has to work with. In July this year, Russia's nuclear icebreaker Yakutia still had to reroute a convoy around heavy ice fields in the Kara and East Siberian seas, a useful reminder that less ice doesn't mean no ice (at least not yet).

The longer horizon points in the same direction. The IPCC assesses that the Arctic is likely to experience at least one September with less than one million square kilometers of sea ice before 2050 under all of the emissions scenarios it considered, while winter ice is very likely to remain through this century. At sustained global warming of 2 to 3 degrees Celsius, practically ice-free September conditions become much more frequent. For shipping, the important change comes before a blue Arctic Ocean appears on a map: a longer summer season, thinner ice over more of it, and a greater chance that ships can plan passage for a particular week rather than wait indefinitely for conditions.

Beijing has been planning for that opportunity for years. China's 2018 Arctic policy called for a "Polar Silk Road" developed through Arctic shipping routes and specifically encouraged Chinese companies to conduct commercial trial voyages "to pave the way for their commercial and regularized operation." The same document directed attention toward Arctic infrastructure, hydrographic surveys and improved navigation and logistical capacity. Sea Legend is a private operator rather than an arm of the Chinese state, but the activity Beijing said it wanted in 2018 is now occurring: Chinese cargo is being booked on dated summer sailings through the Russian Arctic.

The carrier compares about 20 days from Ningbo-Zhoushan to Felixstowe with roughly 40 days through Suez and as much as 50 around the Cape of Good Hope; the Arctic number ends at the first European port, not the final destination. Last year's Istanbul Bridge carried lithium-ion batteries, photovoltaic products and other cargo for which several weeks in transit can matter. This summer's eight sailings put the same proposition into a recurring schedule, while South Korea has also begun a container trial through the Northern Sea Route with PanStar Acro.

The Northern Sea Route recorded 103 complete transit voyages by 88 ships in 2025, during a season that ran from June 30 to November 17; Suez handled 12,758 vessels in the same year despite the Red Sea crisis. Tankers accounted for 34 of the Arctic transits, and total transit cargo was about 3.2 million tons. The route remains seasonal, Sea Legend's ships are small by the standards of the main Asia-Europe trade, and passage along the Russian coast still depends on Russian permits and infrastructure.

Those permits sit inside an Arctic system Russia has spent decades building. The U.S. Coast Guard says Russia currently operates more than 40 icebreakers; eight are nuclear-powered vessels operated by Rosatom, which remains the only operator of a nuclear icebreaker fleet in the world. Four newer Project 22220 ships, Arktika, Sibir, Ural and Yakutia, were already in service by 2026, with additional nuclear icebreakers under construction. Rosatom reported that its nuclear fleet escorted 976 vessels along the Northern Sea Route and into frozen Russian ports in 2024, up from 759 the year before.

That fleet supports much more than through-shipping between China and Europe. According to the Defense Department's Arctic Strategy, more than 80 percent of Russia's natural-gas production and nearly 20 percent of its petroleum production comes from the Arctic; after Russia's invasion of Ukraine increased its economic isolation, Moscow became more reliant on China to finance Arctic energy export infrastructure and purchase those resources. A longer operating season makes the same ports, icebreakers and coastal facilities useful for more commercial traffic, while Chinese capital and demand give Russia customers for an Arctic economy that Western sanctions have made harder to finance.

China is building its own ability to operate there as well. The Defense Department counted three Chinese icebreakers in its 2024 strategy, Xue Long, Xue Long 2 and Zhong Shan Da Xue Ji Di, and described their research work as dual civil-military activity. Across 13 Arctic research expeditions, Chinese vessels tested unmanned underwater vehicles and polar-capable fixed-wing aircraft; People's Liberation Army Navy ships have also exercised alongside the Russian Navy, including operations by Chinese and Russian vessels in international waters off Alaska in 2022 and 2023. Commercial access, scientific operations and military familiarity are developing in the same region, even though China owns no Arctic coastline.

The American comparison is uncomfortable. The Coast Guard currently operates three polar icebreakers, one heavy vessel commissioned in 1976 and two medium icebreakers commissioned in 2000 and 2025; Canada, by comparison, says it has 18 icebreakers of varying capabilities, the second-largest icebreaking fleet in the world. China already has three, while Russia has more than 40. Counting hulls alone obscures differences in size and mission, but it captures the basic disparity in the number of ships available to maintain presence, escort vessels, reach an emergency or put government crews into ice-covered water.

