On the morning of October 24, 2023, Otis was still a tropical storm off Mexico with winds around 70 mph. By 10 p.m., it was a Category 5 hurricane carrying 160 mph winds and sitting about 55 miles from Acapulco. The National Hurricane Center opened its discussion that night with an unusually stark sentence: “A nightmare scenario is unfolding for southern Mexico this evening.”
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.