Newsletter · · Ashutosh Agarwal

AI Drug Discovery Pure Plays Split on Earnings Day - AI Drug Discovery Weekly - Week of July 30 to August 6, 2026

A synthesis of what AI drug discovery podcasts and Q2 earnings calls said for the week of July 30 to August 6, 2026, as Schrödinger and Recursion reported on the same day and drew opposite grades, while the sector converged on the view that proprietary data, not the AI model, is the real moat.

AI Drug Discovery Weekly

Week of July 30 to August 6, 2026: AI Drug Discovery Pure Plays Split on Earnings Day

TL;DR

  • The big event was earnings. All three companies we track reported on Tuesday, August 5. The two AI-native "pure-plays" printed within hours of each other and the market gave them opposite grades. Schrödinger (SDGR) beat and raised, signed a deal to put its new AI assistant across Bristol Myers Squibb's entire research group, and jumped. Recursion (RXRX) missed on revenue, cut its spending, collected another partner milestone, and barely moved, still stuck near its 52-week low.

  • Eli Lilly (LLY) delivered a blockbuster beat-and-raise on the back of its weight-loss franchise. Its relevance to this newsletter is mostly indirect: Lilly's $2.3B purchase of a startup called Ajax, a company built on Schrödinger's software, is exactly what handed Schrödinger a one-time gain and a $10M milestone this quarter. The AI-drug story and the GLP-1 story are now financially wired together.

  • The single loudest theme across the podcasts this week: the AI model is no longer the hard part. Owning data you generated yourself is. A thoughtful investor at Braidwell laid out why, and the founders of Chai Discovery, who are partnered with Lilly, Novartis, Argenx and Pfizer, described the "flywheel" that turns that data into an advantage. This is the third straight week the "data, not the model" idea has dominated, and it is now the most durable claim in the sector.

  • At the frontier, Google reshuffled its entire AI leadership. Demis Hassabis was moved "upstairs," and legendary engineers Jeff Dean and Sanjay Ghemawat walked out to start a company, Discovery Loop, whose entire pitch is that "AI is the researcher", with drug discovery explicitly on the roadmap. This is a direct continuation of last week's AlphaFold-team break-up story: the people who built AI-for-science at Google keep leaving to do it somewhere else.

  • A smaller but real coverage-universe item: Vertex signed a $28M collaboration with AbCellera (ABCL) on next-generation antibody-style cancer/autoimmune drugs, another vote for the "partner with the AI-antibody specialist" model.

What's new

1. The AI model is not the bottleneck. The data is. (This is now the whole thesis.)

If you read one thing this week, make it the Biotech 2050 Podcast (Aug 3) interview with Nick Myerberg, Head of AI at the crossover fund Braidwell. It is the clearest articulation yet of where the smart money thinks the edge lives.

His core argument, in plain terms: the fancy AI model everyone talks about is becoming a commodity. He put it directly: "model architecture... is likely already is and will continue to be more and more commoditized. Of course, the value isn't so much in the model architecture. If you look across the frontier labs in the LLM space... models, by and large, look the same." What actually separates winners, he argues, is a company's ability to generate its own proprietary data, experiments you run yourself, in your own labs, that a competitor can't scrape off the internet, and to feed that data back into the model so that "the next piece of data that you generate is more informative than the last." He called the exhausted alternative bluntly: "the value that was latent in publicly available data sets has been exhausted, at least in certain domains, and... it didn't get us particularly far."

Myerberg also offered a memorable reframing of AlphaFold, the protein-structure breakthrough that won a Nobel Prize. He argued its success wasn't only Google's model, it was "the harvest of this scientific institution that was built over generations of scientists," namely the roughly half-century and "an order of magnitude, say, $50 billion" that structural biologists spent building the public database of experimentally-determined protein structures the model trained on. His warning to founders who think one model will "solve all of human disease in 10 years": AlphaFold solved one very specific, very well-defined problem because it sat on top of a rare, pristine dataset. Most of biology has no such dataset waiting.

He named the companies Braidwell has backed and why:

  • Proxima: works on "molecular glues" and other proximity modulators (a class of drug that pulls two proteins together rather than simply blocking one). Historically these have been found "through serendipity"; Proxima's edge is generating the interaction data that doesn't otherwise exist.
  • Lila Sciences: building "autonomously driven laboratories" so the AI's predictions "confront nature" quickly and cheaply. Braidwell co-led its Series A about a year ago.
  • An unnamed antibody-design company that pairs a strong model with a "lab in the loop," which he says is racing ahead because the public antibody datasets are unusually good.

