The goal of AI drug discovery has stopped being abstract: medicines that are safer and more likely to work, reaching the patients who need them in years rather than decades. The first medicines designed with AI are in late-stage trials, the models can now design molecules that bind, and the people building the models have concluded that the next gains will come from data that does not yet exist, or exists only inside companies where no outside model can read it. The public data teaches a model to find a molecule. What happens to that molecule inside a person, it has mostly never seen.

In September, Insilico’s rentosertib, a lung-fibrosis drug whose target and chemistry were both generated by AI, started a 320-patient Phase 3. Generate’s GB-0895, an AI-designed antibody for severe asthma, is in two Phase 3 trials of nearly 800 patients each. Nobody has an approval yet, and the first readouts are expected from 2029. But the question has moved from whether this works at all to how much of the journey to a patient it can shorten.

The capital followed, and AI drug discovery is now an investment category in its own right. Biopharma remains healthcare’s largest venture sector, with $12.6 billion raised in the first half of 2026 by Silicon Valley Bank’s count. The AI-native companies take a growing share of that money. In the twelve months to last November, investors put $3.2 billion into AI drug development across 135 deals, and those companies raise at roughly twice the valuation of their peers, according to PitchBook. This year the rounds grew larger still: Isomorphic’s $2.1 billion in May was close to a fifth of all biopharma venture money in the quarter, by PitchBook’s count; Chai raised $400 million in July at a $3.8 billion valuation, taking its total past $600 million; Xaira, launched with more than $1 billion in 2024, revealed its first candidates in October; Enveda added $311 million in September. And the buyers signed. Lilly, Pfizer, Novartis, GSK, Takeda, Bristol Myers Squibb, AbbVie, Johnson & Johnson, argenx and Incyte each hold at least one agreement with a platform partner, and most hold several.

The models justified it. A year ago, designing a molecule that binds a chosen protein still took a dataset of known binders for that particular protein. It no longer does. GSK ran Chai’s models blind across a panel of its own targets and found binders to every one, which is why it expanded the relationship on 9 October; Christopher Austin, Senior Vice President of R&D Technologies at GSK, was careful to add that “their technology complements our own internal models that we continuously retrain against our own wet-lab data.” Xaira went from a target to a 17-nanomolar antibody lead in seven weeks. David Baker, Professor of Biochemistry and Director of the Institute for Protein Design at the University of Washington, the Nobel laureate who co-founded Xaira, said on 2 October: “For decades, some of the proteins that drive disease have been off-limits to antibody drugs because traditional discovery approaches failed on them. X-Design shows that AI is beginning to design those antibodies from scratch.”

A binder is not a drug, though, and the people closest to the work are the first to say so. Most medicines fail in the clinic because the mechanism was wrong or the molecule was unsafe, and no model of how molecules bind speaks to either. Daphne Koller, Founder and Chief Executive Officer of insitro, wrote in August: “We are doing a pretty good job at manufacturing keys, but they are often for the wrong locks.” Dave Ricks, Chair and Chief Executive Officer of Lilly, said in February: “We’re sort of like a toddler in the language of biology,” adding that large language models, “trained on the human language, not on the language of chemistry, physics, and biology,” are “not particularly good” at it. Neither is an argument against the technology. They describe where the technology’s data runs out, which is exactly where this year’s money went.

Every company here is working towards the same thing: a model that predicts biology well enough to skip experiments. Every one of them has also run into the same limit: data, not compute.

Biology has no internet to learn from. That is why, for the first time, this year’s investment went into measuring cells on the same scale as it went into the models.

Language models worked because the text was already there. Biology’s textbooks hold conclusions, not the measurements behind them, and a model needs the measurements. Evan Feinberg, Co-Founder and Chief Executive Officer of Genesis Molecular AI, described the difference on the Latent Space podcast in July: “we can’t just download Reddit posts … and train a model and voila.” The closest thing biology has is the Protein Data Bank, about 250,000 structures gathered over fifty years, and only about 10,600 of them contain an approved or investigational drug. Marc Tessier-Lavigne, Chief Executive Officer of Xaira Therapeutics, in January: “In AI, you need three things: algorithms, compute, and data. Compute isn’t limiting … What’s really limiting in biology is data.”

Even Google DeepMind, with AlphaFold behind it, found the wall. Pushmeet Kohli, Vice President for AI for Science at Google DeepMind, said, in a conversation published on 10 October, that his team had asked what could be done with the cell data already in the world: “After a lot of work, it was very clear that the data was not there yet to pursue that grand ambition of building the virtual cell.” Hence the $300 million that Google, Isomorphic and Meta put into Biohub’s cell-measurement programme that same week, the first row of the table in Section IV. Max Jaderberg, President of Isomorphic Labs, said the same thing in one sentence on Sequoia’s Training Data podcast last year: “In my mind, the data for machine learning in biology hasn’t actually been created yet.” Better methods still matter; John Jumper, who led AlphaFold at Google DeepMind and announced in June that he is joining Anthropic, likes to point out that AlphaFold 2 learned from the same data as AlphaFold 1, and the leap was method. That is also why structure was the problem that fell first: the data for it already existed.

A drug has to answer six questions in order, and each needs its own kind of data. The public has good answers to the first two. The rest live inside companies, and until this year almost nobody was building them.

The last row is the one the people building the models mention first. Diogo Rau, Chief Information and Digital Officer at Lilly, put it plainly last year: “Public datasets mostly contain molecules that do work.” Aviv Regev, Executive Vice President and Head of Genentech Research and Early Development, told investors on 28 September that the programmes that never made it out of research are “a closed garden that models have never seen, that the knowledge is only ours.” Tony Wood, Chief Scientific Officer of GSK, was asked on the company’s July earnings call why he talks about AI so little: “It’s pointless without data. So we focus very heavily on the acquisition of data and that really underpins the approach that we’ve taken.”

“For every molecule that worked, we had millions that failed, and those failures are exactly where the real learning signal is.”

Thomas Fuchs, Senior Vice President and Chief AI Officer, Lilly · August 2026

Quality matters as much as volume, and the people who run the archives say so. Emmanuel Frenehard, Chief Digital Officer of Sanofi, described what he inherited: “when I started this job, we had what I call data dumps. Everybody had one,” the legacy of “53 years of acquisitions that were never integrated successfully.” Wafaa Mamilli, Chief Digital and Technology Officer of Roche, has split her AI and data teams into two roles for the same reason: “There’s no AI without data, and we have foundational work to do there.”

Kim Branson, Chief AI Officer at GSK, said it earliest: in 2023 he told Pauline Yang, Partner at Altimeter Capital, in an interview, that fifty years of the company’s solubility data came from “six different versions of an instrument with vastly different errors,” and that the answer was to “generate data with the express purpose of fitting into the model.” In January 2026, when GSK licensed Noetik’s virtual-cell models, he put it in one line: “Foundation models are only as good as the underlying training data they are built upon.” Alex Zhavoronkov, Founder and Chief Executive Officer of Insilico Medicine, whose rentosertib is now in Phase 3, added the caution in August: “‘more data’ is not a universal solvent. Ten times more of the wrong assay buys you ten times more confidence in the wrong abstraction.”

About $3 billion was committed this year to measuring biology for AI. It is a different kind of investment from the venture rounds: it pays for the measurements themselves, cells, tissues and patients recorded at a scale biology has not had before, and it will outlast any single company. For each programme, the question that matters most is who gets to use the data, and when. Whether measurement at this scale pays off is not yet proven. The strongest evidence so far is at the molecule level, where five pharmas’ pooled structures lifted a model’s accuracy by eighteen points. At the cell level, where most of this money is going, the first blind test, Arc’s 2025 challenge, found no model that beat a simple baseline consistently, and almost none that beat it on one of the three scoring measures; the second challenge closes in November. The money is a bet that more data, made on purpose, will change that.

Most of these programmes measure cells or patients rather than molecules, so what they build serves every kind of medicine, not only the ones designed by today’s platforms.

“What we really need in biology right now are foundational data sets that are open and standardized and that everyone has access to in order to do their research and build upon.”

Priscilla Chan, Co-Founder, Biohub · Stanford’s Future of Medicine podcast · April 2026

The programmes differ most in one thing: how the data reaches a model. There are three main ways.

The public programmes and the members-only pools are learning from each other. Alex Rives, Head of Science at Biohub, explained on 7 October why commercial funders get a year’s head start on data the public will eventually own: “We have to have some incentive for commercial players to be a part of this, and the embargo period creates that.” It is the model the Broad Institute used in 2020, when ten pharmas pooled cell-imaging data on 116,750 compounds, took a one-year head start, and then released all of it free. If Biohub delivers, data on how an average cell responds becomes a public good, as protein structures did, and owning it stops being an edge. What stays scarce is what no open programme will produce: cells from individual patients, measured over time and linked to what happened to them.

The buyers have their own answer to the aggregation question, and it is to keep a loop of their own. GSK retrains its internal models on its own wet-lab data and treats a platform partner’s model as a starting point that a small, unique dataset can adapt; Branson, on the first milestone of the Noetik agreement in August, said the models “using a small amount of GSK’s unique data can be fine-tuned to generate high accuracy and specific predictions.” Regev calls the same idea owning the loop: “You need to own the loop. We need proprietary AI models, the ability to improve them, not just parameters, but architectures.”

The platform partners have answered by building each buyer a private version. Chai is building Lilly a model “exclusively for use by Lilly, trained on large-scale proprietary Lilly data” (January), and Pfizer’s licence in June includes a custom model built on Pfizer’s own data; Boltz fine-tunes its open and proprietary models on a partner’s archive, on-premises if needed, and Cradle trains on each customer’s data inside that customer’s own tenancy.

Two numbers explain why the money is shaped the way it is. The first is what private data is worth: fine-tuning an open structure model on more than 20,000 drug-bound structures pooled across five pharmas, without any of them moving, raised the share of drug molecules placed correctly in 1,056 held-back structures from 28.9 per cent to 46.8 per cent. Eighteen points from twenty thousand examples is the best public measurement yet of what an archive teaches a model. The second is cost. A perturbed cell costs cents to measure, which is why there are now hundreds of millions of public cells and a $1.8 billion programme to make billions more. A single animal study costs tens of thousands of dollars and takes months. The categories with the least data have the highest marginal cost, and no amount of compute changes that; only a lab does. Or, as Regev put it to investors: “AI makes reasoning a commodity. That actually means that the scarcity becomes lab and clinical reality.”

The clearest gains so far are inside research organisations, where better data is already changing which programmes live, which die, and how fast the decision is made.

In discovery. Genentech reports, by its own count, that AI and computation contributed to 40 per cent of its tracked research decisions over the last three quarters, and that the time it takes to find a lead molecule roughly halved in three years with resources held roughly flat. Greg Meyers, Executive Vice President and Chief Digital and Technology Officer at Bristol Myers Squibb, said last November that the company had “already cut the time it takes from deciding to optimize a lead to doing toxicology work by about 20 percent.” Andy Plump, President of R&D at Takeda, described a preclinical antibody the company was about to abandon because it clumped at scale, until a platform partner redesigned it: “AI gave us this rebirth, like a phoenix rising from the ashes.” Could Takeda have found the fix with traditional methods? “Never in a million years.”

In medicines. Najat Khan, Chief Executive Officer of Recursion, keeps a number in front of her team: “Eighty percent of the diseases that exist in the world don’t have disease-modifying medicines.” The purpose of better data on targets, tissues and failures is to aim at the right mechanism before the molecule is made. Koller again: “For the hundreds of millions of patients for whom no meaningful medicines exist, the difference between a wrong mechanism and the right one is the difference between another devastating clinical failure and a life-altering breakthrough.”

On the clock. By the industry’s own long-run averages, a new medicine takes ten to fifteen years from first idea to approval, and roughly nine in ten candidates that reach human trials fail. The early stages are where AI has already changed the timelines; the clinical years have barely moved, and that is where the next decade of data has to do its work.

