After a hype-laden first wave, pharma bets that 2025’s bigger models, better data and bolder deals can finally move molecules from code to clinic.

In the mid‑2010s, a wave of start‑ups promised to turn the slow, labor‑intensive search for new medicines into something closer to software engineering. Algorithms would comb vast chemical spaces, learn from biomedical data and propose fresh drug candidates in months instead of years. Money flooded in. Partnerships with blue‑chip drugmakers proliferated. And then—biology didn’t cooperate. As of 2025, no medicine designed by artificial intelligence has won regulatory approval, and only a handful have made it beyond early‑stage trials. The mismatch between marketing and measurable impact has been stark, exposing just how hard it is to tame physiology and the long, expensive grind of clinical testing.

That comedown reshaped the sector. Several high‑profile players retrenched, merged or delisted. BenevolentAI, once a poster child for the field, has pivoted and quit public markets. Exscientia, another early standard‑bearer, agreed to combine with Recursion, pooling compute‑heavy cell imaging and model‑driven chemistry under one roof. Layoffs and pipeline cuts have been common as investors demanded evidence, not promises. Yet the story didn’t end there. A second act is underway—and the bets are bigger.

The change starts with tools. In 2024, researchers unveiled AlphaFold 3, a step‑change model that predicts not just protein structures but how proteins, ligands, nucleic acids and ions interact. In parallel, a new class of generative “foundation” models such as ESM3 emerged for protein design, trained on billions of sequences and increasingly conditioned on structure and function. These systems don’t magically yield drugs, but they improve the fidelity of early hypotheses and narrow the search space. That is helping executives reframe AI not as an oracle, but as a way to stack the odds earlier in discovery.

Compute is the second pillar. The most ambitious platforms now run on supercomputers that rival academic HPC centers. Recursion’s BioHive‑2, for example, stitches together hundreds of H100 GPUs into a pharmaceutical‑grade DGX SuperPOD that earned a place on the TOP500 list of the world’s fastest systems. That horsepower, coupled with industrial‑scale microscopes and automated wet labs, lets companies train multi‑modal models spanning cell images, chemical graphs and patient‑level outcomes—exactly the kind of heterogeneous signals that tripped up the first wave.

Money and alliances form the third leg. Alphabet’s Isomorphic Labs has signed multi‑target discovery deals with Eli Lilly and Novartis that together top out near $3 billion in potential milestones. Insitro deepened its collaboration with Lilly to build models that steer small‑molecule design and siRNA delivery. In China, XtalPi has parlayed AI‑enabled chemistry into blockbuster agreements and a public listing. The tenor of these pacts has shifted from one‑off pilots to multi‑asset pipelines, often with shared risk through later stages.

Meanwhile, the first clinical green shoots have appeared. Insilico Medicine reported Phase 2a data in idiopathic pulmonary fibrosis for rentosertib (formerly ISM001‑055), a TNIK inhibitor where both target and compound were identified using generative AI. Over 12 weeks, patients on the highest dose saw a mean improvement in forced vital capacity compared with a decline on placebo—encouraging, if still early. It remains true that no AI‑designed drug has yet reached Phase 3, let alone approval. But executives argue that this is precisely what a reality‑based ramp looks like in drug R&D: patient results first, headlines later.

Regulators are also catching up. In January this year, the U.S. Food and Drug Administration published draft guidance on using AI to support regulatory decision‑making for drugs and biologics, laying out a risk‑based framework for establishing model credibility in a stated context of use. In Europe, the EMA adopted a reflection paper in 2024 on AI across the medicines lifecycle. Neither document green‑lights algorithmically designed molecules; rather, both demand transparency about data provenance, validation and monitoring. That may raise the bar, but it also gives sponsors a clearer path to compliance.

So why did the first wave stumble—and why might this one fare better? Start with data. Many early start‑ups trained on sparse, biased or heterogeneous public datasets, then overfit to narrow tasks such as virtual screening on curated targets. As models met fresh biology, performance crumbled. Today’s platforms ingest richer modalities—live‑cell microscopy, single‑cell multi‑omics, proteomics and increasingly real‑world evidence—and link them with causal inference rather than pure pattern‑matching. That doesn’t solve biology, but it reduces the odds that models chase misleading correlations.

Next, incentives. In 2017 or 2018, it was tempting to chase press‑friendly “me‑too” programs, hoping to show cycle‑time wins. The new bets are bolder and more target‑centric: protein‑protein interactions, molecular glues, degradation pathways, RNA therapeutics, immune synapse tuning. They also come with crisper decision rules set upfront with partners, limiting zombie programs. Crucially, the economics have matured. Where earlier collaborations paid small option fees for research widgets, 2025‑era deals fund multi‑year platform access, data generation at scale and milestone‑laden development plans.

None of this removes risk. The sector’s most hyped models are increasingly closed, spurring debates over reproducibility. AlphaFold 3’s server‑only access, capped for many users, has already created friction in academia. If generative models become proprietary black boxes, regulators and peer scientists will demand more wet‑lab validation, slowing the loop. Biobanks and patient datasets raise privacy and consent questions that open‑source zeal cannot wish away. And even perfect target‑binding predictions can’t foresee idiosyncratic human toxicity or clinical behavior. As the first wave learned, statistical elegance does not guarantee pharmacology.

What would count as proof that this second act is different? Three signals to watch over the next 24 months. First, a genuine, independently replicated Phase 2 success with disease‑modifying impact, not just biomarker nudges. Second, evidence that foundation models are compressing attrition—fewer dead ends per candidate—showing up as leaner pipelines that still deliver steady readouts. Third, procurement: whether big pharmas build BioHive‑like stacks in‑house or double down on external platforms. Where compute and data sit may decide who captures value.

The audacious vision—learning systems that map human biology well enough to design drugs as predictably as chips—remains distant. But the second act looks less like a magic trick and more like engineering: big models disciplined by big experiments, run on big computers, governed by clearer rules. The first wave overpromised and under‑delivered. The current one is making quieter, more measured bets that, taken together, may finally bend the curve on time, cost and probability of success. If AI’s drug‑discovery revolution arrives, it will look like a thousand sober decisions that add up to something transformative.