The same "won't pay for the data" complaint anchored a whole episode title this week, Data in Biotech (Aug 5), "Why Biotech Talks About AI But Won't Pay for the Data It Needs." The tension it describes is the flip side of Myerberg's thesis: everyone now agrees data is the moat, but generating it is expensive and unglamorous, so it stays underfunded.

Why it matters: this is the lens to judge every company in the space by. Do they generate their own differentiated data, and does it compound? Recursion's entire pitch (a giant in-house data factory) and Schrödinger's (physics-based simulation that generates "ground-truth" data) are both, in effect, answers to this exact question.

2. Chai Discovery: "drug design is another scaling problem"

The week's other must-listen is Training Data (Aug 4), a long interview with the two co-founders of Chai Discovery, Josh and Matt. Chai is one of the most-watched private AI-drug companies, and the title captures their belief: that designing a drug will become an engineering discipline, the way writing code is, rather than the "trial and error" and "serendipitous" process it has always been.

Josh's background is telling: he was on the early team at OpenAI (GPT-1, GPT-2, and the original "scaling laws" work) and framed the whole company around one question: "if the models can learn to speak English, German, French, why can't they learn to speak DNA and protein?" Matt came from pure math and theoretical computer science, not biology, and made the point that the field is less scary than it looks: an antibody, a mini-protein, "these are all just sequences of amino acids... different types of prompts for the model."

The most investable part was how they described their business model and the data flywheel. Chai partners with big pharma rather than building its own drug pipeline, and they named the partners: "Eli Lilly, Novartis, Argenix, Pfizer." Crucially, they argued the partnership model forces rigor a captive pipeline doesn't: "when we ship models at Chai, they really have to work... those partners are not going to come easily." And they pushed back on the cliché that pharma can't use AI: "a lot of people told us that pharma doesn't know how to use AI... to be honest, that hasn't really been our experience... when they see the data, they go all in." Their reason pharma has no choice: "Eli Lilly is a trillion-dollar pharma company right now. If they don't get more blockbuster drugs, they will not be a trillion-dollar pharma company forever."

They were candid about the data problem too. Their gold-standard training data is the same public protein database going back to 1970 that Myerberg described, "legit lab scientists who since 1970 have just been depositing crystal structures... without that, structure prediction and design wouldn't have been a thing", and their answer to its limits is the same flywheel: use the model in-house, generate new data on what it can and can't do, and retrain. One counterintuitive prediction: better AI might mean more lab work, not less, "the same way there's more demand for software engineers now that they've become more productive... there might be more demand for the lab."

Why it matters: Chai is the clearest private-market benchmark for the public pure-plays. It is partnered with the same trillion-dollar customer (Lilly) that just posted a blockbuster quarter, and it validates the exact "partner-and-compound-your-data" model that both Schrödinger and, to a lesser degree, Recursion are pursuing.

3. Google's AI brain trust walks out, and drug discovery is on their whiteboard

Last week the story was that Google was breaking up its Nobel-winning AlphaFold team. This week the shake-up went to the very top, per Tech Brew Ride Home (Aug 5):

  • Demis Hassabis, the head of Google DeepMind, was moved to chair of Google DeepMind and chief scientist of Alphabet, and will keep leading Isomorphic Labs, Alphabet's AI-drug-discovery arm. He framed it around health: "I've always believed the number one application of AI should be to improve human health... what better way to demonstrate that than to help finally cure diseases like cancer." Several commentators read the move less charitably, as Hassabis being "kicked upstairs" so he can pursue his preferred long-term research bets (like "world models") while Google races to catch up on more commercially urgent coding AI.
  • Jeff Dean (Google's ~30th employee and one of the most revered engineers in the industry) and his longtime collaborator Sanjay Ghemawat are leaving, along with two more top DeepMind scientists, Oriol Vinyals and Quoc Le, to found a public-benefit company called Discovery Loop. Google is a founding investor; other backers include Radical Ventures and Vinod Khosla's firm.

Discovery Loop's entire premise is worth understanding because it points straight at drug discovery. Its pitch is to automate the scientific method itself: "you propose an experiment, you implement what you need to run the experiment, you evaluate the experiment, and you get results", running thousands of these loops to make "superhuman advances" in, explicitly, "chip design, biology, drug discovery, and material design." Khosla drew the distinction that makes it different from today's tools: "Humans have been using AI to do research, not AI to be a researcher... The fundamental thing in Discovery Loop is that AI is the researcher."