For patients. No approved medicine yet, and it is better to say so plainly, though patients in trials have already felt it: Insilico’s Phase 2a in pulmonary fibrosis improved lung function at the top dose in a 71-patient, twelve-week trial, the strongest human result from any AI-native company so far. No medicine designed with this year’s new datasets could have reached a trial yet. The first Phase 3 readouts from AI-era programmes are due at the end of 2028 and in 2029 (the Generate and Insilico trials list primary completion in December 2028, January 2029 and October 2029), and approvals, if they come, follow no earlier than 2030. Lotte Bjerre Knudsen, until this year Chief Scientific Advisor at Novo Nordisk, who led the science behind its GLP-1 medicines, set the realistic floor at the end of June: “I think we really will see a large compression, but we still need the clinical trial data… So it’s probably hard to imagine that you can go below five years.” Fuchs said the same in January from the other side of the lab bench: “You cannot accelerate a cancer trial or a neuro trial, the bottleneck is biology.” This year’s money is infrastructure, the way genome sequencing was in 2001: the first draft of the human genome was published that year, and the medicines built on it came a decade later. The question patients can fairly ask now is whose biology the models learned from, because medicines will only be as good for everyone as the data was. Zhavoronkov’s measure is the right one: “The final score is years of healthy life created, suffering prevented, disability delayed, and the number of people who can actually get the medicine.”

Five gaps remain, and each is something that helps everyone and belongs to no one. The encouraging part is that the pieces needed to help close them appeared in 2026.

Failures. There is no registry of failed programmes the way there is a registry of protein structures. The only funded effort is a small pilot recovering records from bankrupt biotechs, and the OpenAI Foundation’s grant to it is, as far as we can find, the first money aimed squarely at negative data.

Living systems. The step between a molecule and a patient, animal and tissue testing, has the least data of all, because animal studies are slow, expensive and were never recorded for machines to read. Three things are changing it.

• New data, made on purpose. Lilly and NVIDIA’s lab is built to generate wet-lab data for models. ARPA-H’s CATALYST programme (eight teams, up to $125 million over four and a half years, Inductive Bio and Draper among them) is building human toxicity models from organ-chips, patient-derived cells and in silico physiology, with the stated aim that approval to begin a first human trial could one day rest on in silico safety data.
• A regulator moving. Since its April 2025 roadmap the FDA has made its ISTAND pathway for new methods permanent, proposed replacing the six-month monkey study for monoclonal antibodies with a three-month study and a weight-of-evidence assessment, and opened a database of accepted non-animal methods, while still saying animal models remain necessary for many drug classes. The first credible dataset here is a regulatory asset as well as a scientific one: it could support a filing.
• Stand-ins for the body. Vivodyne grows more than twenty types of human organ tissue and doses more than 10,000 of them at a time by robot, producing human datasets a model can learn from. Emulate’s organ-chips are the kind of system the FDA’s roadmap points to. Parallel Bio grows human lymph-node organoids and has eight pharma partners testing more than fifty drugs on them.

In theory that road ends with far fewer animals and smaller, better-aimed human trials. In practice, as Koller says, the only true ground truth is still a human clinical trial; what these companies can do is make sure fewer of them fail.

People. Under 2 per cent of sequenced human genomes come from people of African ancestry, as Michèle Ramsay, Director of the Sydney Brenner Institute for Molecular Bioscience at the University of the Witwatersrand, noted in January. A model trained on European genomes tends to make medicines for European patients. The most diverse large cohort open to industry, the NIH’s All of Us, has lost most of its funding since 2023, and the VA’s larger cohort does not accept industry proposals. The one effort aimed at the right scale is the Truveta Genome Project, launched in January 2025 by seventeen US health systems with Regeneron and Illumina, alongside a $320 million investment in Truveta itself: ten million exomes from consented patients, linked to their records, promised as “the largest and most diverse database of genotypic and phenotypic information ever assembled”, sequenced by Regeneron and to be added to Truveta Data for biopharma and academic research. Twenty-one months on, it has not published how many have been sequenced or on what terms anyone other than Regeneron may train on them.

Scorekeeping. Almost every performance claim is self-reported, and the groups running benchmarks are usually building models too. The field knows it. When Arc ran its first blind virtual-cell challenge, its organiser, Hani Goodarzi, Core Investigator at the Arc Institute, said simply: “We don’t think we have a virtual cell yet.” OpenADMET’s blind challenges now draw more than 350 teams each, and Pat Walters, Chief Scientist at OpenADMET, reported the honest result in July: “fine-tuning on project data beats training on the public corpus alone, with few exceptions.” No benchmark yet runs at the scale of a pharma archive.

Terms. “Open” means anything from public domain to research-only. The largest data programme of the decade, Biohub’s, has no published licence, governance or price for early access; for now its terms are known from interviews. Vijay Pande, Co-Founder of VZVC, who built a16z’s bio and health practice before starting his own fund, said in August that when investors hesitate, “it’s about doubt of the data.” Publishing the terms would remove most of that doubt.

1. Isomorphic doses its first patient, and Xaira discloses its first partnership and the shape it takes: the “preserve upside” structure its CEO has described, or something closer to a conventional licence.
2. A medicine designed on a platform for a partner enters the clinic: Chai, Boltz, Nabla or Cradle on the design side, Noetik on the target side, Isomorphic for Lilly, Novartis or Johnson & Johnson, Xaira with the partners it is now choosing. Every platform in the table is working towards that moment. None has announced one since Exscientia’s DSP-1181 for Sumitomo Dainippon entered the clinic in 2020 and was later dropped, and the next will show the field what a partnership built on a platform can become.
3. Biohub publishes a licence and governance for the Virtual Biology Initiative, at least one pharma takes a seat on the early-access list, and the first large dataset is released, on the current plan in late 2027; and the Truveta Genome Project reports how many of its ten million exomes exist and who may train on them.
4. A second pharma invests in a dataset as well as the models built on it, the way GSK did with Relation, and the first failure-data registry, the CTD Commons pilot the OpenAI Foundation is funding, releases its first records.
5. An IND is cleared on non-animal safety evidence alone, with tissue data from a body-replica company or from the Lilly–NVIDIA lab.
6. A measured cut in clinical time: BMS or Genentech reports a trial that ran shorter or smaller because of simulation, against the two-to-three-year aim BMS has set.
7. The first custom-model agreements show what they deliver: whether partners renew, expand or bring the work in-house, and, if the terms are ever disclosed, who owns the fine-tuned model, which would set a precedent for every later deal; and an independent, blinded benchmark runs at pharma scale, so the platforms can be compared on common ground.

The medicines are coming, later than the headlines say and sooner than the history of the industry would predict. This year, more than ever before, the money went to the measurements that decide whether they work. Whether they work for everyone is the measurement still waiting to be funded.

Figures as of 15 October 2026, from company releases, SEC filings, licence files, earnings-call and conference transcripts, Biohub’s and the OpenAI Foundation’s announcements, PitchBook and Silicon Valley Bank venture data, ClinicalTrials.gov, and, for the cost of a new medicine, Deloitte’s 2026 analysis of pharma R&D returns and a 2020 JAMA study of development costs. Several quotations, including Feinberg’s, Jaderberg’s and Chan’s, are from podcast interviews, and Jumper’s point is paraphrased from one.

Written by Tonya Brody. Deal terms, dates and figures are researched against primary sources, cross-checked across three LLMs to surface any discrepancies, and then, wherever possible, confirmed directly with the founders and others involved. Editorial judgment, opinions and anything in the first person are mine alone. I dictate edits with Wispr Flow and use Claude, Codex or similar tools for proofreading and, where useful, for illustrative images, which are occasionally AI-generated. Any AI-assisted edits or images are reviewed and approved by me before publication.
Originally posted 12 October 2026. Updated 15 October 2026 with newly announced partnerships and Iambic’s IPO.

II wrote Part I of The Data Behind the Drug for people inside the industry and outside it: where AI drug discovery stands today, and what it could make possible. The questions it drew mirrored the ones we hear in the work, and they all came down to two sentiments: what could this look like, and what should it look like? This one is for the people on both sides of the table: founders in AI drug discovery and the teams who lead their partnerships, and the discovery, data and compliance leaders at biopharma who have to decide whether to say yes, and whom to take a chance on.

AI drug discovery has had years when it did not deliver on its promise. Most people in the field are optimistic that this has changed and that the moment is now. It is still hard to see through the hype. Most of what I cover here are the hard, uncomfortable questions both sides ask and are asked every day, and most of them get decided in rooms with the door closed. Who owns a model after it has learned from your data. What happens to a good molecule when the programme that would have carried it is cut. What a platform is worth when the two sides cannot agree on what is being sold. These high-stakes questions are not new because someone is being difficult. They are new because the category is new. When Noetik signed with GSK in January, Ron Alfa, Co-Founder and Chief Executive Officer of Noetik, said “there was no prior template for this deal or deployment,” and his investor James Hardiman, General Partner at DCVC, named the problem in one line.

“The quandary, up to now, has been knowing how to price such deals.”

James Hardiman, General Partner, DCVC · on the GSK and Noetik agreement · January 2026

Zoom out and the picture is simpler than the deal terms suggest. For forty years this industry built its agreements around molecules. What partners are building together now is learning: a system that gets better every time it is used on a real problem, and keeps getting better after the agreement is signed. That is new, and it is good news for anyone who wants safer, more effective medicines and cures to arrive sooner. Every question in this note, about IP and ownership, partnership structures, open weights and who survives, is a version of one: how do two sides share something that keeps improving, so that both of them, and the patients they serve, benefit from it?

One belief runs through all of it. This industry moves at the speed of trust, and trust is earned in drops. The agreements still standing in three years will be the ones where the pharma partner, the platform partner and the patient all come out ahead, and where each side could check the other’s claims. When those interests look like they are in tension, it is usually the structure that is wrong, not the people. If you are asking the hard questions about how to build mission-driven partnerships that scale in AI drug discovery, this conversation is for you.

The patient sets the clock. Every structure here comes back to one outcome: safer, more effective therapies that reach the clinic sooner and are more likely to succeed once they get there.

Price the thing you are transferring. A target is not a model. A model is not a molecule. A fine-tune is not a platform. When a deal feels hard to structure, the two sides are usually pricing different objects. And nobody is bound to price the way this industry always has. What is being offered is different, the people writing these terms are deciding what the category becomes, and that is room to lead rather than a reason to copy.

Whoever compounds the learning should own the loop. Data can be shared, access licensed, compute rented. The thing that gets better every cycle is the asset, and it should sit with the party doing the work of improving it, by agreement rather than by accident. There is real pressure on pharma to own all of it, and the pressure is understandable: the data was entrusted to them, the risk sits with them, and so does the regulator. But there can be a middle ground where both sides keep learning inside the rules, and each pharma’s proprietary data stays protected throughout: it never leaves the company’s own environment, no other party sees it, and what is shared is general learning, not molecules, targets or patient records. The working examples now exist. Lilly’s TuneLab lets members fine-tune its models on their own data through federated training no other member can see, and delivers them inside Benchling and CDD Vault, where the work already happens. Everyone improves, nobody’s data moves, and the patient gets a better medicine sooner. That is worth holding on to on both sides of the table, because we are all patients.

A no from compliance is a promise to a patient. When a pharma team declines a data request, it is usually honouring consent given by a person who volunteered for medical research. That is not a wall to argue with. It is the specification, and the platforms closing fastest this year were built to it.

Independent evidence is leverage for both sides. The strongest public evidence is a blinded benchmark run by a third party with no stake in the result, the kind CASP set for protein structure and that OpenADMET, the Arc Institute’s Virtual Cell Challenge and the AIntibody study now run for drug discovery: the model is scored on data it has never seen, by people who did not build it. The strongest private evidence is the partner’s own test. The platforms that moved fastest this year let a partner test their designs in its own wet lab before expanding, and priced on the result. That evidence lets a founder ask for more than a CRO rate, and lets a discovery lead say yes without staking a career on a vendor’s slide. A partner that wants to reprice around performance owes the platform a fair test. A platform paid on performance owes the partner a way to check.

