Weekly intelligence on AI reshaping pharma and biotech — for the people making the decisions. Short, opinionated, written from inside the field.
Nobody is asking whether AI works in pharma anymore — they're asking what would count as showing it. EMA's draft Annex 22 bars generative AI from critical GMP work, and four of its own controls are unenforceable against a foundation model. Plus: Schrödinger trims headcount in the quiet repricing of platforms that actually have revenue, Nature Medicine publishes the digital-organism blueprint with every author holding equity, Insilico's CEO argues deal value is the wrong scoreboard, and the Onion Desk.
EMA's draft GMP Annex 22 says generative AI should not be used in critical manufacturing applications — and four of the annex's own controls turn out to be difficult or impossible to apply to a foundation model, not through bad drafting but because the models lack the properties the text assumes. FDA, meanwhile, has already skipped from guidance to enforcement. Also: Schrödinger appears on the layoff tracker while pre-clinical platforms raise billions, a Nature Medicine Perspective proposes an AI-driven digital organism with all three authors declaring equity in the company that would build it, Insilico's CEO draws a hard line at discovery, and the Onion Desk.
The headline deal of the week wasn't about molecules — it was about cloud contracts, agent frameworks, and 25,000 employees. Novo Nordisk named AWS its preferred cloud and strategic AI partner and opened a London co-innovation hub. Plus: a USC model maps brain ageing region by region, FDA's first AI-citing warning letter gets its sharpest industry reading yet, the week's partnership flow was almost entirely AI and nobody noticed, and the Onion Desk.
Novo Nordisk made AWS its preferred cloud provider and strategic AI partner, anchored in a London co-innovation hub — and the one hard number it chose to publish was 25,000 employees and reduced clinical documentation time, not a molecule. That's an operations programme wearing a discovery frame, and it lands the same week FDA's first AI-citing warning letter gets its sharpest industry reading yet. Also: a USC model replaces the single brain-age number with region-by-region maps, a trade publication files four AI partnerships under routine, and the Onion Desk.
Pathos AI spent $125 million buying a clinical asset instead of selling access to its models — and took an AstraZeneca PROTAC off the shelf for undisclosed terms. Plus: Relation launches MORGAN while GSK pays $110M for the fuel rather than the engine, the EU AI Act's transparency rules go live globally, European biotechs aim AI target discovery at respiratory, and the Onion Desk.
Every marquee AI-biotech deal of 2026 has run the same direction — the AI company sells upstream capability and pharma pays for it. Pathos AI just ran it backwards, licensing a TROP2/HER3 bispecific ADC from Alphamab for $125M upfront and taking on AstraZeneca's ERα PROTAC, absorbing full development cost on molecules other people invented. Also this week: GSK pays Relation up to $110M for perturbation data rather than model access, the EU AI Act's Article 50 transparency layer becomes enforceable with an output-based territorial hook, European biotechs push respiratory past broad immunosuppression, and the Onion Desk.
A-Alpha Bio's Atlas Consortium convinced GSK, Boltz, Cradle, and Dyno to jointly fund and share standardized antibody-antigen data — a bet that binding was never the moat. Plus: capital rotates into Chinese biotech as the AI rally stalls, a House-backed trade probe puts the sequencing supply chain in scope, the DOE's Genesis Mission hands national-lab compute to a nano-cap, and The Onion Desk.
A-Alpha Bio launched the Atlas Consortium with GSK, Boltz, Cradle, and Dyno Therapeutics — competitors agreeing to co-fund and pool standardized antibody-antigen binding data rather than each build a private moat. The tell is that two rival model-builders signed on: binding data is being reclassified as shared infrastructure, pushing scarcity downstream to developability and hit-to-lead. Also this week: capital rotates into Chinese drugmakers as the EM AI rally stalls, 10+ House members back a trade probe that would sweep in sequencing and APIs, the DOE's Genesis Mission gives a nano-cap national-lab compute, and a ProteinMPNN paper turns a 96-well plate into a reliable source of better-than-nature enzymes.
