When 60 Minutes asked Google DeepMind CEO Demis Hassabis whether artificial intelligence could eventually mean the end of disease, he didn’t treat the question as science fiction. Hassabis said curing all disease with the help of AI was within reach, perhaps within the next decade.
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.