Newsletter · 2026-07-30 · Ashutosh Agarwal

The AlphaFold Team Breaks Up as AI Labs Start Making Their Own Drugs - AI Drug Discovery Weekly - Week of July 30, 2026

AI drug discovery newsletter for the week of July 30, 2026. Google is dismantling the AlphaFold team, moving researchers to Gemini and Isomorphic Labs; Anthropic disclosed it will develop its own preclinical drugs; and Regeneron, Bristol Myers Squibb and Eli Lilly showed that the proprietary patient data now seen as the real moat sits with the incumbents, all ahead of August 5 earnings from Recursion, Schrodinger and Lilly.

AI Drug Discovery Weekly

Week of July 30, 2026: The AlphaFold Team Breaks Up as AI Labs Start Making Their Own Drugs

The team that solved protein folding is being broken up, and the AI companies have decided to stop selling shovels and start digging.

TL;DR

• Google is dismantling the AlphaFold team. Researchers from the group that won a Nobel Prize for predicting protein structures have been moved onto Gemini or reassigned to Isomorphic Labs, Alphabet's drug discovery arm. Around a quarter of the team has already left the company outright. DeepMind's VP of research said the group's "longstanding grand challenges strategy has now evolved." Translation: the single most celebrated AI-for-science project of the decade lost an internal budget fight to a chatbot.
• Anthropic is going to develop its own drugs. Not tools for drug companies. Actual preclinical programs, for rare and neglected diseases. This was disclosed on BioCentury's podcast this week, and it changes the competitive question for every pharma company thinking about subscribing to its platform.
• Last week the consensus was "the model isn't the bottleneck, the data is." This week we found out who owns the data, and it's the incumbents. Regeneron has sequenced three million people and found over 100 protective genetic mutations, roughly 50 of which became drug programs. Bristol Myers Squibb expanded its own AI supercomputer to cut target identification from weeks to days. Eli Lilly is licensing access to a billion dollars' worth of its historical research data.
• Reid Hoffman is leaving the Microsoft board to run an AI chemistry company full time. He says Manas AI has produced cancer compounds that "the humans hadn't discovered," and that his computational chemists had never seen before. He calls it internally "Move 37 for chemistry."
• Schrodinger shipped the one real product launch of the week, an agentic AI assistant called Bunsen, bundled with NVIDIA and Google Cloud infrastructure. Notably, no podcast discussed it.
• Recursion had a genuinely silent week. No news, no analyst actions, no podcast coverage. Its only appearance was its CEO describing a laboratory that grew one trillion neurons.
• All three of our named stocks report earnings on the same day: August 5. Everything this week is pre-print positioning.

What's new

The AlphaFold team is being taken apart

This is the story of the week, and it deserves to be stated plainly because it is easy to under-react to.

AlphaFold is the reason most people outside biology have heard of AI in drug discovery at all. It predicts the three-dimensional shape a protein will fold into from its underlying genetic sequence, a problem that had resisted the field for fifty years. Its creators won a Nobel Prize. Before AlphaFold, humanity had experimentally determined about 170,000 protein structures over five decades. Afterwards, the count went to 200–300 million.

On the Daily Tech News Show (July 29), the hosts reported that many of the researchers from that team, "some of them, by the way, Nobel recognized for the work that they've done with AlphaFold", have been moved "onto Gemini or reassigned to Isomorphic Labs, which is Alphabet's drug discovery company that spun off from DeepMind." And the number that matters: "around a quarter of the team has apparently already left the company."

DeepMind's VP of research, Pushmeet Kohli, framed it as strategic evolution, the group's "longstanding grand challenges strategy has now evolved" toward a tighter focus on Gemini, which the hosts noted "is apparently Google's biggest priority. They're working really hard to compete at that frontier level."

The hosts read this as a win for Google, and there is a defensible version of that argument. As one put it: "DeepMind AlphaFold at one point was its own thing because it had to be its own thing. And folding it, sorry for the bad pun, into Gemini and having that be part of scientific research that Gemini is already capable of doing and is being used in research already. It just makes the whole thing feel much stronger." The AlphaFold tools remain open, and Isomorphic Labs continues.

But strip away the framing and here is what happened: the most scientifically celebrated AI project in the world was competing internally for talent against a general-purpose chatbot, and it lost. A quarter of the people have walked. For a sector whose entire investment case rests on "AI will transform biology," the fact that the flagship proof of that thesis was not defensible as a standalone priority inside the company that built it is worth sitting with.

There is a related signal in the same week. On Tech Brew Ride Home (July 29), the discussion of how AI labs are racing toward self-improving systems noted that Demis Hassabis is allegedly pursuing a different path, through "world models" rather than through coding agents that recursively improve themselves. Last week we covered Cythera's Woody Sherman arguing for physics-based "nanoscale world models" for molecules instead of language models. These threads may be converging: the bet inside DeepMind may be that molecular science comes back through world models, not through a dedicated structure-prediction team.

Anthropic is becoming a drug developer

BioCentury's July 28 episode contained a disclosure that deserved more attention than it got.