Pharma will build and buy at once, and adoption decides who wins. Every large pharma is building its own AI team and buying from platforms in the same year, and the platforms that earn a place inside that decision are the ones with scientists and engineers working on the partner’s site, inside its data and its workflows. That is how trust is built, and it does not scale on its own: every deployment drifts toward custom, and a service grows by headcount while a platform grows by use. Matt Gline, Chief Executive Officer of Roivant, explains the pull toward custom: drug discovery is “a composite of 150 different difficult problems, and each … requires bespoke AI solutions.” The partnerships that last agree early on what is product, what is service and what can scale, and price each for what it is.

Partnerships trade more than money, and the other currencies deserve the same care. A Lilly logo on a startup’s deck is credibility that helps it raise the next round and keep building. A pharma team’s compliance culture is trust earned over decades of consent work, and it is what lets a partnership touch patient data at all. An open-weight release is standing with the researchers who will use the tool and improve it. Name what each side is really bringing, and what it most needs back, write it into the agreement with the same care as the royalty, and the cash conversation gets shorter and easier. It helps to remember where the cash comes from. Most of these agreements draw on a budget line that does not exist yet; the value the work proves is what creates it, and that is a case both sides can make together.

Nobody has this figured out, and the founders who say so are the ones to trust. Every founder we talk to is unsure whether the terms they signed are the right ones, whether those terms serve where the industry or their company is going, and whether to follow the competition or set their own course. Their partners are unsure too. That is the honest state of a new category, and it is why the agreements that will last are the ones built so both sides can learn and scale together.

Price the learning, not the molecule. The agreement you sign this year sets the price every later one is measured against, and decides who owns the model your data will make better. The friction is structural, not personal. Pharma procurement has a line for CRO work and a line for assets, and most platforms are neither. Founders are asked to prove performance before the clinic can; partners are asked to pay for it before it is proven. What follows is a way to take the agreement apart so that both sides are pricing the same thing, and so that pharma can ask for a budget it had not known to allocate.

For a decade an AI discovery deal looked like a biotech licensing deal: a modest upfront, a ladder of clinical milestones, royalties at the top, sometimes equity. The headline number is the sum of every rung on the ladder, and the ladder assumes the programme survives. Most do not. In the asset deals disclosed this year, the cash paid at signing is 1.5 to 10 per cent of the headline; the rest is owed only if the molecule clears the stages where roughly four in five programmes fail. Insilico, the most prolific dealmaker in the category, had realised $308 million against $10.6 billion of outstanding milestones by the end of June, under 3 per cent. For a founder, that is a runway priced on a number that is mostly notional and a revenue line owned by someone else’s clinical calendar; Iambic’s prospectus, filed this month, says it in the language of risk factors, its “potential to receive milestones and royalties is dependent on our partners’ efforts over which we have no control.” For a biopharma partner, it is paying for an asset’s risk-adjusted value while learning nothing about whether the platform that produced it is any good. The structure is not wrong. It is one point in a much larger space.

Both sides are better served pricing what they can control. A clinical milestone is paid on biology and on a calendar that neither party owns. A deployment that passes the buyer’s tests, a blind run on the buyer’s targets, a model that beats the buyer’s baseline: these are events the two sides can bring about together and observe this year. Price those, and the agreement pays for the partnership rather than for the lottery ticket attached to it.

What are we pricing against?

• Cost-plus, the CRO logic, and the starting point for Recursion’s partnerships: Ben Taylor, Chief Financial Officer of Recursion, told investors in August that “All of our partnerships are designed to be breakeven or profitable on a direct cost basis from the start.”
• Cost displacement, what the buyer would otherwise have spent, in money and in time. Boltz’s WRN helicase work found three functional inhibitors from 27 compounds for roughly $10,000, against the nearly 350,000 compounds Novartis screened. Time is the saving most often left off the sheet: a quarter saved at the start of a programme is a quarter of exclusivity kept at the end, and it is the one saving a patient can feel.
• Value, the asset’s risk-adjusted NPV, which still fits when the platform partner brings the biology. Alex Zhavoronkov, Founder and Chief Executive Officer of Insilico Medicine: “Assets are like diamonds; assets are forever.”
• Capacity, something in short supply, which today is experimental data: GSK has agreed to pay Relation up to $110 million to generate perturbation data at scale. Birgit Girshick, Chief Executive Officer of Charles River, expects AI-designed molecules to “drive demand,” because every design still has to be tested.
• The free alternative, which now sets the floor under every price: open weights, TuneLab, a Boltz API at two and a half cents a prediction.
• Improvement rate, new this year. Hansa chose Cradle because it wanted a platform that learns from its own data, in the words of Stef van Grieken, Co-Founder and Chief Executive Officer of Cradle, “so that round 3 is better than round 1 in their hands, not in a benchmark.”

When does money move? The mechanism. Four are familiar and two came to the fore in 2026.

• Subscription and consumption, borrowed from enterprise software and cloud metering: the partner pays for access or for use, the way pharma already buys software. GSK’s agreement with Noetik includes a subscription; Boltz meters its API by the prediction.
• Option, milestone and royalty, the biobucks ladder, which still fits when an asset changes hands, as in insitro’s agreement with AstraZeneca this month.
• Outcome, payment on an event both sides can observe, which can now be a deployment rather than a clinical result. GSK paid Noetik’s first milestone when inference and fine-tuning were deployed and met testing criteria. This is a principle I had wanted to test since 2020, and seeing it written into a signed agreement this year was a quiet milestone of my own. Shafique Virani, Chief Business Officer of Noetik, named the shift in the announcement itself: “We are moving the industry from AI services collaborations to licensing AI infrastructure.”
• Gated expansion, new this year as a named mechanism: a small pilot that grows only when it works, which matches how pharma builds trust. Takeda and Nabla had already run it from a 2022 pilot to a billion-dollar expansion, smaller pilots getting larger, in the telling of Andy Plump, President of R&D at Takeda. GSK tested Chai’s zero-shot designs in its own wet lab, found binders to every target it tried, then expanded.

What do we pay in? The currency. Four are familiar and two are new.

• Cash.
• Equity, as Lilly holds in Absci and BigHat.
• Compute, which Incyte funds for Genesis.
• Data, where TuneLab’s price of entry is 1,000 readouts.
• Distribution, new this year. Lilly put TuneLab inside Benchling and CDD Vault because that is where scientists already work.
• Validation, new this year. When a pharma tests a vendor’s designs in its own wet lab, the result is worth something no public benchmark can replicate. Few contracts price it. They should.

Where 2026 Partnerships Landed to Date

The classic structure is not dead: where the platform partner supplies an asset, value anchors with milestones and royalties still fit, though some newer players are challenging them, and insitro signed one with AstraZeneca the week this was written. Where it supplies a model or a learning loop, the anchors move to cost displacement, capacity and improvement rate, and the mechanisms to subscription, consumption and gated expansion. And GSK alone priced five layers five ways in nine months: the buying side has already stopped treating this as one market.

The anchor also decides what IP moves. Price against the asset and the asset’s IP transfers. Price against cost, capacity or improvement rate and what transfers is access, while the platform and its improvements stay home. Both sides are still working out what a durable version of this looks like, and the honest position is that nobody has settled it yet. That is the opportunity: the terms being written this year and next will decide whether the value created here is shared well enough that both sides want to keep building. Negotiating IP, the next section, takes up the question underneath it, which is ownership.

Price the Thing You Transfer

The structure follows what is being transferred. insitro and AstraZeneca, and Profluent and Lilly, are value-anchored because what the platform partner delivers is a development candidate or a designed molecule, against AstraZeneca-nominated targets in insitro’s case. GSK pays Noetik on engineering milestones because Noetik supplies a model and the loop that improves it. A model’s value shows up when it ships and beats a baseline, not at Phase 2. Jack Dent, Co-Founder and President of Chai Discovery, named the assumption this breaks: “people told us the only way to make money is to make your own assets and become a drug company.” Chai has signed a run of pharma agreements this year, none of them asset deals, and describes that as a choice rather than a limit. “That’s the dogma we had to challenge,” Jack Dent said. A company that does not make its own assets is not competing with its partners for them, and can put its whole effort into the models, which in his words had to be “literally 100 times better for it to be valuable for real drug discovery programs.” Schrödinger, founded in 1990, has drawn the same line this year: Karen Akinsanya, President of Research and Development, Therapeutics, at Schrödinger, told investors in May that “we will not be taking programs into the clinic,” and Ramy Farid, Chief Executive Officer of Schrödinger, prices access by use, “Our throughput-based licensing model ensures we are positioned to capture the value.”

There is a structural reason capability sellers struggle to capture value, and Jack Scannell, Chief Executive Officer of Etheros Pharmaceuticals, who named Eroom’s Law, the observation that drug discovery keeps getting more expensive rather than cheaper, has said it plainly. Once Phase 1b data is published and competitors learn the target, they can make their own molecules without having paid for the model, so “the economic value from better models is often hard to appropriate.” The people who control the bottlenecks, Scannell adds, “make even more money than they used to.” Jason Kelly, Co-Founder and Chief Executive Officer of Ginkgo Bioworks, has the reason the cost curve has not bent: “Because we do it manually.” Price against the bottleneck you actually hold. Today that is experimental throughput, curated data and a learning loop on the buyer’s targets, not model quality.

This gives the market three tiers, and most companies should be honest about which one they are in: a floor of open and near-free capability that sets the reference price, where, as Jonathan Chauvin, Co-Chief Technology and Product Officer at Simulations Plus, put it in January, “New value sits above the engines, not inside them”; a middle of metered access and subscriptions, where most of the category’s revenue will sit because it matches how pharma buys software and how fast models turn over; and a top tier of asset partnerships for the cases where the platform partner genuinely brings the biology. The mistake we see most is a middle-tier product holding out for top-tier economics, and handing control of its revenue to someone else’s clinical timeline. Vijay Pande, Co-Founder of VZVC: “the go-to-market part is at least as hard or harder than the technology side.” Alex Zhavoronkov, who chose the asset tier: “The actual target market for software is extremely tiny.” Both can be true, and the stakes are concrete: the median company in this year’s census, Artificial intelligence-enabled drug discovery: A growing but largely early-stage clinical pipeline in Cancer Treatment Reviews, reached first-in-human about 6.5 years after founding, having raised roughly $187 million. The pricing decision decides whether a company is still standing at year six.

What Biopharma Partners Are Telling AI Discovery Teams

The buying side has been unusually direct this year. What we keep hearing:

Speed of procurement is part of your product. Here is Chris Arendt, Chief Scientific Officer and Head of Research at Takeda: “If it takes you six months to bring a new vendor on board, that means their technology could be obsolete by that time.” Escrowed weights, on-premise deployment and a standard data processing agreement are worth real money because they shorten diligence.

The buyer wants the system to improve on its targets. Hansa chose Cradle so that round three is better than round one in its own hands. Incyte wrote the same wish into its terms with Genesis in May: alongside the upfront, it funds the compute for model training and inference, and shares its own experimental data “to enhance GEMS’ capabilities,” so the platform gets better on the problems Incyte cares about. Boltz’s agreements with Pfizer and GSK, below, go a step further and make the improved model the partner’s own. That is a relationship, and it should be priced as one.

