A field that spent a decade promising to reinvent how medicines are found has, in the span of three years, produced its first real evidence — and its first record-breaking funding rounds. The gap between those two things is the story.

In May 2026, an Alphabet spin-out called Isomorphic Labs raised $2.1 billion in a single round — one of the largest private financings any drug-discovery company has ever taken. It is led by a Nobel laureate, backed by some of the deepest pockets in technology, and built on the most celebrated AI system in the life sciences. As of that raise, it had not yet given a single dose of any drug to a single human patient.

That sentence captures the strange moment AI-driven drug discovery finds itself in. The technical capabilities are real and, in places, genuinely astonishing. The capital is flowing at a scale the sector has never seen. The biggest names in computing — NVIDIA, Google, Microsoft, Amazon, OpenAI, Anthropic — have all planted flags. And yet the thing that would actually validate the entire enterprise — an approved medicine discovered and designed by AI — does not exist. Not yet.

This briefing traces what changed between the “GPT moment” of late 2022 and the first half of 2026: the leap in what the models can do, the new toolkit that emerged on the way to the so-called AGI era, the extraordinary capital and corporate manoeuvring of early 2026, and the one fragile clinical datapoint that the field’s credibility currently rests on. It is written for two readers at once — the scientist who wants to know what the models actually do, and the investor who wants to know where the money and the moats are.

The defining technical fact of this period is that AI in the life sciences stopped being about predicting the shape of a single protein and became about generating, at atomic resolution, almost any biomolecule and the interactions between them.

AlphaFold2 had essentially solved single-chain structure prediction in 2021. The post-GPT wave extended that foundation in every direction. AlphaFold3, unveiled by Google DeepMind and Isomorphic Labs in May 2024 and published in Nature, replaced the older architecture with a simplified “Pairformer” module and a diffusion-based generative head that outputs 3D coordinates for proteins together with DNA, RNA, ions and small-molecule ligands — reporting at least a 50% accuracy improvement on some protein-interaction classes and large gains on the antibody–antigen complexes that had bedevilled its predecessor. In October 2024, the work was recognised with the Nobel Prize in Chemistry: Demis Hassabis and John Jumper for structure prediction, and David Baker for computational protein design.

Because AlphaFold3 launched with restricted code, an open-source counter-movement formed almost immediately. MIT’s Boltz-1 and Chai Discovery’s Chai-1 reproduced AlphaFold3-class structure prediction for anyone to use. The more consequential step came in June 2025 with Boltz-2, which predicts not just structure but binding affinity — how tightly a candidate drug grips its target — in roughly twenty seconds on a single GPU, approaching the accuracy of physics-based free-energy methods that cost orders of magnitude more compute. Recursion’s chief R&D officer called binding affinity “core to developing a therapeutic start to finish”; getting it cheaply is a quiet turning point.

On the design side, David Baker’s lab at the University of Washington moved from RFdiffusion (de novo protein backbones, 2023) through enzyme active-site design to RFdiffusion3, open-sourced in December 2025 — an all-atom diffusion model roughly ten times faster than its predecessor, deliberately released with training code as an open contrast to the proprietary engines. Protein language models advanced in parallel: EvolutionaryScale’s ESM3, a 98-billion-parameter multimodal model, generated a novel fluorescent protein its creators estimated as equivalent to 500 million years of evolution. And genome-scale models arrived: the Arc Institute’s Evo 2, with 40 billion parameters and a one-million-nucleotide context window, was trained on more than 9.3 trillion nucleotides spanning all domains of life.

If 2023–24 was about structure and design, 2025–26 added two genuinely new categories: models that try to simulate the behaviour of a whole cell, and AI “scientists” that attempt to run the discovery loop themselves.

The virtual cell ambition is to predict how a cell responds to a perturbation — a drug, a gene knockout — without running the experiment. The Arc Institute’s State model was trained on observational data from 167 million cells and perturbational data from over 100 million more. The Chan Zuckerberg Initiative released rBio, a reasoning model that learns from virtual-cell simulators. To benchmark the field, Arc launched a Virtual Cell Challenge in June 2025, explicitly modelled on the CASP competition that drove protein folding; it drew over 1,200 teams from 114 countries. The honest headline from its December 2025 results, though, was sobering: the best perturbation models are not yet reliably beating naive baselines across all metrics. This is a frontier, not a solved problem.