The episode was built around a mapping exercise: reporter Lindsay and her colleague Karen Tkach-Tusman spent time working out how the big AI companies' recent releases fit together across the pharmaceutical research stack. Anthropic released Claude Science at the end of June. Google is building something similar. NVIDIA sells the tools and the compute underneath. Amazon builds agents for specific tasks. OpenAI is working on the model itself.

Then, mid-conversation, a host asked: "Lizzie, I have a question. I thought I heard that is Anthropic getting into preclinical drug development itself?"

The answer: "Yes. So... Anthropic is now going to be developing preclinical programs for rare and neglected diseases. They haven't said much else beyond that."

The follow-up question is the one every pharmaceutical executive will now ask: "So if you're a company thinking about using its tools, does that raise any red flags for you? And now this company is also my competitor?"

BioCentury put this to Anthropic directly and spoke with Jonah Kuhl there, who said the platform is "not training on the user's data or holding their data in any way. And that that is not part of their business model." The stated motive is to understand the difficulties pharma faces in developing drugs. BioCentury did not simply accept it: "I think whether that fully plays out, we'll see if there's something else going on there, too."

Why this matters commercially: the entire pitch of platform AI to pharma is "give us your proprietary data and we'll make you faster." If the platform provider is also running its own pipeline, even in rare and neglected diseases, which are commercially unattractive and therefore a diplomatic place to start, the trust calculation changes. BioCentury predicted the workaround: "an emergence of alternative ways to tap into models. So things like federated learning or other consortia that are grouping or collecting data in a protected way for a specific purpose where you're not submitting your data necessarily to a potential competitor." (Federated learning means training a shared model without any party handing over its raw data.)

For scale context on who is doing this: on Equity (July 22), Menlo Ventures' Matt Murphy noted Anthropic crossed a $47 billion revenue run rate as of May, against $9 billion in 2025, and credited its enterprise focus and products including "Claude Code for life sciences." A company at that run rate entering preclinical development is not a science project.

The incumbents just showed their hand, and their data is enormous

Last week's issue found three independent voices converging on one claim: the AI model is no longer the constraint, the right laboratory-generated data is. This week we got the answer to the obvious follow-up question, who actually has that data, and the answer is uncomfortable for the pure-play AI drug discovery companies.

Regeneron. Aris Baras, who runs the Regeneron Genetics Center, gave what may be the single most substantive interview of the week on Business Of Biotech (July 27).

He starts with the real problem, and it is not modeling. His team analyzed every drug target ever pursued, every approved medicine, everything ever tried in a trial, and found it covered "something like maybe 10% of the genome. 10%. There's 20,000 genes and we'd only been able to really go after about 10%." Hence, in his words, "our industry has a 90% failure rate. I mean, most people don't appreciate that. We wake up every day and most of what we do fails."

Their approach: find people whose genetics protect them from disease, then build a drug that mimics that mutation. The inspiration was PCSK9, the Dallas Heart Study found people carrying a broken version of that gene had "90% lower rates of heart attacks and very low levels of bad cholesterol." Baras's point about why this de-risks everything: "The clinical trial has almost already been run. You can look at thousands of people with their lifetime worth of data missing that gene, and look at the benefits that they have. Okay, well, now all of a sudden, it's not looking like 90% failure rate."

The catch is that these people are rare, "one in 1,000, one in 10,000", and one example proves nothing. "finding one of them is great. But to be absolutely sure, you want to see hundreds of them. You want to see thousands of them." Which is the entire justification for the scale: three million people sequenced.

Thirteen years in, the result: "We have found over 100 of these amazing, kind of mesmerizing, almost unbelievable stories of protective human genetic factors. We've taken about half of them or so, maybe 50, and turned them into new medicines over a decade now into Regeneron's pipeline." Including in the hardest areas, neurodegeneration, and even cancer: "People born with genetic variants that protect them from ever getting a cancer. Imagine the possibilities."

He walked through one live example without naming the target: a gene with "the largest effect size we've seen in terms of reducing liver fat and reducing the risk of liver disease of any genetics, of any gene, of any target." Because the protective version of the gene is broken, "we immediately know we have to make an inhibitor." Because their transcriptomic data showed it sits inside liver cells, "Can't do it with an antibody." So they routed it through their Alnylam partnership into a gene-silencing drug now in late-stage development, with data hoped for later this year. That is the whole loop, from population genetics to modality selection to clinic, and note how little of it is AI.

Where AI does come in is split into two layers, and the second one is the strategically interesting part.

The mundane layer: they use it for "variant calling", sorting which of the billions of genetic variants per genome are meaningless and which actually matter, and for automating annotation across "hundreds of thousands of images." Baras credits early adoption of "certainly Claude Models" to his retiring chief data officer Jeffrey Reed.

The strategic layer, in his own words, is the quote of the week:

it's phenomenal what these companies have done in developing these AI models. And they're only getting better. And more efficient, possibly. But they do tend to start to look the same, right, when they're trained on public data. So Anthropic and OpenAI, they're largely training on a lot of the same data. So they can differentiate on their models. But what about if you could pair those models or even better, train them on the type of data that we're generating here?

