Medicine becomes a design problem
AI made the first step of drug discovery cheap. The money goes to whoever owns the data, the lab loop and the clinical asset, not the model.
Neil Gaikwad
The size of the prize
What the consensus misses
The popular version of this story is that AI designs better molecules and drug development gets cheap. The data says something narrower. AI-discovered molecules sail through Phase I, which tests safety, at 80 to 90%. In Phase II, which tests whether the drug works, they succeed about 40% of the time, which is what drugs have always done. The expensive failure is still biology, not chemistry.
So I do not think the model is the moat. Structure prediction is open, protein generation is crowded, and Lilly now hands biotechs access to models trained on data it says cost more than $1B, in exchange for their data. When the largest buyer gives the model away to get the data, the data is the asset.
The other thing I discount is the partnership headline. Recursion and Exscientia had received about $450M in cash against more than $20B in potential milestones. Isomorphic's first two pharma deals were $82.5M upfront against roughly $2.9B in milestones. The real signal is upfront cash and clinical readouts, nothing else.
Why now
The science got its proof. AlphaFold's creators won the 2024 Nobel Prize in Chemistry, and AlphaFold is used by more than two million researchers.
The first fully AI-discovered drug is in a pivotal trial. Insilico's rentosertib showed a lung function gain in a 71-patient Phase IIa and started a 320-patient Phase III in July 2026. Generate Biomedicines' AI-designed antibody is in Phase 3 for asthma.
Regulators opened a door. In April 2025 the FDA announced a plan to phase out required animal testing for antibodies and other drugs, inviting AI models and organoid data instead. And public markets reopened: Generate raised $400M in a Nasdaq IPO, Insilico about $293M in Hong Kong.
Where the value pools
| Layer | My call | Why |
|---|---|---|
| Foundation models for biology | Commoditizes | Open structure models and crowded protein generation; the largest pharma is giving model access away for data. |
| Proprietary data and autonomous labs | Most value | Differentiated biological data is scarce, and lab throughput is what makes models better. |
| Full-stack AI biotechs with assets | Durable value | Owning the molecule through approval is where the economics of a drug actually land. |
| Clinical trial operations | Underpriced | Trials are the largest cost and time sink, and Phase II success has not moved. Better execution compounds. |
| Big pharma | Incumbents keep it | Late-stage trials, approval and selling still run through them, funded by obesity-scale cash. |
Who is winning so far
| Company | What it does | Stage and last round |
|---|---|---|
| Isomorphic Labs | DeepMind spin-out designing drugs; partners Lilly, Novartis, J&J | $2.1B Series B led by Thrive Capital May 2026 · aap.com.au |
| Xaira Therapeutics | AI discovery and development, team behind RFdiffusion | Over $1B committed, led by ARCH and Foresite Apr 2024 · biospace.com |
| Insilico Medicine | End-to-end generative discovery; rentosertib in Phase III | ~$293M Hong Kong IPO Dec 2025 · pharmaphorum.com |
| Generate Biomedicines | Generative antibody design; asthma antibody in Phase 3 | $400M Nasdaq IPO Feb 2026 · pharmaphorum.com |
| Chai Discovery | Computer-aided design suite for molecules | $130M Series B at $1.3B, co-led by Oak HC/FT and General Catalyst Dec 2025 · builtinsf.com |
| Lila Sciences | AI models paired with autonomous labs | Valued above $1.3B; $550M raised in total Oct 2025 · finance.yahoo.com |
| Profluent | Frontier protein design; open-source gene editor OpenCRISPR-1 | $106M co-led by Altimeter and Bezos Expeditions Nov 2025 · siliconangle.com |
| Recursion | Industrial AI discovery with a phenomics lab; merged with Exscientia | Public; cut 20% of staff, cash runway into Q4 2027 Jun 2025 · biospace.com |
Where I would write a seed check
Perturbation and single-cell data for "virtual cell" models. Models commoditize; data that cost real money to generate does not.
Early examples: Tahoe TherapeuticsRobots plus experiment design close the design, make, test loop that improves models fastest.
Early examples: MedraPatient matching, site selection and trials run inside routine care attack the stage where AI has not yet helped.
Early examples: Paradigm HealthThe FDA's 2025 plan invites computational and organoid data in filings. Whoever packages it for regulators gets paid by every developer.
How it plays out
- 2026
Rentosertib's Phase III starts and Isomorphic raises $2.1B. AI biology moves from demos to pivotal trials.
- 2028
The verdict years. Rentosertib and Generate's Phase 3 read out, and Recursion must show results before its cash runs out.
- 2031
My expectation: the first approvals of drugs discovered end to end by AI, and FDA guidance that accepts non-animal preclinical packages broadly.
- 2036
My expectation: AI-first discovery is standard at every top-20 pharma, and the scarce asset is clinical-grade biological data.
As of Oct 2026
Risks
- No AI-discovered drug is approved yet; the most advanced started Phase III only in mid-2026, with a 52-week endpoint. geneonline.com
- Partnership headlines are mostly contingent: about $450M received against more than $20B in potential milestones at Recursion and Exscientia. sec.gov
- Clinical attrition hits AI platforms too: Recursion dropped programs and cut 20% of staff in 2025. biospace.com
- About 3,500 FDA staff were cut in April 2025, with expected effects on drug review. hlc.com
- Early biotech funding is tight: seed and Series A fell 15% to $4.4B in the first half of 2026. biocentury.com
Sources
- IQVIA, Global use of medicines outlook
- Deloitte, Measuring the return from pharmaceutical innovation (2025)
- Jayatunga et al., Drug Discovery Today (2024)
- Lilly, Q4 2025 results
- GlobalData, The next patent cliff (Jul 2025)
- Google DeepMind, Nobel Prize in Chemistry 2024
- Isomorphic Labs, $2.1B Series B (May 2026)
- GeneOnline, Rentosertib enters Phase III (Jul 2026)
- FDA, Plan to phase out animal testing (Apr 2025)
- BioPharmaTrend, Lilly TuneLab (Sep 2025)
- pharmaphorum, Generate IPO (Feb 2026)
- BioCentury, Seed and Series A in 1H26
Nothing here is investment advice: these are my own notes on public information, written to sharpen judgment, not to recommend buying or selling anything.