Washington has finally moved beyond talking about the gap. The Coast Guard completed contract awards for 11 Arctic Security Cutters in February 2026; four are planned for construction in Finland, with seven follow-on ships built in the United States, and the first delivery is expected in 2028. The first two are slated for Kodiak and a third for Seward. Separately, construction has begun on the first Polar Security Cutter, the first new American-built heavy polar icebreaker in nearly five decades. The medium cutters provide presence and icebreaking capacity across much of the Arctic, while the heavy class is intended to restore the capability now concentrated in the aging Polar Star.

Ships need somewhere to operate from, which is why the work underway at Nome matters alongside the cutters. Alaska and Russia face each other across roughly 53 miles at the narrowest part of the Bering Strait, the Pacific entrance to the Arctic routes now being tested by Chinese shippers. The Army Corps of Engineers awarded a roughly $399.4 million contract in August 2025 for Phase 1A of the Port of Nome expansion, including a 1,200-foot causeway extension; later work is planned to deepen the basin from its current 22 feet to 40 feet so larger vessels can enter. The Corps describes the expanded port as supporting cargo and public safety as well as America's military presence above the Arctic Circle.

The military implications obviously extend well beyond protecting merchant ships. The Defense Department already treats the Bering Strait and the Barents Sea north of Norway as chokepoints becoming more navigable and more militarily significant as sea ice declines. Russia's Kola Peninsula is home to the Northern Fleet and strategic nuclear forces, including its submarine-launched ballistic missile force; Moscow has continued to invest in Arctic bases and refurbish Soviet-era installations. DoD also describes the Arctic as an avenue of approach to the United States and a region from which Russian capabilities could interfere with American power projection toward Europe or the Indo-Pacific.

Less ice doesn't necessarily make military operations easier because weather remains severe, communications above 65 degrees north are difficult, old infrastructure is being damaged by thawing permafrost and coastal erosion, and thinner, mobile ice can be less predictable than a stable frozen surface. It does, however, increase the amount of water where surface ships can operate and the number of actors trying to operate there. DoD's response includes greater maritime and aerospace surveillance, better satellite communications, additional sensing and forecasting, and more Arctic training; by the 2030s, the United States and its Arctic allies expect to have more than 250 advanced multirole combat aircraft available for possible Arctic operations. Seven of the eight Arctic states are now NATO members, with Russia the exception.

Later in the century, continued summer ice retreat could alter Russia's advantage again. Today's Northern Sea Route follows the Russian coast because that is where the infrastructure exists and where the ice is most manageable for regular navigation; a central Arctic route could eventually allow some ships to travel farther from Russian territory and use less Russian coastal support. The IPCC gives strong reason to expect substantially less summer ice, but it doesn't establish when ship design, insurance and remaining ice conditions will make a transpolar commercial route practical.

The United States doesn't need to bet on a specific year for that transition some time in the vague future. It already has Chinese commercial sailings approaching the Bering Strait, more than 40 Russian icebreakers on the other side of the Arctic, three Chinese icebreakers building operating experience, and Russian and Chinese military forces working together in northern waters. Contracts for 11 American Arctic Security Cutters are now in place, but the first isn't expected until 2028; the heavy replacement fleet will take longer. The ice is moving faster than the ships.

On August 28, X said it had identified about 200,000 suspected Chinese inauthentic accounts, including roughly 200 posting about AI data centers and energy policy in ways the company said "could manipulate a legitimate debate." Their material would have looked familiar to anyone following the fight over new data centers: higher household electric bills, pressure on the grid, technology companies making money while ratepayers absorbed the cost. Some posts used AI-generated cartoons, while others borrowed the credibility of legitimate news coverage. The accounts presented themselves as American voices inside the American political fight, and they chose the side most likely to make new compute and the generation required to run it harder to build.

This new discovery by X wasn’t a surprise to people who had been watching the issue of growing Chinese influence, whether directly or algorithmically, against the American data center buildout. OpenAI had already found a similar operation in June. Its "Data Center Bandwagon" investigation identified a cluster likely originating in China whose operators prompted ChatGPT in Simplified Chinese to write as Americans from different backgrounds, then posted the resulting English through likely inauthentic accounts.

