The AI Drug Revolution Has a Phase 2 Problem

Sometime before the end of this year, a cancer drug that no human chemist designed will be injected into a human being for the first time. The molecule was conceived inside Isomorphic Labs — the drug-discovery company Alphabet spun out of Google DeepMind — and it represents the moment the AI drug-discovery story stops being a slide deck and becomes a clinical fact.

The capital has already arrived ahead of it. In early 2026, Isomorphic stacked two partnerships worth nearly $3 billion in combined headline value: Eli Lilly committed $45 million upfront and more than $1.7 billion in milestones, and Novartis added $37.5 million upfront against roughly $1.2 billion in potential payments. In February, Isomorphic unveiled its drug-design engine, IsoDDE — a system one prominent computational biologist described as "a major advance, on the scale of an AlphaFold 4." Demis Hassabis, fresh off a Nobel Prize for the protein-folding breakthrough that started all this, now runs a company valued around $2.5 billion with seventeen programs in oncology, immunology, and cardiovascular disease.

The market has taken the hint. AI drug-discovery names have been bid up on the premise that machine intelligence is about to do to pharmaceuticals what it did to protein structure: collapse a decade of work into months and rewrite the economics of an entire industry.

Here is the problem with that trade. AI has gotten extraordinarily good at the part of drug-making that was never the expensive part.

What AI Actually Compresses

Drug development has two halves, and they are nothing alike.

The first half — discovery and preclinical work — is about finding a molecule that binds the right target. This is where AI lives, and where it genuinely shines. Predicting protein structures, generating novel compounds, screening billions of candidates in silico, optimizing for properties that once took years of wet-lab iteration: these are pattern-recognition problems, and pattern recognition is exactly what these systems were built for. Timelines that ran four to six years are being compressed toward eighteen months. That is real, and it is not hype.

The second half — clinical trials — is where money goes to die. Phase 1 tests safety. Phase 2 asks the only question that ultimately matters: does the drug actually work in sick people? And Phase 2 is a slaughterhouse. Historically, roughly two-thirds of drugs that enter it fail, overwhelmingly for lack of efficacy — the molecule does what it was designed to do at the molecular level and the patient gets no better.

Nothing about training a larger model changes a tumor's biology. A beautifully designed compound that hits its target perfectly is still subject to the brutal, irreducible uncertainty of human physiology. AI can tell you a key fits a lock. It cannot tell you the lock opens the door you actually need opened.

The Industry Already Learned This the Hard Way

This is not a theoretical caution. The first generation of AI-first biotechs already ran straight into it.

BenevolentAI, once a £1.5 billion market darling, watched its lead drug — an AI-discovered atopic dermatitis treatment — meet its safety endpoint and then fail to beat placebo on efficacy. The company cut roughly a third of its workforce. Exscientia, the British pioneer that promised AI-designed drugs at a fraction of the cost, discontinued its lead immuno-oncology program after concluding it couldn't reach a viable therapeutic window; it ultimately folded into Recursion. Recursion itself reported limited efficacy on an early candidate and restructured its pipeline.

The lesson the sector quietly absorbed in 2024 and 2025 — what some now call the "post-hype reality verification phase" — is blunt: AI accelerated discovery, but the probability of clinical failure looks fundamentally unchanged. The robots got the companies to the starting line faster. The race itself is just as long, and just as fatal.

Which raises the question every investor in this theme should be asking right now, with the first AI-designed molecules finally reaching the clinic: if the discovery edge is real but the clinical risk is not solved, where does the money actually accrue?

This is where the analysis gets actionable. AlphaBriefing members get the full investment framework — scenarios, positioning, and the bottom line.

Subscribe to AlphaBriefing — Free, Member, and Paid tiers available.