The agentic thread is further along in narrow cases. FutureHouse’s Robin system, published in Nature in 2026, autonomously generated hypotheses, designed experiments and analysed the results to identify an existing glaucoma drug as a candidate for dry age-related macular degeneration — with humans executing the bench work. Google DeepMind’s Gemini-based AI co-Scientist generated drug-repurposing hypotheses for leukaemia and liver fibrosis. Antibody design also took a striking step: Chai Discovery’s Chai-2 reported a 16% hit rate for fully de novo antibody design in mid-2025 — described as an over-hundredfold improvement on prior methods — producing validated binders in under two weeks for half the novel targets it was tested against.

It is worth keeping these capabilities in proportion. OpenAI’s life-sciences model, launched in 2026, posts a pass rate around 27.5% on one medicinal-chemistry benchmark — meaning it fails roughly three-quarters of the tasks. Impressive trajectory; far from autonomy.

The capability gains of the last three years did not come from one place. They came from an unusual ecosystem in which nonprofit research institutes, university labs and venture-backed startups often publish, compete and borrow from one another in the open. Three groups are worth knowing.

Research institutes

The Arc Institute, a Palo Alto nonprofit co-founded by Patrick Hsu, has become the centre of gravity for genome and virtual-cell foundation models — the Evo and Evo 2 DNA models, the State perturbation model, and the CASP-style Virtual Cell Challenge all originate here. The Chan Zuckerberg Initiative and its Biohub run the parallel virtual-cell effort (the rBio and TranscriptFormer models) and in 2025 absorbed the team behind EvolutionaryScale’s ESM protein models. And the UW Institute for Protein Design, home of David Baker’s lab, remains the world leader in de novo protein design, producing the RFdiffusion family that much of the field now builds on. Public infrastructure underpins all of it: EMBL-EBI co-hosts the AlphaFold database of over 214 million predicted structures.

Startups

The startup layer is where the capital has concentrated, and it sorts roughly into four technical lanes: structure and design foundation models, generative chemistry, phenomics and perturbational biology, and translational prediction. The table below profiles the best-capitalised names and what distinguishes each.

Iambic Therapeutics and Genesis Therapeutics round out the well-funded generative-chemistry cohort. What unites the leaders is less a single algorithm than a strategy: pairing proprietary models with proprietary experimental data, increasingly inside closed design-build-test loops.

University laboratories

Academic labs remain the source of much of the field’s open foundational work — and much of its talent pipeline into the startups above. The Baker Lab at the University of Washington is the standout, but MIT (where Regina Barzilay and Tommi Jaakkola’s group produced the open Boltz models) and Stanford (Brian Hie, a co-developer of Evo) are central to the open-model ecosystem, with further Evo 2 collaborators at UC Berkeley and UCSF. The pattern is circular and deliberate: labs publish open models and train researchers; startups hire those researchers and raise capital to industrialise the work; and the resulting tools flow back into the labs.

The first half of 2026 saw capital and corporate strategy converge on a single pattern: bring together compute (NVIDIA), a frontier AI lab (OpenAI, Anthropic, Google DeepMind) and proprietary pharmaceutical data into integrated, “lab-in-the-loop” platforms. The deals came fast.

The single biggest private raise was Isomorphic Labs’ $2.1 billion Series B in May 2026, led by Thrive Capital alongside Alphabet and sovereign funds, taking its total external capital to roughly $2.6 billion. On the compute-plus-pharma axis, NVIDIA and Eli Lilly announced a co-innovation lab committing up to $1 billion over five years, while Merck signed a deal worth up to $1 billion with Google to deploy agentic AI across its R&D and operations — a collaboration Merck’s digital chief said was expected to run “at least a decade.”