OpenAI traced the activity to operators likely working at a private Chinese technology company serving provincial-government clients and found no meaningful breakout beyond the operation's own activity. Now, having 200 policy accounts isn’t evidence that China secretly controls American zoning politics, and the public record doesn't show that either operation stopped a particular project, but what the findings do show is that China-linked operators twice chose the same live American dispute at a moment when delay itself has acquired strategic value.

The United States is trying to build the physical infrastructure of AI just as electricity and interconnection have become hard limits on how quickly new capacity can reach service. Washington has spent years restricting China's access to advanced semiconductors because usable compute is a strategic asset; meanwhile, American developers can buy the chips and still wait years for enough power to run them. A hyperscale facility needs far more than racks and servers. It needs generation, substations, transmission and a place on an electric system that was planned long before individual customers began asking for loads measured in hundreds of megawatts or even gigawatts. An NVIDIA accelerator sitting in a warehouse because a data center can't get electricity isn't doing much for American AI.

China has treated that physical problem as part of its AI strategy for years. Beijing's "East Data, West Computing" program is concentrating large-scale computing in eight national hub nodes and 10 major clusters, shifting heavy workloads toward regions where land and electricity are easier to obtain. Chinese government reporting said that by June 2026 more than 80 percent of the country's intelligent computing capacity had been built inside those eight hubs; one cluster in Inner Mongolia had more than 30 data centers under construction within a three-kilometer radius. The power system underneath them continues to expand as well, including new coal and nuclear generation. Whatever the environmental costs of that approach, Beijing has been clear about the industrial requirement. The cloud still plugs into a wall, and a national AI strategy that can't produce electricity on time isn't much of a strategy.

American investment has been running into a very different system. Data Center Watch counted at least 75 U.S. projects worth roughly $130 billion blocked or delayed by local opposition in the first quarter of 2026 alone. Carbon Direct, using a different methodology and looking from January 2024 through May 2026, identified at least 46 AI data-center projects across 20 states that had been delayed or canceled following community opposition, representing $170 billion in announced investment. Baird counted 25 outright cancellations in 2025, compared with six in 2024 and two in 2023. Those datasets shouldn't be combined into a single total, and some announced projects are speculative or competing for the same scarce grid capacity. Even after allowing for that, a substantial amount of American compute is spending more time on paper than its developers expected.

The grid queues tell much the same story from the other direction. Across the Midwest, Mid-Atlantic and South, large users, primarily data centers, have submitted more than 700 gigawatts of electricity requests. Many are duplicative or speculative, which is why utilities are trying to separate serious projects from developers reserving several possible places in line. Texas paused new data-center grid connections in August while it audits almost 49.8 gigawatts of projects, a group BloombergNEF estimated at roughly 20 percent of the U.S. data-center pipeline. One of the country's most development-friendly states is now trying to distinguish actual projects from a queue that has grown faster than the system underneath it.

Adding generation isn't fast either. Lawrence Berkeley National Laboratory counted roughly 8,200 proposed power projects seeking U.S. grid interconnection at the end of 2025, representing 1,312 gigawatts of generation and another 749 gigawatts of storage. Most have nothing specifically to do with data centers, although data centers will depend on the same supply. In PJM, projects that entered service in 2025 had spent more than seven years on average moving through interconnection and the work required afterward to get online. AI companies can build a new model, raise another round and order another generation of chips several times over while a power project is still working its way toward commercial operation.

New York has paused discretionary environmental permits for new hyperscale data centers of 50 megawatts or more while it conducts a broader environmental review; Texas is auditing its large-load queue, while Pennsylvania has tightened the relationship between state permitting and local approval. The concerns behind those fights aren't invented. Utilities shouldn't be allowed to dump the cost of serving a new hyperscale customer onto existing ratepayers, and local governments have an obvious interest in what happens when a project consumes enormous amounts of electricity or requires major new infrastructure.

Beijing doesn't need to manufacture an angry ratepayer or invent a zoning dispute from scratch. It can enter an argument that already exists under an American name, borrow facts already circulating in legitimate news coverage and push the interpretation most likely to make another project slower, harder or more expensive. Fake accounts don't need to persuade a county commissioner in one sitting. Once their messages circulate beside real complaints about power bills or permitting, reposts and algorithmic recommendation can carry the argument well beyond the accounts that introduced it. The operator can disappear while the claim keeps moving.

