Newsletter · · Ashutosh Agarwal

A $400 Million Raise for Chai Discovery and a Reality Check on AlphaFold - AI Drug Discovery Weekly - Week of August 13, 2026

How AI drug-discovery podcasts covered Chai Discovery's $400 million raise, a sober AlphaFold reality check, federated learning, and the quiet public pure-plays, for the week of August 6 to 13, 2026.

AI Drug Discovery Weekly

Week of August 6–13, 2026: A $400 Million Raise for Chai Discovery and a Reality Check on AlphaFold

A protein-design startup that refuses to make its own drugs just raised another $400 million, while a leading researcher explained, in plain terms, why AlphaFold's billions of predictions still haven't produced a single blockbuster.

TL;DR

  • Chai Discovery, the marquee this week. The startup that builds AI models to design antibodies (but pointedly does not develop its own drugs) went deep on the Latent Space podcast (Aug 11) and disclosed a fresh $400 million raise. It now has four big pharma partners, Eli Lilly, Pfizer, Novartis and Argenx, and just ~30 employees. Its pitch: be the "neutral software factory for making medicines," the picks-and-shovels layer everyone rents rather than a rival drugmaker.
  • The uncomfortable reality-check. On AI and You (Aug 10), protein-folding researcher Mohammed AlQuraishi (creator of the open-source OpenFold) explained exactly why AlphaFold's hundreds of millions of predicted protein shapes have not translated into a wave of new drugs, and why the parts of biology that matter most for medicine are precisely the parts AI is still weakest at.
  • "Data, not the model" is now the settled view, and the industry has picked its answer. BioCentury This Week (Aug 11) laid out federated learning, a way for rival companies to pool the lessons from their private data to train a shared model without ever handing over the data itself, as the emerging fix for biology's data shortage.
  • The two public pure-plays had a quiet week. After their August 5 earnings, Recursion (RXRX) and Schrödinger (SDGR) generated no material news. Both stocks drifted up modestly (RXRX +4.1%, SDGR +3.7%). Still zero podcast coverage of either by name.
  • Eli Lilly's news was mostly not about AI, but one item was: its once-daily weight-loss pill orforglipron (Foundayo) won its first European approval, in the UK, and Lilly committed to a global AI-driven "agentic" commercial software rollout with Veeva. Stock +1.4% to a fresh all-time-high zone.

What's new

Chai Discovery: $400 million, four pharma partners, 30 people, and a refusal to make its own drugs

The single richest source this week was a long, unusually candid conversation with Matt McPartlon (co-founder) and Neil Patil (who leads platform and product) of Chai Discovery on the Latent Space podcast (Aug 11). Chai is roughly two-and-a-half years old. It builds AI models that design antibodies, the Y-shaped proteins your immune system uses to latch onto and neutralize threats, and one of the most important classes of modern medicine.

The headline number: "we've now raised another $400 million," Patil said, adding almost as an aside, "I have to go buy another compute cluster." The company now works with four large drug companies, Eli Lilly, Pfizer, Novartis and Argenx, and, remarkably, does all of this with about 30 employees, a research team of only ~10.

What makes Chai worth understanding is its business model, which is the opposite of the public pure-plays this newsletter usually tracks. Recursion and Schrödinger increasingly develop their own drug candidates. Chai deliberately does not. "We see ourselves as almost a neutral software factory for making medicines," Patil said. It licenses its models to the drugmakers and shares in their success rather than competing with them. McPartlon: "Chai definitely has no plans of starting a pipeline… I love the incentive alignment, we make the models better, the partners succeed more."

A few concrete things stood out:

  • The product looks like design software, not a chatbot. Chai deliberately did not wrap its models in a ChatGPT-style text box. Patil: "It looks a lot less like a ChatGPT and a lot more like Autodesk or SolidWorks or Figma… there's this almost Photoshop-esque design suite. You have this equivalent of a paint tool to kind of paint your epitope [the exact spot on a target you want to grab onto]. You have this equivalent of a content-aware fill tool to get your binders generated." The goal, in the company's words, is to turn drug discovery "from a scientific experiment to an engineering discipline."
  • The results are genuinely good, but this is still hard. For its flagship "Chai-2" work, the team set itself a challenge: design antibodies against 50 different targets. It got working binders to about half, at roughly a 20% hit rate. In one validation, a physical measurement of a designed protein's 3D shape came back within 0.33 ångström of the prediction, about one-third the width of a single atom, so accurate the team's first reaction was that the lab must have mistakenly sent back their own design. (An ångström is a unit of length used for atoms; for scale, a single atom is a few ångströms wide.)
  • A candid limit, from the people building it. Even Chai concedes the field's showpiece model is far from solved on the hard problems. McPartlon noted that AlphaFold2's version for predicting how two proteins stick together "got like 11% of antibody-antigen prediction cases correct. That means 90% of the time it's wrong." Antibodies are, by design, the one class of protein that cannot rely on evolutionary look-alikes, every person's antibodies are custom-built for the threats they've encountered, which is exactly why they are so hard to predict and so valuable to design well.
  • Compute is a genuine chokepoint, even for the hot startups. Patil, whose job includes buying chips, described a brutal squeeze: of the newest top-end GPUs shipping, "the hyperscalers and the biggest AI labs are buying 95-plus percent of it… and then you have the startups fighting over the scraps." He also argued the entire compute stack has become "LLM-pilled," optimized for chatbot-style models, not the geometry-heavy models biology needs.