The internal AI team is your competitor and your necessary ally. Here is Christopher Austin, Senior Vice President of R&D Technologies at GSK, announcing the Chai agreement: “Their technology complements our own internal models that we continuously retrain against our own wet-lab data.” GSK is not unusual. Albert Bourla, Chairman and Chief Executive Officer of Pfizer, told investors in August that “Our ambition is to build an AI-native R&D organization.” Jay Bradner, Executive Vice President of Research and Development and Chief Scientific Officer at Amgen, said that month that Amgen had “recently established a frontier AI laboratory that combines advanced models with proprietary data.” Roche’s Pharma Day deck in September reported that 40 per cent of its pipeline decisions from late 2025 to mid-2026 carried a tracked AI or computational contribution, and Martin Holst Lange, Chief Scientific Officer of Novo Nordisk, said in September that “We have identified not only new targets but also drugs with the help of AI.” Every large pharma now has a group that can retrain away a vendor’s edge. The deals that last make that group a beneficiary: co-publication, capability transfer, a named internal owner whose own work gets better because the vendor is there.

The channel is becoming the moat. Lilly distributes TuneLab through Benchling and CDD Vault, and Dave Ricks, Chair and Chief Executive Officer of Lilly, has compared it to an app store for scientists. Ashu Singhal, Co-Founder and President of Benchling, explained why the channel chose itself: “Lilly came to us to distribute it because we are where scientists already work and we hold the wet lab data that federated approach needs.” For proprietary models Benchling now runs “revenue-share-style arrangements with the model makers,” while “Open source runs entirely through us.” For a company without a sales force, distribution is worth as much as the model.

Metrics matter, and you are being measured against everyone. Independent evaluation is the ideal, and it is not yet the market norm. Chai publishes its own numbers and has signed GSK, Lilly and Pfizer on the strength of them, which tells you self-reported metrics still clear the bar when the team is credible and the claim is specific. What actually closed the GSK agreement, though, was not only the benchmark: GSK ran Chai’s zero-shot designs in its own wet lab and found binders to every target it tested. That is the pattern worth planning for. The Apheris federation did the same thing in public, lifting high-quality interface predictions from 36 to 52 per cent on 1,056 held-back structures from live programmes, about eleven points clear of the best public model. The counter-evidence is just as instructive: in the blinded AIntibody study successes did not transfer across tasks, and on one task all but one model ranked clones worse than picking at random. The people who build benchmarks for a living are blunter still. Evan Feinberg, Co-Founder and Chief Executive Officer of Genesis Molecular AI: “If your model is sitting at 1.8, 1.9 Angstrom RMSD, that’s slop, most likely.” OpenADMET, in September: “‘Model A beat Model B by 0.02’ is not a useful result unless you can attach a confidence to it.” The Arc Institute, launching its Virtual Cell Challenge: “the only way to establish how far away it is, is to measure it in public.” So report your own numbers if you must, but offer the buyer’s test before they ask for it, and be first to say where your models have not been proven. The buyers’ patience for numbers without medicines is also shorter than it was. Mike Nally, Chief Executive Officer of Generate Biomedicines, repeated the refrain he hears from investors: “I hear a lot about these AI, [machine learning] biotech companies, but where are the medicines?” Najat Khan, Chief Executive Officer of Recursion, set the standard in a sentence: “It’s not going to be how you use AI. It’s going to be what value you’ve created using AI.” And Rohan Ganesh, of Obvious Ventures: “A model that’s accurate but doesn’t change the pace or probability of success in the clinic is meaningless.”

And you are not being measured alone. GSK bought five layers five ways in nine months, and ran Chai’s designs against its own targets in its own wet lab before expanding. The 2026 census counts 69 companies with AI-enabled assets already in human trials. Assume a bake-off, ask what you are being measured against and on whose targets, and ask for the protocol in writing.

Attention compounds in this market too. A first named partner shortens the second conversation, a result a buyer can check travels further than a deck, and a well-timed announcement can shift what the whole category is assumed to be worth before any trial reads out. Jorge Conde, General Partner at a16z Bio + Health, greeted Boltz’s GSK agreement in July with six words, “Another top 20 pharma AI partnership,” and the counting is the point: in a young market a named partner is evidence in itself. That is how a market prices what it cannot yet measure, and every side of the table is navigating it in good faith. The practical lesson is to treat your wins as evidence and publish them properly: what was tested, on whose targets, against which baseline, and what decision it changed. Tom Miller, Co-Founder and Chief Executive Officer of Iambic, said it the week his company listed: “Iambic has never done it just for the logos on the slide,” and “The point of any platform is to execute on what comes off of it.” The teams getting the best terms this year are the ones whose news a partner could verify in an afternoon. Then plan for the other direction, because attention reprices both ways and the quarter after a wave is always quieter; Diogo Rau, Chief Information and Digital Officer at Lilly, went further in February: “The hype is actually a serious threat to the research itself.” Build the agreement so it holds when the news cycle moves on. And know that the largest buyers understand exactly what a first named partner is worth to a young company, and will bring that to the negotiation. That is not bad faith; it is leverage, and it belongs in your model before you walk in.

The comparison set is not fair, and it is still the comparison set. A team eighteen months old on a seed round is put on the same slide as a company five years in with several hundred million dollars raised, and both are judged on the same benchmark and the same unspoken question: will you still be here in three years? For the founder, that is the game, so play it in the open. Say what stage you are at, what you have proven on whose targets and what the next round buys, and ask for the structures that pay as you prove rather than the ones that need you to be big already: a gated pilot, weight escrow, a short first term with the expansion written in. For the buyer, the youngest team in the bake-off is sometimes the best model in it, and the cost of finding out is a pilot, not a programme. Nobody at the table was born large. Marc Tessier-Lavigne, Chief Executive Officer of Xaira Therapeutics, has said the quiet part from the well-funded side: “Someone may be first, but others can catch up.” Iambic’s prospectus says it in the language of risk: benchmark comparisons “are inherently point-in-time, and competing models, including open source and commercially available alternatives, are rapidly advancing.” Capital buys time. It does not settle who learns fastest.

One modality is a feature, not a company. The buyers are expanding across modalities and expect partners to follow. Lilly’s TuneLab runs small-molecule and antibody models side by side; Zealand took Generate’s multispecific proteins into obesity this month; Profluent sold Lilly recombinases for genetic medicine. Isomorphic says the quiet part in its own words, that it is building “from small molecules to antibodies, peptides, and beyond.” If your platform works on one modality, say so plainly and price it as depth. If it generalises, prove it on a modality the buyer cares about rather than the one your benchmark was built on. Boltz went the other way entirely, horizontal into flavours and fragrances with dsm-firmenich, so that pharma’s procurement cycle is not its only clock.

They want a model that becomes theirs. The thing being bought is increasingly a model fine-tuned on the buyer’s own data, held exclusively. Boltz’s GSK agreement is precisely that: its models fine-tuned and retrained on GSK’s proprietary data, with the resulting models exclusive to GSK. Its Pfizer agreement in January read the same way, Boltz refining its models on Pfizer’s historical data to create exclusive models while Pfizer keeps the compounds. Kim Branson, Chief AI Officer at GSK, has said plainly that a model trained on a partner’s data belongs to the partner. Decide before the term sheet what you will hand over, what stays yours, and what you may learn, because this is the question that stalls the most deals. More on this in Negotiating IP, below.

And beyond the model, the pressure to own the targets and the molecules too. The larger buyer often asks for more than a fine-tuned model: exclusivity on targets, a first call on anything the platform designs, a non-compete that reaches past the collaboration. Ask for what you will work. The record on the rest is poor for everyone, buyer included, and Negotiating IP sets it out under When the Molecule Stops Moving: assets locked and not worked are the ones that never reach a patient. Daphne Koller, Founder and Chief Executive Officer of insitro, gave the platform side of it in February, explaining why insitro acquired a capability rather than partnering for it: “partnering turned out to be less easy than we’d thought in terms of having control over the timelines.” A clock protects a pharma partner better than a lock: exclusive while worked, reverting when not, and the platform free to find the next one.

The budget calendar is the real clock. In our experience, a pilot that proves itself after the planning cycle closes waits a year for its expansion, and the budgets are tight: Richie Jain, Chief Financial Officer of Schrödinger, told investors in February that net dollar retention had fallen to 100 per cent after several years averaging above 110 per cent, the clearest public sign that pharma software spend is being held flat. Design the gated expansion so the proof lands before the budget locks, and ask the date out loud. A partner who wants you to succeed will tell you.

Before the First Term Sheet

Most stalled negotiations we have seen are not about the number. They are a mismatch on the anchor: the vendor pricing against the value of an asset it hopes to enable, the pharma team pricing against what the work would cost internally or against the free alternative it already runs. Neither is wrong; they are answering different questions, and no amount of back and forth on milestones closes that gap until someone names it. So before the first pricing conversation, each side writes down its three answers, compares them, and asks three more. If this works, what does the buyer stop spending, and how many days does it save? Count the days, because they have a price: a medicine that will sell a billion dollars a year earns about $2.7 million for every day it is on the market, so each day gained at the front of a programme is a day of exclusivity kept at the end of it, and a candidate that enters the clinic with cleaner properties fails less expensively, which matters because the clinic is where the money goes. That is the cost-displacement anchor, usually the most defensible number in the room. What can only we supply? If it is data, throughput or a learning loop on the buyer’s targets, price against capacity or improvement rate, not a forecast. What is the smallest test that would change the buyer’s mind? Vas Narasimhan, Chief Executive Officer of Novartis, described his own diligence in those terms this year: “often when we’re doing due diligence, we’re looking for those killer experiments,” the preclinical result that proves a target does what the founder says it does. That is the gated expansion, run by the buyer’s scientists on the buyer’s targets, with the next stage agreed in advance. The old structure asked both sides to value a drug that did not exist. These ask them to agree on things they can observe this year.

New Approaches to Test

GSK’s engineering milestone with Noetik matters less for its size than for what it proved: you can pay a platform partner on an event both sides can observe this year, instead of an uncontrollable clinical event five years out. Once you see that, other doors open. Nothing here is as new as it feels. The Protein Data Bank showed that the layer which changes a field is usually data nobody could have collected alone. The Bermuda Principles showed that openness works best as a schedule everyone signs up to. MELLODDY, the ten-pharma federated-learning project, showed that federation works technically and fails commercially unless someone’s business is the federation. Bayh-Dole’s march-in rights, never once exercised, showed that if the remedy for a shelved asset is not written into the contract, it does not happen. New structures worth weighing, to decide whether they should be tested, each a potential win for all parties: the biopharma partner, the platform partner and its investors, and the patient.

1. Speed-to-patient bonuses. Agree a baseline timeline at signing and pay a bonus for every quarter a programme reaches first-in-human ahead of it. Observable, near-term, and tied to the one thing a patient actually feels.
   BiopharmaPays only for time it actually gains, on a date its own clinical team controls; a quarter saved early is a quarter of patent life kept at the other end.

   AI discovery teamIs paid for what the platform is already best at, years before a clinical milestone, in a number investors can model.