Beijing's interest in keeping those arguments alive is unusually direct because American export controls are designed in part to slow China's access to frontier compute while U.S. companies continue moving ahead. Chinese firms are using that time to improve domestic accelerators, expand available computing capacity and reduce the leverage Washington gets from restricting advanced chips. A year matters in a way it wouldn't in an ordinary real-estate cycle. If a large American cluster reaches service a year late, Chinese labs get another year to improve their own hardware and train models before that additional U.S. capacity comes online.

Persistent delay also reaches beyond commercial model development because the same physical infrastructure increasingly supports national-security workloads. The White House AI strategy has called for high-security data centers capable of handling classified computing, placing defense and intelligence users in the same basic world of transformers, generation, substations and secure facilities. Scarce power remains scarce power regardless of who needs it. If the United States repeatedly takes longer to turn chips and capital into operating compute, eventually the consequence isn't merely a delayed ribbon-cutting but less capacity available when the government or private sector wants to use it.

Washington wants foreign customers using American chips, cloud services and AI systems, while Beijing has spent years pushing Chinese digital infrastructure into overseas markets. Customers buy what can be delivered at the price and on the schedule they need. An individual delay in Virginia or Texas doesn't hand Beijing a contract, but persistent shortages can change which supplier is ready when a government or company is choosing its next system.

The public evidence doesn't show that China created America's data-center backlash, killed a particular project or turned local opposition into a foreign-directed movement. OpenAI found no meaningful breakout in the operation it studied, and X publicly tied only a small fraction of the accounts it identified to data-center and energy policy. The more important fact is narrower and harder to dismiss: China-linked operators are impersonating Americans inside a dispute that bears directly on how quickly the United States can add AI capacity.

That dispute will become more valuable to Beijing as electricity occupies a larger place in the technology competition. The United States can tighten chip controls, improve its models and pour more capital into AI, but eventually that advantage has to become operating compute somewhere on the ground. China doesn't need a technological miracle to benefit when American capacity arrives later than planned. It only needs the bottlenecks already inside the system to keep doing some of the work for it.

Treasury has documented cartel extortion of avocado and citrus growers, cattle ranchers, packers and other businesses in Michoacán, where farmers and exporters can face compulsory payments simply to keep operating. Criminal groups have seized land and entered agricultural businesses; CJNG-linked networks have used shell companies and falsified customs documents in cross-border fuel schemes that Treasury says generated tens of millions of dollars annually for the cartel, while fuel theft and smuggling more broadly have cost Mexican authorities billions of dollars in lost revenue. For a legitimate company, the protection payment is only part of the expense. Guards cost money, cargo insurance rises, stolen loads interrupt production and employees still have to travel through insecure corridors. Reporting in August based on data from supply-chain risk firm Overhaul found that 86.3 percent of cargo thefts in Mexico were concentrated in ten states, with the State of Mexico, Puebla and Guanajuato accounting for 46.7 percent of reported incidents; roughly seven in ten attacks involved violence. ConMéxico, representing large consumer-products companies, separately reported more than 250 million pesos in stolen merchandise during the first half of 2026. Across Latin America and the Caribbean, the Inter-American Development Bank estimated the direct cost of crime and violence at 3.44 percent of regional GDP in 2022, including business security spending, government expenditures and lost human capital.

Those costs now sit inside the economics of the USMCA nearshoring push, where U.S. Trade Representative Jamieson Greer and Mexican officials have spent 2026 discussing increased North American production, stronger regional supply chains and limits on non-market inputs from outside the region. After their July negotiating round, Greer and President Claudia Sheinbaum specifically agreed on the urgency of growing North American manufacturing. Mexico already has an industrial base, proximity to U.S. consumers and preferential market access, but none of those advantages removes the road to the factory from a site-selection model. Higher cargo losses or additional security can eat into an advantage that looked much better when the spreadsheet contained only wages, tariffs and distance; a slower or less reliable truck cycle starts giving back part of the benefit before the first shift is hired. Criminal organizations do not have to control an industrial site to make it more expensive to operate.