Why it matters: two weeks running, the people who pioneered AI-for-science inside the company that arguably invented it keep leaving to do it as focused, independent bets. That is bullish for the idea of AI drug discovery and, arguably, bearish for the notion that any one big tech platform will simply own it. It also raises the competitive bar for every pure-play: the next wave of competition is a murderers'-row startup that treats "AI as the scientist," not "AI as a tool for the scientist."

4. A cheaper way to make drugs: rescuing shelved ones, with AI doing the reading

Business of Biotech (Aug 3) featured Annette Bakker (Children's Tumor Foundation) and Andrew Lo (the MIT professor famous for applying financial engineering to drug development) on a different flavor of "AI in drug discovery": using it to rescue drugs that companies abandoned. The AI role is unglamorous but real, chatbots working alongside human experts can "sift through humongous amounts of literature" to find promising shelved assets in days, work that "would have taken literally years."

The financial insight was the interesting part. Lo argued there is now "a very credible economic case for a completely private sector based solution" for ultra-rare-disease drugs, built around the priority review voucher (PRV), a transferable reward the FDA grants for developing certain rare-disease drugs, which can be sold to other drugmakers. Voucher sales are now "north of one hundred million dollars," some "over one hundred and fifty million dollars." Lo's provocative claim: "even if you never charged a single dime for the drug after approval, even if you gave it away for free, investors would still make a very, very handsome return just from the sale of the PRVs." Bakker's practical obstacle: inside big pharma, exiting a shelved drug "is nobody's job", getting one drug (Gomekli) out of Pfizer to SpringWorks reportedly took "200 volunteers." She and Lo are building a vetted database of shelved assets to fix that.

Two threads here connect to prior weeks: the FDA held a drug-repurposing meeting on August 5 (Bakker was invited), and the "Chinese biotech can outrace them" secrecy fear surfaced again, a recurring undertone in the sector's policy conversation.

The debate

Is any of this actually working yet, or is it still a promise? The most useful reality check came from the buy-side roundtable on Biotech Hangout Ep. 191 (recorded Jul 31), a week that set up the earnings. The panel of biotech investors and bankers landed on a nuanced view worth holding onto:

  • The optimistic case: AI is genuinely transforming the process of doing science, writing protocols, designing experiments, writing up results, reviewing papers, turning "weeks and months" of scientist time into "10 minutes." One panelist noted a research-tools company, Cadence, "has added like over 150,000 scientists to their site in just like the last couple months." The space is "red hot," and interest from big pharma in AI is "really strong", there was even (dismissed) speculation that Anthropic could "just snap up Bristol Myers or AbbVie."
  • The skeptical case, and it's important: "making molecules with AI doesn't necessarily solve our industry's bottleneck," one investor cautioned, because "the real cost is actually in the clinical side where AI is helpful, but not as helpful." Another put it as: a platform "can have unlimited power and limited ability... to then reduce that to practice." The conclusion: AI will be "a technology that's going to get embedded in everybody's research," an enabler rather than "the sole driver."
  • And directly on our two pure-plays, a panelist said the quiet part out loud: "recursion and Schrodinger... have been very levered to AI... if you look at those stock charts, they haven't really been all that great."

That last line is the whole debate. This week's earnings were the first real test of whether the "levered to AI" pure-plays can convert the hype into numbers, and they answered very differently.

A second, narrower debate continues on the clinical-trials side. Note to File (Aug 2) stayed skeptical of the FDA's push to weave AI and real-time data into trials, noting that in one trial (TRAVERSE) only 2 of 29 sites had actually implemented the initiative since it launched in December 2023, a "ready, fire, aim" pattern where the problem being solved is still unclear. The reminder: even where AI works in the lab, the expensive, slow clinical stage is where most of the cost and risk still lives.

Stocks in play

All three companies reported Q2 FY2026 on Tuesday, August 5. Prices below are as of Thursday, August 6.