The frontier labs moved decisively into the sector. Novo Nordisk partnered with OpenAI, which launched GPT-Rosalind, its first purpose-built life-sciences model. Anthropic — the maker of Claude — acquired the biology-AI startup Coefficient Bio for a reported ~$400 million, launched Claude for Life Sciences and Claude for Healthcare, recruited Novartis’s chief executive to its board, and signed Bristol Myers Squibb for enterprise-wide adoption. Meanwhile public markets reopened to the sector: Generate:Biomedicines and Eikon Therapeutics each completed IPOs in early 2026, and Insilico Medicine had become the first AI biotech on the Hong Kong main board weeks earlier.

Underneath the headline names sits a steep funding pyramid. A handful of platforms — Xaira (launched in 2024 with over $1 billion), Eikon, Isomorphic, Recursion, insitro, Insilico, Generate and Chai — hold the great majority of the capital, while the median round is far smaller. That concentration is itself a risk: a single high-profile clinical failure could swing sentiment for the whole category.

For all the models and money, the entire edifice currently rests on a remarkably thin clinical base. The landmark is rentosertib, a treatment for idiopathic pulmonary fibrosis from Insilico Medicine — the first drug with both an AI-discovered target and an AI-designed molecule to report a peer-reviewed clinical result, published in Nature Medicine in June 2025.

The Phase 2a result was genuinely encouraging. Over 12 weeks, patients on the 60 mg daily dose showed a mean improvement in lung function (forced vital capacity) of +98.4 mL, against a −20.3 mL decline for placebo — a meaningful divergence in a disease defined by relentless lung-function loss.

But the caveats matter as much as the numbers. This was a small trial — 71 patients across 22 sites in a single country — over just 12 weeks. Insilico’s own chief executive said the result “warrants larger and longer studies.” It is proof that the approach can produce a clinically active molecule. It is not yet proof that AI-designed drugs are safer, more effective, or more likely to win approval than those found the old way.

The backdrop makes the stakes plain. Historically, only about 7.9% of drug candidates that enter Phase I trials go on to win approval; more recent estimates put the figure even lower. Whether AI materially lifts that number is the single most important open question in the field — and rentosertib’s early readout is, so far, almost the only relevant datapoint. The best-funded player of all, Isomorphic, had still not dosed a patient as of its $2.1 billion raise, with first trials pushed toward the end of 2026.

Strip away the hype and a realistic picture emerges. The near-term impact of AI on drug discovery is not miracle cures; it is R&D productivity. Executives describe AI compressing the earliest, most attrition-prone stages of discovery. AstraZeneca’s US oncology head spoke of “more than 50% faster target drug design and validation” where AI has been applied. Insilico reports nominating drug candidates in 12–18 months rather than several years, synthesising a fraction of the usual molecules. Novartis’s research head captured the consensus tone: potentially years saved in discovery, but “not a magic panacea” — AI can replace some bench science, not all of it, and it cannot shortcut the human safety trials where most drugs still fail.

The structural consequence is a hybrid “build, buy and partner” model hardening across the industry. Pharma companies are building internal AI capability, acquiring AI startups outright, and simultaneously partnering with both frontier labs and compute providers. The likely endpoint is consolidation: fewer, better-capitalised, vertically integrated “AI-native” biotechs with their own pipelines, and a small number of platform-and-compute alliances underpinning the rest.

Regulators are adapting in parallel. The FDA issued its first draft guidance on AI in drug development in January 2025, proposing a risk-based “credibility” framework keyed to how the AI is actually used, with final guidance expected around mid-2026; it followed with joint AI principles alongside the European Medicines Agency in early 2026. A pragmatic carve-out matters here: if an AI discovers a drug but conventional methods confirm its safety and efficacy, the regulatory burden specific to the AI may be modest.

The honest summary of AI drug discovery in mid-2026 is that it has crossed from pure promise into partial proof. The models can now design proteins, antibodies and enzymes that work in a dish; the capital to industrialise them has arrived; the biggest companies in technology and pharma are committed for the long haul. What remains unproven is the only thing that ultimately counts — that these tools deliver approved medicines, faster and more reliably than the methods they aim to replace. The next two years, when the first AI-native drugs face controlled clinical trials, will start to answer that. Until then, the models and the money are necessary. They are not yet sufficient.