Industry View · Pharma & Biotech
2025 was the year AI in drug discovery moved from slideware to clinical readouts: the first molecule with both target and compound designed by AI posted a positive Phase IIa, and capital followed at scale, with Isomorphic Labs raising $2.1B in May 2026. But the inconvenient truth is that AI has yet to dent the industry's ~90% end-to-end clinical failure rate, and most of the value is still a 2030 promise, not a 2026 fact.
Discovery-stage signals but Phase II ~40%, no better than traditional; late-stage unproven
Foundation models on omics and structure data nominate and validate targets that were previously intractable.
De novo design of small molecules and antibodies compresses hit-to-lead from years to months.
Surrogate models predict ADMET, toxicity and binding, trimming wet-lab cycles before candidates advance.
AI speeds patient matching, site selection and statistical programming, but has not improved late-stage success.
FDA's Jan-2025 draft guidance on AI in regulatory decisions begins to define how models support submissions.
01 · The thesis
The bull case rests on a real inflection. AlphaFold has predicted 200M+ protein structures, generative models now design antibodies and small molecules de novo, and AI-native pipelines report Phase I success near 80-90% against a historical ~50%. Big pharma is no longer dabbling: Eli Lilly is standing up an AI factory on 1,016 Nvidia Blackwell Ultra GPUs (>9,000 petaflops) and has signed 16 AI deals since 2025, including a $2.75B pact with Insilico Medicine. The bear case is that the bottleneck was never the early stages. Discovery and Phase I are where AI helps most, yet roughly 70% of R&D spend and the bulk of attrition sit in late-stage development, where AI has not moved the ~90% failure rate. The market is also small relative to the hype — on the order of $3-4B in 2025-2026 — meaning today's valuations price a future that still has to clear Phase II and III. The next two years separate platforms with assets from platforms with decks.
Foundation models on omics and structure data nominate and validate targets that were previously intractable.
De novo design of small molecules and antibodies compresses hit-to-lead from years to months.
Surrogate models predict ADMET, toxicity and binding, trimming wet-lab cycles before candidates advance.
AI speeds patient matching, site selection and statistical programming, but has not improved late-stage success.
FDA's Jan-2025 draft guidance on AI in regulatory decisions begins to define how models support submissions.
02 · The two clocks
The proof clock. The first AI-origin drug, Insilico's Rentosertib, cleared Phase IIa with a +98mL forced-vital-capacity gain versus a -20mL placebo decline — but Phase II success for AI drugs still sits near 40%, statistically no better than traditional methods. The next 18 months of Phase II/III readouts decide whether the early-stage edge survives biology. The capital clock. Money is arriving faster than assets mature: Isomorphic's $2.1B raise, Xaira's $1B launch and Lilly's $2.75B Insilico pact are bets on platforms that, in most cases, have yet to register an approval. Cash runways — Recursion's stretches into 2027 — set hard deadlines for clinical validation. The compute clock. Infrastructure is being poured now: Lilly's 1,016-GPU AI factory and Nvidia BioNeMo deals across Roche, AstraZeneca, GSK and Novo Nordisk mean the discovery engine is scaling years ahead of the regulatory framework, which the FDA only began drafting in January 2025.
03 · Public players & exposure
We plot the listed players on two editorial axes — how exposed each is to AI disruption, against how ready its data, brand and position are to be the answer. The figures in the table are sourced; the placement is our read.
04 · Private flagships
The companies attacking this industry AI-first, with disclosed funding where available:
Co-developed AlphaFold 3; aims to design drugs for any target and is moving its own programs toward first-in-human trials.
Delivered the first clinical proof-of-concept for a fully AI-designed drug and listed on the Hong Kong exchange in Dec 2025.
Combines first-principles molecular simulation with ML; its TYK2 program reached Phase III via partner Takeda.
Built on RFdiffusion/RFantibody to design antibodies and proteins from scratch; led by ex-Genentech's Tessier-Lavigne.
Industrialised cell-image screening; post-Exscientia merger it pruned to a focused ~10-program pipeline with big-pharma milestones.
The de-facto compute and model platform under most large-pharma AI factories, turning drug discovery into a buyer of GPUs.
05 · Signals
06 · The exposure read
AI rewards clean, structured advantage and punishes friction. The line runs through who owns the data, the brand and the customer — and who is merely a step the technology can route around.