He then draws the spending comparison explicitly: the AI companies are putting billions into data centers to train on public data, while "the billions we've spent and the decade we've spent building these proprietary data sets that are more specific and unique to healthcare." His view of the prize: pair hundreds of millions of health records spanning decades with genomic and proteomic depth, and "you can finally start to train models to understand human biology and predict it."

His stated endgame leaves no room for ambiguity about whether this program levels off: "I really see a future where we're going to be generating the genome and the proteome on everyone. So there is no diminishing returns. There is no finish line here. We're going all in."

One genuinely candid admission, which cuts the other way: being an early sequencing leader "is becoming much less of a strategic advantage for us." Sequencing costs have collapsed. The moat is the linkage between genomes and decades of health records, not the sequencing itself.

Bristol Myers Squibb. Science News Daily (July 22) reported BMS expanded its AI platform, building on "an initial DGX superpod started about three years ago, aiming to speed target identification from weeks to days." (A DGX superpod is NVIDIA's pre-packaged AI supercomputer.) Leadership framed the goal as "reducing manual work, and improving the likelihood that advancing programs are the right ones", note that second clause, which is about raising the hit rate, not the speed. No financial figures or commercialization timeline were disclosed. BioCentury separately flagged this as part of a trend: "maybe big pharmas that have the capital to make those kinds of deals, as opposed to like a small biotech that maybe can only access the platform itself, will be able to do more faster as well."

Eli Lilly. Lilly is doing the same thing from a third angle, renting out its data as a moat. On July 23, Sprint Bioscience formally joined Lilly TuneLab, the AI drug discovery platform Lilly launched in late 2025. The arrangement gives Sprint access to predictive models trained on Lilly's proprietary research data, described as representing over $1 billion of historical research investment. A separate industry report the same day highlighted Lilly's roughly $1 billion co-innovation lab with NVIDIA and its "LillyPod" supercomputer, live earlier in 2026 with 1,016 Blackwell Ultra GPUs.

Put the three together and a pattern emerges that should worry anyone long the pure-plays: large pharmaceutical companies are building in-house what the AI drug discovery companies sell. They have the data, they have the capital for compute deals, and as BioCentury's reporting put it, "the limiting factor is not the model or access to the platform, but more so the data and the types of assays that they're doing."

Reid Hoffman is going all-in on AI chemistry

On Masters of Scale (July 25), in conversation with Satya Nadella, Reid Hoffman announced he is stepping off the Microsoft board at year-end after ten years to return to "founder mode" at Manas AI, his AI-driven chemistry company.

His reason is a result, and he described it carefully. Manas has an internal name for what they are seeing: "we've got an internal description of Move 37 for chemistry because we're seeing chemistry that might actually in fact take shots at really interesting cancers and other things."

(For anyone who missed the reference: Move 37 was the move AlphaGo played against Lee Sedol in 2016 that no human player would have chosen, that looked like a mistake, and that turned out to be brilliant. It became the standard shorthand for AI producing something genuinely outside human intuition rather than merely faster.)

The validation he offers is human expert surprise: "we've got some of the best computational chemists in the world and they looked and said, that's very interesting, we've never seen it before, and that might work." And crucially: "the kinds of chemistry we're doing, like the humans hadn't discovered this chemistry at Manas."

His framing of why this happens is the useful part of his argument. AI, he says, is "an alien intelligence... a different version of AI that has been built to very much heavily mimic human intelligence, which creates a whole bunch of utility for us, but its patterns of reasoning are not the same as ours. And people frequently encounter that and they go, oh, that's scary. You're like, well, no, actually... it's also wonderful because it enables new things."

To his credit, he did not oversell the timeline: "it's very early, right? And so these things take a while with all the INDs." (An IND is the application a company files with the FDA before it can test a drug in humans, the first real regulatory gate.) A novel compound that impresses chemists is a long way from a drug.

The week's other genuine science

A designed protein cut CRISPR's error rate by more than five-fold. Reported on AI Chat (July 24): a Chinese research team used AlphaFold to redesign the CRISPR-Cas9 gene-editing protein, testing 23 amino acid swaps across 10 key positions. Off-target edits, cuts in the wrong place in the genome, which is the central safety problem in gene editing, fell from 28% to 5%, while accuracy on the intended target held. Published in Nature. A striking mechanistic detail: over 95% of off-target sites showed changes in which amino acids were contacting the guide RNA, even where the protein's overall shape was unchanged. This is a good example of AI structure prediction doing something useful that is not "design a new drug from scratch", it made an existing tool safer.

Scape Bio published a Nature paper on AI-designed mini-proteins for a class of targets that has resisted biologics. On The Bio Report (July 22), co-founder and CEO Christopher Norn laid out the gap. GPCRs, the receptor family behind a large share of approved drugs, and the family the GLP-1 obesity drugs act on, are mostly drugged by small molecules and peptides. Antibodies barely work here: "there are very few examples of approved drugs. They're three in total. And of those three antibodies, none of them have the ability to activate the receptor."