Schrödinger (SDGR): the pure-play that delivered

  • Price: $17.81, up 8.6% on the day (Aug 6), and up roughly 16% from last week's $15.35. This was the standout move of the group. Market cap ~$1.3B; still well below its 52-week high of $23.02.
  • The print (a clean beat and raise on the metrics that matter):
  • Revenue $58.9M vs. ~$47.2M expected.
  • Annual Contract Value (ACV), the best gauge of its recurring software business, was $29.6M, up 27% year over year, comfortably above its full-year guide of 10–15% growth. Software revenue was $32.5M, and the higher-margin "hosted" (cloud) portion jumped to 47% of software, up from 31% a year ago.
  • It raised full-year drug-discovery revenue guidance to $65M–$75M (from $55M–$65M) on a $10M milestone, and reiterated its full-year ACV guide of $218M–$228M (note: a wire service rendered this as a garbled "$21M–$228M", the correct figure is $218M–$228M).
  • Reported EPS of $0.08 looks like a huge swing to profit, but be careful: it was heavily flattered by roughly $49M of one-time other income, mostly a gain tied to Lilly's acquisition of Ajax, plus the $10M Ajax milestone. This is not clean operating profit.
  • The AI substance: the morning of the print, Schrödinger announced a strategic agreement with Bristol Myers Squibb to deploy Bunsen, its new "agentic AI co-scientist", across BMS's research organization at scale, alongside its large-scale chemical-exploration tools and RetroSynth (AI synthesis planning). Chief Scientific Officer Robert Abel: "We are thrilled they are deploying Bunsen at a large scale." CEO Ramy Farid tied it to the week's dominant theme almost word for word: "as biopharma navigates a rapidly evolving AI landscape... the need to generate ground-truth data has never been greater. By integrating our highly accurate, physics-based simulations with AI, and launching our agentic co-scientist Bunsen... [we are] building the definitive computational infrastructure for the future of drug discovery."
  • The tension worth noting: Bunsen is Schrödinger's biggest product story, yet, as last week, it drew essentially zero podcast discussion. The commercial event (a major pharma deploying it broadly) is real; the outside conversation hasn't caught up. Bunsen's full commercial release is still targeted for end of 2026.

Recursion (RXRX): the pure-play that didn't

  • Price: $3.18, roughly flat on the day and up about 5% from last week's $3.03. Still languishing near its 52-week low of $2.77 (high was $7.18). Market cap ~$1.7B.
  • The print (a soft one): revenue of just $7.7M, well below the ~$12.0M expected, with a small per-share loss. The story management chose to tell was about discipline and endurance, not a breakout: it cut full-year cash operating expense guidance by $15M to $375M, and said its cash (about $556.8M at quarter-end) gives it runway "into early 2028." The CFO framed the cut as "doing the same amount or more with less."
  • The AI/partnership substance (the one genuine positive): on the call, Recursion announced that Genentech advanced the Roche/Genentech collaboration's first neuroscience target, described as "previously unexplored", into a joint early discovery program, which management held up as proof its platform can "discover new biology," not just optimize known biology. It has now realized over $216M across the Roche/Genentech partnership and over $500M in cumulative cash inflows across all partnerships, with more than a dozen discovery milestones hit. It also touted platform stats: 50+ petabytes of proprietary data; candidates reached using ~330 compounds over ~1.5 years versus an industry norm of ~2,500 over ~4 years, and hired Hoifung Poon (from Microsoft Research) as Chief AI Officer. On the internal pipeline, its lead program REC-4881 (for the rare condition FAP) has additional Phase 2 data coming in November.
  • The read: measured against the "data, not model" thesis, Recursion is the purest bet on generating your own data at scale, but this quarter it still could not point to the clinical or commercial inflection investors are waiting for. The message was "we have runway to 2028 and milestones keep trickling in," which is why the stock barely moved. It remains a show-me story.

Eli Lilly (LLY): the giant, and the financial plumbing behind the pure-plays

  • Price: $1,192.80, up ~2% on the day and up about 3% from last week's $1,155.27. Market cap ~$1.12 trillion.
  • The print (a blockbuster beat and raise): revenue $22.97B, up 48% year over year, vs. ~$20.7B expected; non-GAAP EPS $8.38 vs. ~$6.58 expected (both including a hefty $3.03 of charges for acquired research). Mounjaro sales were $9.9B (up 91%) and Zepbound $4.9B (up 46%); the new oral GLP-1 pill, orforglipron (brand Foundayo), booked $98M in its first quarter of sales. Lilly raised full-year revenue guidance to $85B–$87B. Its next-generation triple-hormone drug, retatrutide, now has a complete data package, with a US filing planned for Q1 2027. Analysts responded with a wave of price-target increases (Wells Fargo to $1,330; BMO to $1,400; Morgan Stanley to $1,419; Cantor to $1,410).
  • Why it's in an AI newsletter: Lilly doesn't market an AI-discovery platform, but it is the financial anchor of the ecosystem this week. Its $2.3B acquisition of Ajax Therapeutics, a company built on Schrödinger's platform, is precisely what generated Schrödinger's one-time gain and $10M milestone this quarter. Lilly is also the lead-named partner for Chai Discovery. When Chai's founders said a trillion-dollar pharma company "will not be a trillion-dollar pharma company forever" without new blockbusters, Lilly is the company they meant, and it is spending accordingly across AI-native partners.