Sources
Landmark AI-pharma capital commitments (2024-2026)
Disclosed raises and deal values; milestone-laden pacts shown at headline figure. Sources: Forbes, TechCrunch, Fierce Biotech, CNBC, pharmaphorum. ($B)
FDA drafts AI guidance
The FDA published its first framework for using AI to support regulatory decisions across the drug lifecycle, beginning to formalise the rules of the game.
First AI-origin Phase IIa win
Insilico's Rentosertib posted positive idiopathic-pulmonary-fibrosis data in Nature Medicine — the first clinical proof for an AI-designed target and molecule.
Lilly builds a pharma supercomputer
Eli Lilly and Nvidia announced an AI factory on 1,016 Blackwell Ultra GPUs, signalling that incumbents intend to own the compute, not rent it.
Isomorphic's $2.1B raise
The largest single financing in AI drug discovery to date reset the funding ceiling and validated the DeepMind-origin platform thesis.
173 AI programs in the clinic
AI-origin clinical candidates scaled from a handful a decade ago to 173 in 2026, shifting the debate from 'can AI design drugs' to 'will they survive Phase II/III'.
Co-developed AlphaFold 3; aims to design drugs for any target and is moving its own programs toward first-in-human trials.
Delivered the first clinical proof-of-concept for a fully AI-designed drug and listed on the Hong Kong exchange in Dec 2025.
Combines first-principles molecular simulation with ML; its TYK2 program reached Phase III via partner Takeda.
Built on RFdiffusion/RFantibody to design antibodies and proteins from scratch; led by ex-Genentech's Tessier-Lavigne.
Industrialised cell-image screening; post-Exscientia merger it pruned to a focused ~10-program pipeline with big-pharma milestones.
The de-facto compute and model platform under most large-pharma AI factories, turning drug discovery into a buyer of GPUs.
| Company | Stance | The sourced fact |
|---|---|---|
| Eli Lilly LLY | Compute incumbent | Building an AI factory (LillyPod) on 1,016 Nvidia Blackwell Ultra GPUs (>9,000 petaflops) and has signed 16 AI deals since 2025, including a $2.75B pact with Insilico (CNBC, Fierce Biotech). |
| Roche / Genentech RHHBY | Platform builder | Genentech runs a 'lab-in-a-loop' generative-AI platform with Nvidia and is launching a hybrid-cloud AI factory to industrialise discovery (pharmaphorum, MobiHealthNews). |
| Isomorphic Labs pvt-ISO | Best-funded native | Raised $2.1B in May 2026 (Series B led by Thrive, after $600M in 2025) and holds Lilly/Novartis deals worth nearly $3B; preparing first clinical trials of AI-designed drugs (Forbes, Fortune). |
| Schrodinger SDGR | Physics-ML pick | Its Nimbus-originated TYK2 inhibitor zasocitinib (TAK-279) advanced into Phase III, validating physics-enabled design at late stage (ScienceDirect 2025 landscape). |
| Insilico Medicine pvt-ISM | First clinical proof | Rentosertib, the first drug with AI-designed target and molecule, posted +98mL FVC vs -20mL placebo in Phase IIa (Nature Medicine); cumulative collaboration value $4.6B (Insilico). |
| Recursion RXRX | Consolidation test | Merged with Exscientia in a ~$688M all-stock deal but cut/paused several programs to limit 2025 burn, with cash runway into 2027 (Pharmaceutical Technology, pharmaphorum). |
| Xaira Therapeutics pvt-XAI | Show-me money | Launched in 2024 with $1B from ARCH and Foresite, among the largest initial commitments in ARCH's history, but is still pre-clinical (Fierce Biotech, TechCrunch). |
| AbCellera ABCL | Engine, few hits | Its AI platform helped identify Lilly's COVID antibody bamlanivimab rapidly, but it has struggled to convert the engine into a deep wholly-owned pipeline (BioSpace). |
| Nvidia NVDA | Picks & shovels | Its BioNeMo platform underpins pharma AI factories with Lilly, Roche, Genentech, AstraZeneca, GSK and Novo Nordisk, plus a $1B+ Lilly co-innovation lab (Nvidia, Fierce Biotech). |
| Tempus AI TEM | Data flywheel | Supplies multimodal clinical-genomic data and AI to pharma for trial enrollment and target work, monetising real-world data the discovery platforms lack (company filings, 2025 trial reviews). |