Investors have poured billions into using artificial intelligence to discover new drugs, and 2026 is the first real test of whether AI-designed medicines actually helps patients. The boom has genuinely transformed the search for molecules — but that was never the costly, failure-prone part of making a medicine, and there AI has so far had little to add. Capital, and the public subsidies have not yet priced the difference, writes Michael A. Santoro.
This is the year the bill comes due on one of the most heavily funded bets in biotechnology. Since the early 2010s, investors have poured billions into companies promising that artificial intelligence would transform the discovery of new medicines, and the promise was not empty: AI really has changed how drugs are found. But in 2026, for the first time, a meaningful cohort of AI-designed drugs reaches Phase 3: the large, multi-year human trials that determine whether a drug actually works. By the end of the year, we will have the first real evidence of whether the AI drug-discovery boom produces efficacious medicines. The early signs suggest the money has expedited the search for promising molecules but barely touched the harder tests that show if the molecules actually produce medicines that can help patients.
What the money bought
AI has genuinely transformed the front end of drug discovery: the search for a promising molecule. A company like Absci can now design a working antibody almost from scratch—testing fewer than a hundred candidate designs—where the older approach screened physical libraries of millions of chemical compounds. Insilico Medicine reports going from the choice of a biological target—deciding which protein in the body to attack—to a finished drug candidate ready for its first safety tests in about eighteen months for roughly $2.6 million, against the four to six years the same early work traditionally took. By 2022 some 150 firms were applying AI to small-molecule design alone, and the money behind them—a single 2024 entrant, Xaira, launched with $1 billion, and Alphabet’s Isomorphic Labs raised $600 million in 2025—has made AI drug discovery a multibillion-dollar industry. This is where AI works: designing a molecule is, at bottom, a problem of chemistry and physics, with firm rules and vast troves of data, exactly the conditions under which machine learning excels. It is also the low-hanging fruit.
Where the value actually is
The trouble is that designing the molecule was never the expensive part. The roughly $2.6 billion it takes to bring a single drug to market is dominated not by discovery but by clinical trials and the failures within them, and there AI has so far made almost no difference. Follow AI-designed molecules into human testing and a pattern appears.
According to one Boston Consulting Group analysis of about two dozen AI-discovered molecules that had reached clinical trials, in Phase 1—the first time a drug is given to a small group of people, to check that it is safe—AI-designed molecules succeed 80 to 90 percent of the time. This is far above the historical norm of about half. But in Phase 2, the first real test in patients of whether the drug does anything, their success falls back to the industry’s ordinary 40 percent or so. (Phase 3, the large confirmatory trial, is the last stage before approval; the first AI-designed drugs are only reaching it now, so there is no track record there yet—BCG conducts its final analysis assuming Phase 3 success rates continue at their historical pace.) BCG concludes that the net effect is positive but modest: the end-to-end odds that an AI-designed drug reaches the market roughly double, from something like 5–10 percent to 9–18 percent. And every point of that gain is banked in the cheap early stage. It is worth knowing how cheap: a Phase 1 study averages about $4 million, a Phase 2 study about $13 million, and a Phase 3 trial $20 million and often far more. AI is saving money at the front of a process whose costs pile up at the back.
Why the hard half is hard
The reason for the divide is the most important and least appreciated fact about AI in medicine, and it is worth spelling out. Designing a molecule is essentially a problem of chemistry and physics: will this structure hold together, will it lock onto its target, will the body tolerate it? Those are questions with firm physical rules and enormous libraries of prior data. These are also the very conditions machine learning was built for, the same way it learned to recognize faces or predict the next word in a sentence.
Choosing the right biological target is a different kind of problem entirely. To know whether blocking a particular protein will actually change the course of a disease requires understanding the tangled causal machinery of human biology—machinery we have mapped only in fragments, on data that are sparse, noisy, and often missing. A model can learn the rules of chemistry because we know them; it cannot learn the rules of a disease we do not yet understand. Drugs fail in two ways—the molecule is wrong, or the biological bet behind it is wrong — and AI has largely conquered the first while barely touching the second. It has made the cheap end of development cheaper and left the expensive end, where the science is genuinely hard, exactly where it was.
The first returns from the clinic bear this out. The brightest result so far is Insilico Medicine’s rentosertib, an AI-discovered drug for idiopathic pulmonary fibrosis that posted positive mid-stage results in Nature Medicine in 2025. Against it sit the setbacks: Recursion, one of the field’s most prominent companies, discontinued its lead AI-discovered program in 2025 after the efficacy signal failed to hold. The once-frothy field of standalone AI-drug firms has been thinned by mergers and shutdowns. Some of that is the ordinary shakeout of a young, overfunded sector: drug development takes ten to fifteen years, and most of these companies are barely that old. But it is also the market beginning to notice that spectacular success at discovery has not yet produced a single approved medicine.
The bet worth watching
For an investor, this should reframe the whole opportunity. The AI money has crowded into the tractable problem—molecule generation—where returns are competed down precisely because so many firms can now do it, and where, by the numbers above, success was never the binding constraint. The larger prize lies at the other end of the pipeline: using AI to attack the Phase 2 and Phase 3 problem directly: validating which targets actually drive disease, finding biomarkers that predict who will respond, designing smarter and cheaper trials. A company that made even a dent in the efficacy failure rate would capture far more value than another molecule-design platform, because that failure rate is what the $2.6 billion is mostly paying for.
Whether AI can reach into that harder problem is the open question on which the entire bet now turns, and it is a genuinely open one. It may be that better data and new methods let AI begin to predict efficacy, in which case the value there dwarfs anything in discovery. It may be that the biology is a limit of knowledge no amount of computation can shortcut, in which case much of the discovery boom is priced for a breakthrough it will never deliver. No one knows yet. That uncertainty is precisely the point: the smart money should be watching the efficacy frontier, not the crowded molecule-design one, because that is where the question, and value, actually lie.
The subsidies, too
Public subsidies carry a milder version of the same mispricing. The Orphan Drug Act of 1983 pays companies—with exclusivity, a clinical-testing tax credit, and waived fees—to develop drugs for diseases too rare to be profitable, and more than 650 rare-disease therapies have reached patients as a result. Its incentives sensibly target the clinical stage, which is the expensive part and the part AI has not made cheaper. On that score, cheap discovery changes little, as it should.
But it does change two things at the margins. Because generating candidates is now cheap, firms can pursue far more orphan designations, which makes the Act’s familiar gaming— carving a common condition into rare sub-populations to capture exclusivity for what becomes a blockbuster—easier to attempt. And the Act’s economic eligibility test, which lets a sponsor qualify by showing it cannot expect to recover its costs, was written for an era when discovery itself was ruinously expensive. As that cost falls, the test is worth re-examining. Neither point rewrites the economics of the Act. Both are reminders that a subsidy calibrated to one cost structure should be re-checked when the technology changes it.
There is an AI revolution in pharmaceuticals, but so far it has happened where the work is cheapest, not where it matters most. By the end of 2026 the first cohort of AI-designed drugs will have faced the test that counts, and we will know more than we do now about whether machines that are so good at finding molecules can help us find medicines. The value in pharmaceutical AI was never in making the cheap part cheaper. It is in the expensive part nobody has yet cracked. Whether AI can crack it is still, genuinely, an open question.
Author’s Disclosure: The authors report no conflicts of interest. You can read our disclosure policy here.
Articles represent the opinions of their writers, not necessarily those of the University of Chicago, the Booth School of Business, or its faculty.
Subscribe here for ProMarket’s weekly newsletter, Special Interest, to stay up to date on ProMarket’s coverage of the political economy and other content from the Stigler Center.