The reason is physical. These receptors sit embedded in the fatty membrane of a cell and are not soluble, so the standard route to making an antibody does not work: pull the receptor out of the membrane and "they adopt slightly different conformations and then the molecules you create that way end up being non-functional." Scape's workaround is to screen for binding "directly inside human cells, leaving the receptors in the native environment."

Their designed molecules are mini-proteins, "just 40 to 70 amino acids long," against antibodies that are "about 15 times longer", which Norn argues gives them a better physical fit: "we are able to penetrate quite deeply into the pocket of them and lock them into the desired state." He describes them as "selective like antibodies, stable like small molecules," and formulatable to survive digestion for gut-restricted therapy.

The result: working with David Baker's lab, "we have now done for 11 different GPCRs... we can generate modulators to either activate or [inactivate] these receptors." The Nature paper was released the same day as the episode.

There is also a nice detail on where language models fit that has nothing to do with molecular design. Norn used them for target selection, historically "an exercise of many people in a room discussing that specific target... endless discussions" where "everybody would have their own favorite GPCR." Instead: "with LLMs, we can do that systematically. We have used billions of tokens for doing this across, I think it was 3,000 different opportunities of target indication mode of action for GPCRs to land at a list of now around eight." He kept humans in the loop: "we are still humans and we look at it and we have maybe in some cases some industry knowledge that lets us rank differently from what the AI would pick." One more honest note: for generating the protein backbones, they use Baker-lab methods like RF diffusion but also "large libraries of native scaffolds which sometimes express a little bit better than the AI-generated backbones." Nature still occasionally beats the model.

Company is small and early: spun out of the Bioinnovation Institute in Copenhagen in June after two and a half years of technology building, seed raised, raising more.

Genomic language models are moving into manufacturing. Off Script (July 28) covered a corner of the stack this newsletter has under-served: applying language models trained on genomes to strain optimization and biomanufacturing yield, making the engineered organisms that produce antibodies and biologics more productive. Discovery gets the headlines; manufacturing yield is where a lot of the margin actually sits.

Public money is funding the data flywheel. Crain's Daily Gist (July 29) noted Northwestern University received a $20 million NSF award to build a cloud laboratory for protein engineering, where AI proposes protein designs, researchers synthesize and test them, and the results feed back as training data. Modest in size, but it is exactly the "generate your own data" architecture that last week's Xaira and Lila Sciences episodes argued was the whole game, now showing up in federal science funding.

Regulation: the FDA's AI trial pilot is drawing its first real criticism

Note to File (July 22) read through the public comments on the FDA's request for information on its pilot program for AI-enabled optimization of early-phase clinical trials. This is worth tracking because it is where AI in drug development stops being a research story and becomes a compliance story.

The sharpest comment came from Massive Bio, which argued the FDA should design this as "a regulatory grade human governed AI enabled trial orchestration program, not a narrow model evaluation or a sponsor only operational experiment." The hosts agreed with the criticism: "it seems like a sponsor only operational experiment. That's who they trotted up on stage, right? Two sponsors."

They also questioned starting in oncology, where "enrollment generally not super high. There's a lot of data points there that are very messy... It seems like a weird place to start." Though they conceded the counter-argument: if the real goal is connecting trials to electronic health records, oncology makes sense "because the clinical trials become clinical care. And so it should be in the EHR."

There was open skepticism about the private company at the center of it: "entering data straight into the paradigm system is what's getting streamed to the FDA, which is also weird. I mean, obviously Paradigm spent a bunch of time raising a bunch of money being kind of dark. And then all of a sudden they like emerge and they have this relationship."

The tone shift is the signal: "when it first came out, it was just everybody sort of clapping about how much of a huge move towards innovation this is. And I just was like, is it, I don't understand. And now at least people are coming out and being a little bit more critical."

A companion episode from the same show (July 30) supplied the clearest statement of the current rule of the road. Discussing Gleam, a company using AI to help trial coordinators with data entry, the standard was described as two-part: the tool must be "fit for purpose" and there must be "an accountable human on the other end." Enforcement targets companies that deploy AI without human review, or use tools unfit for the intended purpose. That is a low bar in principle and a meaningful constraint in practice, it rules out the fully autonomous version of nearly every trial-automation pitch.

Separately, the same episode read the FDA's restrictions on Chinese clinical data as industrial policy rather than data-quality policy: "are they just trying to signal that they don't want the clinical research work to move offshore?"

The debate

Last week's debate was about a bottleneck. Three separate voices said the AI model was no longer the constraint, and the argument was over what replaced it: more and better laboratory data (Xaira, A-Alpha Bio, Lila Sciences) versus a better way of representing molecules that would need less data (Cythera's physics-based approach).

This week the debate moved somewhere more uncomfortable: if data is the moat, then the moat belongs to companies that spent the last twenty years accumulating patients, not to companies that spent the last five accumulating GPUs.

Here is the case laid out from both directions, because both were argued this week.