Nor does the security picture have to improve nationwide before an investor changes the calculation. One manufacturing corridor can become cheaper to use while cartels remain powerful elsewhere; freight through a port can become reliable enough that companies stop building the same losses and delays into every shipment. A quieter change in operating costs can matter more to a plant manager than a high-profile raid that leaves conditions around the plant unchanged. In Colombia, the same arithmetic reaches into the resource economy: an ELN commander told Reuters in June that the organization intended to keep financing itself through "economic detentions" and by taxing illegal mining and drug trafficking in territory it controls, particularly in Chocó. The quality of an ore body and the validity of a concession do not settle the investment case when access to the surrounding territory carries another unofficial price.

That question of access carries much larger numbers farther south, even where the mining projects themselves have no demonstrated cartel connection. Argentina and Chile approved operating protocols on August 27 for three cross-border copper projects under a mining treaty first signed in 1997; Chile's mining minister said the revived framework could unlock more than $20.7 billion in investment and add roughly 540,000 metric tons of annual copper production, in part by giving Argentine projects access to Chilean ports and mining infrastructure. Chile, Argentina, Bolivia and Peru followed a day later with a declaration to deepen cooperation on strategic minerals, particularly copper and lithium. Washington has been moving in the same general direction on minerals, with more than $2 billion in critical-mining and mining-related projects announced by the White House on August 7 and a July 30 presidential determination invoking Defense Production Act authority after finding that U.S. dependence on certain foreign sources of critical minerals threatened serious and sustained supply-chain disruptions.

The scale of those projects helps explain why reliable access to infrastructure and markets has become a strategic concern in its own right. The 2025 National Security Strategy puts a geopolitical frame around that interest in production and access. Its Western Hemisphere section, titled "The Trump Corollary to the Monroe Doctrine," says the United States will deny non-Hemispheric competitors the ability to "own or control strategically vital assets" in the region, with China plainly central to the concern. COSCO, a Chinese state-owned company, holds a 60 percent stake in Peru's Chancay port and exclusive operating rights; the project has been promoted as a major new gateway between South America and Asian markets. Cartel policy and China policy remain separate, but Washington wants more manufacturing and mineral supply in a hemisphere where criminal organizations still raise the cost of using some of the territory and infrastructure needed to produce and move both.

That burden falls especially hard on an American strategy that expects private capital to carry much of the investment. Chinese capital can move through state-owned enterprises and policy financing, with strategic considerations alongside commercial ones; an American investor still has to decide whether the expected return on a factory, port or mine is worth the risk. Extortion, cargo theft and unreliable transport enter that calculation before the geopolitical argument does. A security improvement that makes an otherwise attractive project cheaper to insure, supply or operate can therefore move private money without requiring Washington to finance the project itself.

Whether the August operations have an economic effect will not be clear from the number of boats sunk or commanders charged. It will show up around particular sites and corridors: what companies pay for protection, whether freight arrives reliably, what an insurer charges and whether a project that failed to clear an investment committee before begins to clear one. Those changes can occur well before anyone can plausibly claim that the cartels themselves have been defeated.

Under Canal Authority Advisory A-29-2026, booking dates beginning September 4 are limited to nine daily Neopanamax slots and 25 Panamax slots, or 34 transits in total. Panamax availability falls again on September 15, bringing the combined daily total to 32; the maximum authorized draft for the largest ships falls to 48 feet on September 2 and, under the current schedule, to 47.5 feet on October 1.

Those reductions are significant without being catastrophic. The canal averaged about 34 oceangoing transits a day in July, so the September measures are better understood as a tightening of available capacity than as anything resembling a shutdown. The more consequential question is what happens after September, because Gatún and Alhajuela Lakes still have to rebuild enough water during the remaining rainy months to carry the canal through the January-to-April 2027 dry season.

That uncertainty goes directly to the canal’s operating model. Ships entering the locks are raised roughly 85 feet to Gatún Lake and then lowered toward the opposite ocean, consuming freshwater from the surrounding watershed in the process; when less water is available, the Canal Authority can principally conserve it by reducing the number of daily transits or by limiting vessel draft. The first creates outright scarcity, forcing ships to wait, reroute, or compete for auctioned passage, while the second allows ships to continue crossing but potentially with less cargo aboard, spreading essentially the same voyage cost across fewer containers, fewer barrels, or fewer tons.