   The patientA medicine that works arrives sooner. Nothing else in a term sheet is this direct.
2. Conversion written into the pilot. Most pilots end in a second negotiation, and a second procurement cycle, at exactly the moment both sides have the proof they asked for. Write the conversion milestone into the pilot contract: the test, the threshold, who runs it, and the terms that apply the day it is passed, from scope and price to exclusivity. Write the renewal terms into the original contract the same way, with any change in price tied to what the platform has delivered rather than reopened from scratch. Write the path across the portfolio in as well, so that once the results are proven and trust is built, the work can grow without starting over. Look past the handful of targets on the table at signing. If the platform delivers on its first neuroscience target, the contract already sets out how far the work can go: which targets come next, on what timeline, and on what terms. That scale is not hypothetical. Roche and Recursion signed for up to 40 programmes across neuroscience and oncology in 2021, and Chai has since designed antibodies against 52 targets in a single round. For the pharma, this is the stronger position. Terms agreed at signing are set before the platform has its proof; once it does, every new negotiation starts at a higher price. Each step is an option, not an obligation, so if the results don’t hold, the pharma simply doesn’t take it. And the team, the data and the scientific context already built carry straight into the next target instead of waiting out another procurement cycle. Align the incentives with each step, so the pharma’s rate improves as its commitment grows and the platform earns a larger, more predictable revenue line each time it delivers. The programme keeps moving instead of stalling while a new agreement is drafted.
3. No-go milestones. Pay the platform partner when it helps the buyer reach a correct no-go early. The Phase 2 success rate for AI-enabled assets in this year’s census is 19.5 per cent, and a correct early stop saves years and most of the cost of a failed trial. The vendor files a time-stamped prediction before the decisive experiment, as Formation Bio already does internally, and payment resolves when the result arrives. Fewer patients are enrolled in trials that were never going to help them.
4. Calibration pay. Before the buyer runs an assay, the vendor registers its predictions, and part of the fee scales with how well calibrated they were on the buyer’s own data. Self-graded benchmarks become buyer-graded ones. Scannell and Bosley showed in When Quality Beats Quantity: Decision Theory, Drug Discovery, and the Reproducibility Crisis in 2016 that a 0.1 change in the correlation between a model’s output and the human outcome can outweigh a hundredfold gain in throughput; almost nobody reports that correlation, and this pays someone to.
5. Difficulty-weighted fees. Price the work by how hard the target is. A well-studied target with known chemical matter pays the standard rate; an undrugged target, a protein-protein interface, a rare disease with no prior chemistry or a complex modality pays more, on a difficulty grade both sides agree at signing. Today 37 targets each carry 50 or more programmes while the hardest biology waits, and a flat fee gives a platform no reason to change that. Pharma already sends its hardest problems to the platforms it trusts most: Vas Narasimhan describes Novartis’s work with Isomorphic as being “on what we call undruggable targets.” Write that into the price, and the platform is paid for taking on the targets that most need it, while pharma pays less for the easy ones.
6. Weight escrow with a change-of-control trigger. Software solved vendor survival risk decades ago with source-code escrow, and licensing lawyers already recommend the same for platforms: a January note for the Licensing Executives Society by Dan Shores and Michael Ellenberger, of Rothwell, Figg, Ernst & Manbeck, advises sequestering or destroying customised models at termination, because the partner’s “commercial program may depend on this safeguard.” Apply it to fine-tuned weights, and add the clause pharma actually worries about: if the vendor is acquired by a competitor, the partner’s weights go to the partner and are deleted from the acquirer. A faster signature is worth more to a startup than a better price.
7. Use it or share it. If a partner does not advance a target within an agreed window, rights revert non-exclusively, or convert into a co-development with a royalty back. A shelved asset is worth nothing to the company that shelved it. Pair it with rescue rights, the way Nabla reworked a Takeda molecule that team could not stabilise.
8. Priced on a clock. Offer the newest model at a premium, as companies do today, because that revenue pays for the next generation. When the next generation launches, the one before it stays available at a lower price, on a schedule both sides agree up front. The partners who help most, by sharing data, building alongside or funding the work, see each new generation first and get the best rates, written into their contract from the start. The company keeps earning from every generation it has built. Smaller biotechs and academic groups get a proven model they can afford. Pharma knows what each generation will cost over the life of the contract and can plan its budget around it, with weight escrow (point 6) covering the risk of the company being sold. And more teams working with strong models is how more medicines reach patients sooner. Biohub has described a similar clock for data: its commercial funders get a one-year head start before the data goes public.

Own the loop, not the weights. When people ask who owns the IP in an AI discovery partnership, they are asking multiple questions at once. Three have settled answers. One does not, and it is the one that decides whether a partnership can be signed at all.

The Molecule, the Target and the Platform: Settled

The industry has handled molecule ownership for forty years and AI has not changed the pattern. The one public contract that shows it clearly is Exscientia and Merck KGaA from 2023, later filed by Recursion: collaboration IP vests in the pharma partner on creation, platform IP and generic improvements stay with the platform partner, and the platform partner agrees not to compete on the collaboration targets. The platform stays with the company that built it, and pharma should not contest this: a platform that cannot improve is worth less to everyone, including the partner that licensed it.

Two things have moved underneath. The first is the record: Patents on AI-derived drugs: empirical evidence and legal challenges, by Janet Freilich, Professor of Law at Boston University, and Arti Rai, Elvin R. Latty Distinguished Professor of Law at Duke, in Nature Biotechnology, finds genuine scientific progress alongside new legal challenges, inventorship among them, and a model’s contribution to an invention is not self-documenting. The second is inventorship. In November 2025 the USPTO rescinded its 2024 guidance; in practice AI is a tool, only people can be inventors, and conception has to come from a person. The two US patents covering Insilico’s rentosertib each name three people. The most practical answer in print is the two-part test in Augmented inventorship: rethinking patent law in the age of generative AI, by Haim Levy, of HVL Legal and Biosciences Strategic Advisory, in the Journal of Law and the Biosciences, directability and traceability: a person exercised substantive judgement in steering the model, and a reviewable record links that reasoning to specific claims. He proposes two short documents, a computational traceability report and a human and machine contribution statement, designed so that neither code nor weights change hands. Agree at signing who produces them. The decision log is not paperwork. It is what makes the patent hold.

The Improved Model: The Open Question

When a pharma’s proprietary data fine-tunes a vendor’s model, who owns the result? Goodwin’s life sciences group said as much in July: “There are no established answers.” Once data is absorbed into a model’s weights its influence persists, so contracts “should explicitly state what the model owner can retain or learn from that interaction.” Nathan Frey, Head of Drug Discovery and Development Research at Anthropic, was asked the same question from the other side at Endpoints AI Day this month, after a researcher suggested Claude might be training on prompts: “it’s very explicit opt-in opt-out,” and “we have not trained on user data or user transcripts.” That is the right default for any platform: nothing learned without an explicit opt-in, and the opt-in written down.

Both sides now describe the same artefact as theirs. Boltz describes its GSK agreement as one in which “Boltz scientists will work with GSK’s teams to fine-tune Boltz directly on GSK’s proprietary data.” Chris Boshoff, Chief Scientific Officer and President of R&D at Pfizer, brought in Eric Schadt, who left Pathos AI in August to become Pfizer’s first AI Scholar, an advisory role. GSK’s Kim Branson stated the pharma view as early as 2023: “a model that’s built on the joint of public and private data is ours, and a model on private data is obviously ours.” The federations have partial answers. Apheris lets a departing member keep the model it helped build, but it stops receiving updates. The terms published with CDD Vault’s TuneLab integration on 15 October say the quiet part plainly: “the influence your data has already had on trained models cannot be removed.”

Here is a split we think both sides can accept.

1. The base model stays with whoever built it. It existed before the partner arrived. Lilly’s TuneLab is the clearest working example, and it runs the usual direction in reverse, with the pharma as the model vendor. Aliza Apple, Vice President of Catalyze360 AI/ML and Global Head of Lilly TuneLab, states the premise: TuneLab “was built on the premise that the learnings accumulated from drug discovery data could be leveraged as a collective, instead of under-utilized behind a single company’s walls.” Lilly keeps the base models it trained on more than two million of its own preclinical data points; members fine-tune on their own data and contribute readouts through federated training no other member can see, and ChemLab v2 was trained on Lilly’s data plus 120 member-contributed datasets. Aliza Apple has drawn the line where we would: pre-competitive parameters like ADMET and developability do not need to be siloed, and the molecules themselves are the IP. Base model with Lilly, improvements shared by everyone who contributed, molecules with the member. Any vendor can offer a partner the same three-way split.
2. Partner-specific weights belong to the partner, or are exclusively licensed to it, and are never used to serve another customer. Nicholas Larus-Stone, Head of AI at Benchling, does not think any biotech “will tolerate cross-customer training,” and we agree.
3. General methods the vendor learns stay with the vendor, and the contract lists them. Architecture changes, training recipes and evaluation tools are fair to keep. Anything that encodes the partner’s chemistry or biology is not. Nicholas Larus-Stone thinks the dispute is smaller than it looks: “there is less to learn cross-customer than you would expect. Some things are universal, and those can usually be found in the public literature. Where companies differentiate is in their science.”
4. Exit and change of control are written on day one. What the partner keeps, what is deleted, what happens if the vendor is bought by a competitor, and an honest sentence that influence already absorbed into shared weights cannot be fully removed.

Daphne Koller offers the frame that makes this easier to agree. Responding to the “Reverse Information Paradox” of Satya Nadella, Chairman and Chief Executive Officer of Microsoft, she wrote that “the model is the swappable layer”; what compounds is the loop of data, tools, evaluations and people around it. Accept that, and the argument over the weights matters less than the agreement over the loop, and the loop is much easier to divide fairly.

Founders Need to Learn Too

A platform that cannot improve from its partnerships has no flywheel, and the discovery and compliance leaders we talk to understand that. What they cannot do is wave it through. Kim Branson put the constraint plainly this year: “Go ask your CMO if you can throw all your clinical trial datasets into a system for anyone to use for any purpose. Then tell me what they say.” It is tricky, he wrote, because “there are consents, approvals for some sorts of analyses and not others, rules about combinations,” and because of where the data comes from.

“Every one of our data points is from a human who, in many cases, has decided to contribute to medical research.”

Kim Branson, Chief AI Officer, GSK · LinkedIn · June 2026

The people whose data it is have said what they want, and it is not silence. Dana Lewis, patient researcher and member of the Clinical Trials Transformation Initiative’s Executive Committee, put both halves in April: “Participants can say yes or no. Both responses are valid,” and “We often default to saying, ‘It’s too hard to do anything with that data,’ even when patients are explicitly asking us to use it.” The same month showed what the promise protects against: de-identified UK Biobank participant data was found listed for sale online, 276 participants asked to withdraw, and Lord Clement-Jones told the House of Lords that “contractual promises are not an adequate safeguard for data of this sensitivity.”

The mechanisms that respect those rules already exist. Federated training lets a model learn from a partner’s readouts without the readouts leaving the partner’s environment. MELLODDY, a three-year project funded by the EU’s Innovative Medicines Initiative, proved the mechanics across ten pharma companies: 2.6 billion confidential activity data points on 21 million compounds, never pooled, and every company’s own models improved. TuneLab now runs the same idea as a service. A contribution clause names what the platform may learn, general methods and pre-competitive endpoints like ADMET and developability, and what it may never learn, anything that encodes the partner’s chemistry or targets, and an audit right lets the partner check; Katherine Leibowitz, of Leibowitz Law, advises that “Each party should retain the right to audit the other party’s use of AI tools.” The ask is reasonable. The mistake is leaving it implicit. A quality team will ask where the data lives, who can reach it, how long it is kept and what happens to the weights after training; Gourav Pandey, R&D Quality Lead at Takeda, listed exactly those questions this month, and they are easier to answer when the design already does. Deployment inside the partner’s own environment is not yet the norm (most platforms still serve from the cloud), but Richard Jaenisch, of Open Biopharma, expects it to become one: “You do it in our house, and maybe we have more control over it.” Vendors who design around consent, with federation, zero-retention terms and, where a partner needs it, deployment in the partner’s house, close faster than vendors who argue with it.

When the Molecule Stops Moving

Ownership also has to answer what happens when a partner stops. In pharma acquisitions, Killer Acquisitions in the Journal of Political Economy found, projects bought by an overlapping incumbent are 23.4 per cent less likely to continue, and 5.3 to 7.4 per cent of acquisitions in the sample were made to shut a competing programme down. Iambic’s IPO prospectus, filed on 15 October, warns that partners “may not pursue or prioritize” partnered candidates. We could not find a 2026 AI discovery deal with a published use-it-or-lose-it clause. That is an opening, and the use-it-or-share-it structure set out earlier is one way to take it.

Decide per layer, and say how long open stays open. Open models set the floor under every price in this market, and the floor rises every quarter. The founders who release weights and the founders who do not are both right, for different layers of the stack. The safety argument has moved from theory to a concrete red line. And the artefacts most people cite are no longer the ones anyone maintains.