One human moment cut through the jargon. Patil recalled showing results to a pharma partner's scientist: "One of the scientists in the room started tearing up and crying… She said, 'I've literally spent 10 years trying to get an initial binder to this thing, and you guys were able to help me do it.'"

To put the stakes in perspective, McPartlon offered a striking frame on why this is worth billions: the two leading GLP-1 weight-loss drugs together are roughly a trillion-dollar asset, and until about three months ago, he claimed, GLP-1 drugs' total revenue was larger than all the AI labs' revenue combined. In pharma, "I don't know if there's another domain where the downstream value of a token is as valuable."

An AI-drug reality check: why AlphaFold hasn't produced a wave of new medicines

The best counterweight to the excitement came from Mohammed AlQuraishi, a Columbia researcher who built OpenFold (an open-source answer to AlphaFold), on AI and You (Aug 10). It was the clearest plain-English explanation yet of a question that has quietly dogged the sector: if AlphaFold predicted the shapes of hundreds of millions of proteins and won a Nobel Prize, where are all the drugs?

AlQuraishi gave two reasons.

  1. Most of those predictions have no commercial value. The billions of protein structures now catalogued mostly belong to exotic organisms, "some exotic thermophile that lives in a volcano." The proteins that matter for medicine, human proteins, or those of pathogens we want to target, are a far smaller pool. "From the perspective of medicine, that huge pool is much smaller."
  2. The predictions that matter for medicine are exactly the ones AI is worst at. AlphaFold is excellent at "vanilla" protein shapes. But the therapeutically important cases are what he called "corner cases", a protein after a drug binds to it (which changes its shape), or a protein carrying a disease-causing mutation. "It's precisely under those perturbed conditions that disease arises and that therapies arise… and it just so happens that these side quests are actually the hardest things for AlphaFold to predict."

He also delivered the week's most quietly important technical point, one that reinforces the sector's dominant theme. AlphaFold does not really learn to fold a protein from scratch. It leans heavily on co-evolution: it looks at many related protein sequences from across evolution to build a very good guess about the final shape, then refines it. To prove the point, his lab shrank AlphaFold's training set from ~190,000 known structures all the way down to ~3,000, and still got about 90% of the performance. His conclusion: "If that step weren't present… AlphaFold would not have worked." In other words, the magic was never mostly in the size of the dataset or the model, it was in a clever use of the signal already hiding in evolution.

The industry has picked its answer to the data shortage: federated learning

Last week the sector converged on a single idea: with model architectures becoming commodities, the real edge is proprietary data. This week BioCentury This Week (Aug 11) devoted a segment to the structural fix the industry is coalescing around: federated learning.

The plain-English version: rival companies want the benefit of a model trained on everyone's data, but none will hand over its crown-jewel data. Federated learning squares that circle. Each company trains a small local model on its own private data; only the general lessons, "the general chemistry," as the analysts put it, get passed up to a shared "trunk" model, never the underlying data. Everyone's shared model improves; nobody's secrets leave the building.

BioCentury's analyst restated the now-consensus view bluntly: "As AI models for discovery are commoditized, the models themselves become less important, while the underlying data become more important for differentiating." The discussion named a federated-learning vendor, Aferis, and mapped where the approach works today (data-rich areas like small molecules and protein structures) versus where it can't yet reach, real-world drug behavior in the body, where even a well-funded biotech "may only have 20 endpoints… simply not enough." The honest bottom line: "The limit really is the amount of data being generated, not the ability to combine it." And who won't join a data consortium? Companies whose entire premise is building their own proprietary foundation model, they have every incentive to keep their data to themselves.

Eli Lilly: a UK pill approval and a quiet AI-commercial deal

Lilly's week was busy but mostly about weight-loss drugs, not AI. The two items worth flagging:

  • Orforglipron (branded Foundayo), Lilly's once-daily weight-loss and diabetes pill, won its first European approval, from the UK's MHRA on Aug 10, making the UK the first country in Europe to clear it. It's authorized for adults with obesity (or overweight with a related condition) and for type 2 diabetes, taken daily with no food or water restrictions, escalating from 0.8mg to 17.2mg (UK government announcement). A pill is a much easier product to scale and distribute than the weekly injections that dominate today.
  • The AI angle: Veeva Systems said (Aug 11) that Lilly committed to its Vault CRM globally, software Veeva positions as the foundation for "agentic commercial," meaning AI agents that help route the right medicines to the right patients. It's a small but concrete example of a theme that came up elsewhere this week (see below): AI's nearest-term payoff in pharma may be less in inventing drugs than in selling and delivering them.