Healx dosed its first patient in a Phase 1/2 sarcoma trial, and the bottleneck moves from discovery to trial operations. Plus: Chai Discovery's $400M Series C at $3.8B, Insilico takes generative AI onto the factory floor, Doudna's lab builds a nuclease evolution never made, and The Onion Desk.
Healx dosed its first patient in a Phase 1/2 trial of HLX-4310 for metastatic sarcoma — AI's contribution here was the combination rationale, not a novel molecule. The funnel tells the story: AI-discovered candidates clear Phase 1 at 80–90% against a historical ~50%, then revert to the usual ~40% at Phase 2, which locates the constraint in trial operations rather than discovery. Also this week: Chai Discovery's $400M Series C at a $3.8B valuation, Insilico moving generative AI downstream into manufacturing, and a Doudna-lab nuclease built outside evolved sequence space.
Biomni, Stanford's general-purpose biomedical AI agent, lands in Science with a live user base and a benchmark score nearly double general LLMs. Plus: a universal coordinate system for cell biology in Nature, investors split on whether AI-bio valuations still track clinical reality, and Anthropic's own drug-discovery bet comes into sharper focus.
Stanford's Biomni — a general-purpose biomedical AI agent already running in 10,000+ labs — cleared peer review in Science, executing full research workflows from literature review through experimental design and nearly doubling general-LLM accuracy on a 443-question benchmark. Also this week: Nature's Universal Cell Embedding sets a zero-shot cross-species baseline for cell biology, investors debate whether AI-bio valuations have detached from clinical proof, and Anthropic's own preclinical drug programs come into sharper focus.
Claude Science launches as a research workbench orchestrating 60+ scientific databases and tools, while Anthropic discloses its own internal drug-discovery pipeline for neglected diseases. Plus: Takeda's ~$600M bet on Insilico, an economist's case that AI drug discovery solved the cheap problem, and The Onion Desk.
Anthropic launched Claude Science as a standalone research workbench coordinating specialist agents across 60+ curated scientific databases — and disclosed it's running its own drug-discovery programs for neglected diseases. Early beta results show dramatic time compression: a two-year literature review down to weeks, a germline glioma analysis cut to a tenth of its prior time. Also this week: Takeda's ~$600M platform deal with Insilico, an economist's argument that AI drug discovery bet on the wrong bottleneck, and Bioptimus's CEO on what's actually been delivered.
Medra's Physical AI Lab pairs a multi-agent reasoning layer with DARPA funding and round-the-clock autonomous operation, reframing who generates AI training data in biology. Plus: RQ Bio's $115M flu antiviral Series A, why China — not the FDA — rattled BIO 2026, virtual-cell scaling laws, and The Onion Desk.
Medra launches a reasoning layer atop its autonomous lab, reframing drug discovery as a race to own the data-generation loop, with DARPA already collaborating. Around it: a $115M antiviral raise, a skeptical turn in virtual-cell benchmarking, record biotech M&A, NVIDIA's infrastructure play, and satire on longevity runways and shifting EU AI Act deadlines.
At Fortune Brainstorm Tech, NVIDIA's healthcare chief reframes AI drug discovery's core constraint as regulatory acceptance, not scientific capability. Plus: Novartis expands its molecular-glue AI deal to $1.4B, Sanofi locks in Owkin for five years, and The Onion Desk.
NVIDIA's Kimberly Powell tells Aspen that the real bottleneck for AI drug discovery is regulatory trust in simulated biology, not model capability. Novartis and Sanofi deepen their AI-platform bets, while the sharper signal points toward manufacturing over discovery. Plus the Onion Desk on pharma finally asking its own data what's in it.
Sanofi becomes the second major pharma in three weeks to license Owkin's K Pro. BMS is now counting the bill on its 30,000-seat Claude rollout. Plus: 'first AI drug in the clinic' unpacked, the AI-funding bifurcation, and the Onion Desk.
