Last week, Insilico Medicine and Eli Lilly announced an AI-driven drug development partnership worth up to $2.75 billion — one of the largest deals yet between a pharmaceutical giant and an AI-native drug discovery company. Insilico's shares surged. The announcement was treated as a milestone. What it actually represents is harder to assess: another major bet on a technology whose clinical track record is still being written.
The AI drug discovery sector has been accumulating commitments at a pace that would be striking even for a field with proven returns. Chai Discovery, a startup that spun out of OpenAI's offices, closed a Series B last year at a $1.3 billion valuation and signed its own partnership with Lilly, projecting first-in-class medicines in clinical trials by end of 2027. Proxima raised $80 million to design drugs targeting protein interactions. Earendil Labs is reportedly considering a Hong Kong IPO. Isomorphic Labs, Google DeepMind's pharmaceutical spin-off, published a technical report in February on IsoDDE, its new drug-discovery engine, drawing comparisons to AlphaFold in the precision of its predictions about how proteins interact with candidate drugs.
The fundamental promise has not changed since AlphaFold's protein structure predictions electrified the field in 2021: AI can navigate the combinatorial space of potential drug candidates faster and more cheaply than traditional methods, flagging the molecules most likely to bind to a target while avoiding the ones that will fail in toxicity screening. The computation that once took months can now take days. The question is what happens next.
The clinical record is thin and should be read carefully. Insilico's most advanced compound, an AI-designed drug for idiopathic pulmonary fibrosis, entered Phase II trials in 2023. That is genuinely notable — it is among the first AI-designed small molecules to reach human trials. It is also still in trials. The gap between a molecule that looks promising in silico and one that improves outcomes in patients remains the same gap it has always been: biology is hard, clinical trials are long, and most drugs fail.
The IsoDDE paper from Isomorphic Labs illustrates both the progress and the framing challenge. The model's predictions of protein-drug and antibody interactions are measurably better than prior AI approaches, which is real scientific advance. But the measure is prediction accuracy on known structures — a step removed from the harder question of whether a predicted interaction will produce a safe, effective therapy. Scientists who called it "an AlphaFold 4 moment" are expressing genuine excitement about the underlying capability. Whether that capability translates to approved drugs in the 2020s or the 2030s is a separate question.
What the current wave of deals does establish is that pharmaceutical companies are treating AI-native drug discovery as a serious commercial option rather than a research curiosity. The Insilico-Lilly partnership is structured with milestone payments contingent on clinical progress — a structure that puts real financial stakes on the AI's ability to produce drugs that work in humans, not just drugs that look good in models. That is a meaningful test. Sam Altman has said OpenAI may consider taking royalties from drug discoveries made using its AI, signalling that the major labs see AI pharma as a potential revenue stream, not just a charitable application.
The honest version of where this stands: the infrastructure for AI drug discovery has been built, the money has arrived, and the first generation of AI-designed compounds is now in human trials. The next three to five years will produce real data on whether the fundamental promise holds up in the clinic. Until then, the sector is running on well-funded anticipation — which is not the same as proof.