Read-throughs

  • AbCellera (ABCL): Vertex signed a $28M-upfront collaboration with AbCellera on next-generation, multi-specific T-cell engagers (antibody-style drugs) for autoimmune disease, per Telltales (Aug 2). (To be precise: Vertex's big move that week was a separate $10B all-cash acquisition of Crelenetics; AbCellera was the small science partnership alongside it, not an acquisition.) It's a modest dollar figure but a meaningful vote of confidence in the "partner with the AI-antibody specialist" model, the same antibody-design thesis Braidwell and Chai spent the week championing.
  • The buy-vs-build framing is now everywhere. Telltales, an AI-produced markets podcast, tied the week's pharma deals to the hyperscaler capex debate: a company "with real free cash flow buys" a capability instead of building it, and "the price of that shortcut gets set by whoever else has cash that week." That is exactly the dynamic playing out in AI drug discovery, Lilly and BMS have the cash to buy or deploy platforms at scale; small biotech does not. It reinforces last week's point that big pharma, not the cash-strapped pure-play, may capture most of the value.
  • NVIDIA and Google Cloud remain the picks-and-shovels layer. Schrödinger's Bunsen early access is bundled with NVIDIA and Google Cloud compute and tooling, a reminder that the "sell shovels" trade (compute and cloud underneath everyone's AI biology) keeps compounding regardless of which drug company wins.
  • Priority review vouchers as a financing engine (from the Andrew Lo discussion) are worth watching as a real, non-hype catalyst for rare-disease developers, voucher prices north of $100M can make an otherwise-uneconomic program profitable on their own.

What changed vs. last week

  • Last week's cliffhanger got answered, and it split the two pure-plays. Last week we asked, with frontier AI labs moving into drug development and big pharma building AI in-house, "what is left for the public pure-plays?" and flagged the Aug 5 prints as the test: is Schrödinger's software still compounding above its 10–15% guide, and has Recursion converted its data-generation capacity into a real result? The answers: Schrödinger, yes (ACV +27%, a beat-and-raise, and a marquee at-scale Bunsen deal with BMS). Recursion, not yet (a revenue miss; a new partner milestone and an opex cut, but no clinical or commercial inflection). One pure-play earned its "AI-levered" multiple this quarter; the other is still asking for patience.
  • The Bunsen traction question got its first real data point. Last week Bunsen had just launched into early access with zero outside coverage. This week the substance arrived not as podcast buzz but as a commercial event: Bristol Myers Squibb deploying Bunsen across its research org. (Podcast silence on Bunsen persists, the mindshare gap is still open.)
  • The AlphaFold break-up became a full leadership exodus. Last week: a quarter of the AlphaFold team had left and the group was being folded into Gemini/Isomorphic. This week: Hassabis moved upstairs, and Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le left Google to found Discovery Loop, an "AI is the researcher" company with drug discovery on the roadmap. The trend is now unmistakable.
  • The "data, not the model" thesis hardened into consensus. It was the through-line last week (Regeneron, BioCentury's "negative data" point) and this week it was stated even more plainly by Braidwell, Chai, and a whole episode title from Data in Biotech. It is now the single most durable claim in the sector.
  • Lilly's mysterious July 30 drop resolved itself the boring way. Last week we flagged an unexplained ~4.5% one-day fall in LLY with no clear driver. It didn't matter: the Aug 5 beat-and-raise sent the stock back up, and it finished the week ~3% higher.

Open questions to carry into next week:

  1. Does Recursion's Roche/Genentech neuroscience milestone and its Phase 2 FAP data in November finally give it a catalyst, or does the "show-me" discount persist?
  2. Will Schrödinger's Bunsen deals with BMS (and its earlier NVIDIA/Google bundling) start showing up as accelerating ACV, and will the podcast/outside conversation finally engage with Bunsen?
  3. Does Discovery Loop, or another "AI-as-scientist" startup, name a first scientific or drug target, and how do incumbents respond?
  4. Anthropic's preclinical programs for "rare and neglected diseases" (flagged two weeks ago), still no named program or partner. Watch for the first one.
  5. Do priority-review-voucher-funded rare-disease portfolios (the Lo/Bakker model) actually get stood up with real capital?