The market has already demonstrated what the first kind of scarcity can be worth. South Korea’s SK Gas recently paid a record $5.3 million to secure a September 1 northbound transit for the LPG carrier *G. Spirit*, surpassing a previous record of $4.6 million set only weeks earlier; canal auction prices averaged roughly $1.1 million in August, more than 16 times the comparable average a year before.

El Niño didn’t create those prices. Demand for urgent Panama passage had already surged after the Iran conflict disrupted Middle Eastern shipping and increased the value of moving U.S. Gulf energy cargoes toward Asia without spending several additional weeks at sea. Seven-figure bids were appearing while the canal still had greater operating flexibility; what has changed is that a geopolitical premium on the shortcut is now colliding with a developing water constraint on how much of that shortcut Panama can offer.

For an LNG tanker traveling from the U.S. Gulf Coast to Northeast Asia, avoiding Panama can add roughly 20 days to the voyage, turning a trip of about 30 days through the canal into something closer to 50 around the Cape of Good Hope. The additional fuel and charter expense matters, but so does the time the vessel remains occupied; when enough ships spend extra weeks at sea, effective fleet capacity tightens even though the number of ships in the world hasn’t changed.

That dynamic was already visible in July, when 11 laden LNG tankers used the canal, up from six in June and the highest monthly total since November 2023; nine were headed for Japan or South Korea. Higher Asian energy prices had improved the economics of the shorter Panama route at roughly the same moment the canal’s water position began to deteriorate, placing rising commercial demand against declining hydrological flexibility.

Panama also isn’t tightening in isolation. The Strait of Hormuz remains disrupted by the Iran conflict, while security concerns around the Red Sea and Bab el-Mandeb continue to influence routing decisions elsewhere in the system. Those chokepoints aren’t interchangeable in every trade, but they all draw on a finite global pool of ships, crews, charter time, and port capacity; a tanker or container vessel that spends another two or three weeks on a longer route isn’t available to move another cargo during that period.

Container markets are beginning to reflect some of that strain. On August 21, S&P Global assessed the spot cost of moving a forty-foot container from North Asia to the North American East Coast at $11,000, compared with $7,700 to the West Coast, leaving a $3,300 gap that had widened substantially since mid-July. Panama isn’t responsible for all of that spread, since Asian weather disruptions and limited vessel availability are contributing as well, but tighter drafts and fewer transit opportunities make East Coast routing relatively more expensive and increase the attractiveness of moving cargo through Pacific ports instead.

For American importers, the practical result is less a single Panama-related surcharge than a choice among different forms of higher cost. Cargo can continue to Savannah, New York, or another East Coast port at a higher freight rate; move through Los Angeles or Long Beach and incur additional rail or trucking expense; accept longer delivery times; or be supported with larger inventories against the risk of future disruption. Nearly 70 percent of Panama Canal cargo in fiscal 2025 either originated in or was destined for the United States, particularly through Gulf and East Coast ports, which makes even a moderate deterioration in canal economics unusually relevant to American supply chains.

The inflation implications should not be overstated. An IMF study covering 143 countries found that a doubling of global freight rates was associated with roughly 0.7 percentage point of additional inflation, but Panama’s current restrictions are far narrower than the global freight shock examined in that research; the relevant concern is persistence, not the September slot count by itself. Several quarters of higher freight costs, longer voyages, tighter vessel availability, and larger inventory requirements can eventually migrate through manufacturers, distributors, and retailers even when no single disruption is large enough to dominate the economy.

Whether that becomes necessary depends in part on what the Canal Authority can deliver institutionally. ACP has already been using operational water-saving measures, including tighter lock management, restrictions on special lockages, leakage controls, and reduced hydroelectric generation, and Panama entered 2026 with stronger reserves than during the worst of the 2023 drought. Over the longer term, the Río Indio reservoir project is intended to provide a much larger buffer for both canal operations and national water supply, but construction is expected to begin only in 2027; it will not add meaningful new storage capacity in time for the coming dry season.

That leaves the next several months unusually important. If late-season rainfall improves enough to rebuild Gatún and Alhajuela, the current reductions may prove to be an expensive but temporary adjustment, and ACP could enter 2027 with enough water to avoid much deeper restrictions. If reservoir levels remain weak into the end of the rainy season, however, the October draft schedule, auction clearing prices, LNG transit counts, and the widening freight gap between East and West Coast routes will become increasingly useful signals of how much pressure is building.