Both Sides Are Right, Per Layer

Boltz is the clearest working example, and its founder has been clear about both halves from the start. Gabriele Corso, Co-Founder and Chief Executive Officer of Boltz, has said the company “will continue to put a lot of our models open source,” and also that “putting a model on GitHub is definitely not enough.” The open line is the base models; the commercial line is agents, fine-tuned models, enterprise deployment and a metered API. Chai took the other road: its only open weights, chai-1, were released in late 2024 and last updated in February 2025; the code around them is still maintained, but every model since ships through the API, and the GSK agreement is platform access. Both signed pharma this year. So the question is not which side wins. Open models set the floor and move the science; closed, fine-tuned models capture the revenue. The useful decision is per layer, data, model and product, and conflating them is most of why the argument goes in circles. Recursion, which builds its own models and uses open ones, says the same from the operator’s chair: its strategy, in Najat Khan’s words to investors in August, “isn’t to depend on any one model … whether they’re developed at Recursion or by the broader Open Source community,” because “no single model will remain state of art forever.”

What Open Weights Do for Pharma

The most important thing open weights do in this industry has little to do with ideology. They let a pharma company keep its data. Apheris can fine-tune OpenFold3 inside AbbVie’s and Johnson & Johnson’s own environments only because the weights can travel to the data, and Roche’s Pharma Day deck in September put “Data generation and AI independence” on the same slide as “Proprietary Foundation models.” Pharma funds open models because it does not want its science to depend on a single vendor’s roadmap. Mohammed AlQuraishi, Associate Professor at Columbia University, who leads OpenFold3, named the asymmetry: “The data most useful for binding prediction sits inside pharma and rarely reaches academic models.” The closed alternative shows the cost. When Isomorphic published its IsoDDE benchmark in February, Diego del Alamo, of Takeda, noted that it “comes after extensive efforts to partner with industry and potentially access their private structural data … so we don’t know how impactful that extra data is,” and Mohammed AlQuraishi was blunter: “The problem, of course, is that we know nothing of the details.” Vijay Pande has the shortest version of why pharma cares: “Open weights are what can make them owners.”

The clearest measurement of what that buys came from the Apheris co-folding work: a federated model trained across pharma partners learned from more than 20,000 additional private structures, and on 1,056 held-back structures from live programmes the share of high-quality interface predictions rose from 36 to 52 per cent, about eleven points ahead of the strongest public baseline. Robin Röhm, Co-Founder and Chief Executive Officer of Apheris, is careful about what is new: “This is not model architecture innovation.” The architecture is Boltz or OpenFold. The data is the contribution. Payal Sheth, Senior Vice President of Therapeutic Discovery Sciences at Bristol Myers Squibb, set the condition for joining: “Our data and IP must stay protected, the governance has to work for all members.”

Who Is Still Working on It?

Open is usually discussed as a binary. The more useful questions are two: is the code still maintained, and are new weights still being released? On the first, the commercial labs do better than their reputation. Boltz’s BoltzGen repository took 138 commits this year and was updated this month, and Chai and Profluent both still fixed their open code in 2026, even though neither has released new open weights this year. On the second, the pattern is consistent: the last new open weights from each venture-backed lab date from the generation before its commercial product, as the exhibit shows. Boltz is the commercial lab closest to still shipping open, which is consistent with its founder’s commitment to keep releasing open models. Meta’s ESM2 is the extreme case the other way: weights unchanged since March 2023, the code archived in 2024, and still millions of downloads a month. The labs that released new open weights in 2026 are institutes and consortia: Arc, the OpenFold consortium, which posted new OpenFold3 checkpoints in May, and Biohub, which acquired EvolutionaryScale and put ESMFold2 and ESMC-6B out under an MIT licence in May, in the words of Alex Rives, Head of Science at Biohub, because “It’s going to be open source, it’s going to be MIT licensed, and we want people to use it.” The Baker lab’s head was asked at Endpoints AI Day this month whether open would hold.

“At this point, releasing models open source is best for the world because the more people have it, the more people can do good with it.”

David Baker, Institute for Protein Design, University of Washington · Endpoints AI Day · 14 October 2026

Two footnotes to that record. Boltz-2 itself remains a preprint, and the published assessment of its performance, Reliability of AI Methods in Drug Discovery: Evaluation of Boltz-2 for Structure and Binding Affinity Prediction in the Journal of Chemical Theory and Computation, came in July from a group with no commercial stake in the result; the model and the evidence about it travel separately, so ask for the independent evaluation by name in diligence. And Profluent is the partial counterexample: it released E1 and Protein2PAM for commercial use and still signed Lilly and BioMarin this year, because, in the words of Ali Madani, Chief Executive Officer of Profluent, “the question is who can actually curate that dataset effectively.” Open weights cost it little because the thing being sold was never the architecture.

The institutes are not only still shipping, they are shipping the hardest results. RFdiffusion3 was published in Science on 15 October as Atomic functional site scaffolding in biopolymers with RFdiffusion3, and two companion papers in Nature the day before, led by Computational design of metalloproteases, reported de novo enzymes that break the amide bond, with 36 per cent of 135 metalloprotease designs cleaving at the targeted site. The Baker lab’s own review, The past, present and future of de novo protein design in Nature, written six months before those papers, states the consequence for binders and assemblies: “the key current questions in these areas are not how to design, but what to design.” Catalysis it lists as a live challenge, which is what the October papers began to answer. So the floor is real, and it is a 2024 and 2025 floor held up by institutes and mirrors.

The Friction Is Now Safety

The objection we hear most is no longer commercial. As these models get stronger, open release starts to look like a risk, and the people closest to it are more specific than the debate suggests. It has come up in every conversation we have had this year with an organisation that also releases open models, and in nearly every public interview on the subject.

At Endpoints AI Day this month David Baker, Professor of Biochemistry and Director of the Institute for Protein Design at the University of Washington, was asked how seriously he takes keeping models closed for safety rather than for proprietary gain. His answer was a process, not a slogan: “before we release it, we share it with experts in the field and groups who are concerned about biosecurity well in advance.” His judgement today is that “releasing open source is the way to do it. We will continue to reevaluate this.” He named the red line: “when you go beyond just designing individual proteins to really understanding how they’re affecting host-pathogen relationships.” And he proposed a control that closes nothing: “every time someone synthesizes a DNA sequence, it gets uploaded to a database in an encrypted form,” consulted only if a new virus appears. Two weeks earlier a Google DeepMind team published Function-preserving watermarking of AI-generated proteins in Nature. Both controls sit downstream of release. They let a lab stay open at the weight layer while making a synthesised sequence traceable, and they are cheaper than any access regime.

The sharpest caution came from an open lab. Eric Nguyen, Co-Founder and Chief Executive Officer of Radical Numerics, who helped lead the Evo models at the Arc Institute, said that when people used them to generate viruses “it scared a lot of folks,” and built a biodefence mandate alongside the capability work; his forecast is that once every modality reads into one system, the control it gives over biological sequence “will be something that we just won’t be prepared for.” His colleague Christine Parthemore, Head of Biodefense at Radical Numerics, said at Endpoints AI Day that this work needs “loops of really tight cooperation across the public and private sectors” that do not exist yet, though everybody is working toward them, and that “the single biggest bottleneck I think is the humans.” Nobody on either stage argued for withholding weights.

Which leaves a conclusion both sides can hold. Today the commercial answer and the safety answer point the same way, keep the frontier closed and open the floor, so no founder has yet had to choose between them. When a model does cross the line Baker named, secrecy by default will be the wrong reflex, because openness is how most of the good gets done: as Baker argues, the more people who have these models, the more people can design countermeasures to a new virus or new answers to drug-resistant bacteria. The responsible version of open has three parts, and all of them exist today: review by outside experts before release, traceability after it through synthesis logging and watermarking, and defensive work funded on the same schedule as the capability. Write those principles down now, with people outside the company, while the stakes are low enough to discuss calmly. The labs that do will keep both the openness the science needs and the trust that lets them keep it.

The model is the swappable layer, so survive on what compounds. Sixty-nine companies have put AI-enabled assets into human trials, three frontier labs have arrived with enterprise contracts, and free capability improves every quarter. Capability is arriving faster than obligation, and that asymmetry is what every platform, and every pharma partnership, now has to price around.

The Arithmetic

This is a hard business regardless of who else enters. The 2026 census, Artificial intelligence-enabled drug discovery: A growing but largely early-stage clinical pipeline, found 129 AI-enabled assets across 69 companies, Phase 1 success at 77.8 per cent and Phase 2 at 19.5 per cent. BioXcel, a company built on AI with an approved product on the market, filed for Chapter 11 at the end of August. A separate 2026 review, AI in Drug Discovery: Clinical Failures, Regulatory Reality, and the Validation Crisis Behind the Hype in Pharmaceuticals, holds that AI-assisted approval probability still sits inside the industry baseline of roughly 8 to 12 per cent, even where early timelines fell by 30 to 70 per cent, and lists the casualties by name: Recursion’s REC-994, Exscientia’s A2A antagonist, DSP-1181. Faster is established. Better is not yet. Robert Plenge, Executive Vice President and Chief Research Officer at Bristol Myers Squibb, drew the line in July: “The goal isn’t speed for its own sake; it’s raising the probability that each program we advance is the right one.” A company whose entire pitch is speed is selling the part buyers can already verify. The capital side has done the same sum. Peter Hebert, Co-Founder of Lux Capital, in January: “Investor expectations might have been too large.” Matt Gline, on his own company’s AI discovery bets: “As an investment strategy for Roivant, it was not particularly successful.” And Alex Zhavoronkov, who has outlasted most of his peers: “Our cohort is now a graveyard.”

What the Frontier Labs Change, and What They Do Not

Sajith Wickramasekara, Co-Founder and Chief Executive Officer of Benchling, gave the honest version of the anxiety: “In the AI era, the traditional biopharma industry is the underdog. Big tech is hiring scientists and building wet labs.” That is a claim about capability, not obligation, and the labs furthest in have chosen different paths. We are glad Anthropic is here. Its stated reason for building a wet lab and running its own preclinical programmes in rare and neglected disease is the one this industry should want to hear. Eric Kauderer-Abrams, Head of Life Sciences at Anthropic, said at the launch in June that “there’s no substitute for having our own experiences alongside you all in the trenches,” and “We believe in the power of tight feedback loops”; the diseases it chose are the ones commercial pharma has least reason to pursue. Its boundary is also public: no clinical trials for now, and “not competing with pharma and biotech companies that make their business in bringing drugs to market.” Take that as a commitment and as a description of the landscape at once. A frontier lab now has a Head of Drug Discovery and Development Research, preclinical programmes, a Bay Area wet lab, a life-sciences platform sold into pharma, Novartis’s Chief Executive on its board and, by one report, a drug-discovery team acquired for about $400 million in stock. Whatever it decides about trials, every platform company now shares a market with it, and every pharma partner should know which parts of its stack a frontier lab can supply. That raises the floor for everyone, and it means a founder’s answer to “what gets better here that a frontier lab could not replicate” now has to be true. OpenAI is asking the same questions in public; Greg Brockman, President of OpenAI, framed its approach to biology on No Priors on 15 October as “Do we have the data? Do we have the ability to get the data?” Isomorphic has chosen the other path in its own words, “a vertically integrated company” that will “take those potential medicines to patients in the clinic,” and is reported to be raising at a valuation of at least forty billion dollars before a trial has started. Max Jaderberg, President of Isomorphic Labs: “The programs we are going after are not fast-follower programs.” Demis Hassabis, Co-Founder and Chief Executive Officer of Google DeepMind and Isomorphic Labs, has described the goal as doing “almost all the exploration, which is 99% of the work and the time in silico.” Capability is arriving in all three; the duty to run a trial, carry a safety database, answer a regulator and stand behind a label for a decade is so far being taken on by one. The platforms that have been here longest are not conceding the point. Ramy Farid told investors this year that Schrödinger is “uniquely positioned to leverage AI in a way that cannot be replicated or replaced by a global AI model,” and Alex Zhavoronkov gave the opposite forecast in August, that platform companies “will either be acquired by frontier AI labs … or they will switch to assets.” Andy Plump was blunt about the pitch he keeps hearing from big technology companies: “they feel like they don’t need us... it’s all data, give us data and we’ll put this into our algorithms and that’s how we’ll make drugs. I don’t think that’s the case... you still need to run trials.” His forecast: “It’s just not going to happen in the way that it’s being predicted.” Aviv Regev, Executive Vice President and Head of Genentech Research and Early Development, said the same thing more gently at Roche’s Pharma Day: “AI cannot replace clinical trials in humans.”