The incumbents win. Regeneron has three million sequenced genomes linked to decades of health records, and 100 validated protective mutations, 50 of which are already drug programs. That took thirteen years and cannot be bought. BMS has its own AI supercomputer. Lilly is renting out a billion dollars of research history as a platform. BioCentury's mapping put it most crisply: the differentiating middle layer of the stack is where "pharma or biotech can really differentiate, even though they're using a platform or a model that others also have access to", because "the limiting factor is not the model or access to the platform, but more so the data and the types of assays that they're doing, which is not necessarily going to change at the same rate." And when asked directly whether platform access is a leveler for small companies: "I don't know if, say, a small biotech accessing the model is going to put them on the same level as a big pharma using Claude Science."

There is a second-order version of this argument that is worse for the small players. It is not just about data volume, it is about capital for compute. BioCentury noted that big pharma can strike compute-layer deals, like BMS with NVIDIA, "as opposed to like a small biotech that maybe can only access the platform itself." And the cost of running agents is now a real line item: NVIDIA's Kimberly Powell noted that "every time an agent waits and is thinking about what to do next, it's spending tokens." BioCentury drew the analogy to what already happened in tech: "we saw this concept of token maxing, right, for a while. Everybody just wanted to move as fast as they can, encouraging all their employees to just use the tokens. And now they're pulling back because people are like, wow, that's really expensive." Nadella made the same point on Masters of Scale, warning that "people can go crazy on spending lots of tokens in ineffective ways," and arguing that "whoever figures out that I can use tokens more efficiently for an outcome that matters in the world is going to get ahead." Compute cost discipline is becoming a competitive variable in drug discovery, and it favors balance sheets.

The challengers win. The counter-case is that incumbents have the wrong data, and that the interesting results this week came from small teams. Scape Bio is a handful of people in Copenhagen who just published in Nature and cracked eleven receptors that three approved antibodies could not activate. Manas produced chemistry that the best computational chemists in the world had never seen. Chai Discovery's protein foundation models were described by an investor who has held them since seed as "exceptional." None of those needed three million genomes. They needed a specific, well-chosen problem and a good representation of it.

There is also a real argument that the incumbents' data advantage is narrower than it looks. Baras himself conceded that being an early sequencing leader "is becoming much less of a strategic advantage" as costs fell. Human population genetics tells you which targets are worth pursuing. It tells you almost nothing about how to design the molecule that hits them, which is where the AI-native companies actually operate. Regeneron's own liver example proves the point: genetics picked the target, but the drug came from an Alnylam partnership. Target selection and molecule design are different businesses, and the pure-plays are mostly in the second one.

And the awkward third possibility, which the AlphaFold news raises: the frontier AI labs take the whole thing. Anthropic is at a $47 billion revenue run rate and starting its own preclinical programs. OpenAI is building GPT Rosalind, described on BioCentury as "GPT 5.5, but boosted up with scientific information on chemistry, additional data sets that are being used to train the model to be more specialist in science." Google is consolidating its scientific talent into Gemini and its drug programs into Isomorphic Labs. These are organizations with lower costs of capital than the entire public AI drug discovery sector combined, and they have just started competing with their own customers.

Where we come out. The honest read is that the AI drug discovery pure-plays are being squeezed from both ends, and this week made that worse rather than better. Above them, pharmaceutical incumbents with irreplaceable patient data and the capital to buy compute are building the capability internally. Beside them, frontier AI labs with vastly cheaper capital are moving down the stack into drug development itself. The pure-plays' claim to the middle, that they uniquely combine computational skill with biological insight, is exactly the claim both neighbours are now contesting.

There is one asset the incumbents genuinely cannot replicate quickly, and it is the same one that came up last week and again this week: negative data. BioCentury made the point directly, "you need the negative data. And that's not something that's in [Claude] Science or the protein data bank that you could tap into with the connector. That's something that comes with experience and the proprietary data from the company." Knowing what failed, and why, is not in any public database because nobody publishes failures. A company with a decade of its own failed experiments has something no model can download. That, rather than model quality, is the strongest remaining case for the specialists.

And one reassuring note for anyone who worried this all makes expertise obsolete. Asked whether a PhD with years of experience still has an edge, the answer on BioCentury was unhesitating: yes, "because you have to be able to kind of know what you should be asking and know what you should not be asking. And so you need the context around the science and the experience to guide the agent." The analogy offered was to programming: "people are saying, oh, nobody's going to program anymore. But right now, the most effective users of [Claude] code are people who know how to program."

Stocks in play

All three of our named stocks report second-quarter results on the same day, August 5. Recursion reports before the open, Schrodinger after the close, Lilly that day as well. Everything below is pre-earnings positioning inside a quiet period.

Schrodinger (SDGR), $15.35, up 1.5% on the week. The only one of the three with a real product event. On July 27 Schrodinger launched the early access version of Bunsen, which it describes as "its new agentic AI co-scientist that helps researchers understand scientific objectives, develop computational strategies, execute sophisticated molecular discovery workflows and interpret results."