For Gulf energy exporters, East Coast ports, importers, and inventory planners, that distinction matters more than whether the canal moves 32 or 34 ships on any particular day. A meaningful reservoir recovery would argue for treating the present freight premium as temporary; a weak close to the rainy season would make it prudent to plan for tighter canal capacity, longer voyages, and higher inventory buffers before the 2027 dry season reaches its worst point.

The Panama Canal’s value has always been simple: it eliminates enormous amounts of distance. Panama can manage moderate water shortages, shipping companies can reroute vessels, and global trade will keep moving; what becomes difficult is doing all of those things cheaply when several major chokepoints are under pressure at once.

The world’s ships can always take the long way around. A record $5.3 million bid to avoid doing so is already a pretty good indication of what the long way is worth.

Key Findings

• The more consequential question is what a service does differently once it has reason to believe the person on the other side of the screen is 12, 15 or 17 rather than an adult.
• The emerging regime isn't one rule for "screen time." It is a stack of controls applied at different points in the product.
• The new youth regimes increasingly separate attention, communication, commerce and age-based access instead of assuming that every minute of digital use creates the same risk.
• Age is becoming a product variable rather than merely an eligibility question; once a service identifies a minor, more of the product is expected to change automatically.

Executive summary

The 2026-27 school year opens with a youth internet that is being segmented by age in ways that were mostly aspirational only a few years ago. The debate isn't simply about whether minors should have access to digital services; teenagers already use social media, gaming platforms and AI at scale. The more consequential question is what a service does differently once it has reason to believe the person on the other side of the screen is 12, 15 or 17 rather than an adult.

That shift is visible across products that were once treated as separate categories. Meta's proposed multistate settlement would impose default daily limits, overnight restrictions, school-hour notification limits and stronger age assurance for users under 18 on Facebook and Instagram; Roblox has reorganized accounts by age and requires an age check before chat; New York has finalized rules restricting algorithmically personalized feeds and nighttime notifications for minors without parental consent; and OpenAI began rolling out ChatGPT for Teens on August 18. Schools are drawing a parallel boundary around the device itself, with the National Conference of State Legislatures reporting as of July 15 that 42 states, Washington, D.C., and Puerto Rico had enacted legislation related to cellphone use in schools.

These rules govern different layers of the same experience: the device, age assurance, the teen account, parental override, and then chat, recommendations, purchases and notifications. A phone pouch can take the handset out of a student's hands without changing the account; an age check can route that student into a different product without changing the school rule; and a parent can be given authority over a time limit without being allowed to read a teenager's messages. The emerging regime isn't one rule for "screen time." It is a stack of controls applied at different points in the product.

Teen use is high enough that these distinctions matter. Pew Research Center reported in December 2025 that roughly nine in ten U.S. teens use YouTube, while about six in ten or more use TikTok and Instagram and 55% use Snapchat; about one in five say they use TikTok and YouTube almost constantly. Sixty-four percent report using AI chatbots, including about three in ten who use them daily, while Pew has separately found that 85% of U.S. teens play video games and 41% play at least daily.

The evidence on health effects remains more complicated than the political debate sometimes suggests, and the platform differences are substantial. Research has found associations between some forms of digital-media use and depression, self-injury, behavioral problems and other outcomes, while teenagers also report benefits involving friendship, entertainment and connection. In Pew's 2026 platform study, 37% of teen TikTok users said the service hurt their sleep, compared with 26% of Snapchat users and 24% of Instagram users; 44% of Snapchat users said the service helped their friendships, and most users of all three platforms said the services neither helped nor hurt their mental health.

The regulatory response therefore doesn't require a single verdict on whether social media is good or bad for adolescents. Legislatures, regulators, schools and companies are making narrower choices about time limits, private contact, age verification, recommendations, purchases, overnight access and parental authority. Age is becoming a product variable rather than merely an eligibility question; once a service identifies a minor, more of the product is expected to change automatically.

View or download full report (PDF)

Key Findings

• The danger is not Chinese military dominance; it is Chinese military sufficiency arriving before American and allied readiness.
• Xi does not need Trump to abandon Taiwan; he needs to consume enough time for Chinese military action to change the problem Washington is deciding.
• Artificial intelligence gives China another way to attack that decision process.
• The most dangerous Chinese judgment is simpler: we are ready enough, they are not ready enough, and waiting helps them more than it helps us.

