Can AI cure cancer by 2036? Anthropic’s chief executive told investors it is close to certain, and the men who actually run the trials answered him in public this week with two words: not likely. That answer is worth more than every benchmark in the roadshow, because it was paid for in years and in patients rather than in tokens.

Dario Amodei put a deadline on the miracle in a 42-minute investor video released November 10, ahead of Anthropic’s listing. Cancer, Alzheimer’s, heart disease, mental illness — in his telling they become relics of the past in under a decade. He has made the intellectual version of that argument before, in an essay that imagined AI compressing fifty to a hundred years of biology and medicine into five or ten. OncoDaily’s account of the video sets out the claim and the caveat a pitch tends to leave out: cancer is not one disease, and a laboratory result is not a patient outcome.

What changed this week is that the deadline met its opposition on the record.

Dario Amodei — A Decade Is a Deadline Nobody Has to Ship

There is a structural reason to handle AI cure timelines with care, and it has nothing to do with Amodei’s intelligence. It has to do with the calendar on which his claim gets judged. A decade is longer than the fund cycle, longer than the lockup, longer than the tenure of the analyst who writes the note. His own essay is careful and conditional; the roadshow version arrives with a number attached and a share price behind it.

Nobody gets a margin call on a decade. That is the entire appeal of promising one.

Dave Ricks — A Blank Page Is Not a Business Plan

The reply came out of Indianapolis. Eli Lilly’s chairman and chief executive, Dave Ricks, answered on X that the vision would be amazing and that he does not expect it, adding that the map of biology remains mostly unmarked and that randomized proof in humans is the longest pole in the tent (his post). That is a polite sentence doing the work of a demolition order.

Ricks has been saying a version of this for months. In an earlier interview with Fortune he described current models as trained on human language rather than on the language of chemistry, physics and biology — useful at prediction, close to helpless at the part where a molecule has to behave inside a body. His estimate of the ground so far covered: something like a thousandth of the problems in drug discovery. His image for the state of the field was a toddler learning to speak.

He is not a neutral witness and he is not pretending to be. Lilly sells treatments people take for years, and a cure is a worse business than a refill. But an incentive to doubt does not make the doubt wrong. It makes it expensive to wave away, because the company betting against the deadline is also the one that has to pay for the trials either way.

Jeff Leiden — Can AI Cure Cancer by 2036? His 10,000 Compounds Say No

The sharper answer came from Vertex Pharmaceuticals. Executive chairman Jeff Leiden told the Wall Street Journal that his company had worked with roughly fifty AI vendors across a library of about 10,000 compounds and that none of them produced a working drug on its own. His verdict on the cure timelines, as CH HealthTech recorded it, was blunter than Ricks’: complete nonsense.

Fifty vendors. Ten thousand shots. Zero drugs. That is not a rounding error. That is a result.

Leiden’s second point is the one that actually kills the calendar, and it is arithmetic rather than ideology. A trial requiring two years of patient follow-up still requires two years. Speed at the front of the pipeline — finding a candidate faster, ranking it better, designing a cleaner molecule — does not compress the back of it, where human beings swallow a pill and get watched. The bottleneck is not compute. It is consent, enrollment and time.

George Yancopoulos — Fifteen Years to Five Is Wishful Algebra

Regeneron’s co-founder has made the same argument from the other end of the industry, questioning whether AI takes drug development from fifteen years down to five. Strip the sales language off the optimistic case and the compression has to happen twice: once in discovery, once in development. The first is real and already underway. The second is where the years live, and no model has been cleared to shorten it.

Which is the part the cure discourse keeps dropping. AlphaFold and its descendants genuinely changed structural biology. Screening systems genuinely find tumors earlier. AI is doing work in medicine right now, and pretending otherwise is its own kind of dishonesty. The argument is not whether the tools help. It is whether help compounds into a cure for most cancers inside ten years — and the people closest to the data keep landing on the same answer: not on that clock.

The Patients in the Queue — They Pay for the Optimism Either Way

Here is where a timeline stops being a debate between executives. A promise of a cure within a decade changes what a sick person decides today. It changes whether a family borrows against a house for a treatment that exists, or holds out for one that does not. It changes how an eligible patient reads a consent form, and how a hospital budgets its oncology beds for the next five years. When the promise runs early, the cost of the error lands on the people with the least room to absorb it.

The roadshow does not have to be a lie to do damage. It only has to be early, confident, and attached to a valuation.

Amodei may well be right about the shape of the next twenty years. Physics-trained founders have been right about harder things, and the compounding of research capability is not a story anyone should dismiss for comfort. But a decade is a claim tested by patients, not by benchmarks, and the test that matters has not begun: an AI-designed drug clearing approval, a trial arm beating the standard of care, a scan coming back clean because of something a model proposed.

Until that scan exists, 2036 is a slide. The pharmaceutical industry has now said, politely and in writing, that it will be reading the data.

Sources: OncoDaily, Fortune, CH HealthTech, X.