The positioning is deliberate and worth quoting, because it is precisely the "physics plus AI rather than AI instead of physics" argument this newsletter has tracked for weeks: "By combining AI with physics-based simulation, Bunsen allows researchers to apply their expertise at greater scale... Unlike general-purpose AI systems that only retrieve and summarize existing scientific information, Bunsen is optimized to execute Schrodinger's validated computational methods and leverages decades of the company's physics-based molecular discovery expertise."

The second half of that claim is the commercially interesting one: "Bunsen makes advanced computational methods accessible to experienced drug hunters who are not computational chemists." Schrodinger's structural problem has always been that its software requires a computational chemist to drive it, which caps the addressable user base inside any given customer. If a natural-language layer genuinely removes that constraint, it expands seats per account. CEO Ramy Farid framed it as removing friction from the discovery cycle and expanding the platform's user base. Full commercial release is targeted by the end of 2026.

It also arrives bundled with serious infrastructure partners. As part of collaborations with NVIDIA and Google Cloud, both "will provide a co-engineered full stack AI platform to support early Bunsen customers," including Google Cloud's elastic compute and integration with the NVIDIA BioNeMo Agent Toolkit and NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. NVIDIA's director of digital biology, Anthony Costa, commented publicly on the stack.

Two cautions. First, this is early access, not revenue, full release is five months out, and the August 5 call will be the first chance to hear anything about early traction. Second, and more telling: not a single podcast in our sweep discussed Bunsen. For a company whose narrative depends on being seen as an AI leader, launching an agentic product in the same week the sector's practitioners were talking about Regeneron's genomes and DeepMind's reorganization, and getting no airtime at all, is itself a data point about mindshare.

Context on the print: consensus looks for a loss of roughly $0.41 per share on about $47.2 million of revenue. Standing full-year guidance from the May quarter was software annual contract value of $218–228 million (10–15% growth over 2025) and drug discovery revenue of $55–65 million. Street price targets cluster in the high teens to mid-twenties.

Recursion (RXRX), $3.03, up 0.7% on the week, still close to its 52-week low of $2.77. A genuinely empty week. No company announcements of substance, no analyst rating changes or price target revisions, and no podcast coverage. The stock has now been effectively flat for three consecutive weeks in the $3.01–$3.03 band, having traded as high as $7.18 in the past year.

Its only appearance anywhere was CEO Najat Khan speaking at the U.S. Science Summit on July 23 about the Roche/Genentech neuroscience alliance, where she described the partnership generating one trillion iPSC-derived neuronal cells, described as roughly the cellular equivalent of twelve human brains, inside an autonomous laboratory, to run a whole-genome knockout study systematically disabling all 20,000 human genes. She tied that output to Recursion's Salt Lake City supercomputer, BioHive.

Two ways to read this, and they matter for how you position into August 5. The generous read is that this is exactly the data factory that last week's episodes argued is the actual moat, and Recursion is one of very few public companies that owns one outright. Note the symmetry with Regeneron: 20,000 genes is the same denominator Baras used when he said the industry had only ever addressed 10% of them. The skeptical read is that Recursion has been describing impressive-sounding scale for years while the share price has fallen by more than half, and conference commentary is not a deal, a milestone, or revenue. A trillion cells is an input. Investors are waiting on outputs.

Also worth noting for ownership: BlackRock filed an amended passive stake of 8.5%, about 44.54 million Class A shares.

Eli Lilly (LLY), $1,155.27, down 2.6% on the week, and down 4.5% on Thursday alone. Lilly had by far the busiest week, and almost none of it was about AI. Two things to separate here: what happened during the week, and what happened Thursday.

During the week, the headline was retatrutide, Lilly's triple-hormone obesity drug, reporting topline results on July 23 from two Phase 3 trials. Both met their primary endpoints. In TRIUMPH-2 (type 2 diabetes plus obesity), patients on 4 mg, 9 mg and 12 mg lost an average of 29.8 lbs, 45.4 lbs and 49.6 lbs respectively at 80 weeks, with blood sugar reductions of up to 1.6%. In TRIUMPH-3 (severe obesity with established cardiovascular disease), patients lost up to an average of 55.8 lbs at 80 weeks. Secondary cardiovascular measures at the highest dose were strong: triglycerides down 37.0%, non-HDL cholesterol down 16.5%, systolic blood pressure down 9.3 mmHg, waist circumference down 7.5 inches, and high-sensitivity C-reactive protein, an inflammation marker, down 51.2%.

The caveats are what moved the stock. On hard cardiovascular events, the results were not clean: 44 events on retatrutide versus 52 on placebo on the broader five-component measure (a hazard ratio of 0.82, suggesting benefit), but 27 versus 23 on the narrower three-component measure (a hazard ratio of 1.12, suggesting none). Lilly noted such events "occurred less frequently than anticipated in both retatrutide and placebo arms," which is the standard explanation for an underpowered result, but it leaves the cardiovascular case unproven. Tolerability was a real issue at high doses: discontinuations due to adverse events reached 11.6% at 9 mg in TRIUMPH-2 and 13.5% at 12 mg in TRIUMPH-3, against roughly 4.8–4.9% on placebo. And the filing timing slipped: Lilly plans to submit for U.S. approval in Q1 2027, a delay from previously expected late-2026, while it completes the manufacturing data package. Secondary coverage also noted peak weight loss fell short of the 28.7% seen in the earlier TRIUMPH-4 trial. BMO viewed the data favourably on overall benefit-risk and best-in-class efficacy potential.