Executive summary

April 2027 could present Beijing with a temporary opportunity around Taiwan: the People’s Liberation Army will be approaching the capability benchmark set for its 2027 centenary while the United States and its allies remain partway through a buildup intended to correct known weaknesses. The central risk is relative timing. China may conclude that it is ready enough before the opposing coalition becomes ready enough.

That military transition intersects with political decision time. If China initially confines its campaign to Taiwan and offers direct negotiations while operations continue, Beijing may be able to consume the hours or days in which American intervention would be most effective. Artificial intelligence compounds the problem by making it easier to distribute contradictory narratives, fabricate persuasive evidence, and attack trust in authentic information during a fast-moving crisis.

The United States will not be helpless and China will not be assured of victory. American weapons inventories, penetrating bombers, resilient Pacific bases, attack-submarine capacity, logistics, and allied posture are all improving. The vulnerability window exists because those improvements take time, giving Beijing an incentive to evaluate whether acting sooner presents a better relative opportunity than waiting for a stronger coalition later in the decade.

View or download full report (PDF)

Key Findings

• AI offers a genuine capacity dividend because useful administrative applications no longer require a city or county to build its own software system.
• The relevant question is therefore not whether a model is impressive, but whether it improves a specific government task after correction time and error risk are counted.
• The correct model for a small jurisdiction is broad assistance and narrow authority.
• The objective is not to automate local government. It is to increase the amount of competent work a limited public workforce can perform while keeping governmental judgment, legal accountability and control of the public record where democratic government requires them to remain.

Executive summary

Artificial intelligence matters especially to small local governments because administrative capacity is often constrained by staffing. Research, drafting, document review, permitting support, resident communication, records work, and meeting preparation consume limited professional time, while many jurisdictions cannot create a specialized technology office or build custom systems for each task.

AI can create a real capacity dividend when it is applied to bounded, document-intensive work that employees already understand and can independently verify. The evidence supports measured adoption rather than blanket enthusiasm: public-sector studies show meaningful improvements on some writing and document tasks, but weaker performance on others. The proper unit of evaluation is therefore the government task, including correction time and error risk, not the model or vendor demonstration.

The recommended operating model is broad assistance and narrow authority. Approved tools can be widely available for low-consequence work, while controls tighten when a system receives nonpublic information, searches internal repositories, communicates with residents, influences consequential decisions, or acts on government’s behalf. Human judgment, legal accountability, and control of the public record must remain with the institution.

View or download full report (PDF)

Key Findings

• The United States no longer lacks a maritime strategy. It lacks a production system capable of turning strategy into ships at a rate, cost, and level of reliability that matters.
• America does not need to reproduce China's commercial fleet or outbuild China ship for ship. It needs enough sovereign capacity to build, repair, crew, and surge the vessels its economy and military cannot prudently outsource, supported by a production base that can improve through repetition and expand under pressure.
• The immediate danger is that Washington mistakes policy activity for industrial recovery.
• The next four years should be treated as a production-learning campaign.

Executive summary

The United States no longer lacks maritime policy; it lacks a production system capable of converting strategy into ships at a relevant rate, cost, and level of reliability. China’s advantage is industrial rather than merely numerical. The American objective should not be to match Chinese commercial output ship for ship, but to restore enough sovereign capacity to build, repair, crew, and surge the vessels the economy and military cannot prudently outsource.

The immediate risk is mistaking policy activity for industrial recovery. Financing tools, tax incentives, trade measures, workforce programs, and new federal authorities can support a rebuild, but none substitutes for a multiyear orderbook organized around a small number of repeatable vessel families. Repetition gives shipyards, suppliers, lenders, and workers a reason to invest and creates the learning curve required to lower labor hours, rework, and schedule risk across successive hulls.

Allied investment, mature foreign designs, and reflagged vessels can bridge the period before domestic large-ship construction recovers. Those tools create durable American capacity only when they are tied to U.S. facilities, workforce training, licensed production methods, domestic supply chains, and the movement of subsequent construction into American yards. Success should be judged by delivered ships and measurable production learning rather than the number of programs announced.

View or download full report (PDF)