The reasoning layer is also becoming cheap and sovereign. Thomas Fuchs, Senior Vice President and Chief AI Officer at Lilly, runs the largest open-weight language models on-premise, with no token limits, and stated the reason without hedging.

“No private company should have the right to restrict what AI research needs to be done to help these patients.”

Thomas Fuchs, Chief AI Officer, Lilly · Fierce Biotech · February 2026

When a strong general model can run inside a company’s own walls, pharma can work with the frontier labs by choice rather than by necessity: as partners, on terms that keep its data its own. That is a healthier footing for both sides.

The floor for reasoning tools rises for everyone: Nature published two multi-agent systems on the same day in May, Robin from FutureHouse (A multi-agent system for automating scientific discovery) and Co-Scientist from Google (Accelerating scientific discovery with Co-Scientist); Science followed with Biomni (Autonomous biomedical research with an artificial intelligence agent) and the Virtual Biotech (The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development), and all four are now a baseline a product has to beat. The open tools are being wired into those agents directly. David Baker said at Endpoints AI Day this month that for the software his lab is about to release, “we’re actually working with both OpenAI and Anthropic to have them well set to use them when the software comes out.” A founder whose product is a usable wrapper around open tools should assume the wrapper is being commoditised from above. Distribution changes shape: Bristol Myers Squibb has put Claude in front of more than 30,000 employees and Merck committed up to a billion dollars to Google Cloud, which makes the frontier-lab relationship an enterprise agreement, a different motion and budget from a discovery partnership. And the bar for quality rises with the volume. When answers are abundant, judgement becomes the scarce part. Kim Branson, who runs an AI scientist at R&D scale at GSK, has named the risk every team using these tools shares: “science slop,” output that “to a scientist who is not skilled in that domain, but wants to use the results, looks entirely plausible.” His answer raises the standard for everyone: the organisation’s best experts, encoded as scoring rubrics, so every result is judged the way they would judge it. That is the part only the organisation itself can supply, and it makes every model it uses, open or frontier, more useful.

If You Are Biopharma: Build a Portfolio, Not a Bet

The two most instructive buyers of 2026 made opposite moves and reached the same place. GSK bought five layers five ways. Lilly built one layer and opened it to more than 125 members.

The lesson is not that one is right. It is that the decision is per layer, buy, build, federate or open, and the answer differs by layer inside the same company. Make the internal AI team the owner of the portfolio rather than the rival of the vendors in it, and write their benefit into each agreement. Put idle validation capacity to work as currency; it buys more evidence than cash. Publish the blind-test protocol, not the results, so every vendor is measured the same way and your diligence compounds too. And treat your own consent commitments as a design spec for vendors rather than a reason to say no; the ones worth signing will build to it.

If You Are a Founder: Own What Cannot Be Copied

The companies we would bet on own something a generalist cannot copy. Some own a wet-lab loop no one else can run, which is why Surge Biswas, Co-Founder and Chief Executive Officer of Nabla Bio, argues “it’s a bad idea to outsource your feedback loop.” Some become the system of record inside a customer’s workflow, where improvement compounds on the customer’s own targets. And some go horizontal, as Boltz did into flavours and fragrances with dsm-firmenich, so that pharma’s procurement cycle is not their only clock. Dimension put the phase change in one line when Chai raised its Series C: “2025 was the year of breakthrough research. 2026 is becoming the year of deployment.”

Two people, an operator and an investor, converge on the same test. Abbas Kazimi, Chief Executive Officer of Nimbus Therapeutics, titled an essay in January “Why Owning the Learning Loop Matters More Than Owning the Lab.” Simon Barnett, Partner and Head of Research at Dimension: “Platform companies need strong, informed views on whether frontier AI labs may eventually subsume their technology.” Every founder should expect one question in diligence and be able to answer it in a sentence: what gets better here that a well-funded generalist could not replicate in eighteen months? If the answer is the model, it is not an answer.

The decisions that follow from that are few. Decide which tier you are in, and say it before your investors decide for you. Marc Tessier-Lavigne chose his on day one: Xaira can “take the time to find partners who are willing to preserve upside, like a 50–50 deal,” because “One of the moats will be data.” Alex Zhavoronkov chose the other end just as deliberately: “We decided to become a coach instead of a player,” and “I will license to them,” he said, “because I know they can develop it faster, cheaper and better.” Prove it on their targets, in their lab, and price on the result. Price the bottleneck you hold, which is data, throughput and a learning loop, not model quality. Write the exit, the change of control and what you may learn on day one, so compliance can say yes. And build a channel, because reaching scientists where they already work matters as much as the model.

What Is Being Missed

The category has done the hard part: it has put 129 assets into human trials and cleared Phase 1 at a rate the industry has never seen. What follows would turn that progress into proof, and every item is within reach this year.

Phase 2 is a scoreboard that counts, and the category is not winning it yet. The 2024 analysis that first made the case, How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons in Drug Discovery Today, found AI-derived molecules clearing Phase 1 at 80 to 90 per cent against a historical 40 to 65, and Phase 2 at about 40 per cent, no better than the industry average. Two years and a larger sample later the Phase 1 advantage holds and the Phase 2 figure reads 19.5 per cent. The samples are small and the cohorts differ, so this is not yet a trend. Artificial intelligence in drug discovery: what it is, where we stand and the path forward, a sixteen-author Perspective in Nature Reviews Drug Discovery in August, found “evidence of their clinically relevant impact is, so far, disappointingly limited,” and traced it to benchmarks that test models rather than decisions. Say it yourself before a sceptic says it for you. It is the only way the eventual win will be believed.

Benchmarks are self-graded. In the blinded AIntibody study, successes did not transfer across tasks, and in one task all but one model ranked clones worse than picking at random. Submit to blind external benchmarks, and price calibration into contracts.

The decision record. Papers and patents keep the successes. The reasoning, the dead ends and the changed minds are what teach judgement and what make inventorship defensible. Log decisions with timestamps from now, and put co-ownership in the agreement.

Representation decides who the medicines work for. Models trained on the available data make medicines for the people in that data. Michèle Ramsay, Director of the Sydney Brenner Institute for Molecular Bioscience at the University of the Witwatersrand, put the number on it in January: “less than 2% of the human genomes sequenced came from people of African origin,” and “If we don’t feed the AI algorithms with African data, what we get is going to be biased.” Priscilla Chan, Co-Founder of Biohub, stated the general case in April: “Before AI can simulate biology, it needs to see biology.” Treat cohort diversity as a scientific requirement of partnership design.

Policy risk is being carried unpriced. The BIOSECURE Act was enacted on 18 December 2025, the list of biotechnology companies of concern is due 18 December 2026, and counsel is already advising life sciences companies to screen AI and data vendors for exposure. We could not find an executive addressing it on a 2026 earnings call. The counterargument from Bruce Booth, Partner at Atlas Venture, belongs in the room too, that Chinese biopharma’s engagement is “more opportunity than threat,” which is a defensible position and a different question from procurement eligibility. David Baker put the scientific version of it at Endpoints AI Day: “the best way ahead is collaboration, not trying to cut things off.” Ask in diligence, on both sides, and write the contingency in.

1. Deal terms are written for this moment, not borrowed from the last one. Expect more creative structures, designed around what a platform actually delivers today rather than the licensing templates of the last decade. Engineering milestones in place of clinical ones are the clearest example: GSK and Noetik showed you can pay on an event both sides can observe this year. By next October we expect a growing share of new pharma and platform agreements to pay on deployment, blind-test performance or calibration rather than on clinical events, and the first calibration-pay clause to appear in a disclosed deal. The rest of the terms will change shape with the trigger: shorter agreements with renewal built in, milestones defined as tests the buyer runs, upfronts that look more like a subscription than a licence, and exclusivity measured in months on a model generation rather than years on a target. Peter Gassner, Chief Executive Officer of Veeva, described the same shift in the software next door in August, with pricing “more like modeled off of Amazon Web Services,” because “The customers want to be out of the experimentation phase.”

2. Learning together becomes how the boldest partners work, and it will be the most sensitive item on this list. TuneLab has more than 125 members, 675,000 contributed data points and two CRO networks running Lilly-approved protocols. We expect at least two more large pharma to open their own models to partners on contribution-for-access terms, and the first federation to pay members in proportion to what their data contributed, rather than in access alone. The sensitivity is about the data, so be precise about what moves. In a federation of this kind the readouts never leave each member’s environment, the model learns from them without their being pooled, and what is shared is the general, pre-competitive learning, endpoints such as ADMET and developability, not molecules and not targets. Payal Sheth described the test BMS ran before joining one: “Federation let us test whether combining data across companies helps, without any data leaving our environment.” The members who join will be the ones who can see that line written down.

3. The hands-on phase gives way to the product. The scientists and engineers platforms have placed inside their partners this year are how trust was built, and they will become the hardest part to scale. Every new partner wants the same attention, and a company that grows by headcount runs out of people before it runs out of demand. The agreements are the same. Landing a top-20 pharma partner usually takes custom terms, custom models and custom work, and that is necessary: those first partners are what make every later conversation possible. It does not scale either. As partner teams come up to speed and these tools become a normal part of how discovery is done, the work will shift from doing it for the partner to helping the partner do it: training, shared playbooks, and workflows that live in the partner’s own systems. The custom terms of the first agreements will settle, in the same way, into standard ones that the next partner can sign in weeks. We expect the platforms that plan that transition early, and write it into their agreements from the start, to be the ones that scale: every partner leaves the hands-on phase able to do this work well on its own, and the platform can bring the same to many more partners at once.

The repeated objection to everything above is that it concentrates on the cheap end of the pipeline. Derek Lowe, a Director in Chemical Biology and Therapeutics at Novartis, who writes In the Pipeline, put it bluntly in August: “All that stuff in early-stage discovery work is a roundoff error in time and money compared to the clinical trials.” Michael Santoro, of Santa Clara University, made the economist’s version of the point in July: “The value in pharmaceutical AI was never in making the cheap part cheaper.” They are right about where the money goes. Where the failure begins is a different question.

Need for NAMs: A systematic evidence synthesis revealing over half a century of drug development failure, a synthesis of 32 studies published in NAM Journal in July, found clinical attrition running between roughly 80 and 91 per cent by decade, worst in the two most recent, with about three-quarters of failures traceable to biology rather than trial execution. Richard Harrison’s Phase II and phase III failures: 2013–2015, in Nature Reviews Drug Discovery, put efficacy at 52 per cent of the causes and operational problems at 3. Minikel and colleagues’ Refining the impact of genetic evidence on clinical success, in Nature in 2024, found mechanisms with human genetic support are 2.6 times more likely to succeed, yet most programmes do not have it. Daphne Koller, citing L.E.K. analysis, counts the novel targets advanced per year falling from roughly 100 in 2015 to about 30 in 2024, while 37 targets each carry 50 or more programmes. Trials are where failure is detected. Target choice is where it is created, years earlier. As mentioned in Part I, in Koller’s words: “We are doing a pretty good job at manufacturing keys, but they are often for the wrong locks.”