The second pressure point is commercial. Foundayo (orforglipron), Lilly's small-molecule GLP-1 pill approved in April 2026, has had a soft launch, with weekly prescriptions stalling for five consecutive weeks. Goldman Sachs cut its second-quarter Foundayo sales estimate to $40 million against a street consensus near $100 million, and cut its 2026 U.S. revenue estimate for the drug from $1.1 billion to $755 million. Goldman nonetheless maintained a Buy and a $1,283 target, arguing strong international Mounjaro sales and a likely guidance raise could offset the slow start. Other targets moved up through late July: JPMorgan to $1,400, UBS to $1,425, Bernstein to $1,385, Truist to $1,370, Bank of America to $1,334.

We do not have a confirmed single driver for Thursday's 4.5% drop, and we are not going to invent one. What we can say is that the stock ran up to $1,210 during the week and then gave back more than it gained in a single session, six days before an earnings print where consensus for a key new launch may need to come down by 60%. That is a setup where positioning matters more than news.

Also on Lilly's week, all non-AI: a $750 million joint investment with contract manufacturer Resilience to expand U.S. capacity, upgrading Ohio facilities to increase production of Lilly's KwikPen injectable devices and creating 400 jobs, part of a broad industry response to the threat of 100% tariffs on branded drugs unless companies reshore or cut prices. Novo Nordisk sued Lilly in New Jersey federal court under the Lanham Act, alleging Lilly's Zepbound and Mounjaro television advertising misleads patients by citing older SURMOUNT-5 data that compares Lilly's highest doses against lower doses of Novo's drugs, ignoring Novo's newly approved 7.2 mg formulation; Lilly says its ads are grounded in direct scientific evidence and will defend vigorously. A former executive medical director in Lilly's neuroscience unit filed a discrimination and retaliation suit in the Southern District of Indiana. And retrospectively, Lilly's $6.3 billion acquisition of Centessa closed during the quarter, including up to $1.5 billion in contingent value rights, securing a pipeline of orexin receptor 2 agonists for sleep disorders, a deal whose read-through drove Piper Sandler to raise its Alkermes target to $65 from $43 on July 28, noting Alkermes is up 76% since the Centessa announcement.

Read-throughs

For NVIDIA (NVDA). NVIDIA appeared in this week's material more often than any pure-play AI biotech, and never as a competitor, always as the toll booth. It supplies "the tools for the workbenches and models that scientists can use, and the infrastructure and compute that are needed to run them." BMS expanded a DGX superpod. Lilly runs 1,016 Blackwell Ultra GPUs. Schrodinger's new product ships co-engineered with NVIDIA's BioNeMo Agent Toolkit on RTX PRO 6000 Blackwell servers. BioCentury expects more compute-layer deals. Whichever side of the data-versus-model debate wins, NVIDIA sells to both, and the emerging token-cost discipline arguably pushes customers toward buying more efficient compute rather than less of it.

For Alphabet (GOOGL). Two contradictory signals in one week. The AlphaFold team's dispersal says scientific moonshots lost internal priority to the core model business. Meanwhile Google Cloud is one of the two infrastructure partners behind Schrodinger's Bunsen launch, and Isomorphic Labs is absorbing some of the AlphaFold talent. The coherent reading is that Alphabet is consolidating: research talent into Gemini, drug programs into Isomorphic, commercial biology into Cloud contracts. That may be a better-organized business. It is a worse story for anyone who valued Alphabet partly on breakthrough science optionality.

For the CRO and clinical trial infrastructure complex. The FDA pilot criticism cuts against the most aggressive AI-in-trials narratives. The operative standard, tool must be "fit for purpose," with "an accountable human on the other end", rules out the fully autonomous versions being pitched, and preserves the human-services layer that contract research organizations sell. At the same time, Craig Lipset's argument on The Digital Health Roundtable points the other way over a longer horizon: if AI genuinely lowers development costs enough that nonprofits and patient foundations can act as drug sponsors, that expands the number of trial customers even as it compresses per-trial spend. Also relevant to the geography of this business: the FDA's restrictions on Chinese clinical data appear to be onshoring policy, which raises U.S. trial volumes and costs together.

For Certara (CERT) and the simulation software group. Schrodinger's move is the template, wrap validated physics-based simulation in a natural-language agent to reach users who are not specialists. Anyone selling technical scientific software with a steep learning curve now faces the same question from customers: where is your agent layer? This is an opportunity for whoever ships credibly and a threat to whoever does not, because the alternative for the customer is to build the middle layer themselves on top of a general platform, which BioCentury explicitly identified as the place companies can now differentiate.

For the gene editing names. The AlphaFold-redesigned Cas9 result, off-target edits down from 28% to 5%, published in Nature, is a reminder that AI's near-term contribution to gene editing may be improving the tools rather than picking the targets. The clinical view from the sickle cell discussion pointed the same direction: better editing targets, prediction of combinatorial editing effects, and finding a safer replacement for busulfan in the conditioning regimen. Those are unglamorous improvements that would each meaningfully de-risk trials currently taking 9–12 months to accrue 50 patients.