Sanofi becomes the second top-tier pharma in weeks to license Owkin's agentic K Pro platform, signaling a shift from buying AI companies to renting capabilities. Meanwhile Bristol Myers starts counting the cost of 30,000 Claude seats, a headline "first AI drug" turns out to be AI-formulated not AI-discovered, and the week's largest biopharma deals carry no AI at all.
A $200M Brain Health Accelerator division applies the institute's cell-type atlases to gene therapies for Alzheimer's, Parkinson's, Huntington's, and ALS. Plus: Mayo AI at ASCO, an ex-Palantir team building pharma AI orchestration, and the week's autonomous-research debate.
The Allen Institute launches a $200M unit to convert two decades of brain-cell atlases into precision gene therapies, while ex-Palantir founders raise $12M betting pharma needs orchestration over another molecule model. Across the issue, AI moves upstream — into diagnostics, research workflows, and the governance debate over autonomous "AI scientists."
BMS deploys Claude Enterprise across 30,000+ employees as the unified agent layer. DeepMind's Co-Scientist and FutureHouse's Robin clear Nature peer review the same week Edison Scientific lands at Incyte. Plus: the EU AI Act high-risk draft, Helio × Syneos, and a Benchling adoption snapshot.
BMS makes Claude its enterprise "shared intelligence platform" for 30,000+ staff, pushing pharma's AI race onto the agent layer. Plus: AI-scientist papers hit Nature as Edison deploys at Incyte, EU AI Act high-risk guidelines drop, and a Quick Signals sweep from Karpathy-to-Anthropic to Owkin's Waiv spinout.
Owkin's K Pro goes live inside AstraZeneca's competitive intelligence workflows. The FDA finishes consolidating 40+ systems into HALO and ships Elsa 4.0. Plus: the cell-free expression race, UVA's open drug-design suite, and four dispatches from The Onion Desk.
AstraZeneca licenses Owkin's K Pro to put agentic AI in its executive decision layer, and the FDA ships Elsa 4.0 atop a consolidated submission platform that makes AI-first review the default. Beneath the marquee deals, the real contest is over validation economics — cell-free triage, open design suites, and cheaper generative chemistry all attacking the sequence-to-real-molecule gap.
The UK's OpenBind initiative takes aim at the structural data gap that's quietly limited AI drug discovery since AlphaFold. Also: the FDA's AI-triage inspection pilot, Novo's enterprise OpenAI bet, and Insilico's empirical case against general-purpose models.
The UK's OpenBind releases the first tranche of an open binding-data resource aimed at becoming the Protein Data Bank of drug-ligand interactions, reshaping who can compete in AI drug discovery. Plus the FDA's AI-triaged one-day inspections, Novo Nordisk's enterprise OpenAI deal, Insilico's case against general-purpose models, and CellCentric's $220M raise.
Isomorphic Labs gears up for human trials, Owkin's "AI scientist" lands at three top-ten pharmas, and Bessemer makes the case that biology-native data is the only durable advantage.
Isomorphic Labs prepares its first AI-designed molecules for human trials, ending the benchmark era, while Owkin's agentic AI scientist deploys across Sanofi, BMS, and Merck. Bessemer's new thesis argues moats are migrating from models to biology-native data. Plus: Alloy's $1B-valuation raise, PitchBook's Phase 1 success-rate claims, NVIDIA's BioNeMo expansion, and the EU AI Act's August deadline.
Eli Lilly's record partnership with Insilico Medicine is the largest AI drug-discovery deal in history. Why now, what it actually buys, and what the second-derivative effects are.
Eli Lilly's up-to-$2.75B licensing deal with Insilico Medicine — a platform bet, not a single-molecule buy — signals that AI drug discovery is now core pharma strategy. This issue tracks the aftershocks: Insilico's agentic PandaClaw tool, the first joint FDA-EMA AI principles, and a Benchling report arguing data infrastructure, not models, is the bottleneck.