David Shaywitz, of Harvard Medical School’s Department of Biomedical Informatics, who writes for the Timmerman Report, adds the caution that applies to every claim here: generating exquisitely reproducible data around a weak proxy, he wrote in August, “mostly allows us to characterize the wrong thing with greater precision.” That is the bar. Better molecules becoming better medicines is still a claim, not a result, and the clinic is where it will be settled. Vas Narasimhan gave the timeline in February: “the first purely AI generated candidate, by definition, cannot come out the other end for eight to ten years.” The same learning loop that found the molecule is now being pointed at the patient: which patients to enrol, which controls to borrow, which trials to stop early. Greg Meyers, Executive Vice President and Chief Digital and Technology Officer at Bristol Myers Squibb, has said BMS is “on track to take two to three years off the average clinical trial.” Whether that holds, and what it takes, is Monday’s note: the clinic, where nine in ten medicines still fail, the trusted tools that get trials right, the new ones changing how they run, and the new forms of evidence coming behind them.

If you are building one of these partnerships, on either side of the table, or working with these tools every day, tell us what you are learning: what worked, what stalled, and the progress you’d like to see in this field and in how partners support your work.

Quotations are verbatim from the sources below. Figures are from company disclosures and the peer-reviewed papers named. Research window: January to 18 October 2026.

• Deals: AstraZeneca and insitro press release, 15 October 2026. Profluent and Lilly, Business Wire, 28 April 2026. GSK and Noetik, Business Wire, 8 January 2026; Ron Alfa, Noetik blog, 8 January 2026; James Hardiman, DCVC, 8 January 2026. Relation and GSK, BioSpace, 30 July 2026. Chai Discovery and GSK, Business Wire, 9 October 2026; Christopher Austin, Fierce Biotech, 9 October 2026. Incyte and Genesis Molecular AI, 20 May 2026. Hansa Biopharma and Cradle, 1 October 2026; Stef van Grieken, LinkedIn, October 2026. Lilly TuneLab launch, 9 September 2025; Aliza Apple, LinkedIn, June to October 2026; CDD Vault TuneLab terms, 15 October 2026. A-Alpha Bio Atlas Consortium, GlobeNewswire, 22 July 2026. Boltz and Pfizer, BioPharmaTrend, 12 January 2026. Isomorphic Labs and Johnson & Johnson, 20 January 2026. Chai Discovery agreements, Business Wire: Lilly, 9 January; Pfizer, 4 June; Lilly TuneLab, 18 June; Novartis, 13 July; argenx, 15 July; Bristol Myers Squibb, 20 August; GSK, 9 October 2026. Generate and Zealand, Fierce Biotech, October 2026. Takeda and Nabla, BioPharmaTrend, October 2025. Boltz pricing page and WRN helicase case study, 2026. Iambic Therapeutics prospectus (S-1/A), 8 October 2026. Recursion and Roche (Genentech) collaboration, Recursion press release, 7 December 2021. Chai-2 technical report, Chai Discovery, 30 June 2025.
• Pricing and go-to-market: Joshua Meier, Endpoints News, October 2026. Jack Dent, Forbes, 4 June 2026. Jack Scannell, Decoding Bio, 6 July 2026. Vijay Pande, TechCrunch, 29 August 2026. Chris Arendt, BioCentury, 17 June 2026. Endpoints AI Day pharma panel, 14 October 2026. Thomas Fuchs, Fierce Biotech, 27 February 2026, and LinkedIn, June 2026. Shafique Virani, Business Wire, 8 January 2026. Ben Taylor and Najat Khan, Recursion Q2 2026 results call, 5 August 2026; Najat Khan, BioCentury Show, 19 March 2026. Alex Zhavoronkov, Insilico H1 2026 results call, 26 August 2026, and Fortune, 18 September 2026. Birgit Girshick, Charles River Q2 2026 results call, 5 August 2026. Karen Akinsanya, Schrödinger Q1 2026 results call, 5 May 2026; Ramy Farid and Richie Jain, Schrödinger Q4 2025 and Q2 2026 results calls, 25 February and 5 August 2026. Jason Kelly, Sequoia Training Data, 24 March 2026. Jonathan Chauvin, Simulations Plus investor day, 21 January 2026. Albert Bourla, Pfizer Q2 2026 results call, 4 August 2026. Jay Bradner, Amgen Q2 2026 results call, 4 August 2026. Martin Holst Lange, Axios, 22 September 2026. Dave Ricks, Financial Times, June 2026. Ashu Singhal and Nicholas Larus-Stone, Decoding Bio, 10 June 2026. Evan Feinberg, Latent Space, 2026. OpenADMET, 16 September 2026. Arc Institute Virtual Cell Challenge, 20 August 2026. Mike Nally, BioSpace, 2 March 2026. Rohan Ganesh and Simon Barnett, GEN, 6 July 2026. Jorge Conde, LinkedIn, July 2026. Tom Miller, BioPharma Dive, 16 October 2026. Diogo Rau, Fierce Biotech, 27 February 2026. Marc Tessier-Lavigne, Onyx Newsroom, 16 January 2026. Daphne Koller, What’s Your Number?, 4 February 2026. Vas Narasimhan, a16z podcast with Jorge Conde, 16 February 2026 (auto-transcribed recording). Dan Shores and Michael Ellenberger, Rothwell, Figg, Ernst & Manbeck, for the Licensing Executives Society, 19 January 2026.
• IP: Exscientia and Merck KGaA collaboration agreement, 2023, as filed by Recursion. Freilich and Rai, Patents on AI-derived drugs, Nature Biotechnology, 2026, doi 10.1038/s41587-026-03039-x. Levy, Augmented inventorship, Journal of Law and the Biosciences 13(1), 2026, doi 10.1093/jlb/lsag014. USPTO rescission of February 2024 inventorship guidance, 28 November 2025. Goodwin life sciences, 9 July 2026. Kim Branson, 2023 and LinkedIn article, 26 June 2026. Chris Boshoff, LinkedIn, July and August 2026. Nicholas Larus-Stone, Benchling, 2026. Daphne Koller, LinkedIn, August 2026, and Deep Phenotype, 3 August 2026. Apheris, Bio-IT World, 22 September 2026. Air Street Press, 13 August 2026. Cunningham, Ederer and Ma, Killer acquisitions, Journal of Political Economy, 2021. Iambic Therapeutics prospectus (424B4), 15 October 2026. Aliza Apple, Lilly and insitro release, 6 October 2026. Dana Lewis, Clinical Trials Transformation Initiative, 22 April 2026. Lord Clement-Jones, House of Lords, 28 April 2026; UK Biobank Oversight Committee report, 1 June 2026. Katherine Leibowitz, Clinical Leader, 15 May 2026. Gourav Pandey and Richard Jaenisch, Pharmaceutical Technology AI Pulse Check, 6 and 14 October 2026.
• Open models: Gabriele Corso, Latent Space, 12 February 2026, and R&D World, 23 June 2026. Robin Röhm and Payal Sheth, Bio-IT World, 22 September 2026. Roche Pharma Day, 28 September 2026. Hugging Face model repositories (Biohub, OpenFold, Arc Institute, Chai Discovery, Boltz) and GitHub repositories, read 18 October 2026; Meta’s ESM repository archived 1 August 2024. RFdiffusion3 release in Foundry, RosettaCommons, 22 December 2025. Wan and colleagues, evaluation of Boltz-2, Journal of Chemical Theory and Computation, published online 29 July 2026, doi 10.1021/acs.jctc.6c01334. Ali Madani, Air Street Press, 18 June 2026. Butcher and colleagues, RFdiffusion3, Science, 15 October 2026, doi 10.1126/science.aeh2982. Chen and colleagues, Nature, 14 October 2026, doi 10.1038/s41586-026-11167-2. Yang and colleagues, de novo protein design review, Nature, 29 April 2026, doi 10.1038/s41586-026-10328-7. David Baker, Eric Nguyen and Christine Parthemore, Endpoints AI Day, 14 October 2026. Stutz and colleagues, watermarking AI-generated proteins, Nature, 30 September 2026, doi 10.1038/s41586-026-10965-y. NTI biosecurity scan, spring 2026. Mohammed AlQuraishi, TechTimes, 23 September 2026, and, with Diego del Alamo, Scientific American, 19 February 2026. Vijay Pande, Substack, 14 July 2026. Alex Rives, Latent Space, 27 May 2026.
• The field: Haque and colleagues, AI-enabled drug discovery pipeline, Cancer Treatment Reviews 151:103223, 2026, doi 10.1016/j.ctrv.2026.103223. Khairil and colleagues, Pharmaceuticals 19(6):916, 2026, doi 10.3390/ph19060916. Jayatunga and colleagues, Drug Discovery Today, 2024, doi 10.1016/j.drudis.2024.104009. Bender, Scannell, Shaywitz and colleagues, Nature Reviews Drug Discovery, 7 August 2026, doi 10.1038/s41573-026-01496-2. Sajith Wickramasekara, LinkedIn, 2026. Andy Plump, Fierce Biotech, 13 January 2026, and Averin Capital fireside with David Berry, 28 September 2026 (auto-captioned recording). Ghareeb, Rodriques and colleagues, Robin, Nature, 19 May 2026, doi 10.1038/s41586-026-10652-y. Gottweis and colleagues, Co-Scientist, Nature, 19 May 2026. Huang and colleagues, Biomni, Science, 20 August 2026, doi 10.1126/science.adz4351. Zhang, Zou and colleagues, The Virtual Biotech, Science, 17 September 2026, doi 10.1126/science.aeg6779. Abbas Kazimi, LifeSciVC, 27 January 2026. Simon Barnett, GEN, 6 July 2026. Dimension on Chai Discovery’s Series C, 2026. AIntibody blinded benchmark, Nature Biotechnology, 2026, doi 10.1038/s41587-026-03238-6. BIOSECURE Act status, K&L Gates, 20 January 2026. Bruce Booth, LifeSciVC, June 2026. Robert Plenge, HLTH, 21 July 2026. Peter Hebert, PitchBook, 9 January 2026. Matt Gline, BioCentury Show, 10 July 2026. Eric Kauderer-Abrams, CNBC and Becker’s, 30 June 2026, and Reuters, 18 September 2026. Greg Brockman, No Priors, 15 October 2026 (auto-transcribed recording). Max Jaderberg, Forbes, 13 May 2026. Isomorphic Labs, IsoDDE technical report, February 2026. Demis Hassabis, Sequoia AI Ascent, April 2026 (third-party transcript). Aviv Regev, swissinfo, 28 September 2026. Michèle Ramsay, Business Day, 30 January 2026. Priscilla Chan, TIME, 30 April 2026. Peter Gassner, Veeva Q2 FY27 results call, 26 August 2026.
• The clinic: Filipova and colleagues, NAM Journal, 21 July 2026, doi 10.1016/j.namjnl.2026.100119. Harrison, Phase II and Phase III failures 2013 to 2015, Nature Reviews Drug Discovery, 2016. Hwang and colleagues, JAMA Internal Medicine, 2016. Minikel, Painter, Dong and Nelson, Nature, 2024, doi 10.1038/s41586-024-07316-0. Scannell and Bosley, PLOS ONE, 2016, doi 10.1371/journal.pone.0147215. Heyndrickx and colleagues, MELLODDY, Journal of Chemical Information and Modeling, 2023, doi 10.1021/acs.jcim.3c00799. L.E.K. Consulting, Is Biopharma Doing Enough to Advance Novel Targets?, 15 May 2025. Derek Lowe, In the Pipeline, 10 August 2026. David Shaywitz, Timmerman Report, August 2026. Michael Santoro, ProMarket, 2 July 2026. Greg Meyers, McKinsey, 6 November 2025.

Written by Tonya Brody. Terms, dates and figures are researched against primary sources, cross-checked across three LLMs to surface any discrepancies, and then, wherever possible, confirmed directly with the founders and others involved. Editorial judgment, opinions and anything in the first person are mine alone. I dictate edits with Wispr Flow and use Claude, Codex or similar tools for proofreading and, where useful, for illustrative images, which are occasionally AI-generated. Any AI-assisted edits or images are reviewed and approved by me before publication.