For private AI-bio valuations, and therefore for public comparables. Money continues to arrive. Chai Discovery and Etched landed rounds pushing valuations into multibillion-dollar territory, explicitly drawing comparisons to publicly traded leaders in their sectors (thefly, July 23), and the following week brought $1.7 billion for Atoms plus another Etched round (thefly, July 30). Dimension is deploying $1.65 billion into science and compute with a five-person team. Menlo has backed roughly eight drug discovery model companies including Xaira and Vilya. The uncomfortable arithmetic for public investors: Chai Discovery is privately valued in the billions while Schrodinger's entire market capitalisation is $1.1 billion and Recursion's is $1.6 billion. Either the private marks are wrong, or the public pure-plays are cheap, or, most likely, the market is drawing a real distinction between owning a frontier protein model and owning a legacy software business with an AI story attached.

A caution on "AI-designed drug" claims. One guest on This Week in Startups (July 22) stated that an AI-designed drug has reached Phase 3 regulatory review. We flag it because it matters, and because it conflicts with last week's reality check from Thibault Geoui, who argued no true AI-designed drug has been approved and that the one or two often cited are repurposed existing drugs. Reaching Phase 3 review and being approved are very different milestones, and we have not independently confirmed which compound is meant. Treat as unverified and worth watching.

What changed vs last week

Last week's frame held, and then inverted. The July 23 issue reported a convergence: three independent voices saying the AI model is no longer the bottleneck, the right data is. The swing question we left open was whether value accrues to whoever owns the data factory, or whether physics-based approaches sidestep the data problem entirely and shift the moat back to modeling insight.

This week the data-factory thesis was not refuted. It was confirmed and then handed to the wrong people. BioCentury independently restated it almost verbatim, "the limiting factor is not the model or access to the platform, but more so the data and the types of assays that they're doing", which makes four consecutive weeks of convergence. But the companies displaying the biggest data factories this week were Regeneron, Bristol Myers Squibb and Eli Lilly, not Xaira, Lila or A-Alpha. If data is the moat, the incumbents are further ahead than the narrative allowed.

The negative-data point has now been raised by three independent sources in two weeks. Last week A-Alpha Bio argued the Protein Data Bank is both too small and missing negative results. This week BioCentury raised it unprompted: "you need the negative data. And that's not something that's in [Claude] Science or the protein data bank." This is graduating from an interesting observation into the most durable specific claim in the sector.

Anthropic's role changed category. Last week Anthropic appeared as an enabler, Claude Science, calling AlphaFold and Evo2 in a browser, ordering from IDT, with a low cost of capital meaning "the most shots on goal." This week it appeared as a prospective competitor to its own customers, developing preclinical programs in rare and neglected diseases. That is a different company to have in your supply chain.

The John Jumper thread resolved in an unexpected direction. Last week we reported, and flagged as unconfirmed, a claim that Anthropic had hired John Jumper away from DeepMind, and read it as big AI labs moving toward physics and molecules. This week's news that a quarter of the AlphaFold team has left the company and the remainder has been split between Gemini and Isomorphic Labs is consistent with senior departures, and reframes the story. It is less "AI labs are moving into molecules" and more "DeepMind stopped defending its molecular science group, and the talent dispersed." We still have not independently confirmed the Jumper hire.

The physics-versus-data debate got a commercial product. Last week Cythera's Woody Sherman argued for physics-based representations over language models. This week Schrodinger shipped a product built on exactly that premise, explicitly contrasting itself with "general-purpose AI systems that only retrieve and summarize existing scientific information." The argument now has a paying-customer test attached to it, with a full release due by year end.

Insilico moved from named example to absence. Last week Insilico was discussed as the standout technology-biology company without an approved AI drug. This week it appeared in no new episode.

Token economics are new. This did not feature in any prior issue. Two separate conversations this week, BioCentury on biopharma budgets and Nadella on Masters of Scale, treated the cost of running AI agents as a strategic variable rather than a rounding error. Expect this to become a recurring line item in how these companies are evaluated.

Prices barely moved, which is itself the story. Recursion has printed $3.02, $3.01, $3.03 across three consecutive weeks while sitting near a 52-week low. Schrodinger recovered last week's 3% decline. Lilly gave back its retatrutide gain and more. Nothing in the AI drug discovery narrative, not the AlphaFold reorganization, not Anthropic entering drug development, not Schrodinger's product launch, has moved these stocks in either direction. The market is not trading this theme week to week. It is waiting for August 5.

The swing question for next week. Now that the frontier AI labs are moving into drug development and the pharmaceutical incumbents are building AI in-house, what is left for the public pure-plays? Watch the August 5 prints for the only answer that matters: is Schrodinger's software annual contract value still compounding at the guided 10–15%, and has Recursion converted any of its enormous data-generation capacity into a partnership, a milestone, or a clinical result?