Lilly also filed six new lawsuits (Aug 12) to shut down a black market in retatrutide, its next-generation obesity drug that is still investigational and not approved anywhere. Chief Medical Officer David Hyman: "What is being sold on the black market is not a medicine, it is entirely unverified, unapproved and not worth the risk." Analysts stayed overwhelmingly positive after last week's blowout quarter, with new price-target raises including Jefferies to $1,440 (the Street high), Goldman Sachs to $1,367, and Daiwa to $1,410, though a few skeptics held out, notably Rothschild & Co Redburn at $915 (Neutral). The average target sits around $1,345.

Other new material this week

  • A "simulated cell" progress report. On Citeline's Proof of Concept (Aug 6), Krishna Bulusu of Turbine (ex-AstraZeneca) described building "simulated cells", software versions of diseased cells you can run experiments on before touching a lab bench, to make the earliest, riskiest R&D decision (which target to chase) faster. His reframe: "Drug-discovery decision-making was always seen as a biology problem. But it's actually an information problem disguised as a biology problem." Crucially, Turbine tried the brute-force approach of "dumping all the data in" and found "it doesn't really work"; it deliberately uses explainable, mechanistic models. On the dream of one universal "virtual cell," he was sober: it's "a bit way off" and needs enormous data, "the need of the hour is no longer big data; it's diverse data."
  • An independent, slightly skeptical read on AI antibody design. On The Bioinformatics CRO Podcast (Aug 11), antibody-discovery specialist Adam Woolfe noted that a good binder is not the same as a good drug, antibodies often fail late because they clump together during manufacturing. On the AI-generated-antibody boom he named Chai (Chai-2) and Nabla as companies with "pretty impressive" white papers, but added a caveat worth remembering when reading any of these claims: "No one has actually seen the data, because the field is so competitive… they basically throw the kitchen sink at AI models with as much interaction data as they can."
  • A working investor's grounded take. On Fidelity's Market Sense (Aug 11), a Fidelity biotech portfolio manager (a Harvard neuroscience PhD) argued AI is an accelerant, not a crystal ball: models "are trained on mouse models, cell models… I don't think AI is going to help us predict a priori which drug is going to work." Where it does help, in her view: optimizing a drug once genetics has handed you a validated target, running smarter clinical trials, and, echoing the Lilly/Veeva item, mining medical records to find undiagnosed patients for a drug already on the market. Her north star for what actually de-risks a program is genetics: understanding a disease's genetic basis makes a clinical trial "three to seven times more likely" to succeed.
  • Discovery Loop, the "AI-as-scientist" startup, keeps drawing commentary. The Elon Musk Podcast (Aug 8) revisited the new venture spun out of Google's research ranks. Its stated roadmap runs from machine-learning research into "hardware design, drug discovery, and clean energy," and it's structured as a public-benefit corporation precisely to escape quarterly-earnings pressure: "You need a legal shield that allows you to fail for a decade while pursuing a massive breakthrough." Separately, Limitless (Aug 7) noted in passing that Demis Hassabis is stepping back from day-to-day DeepMind leadership to focus on Isomorphic Labs, its drug-design arm, a continuation of last month's reshuffle at Google's AI-for-science efforts.

The debate

Is AI's contribution to drug discovery real and compounding, or is the hard part still untouched? This week the two sides were stated more clearly than they have been in a while, and, tellingly, both were argued by insiders, not cheerleaders or cynics.

The bull case (Chai): The field has crossed a threshold from "science experiment" to "engineering." Design models now produce working antibody binders at real hit rates, validated to a fraction of an atom's width, and the models keep improving. McPartlon's summary: "The signs of life have been shown. The models work, they're delivering value." And the platform model spreads that value, one improving toolset lifts every partner's whole portfolio of drug programs.

The bear case (AlQuraishi, and the working investor): The impressive predictions are mostly on the easy cases; the therapeutically decisive ones, a protein reshaped by a drug or a mutation, remain largely out of reach. And even a perfectly designed molecule tells you nothing about how it behaves in a human body, which is where drugs actually fail and where the real cost lives. The Fidelity manager's line captures it: AI won't tell you "a priori which drug is going to work."

Notice these two positions barely contradict each other. Both sides agree AI has genuinely automated the front of the pipeline, finding and shaping candidate molecules. Both agree the back, proving a molecule is safe and effective in people, is untouched and still dominated by slow, expensive clinical trials. The disagreement is really about how much the front matters. The most durable specific claim in the sector, now repeated across many weeks and many shows, is the one underneath all of it: the bottleneck is data, not the model, and specifically the right, diverse, self-generated data that a competitor can't simply download. AlQuraishi's 3,000-structures experiment and BioCentury's federated-learning segment were, from opposite directions, both arguments about that same scarce resource.

Stocks in play

Prices as of Aug 13, 2026 (source: FactSet). Week-over-week compares to last week's close.

  • Recursion Pharmaceuticals (RXRX): $3.31, up about 4.1% on the week. Still near the low end of its 52-week range of $2.77–$7.18; market cap ~$1.7B. A quiet post-earnings week, no material company news and, once again, no podcast covered Recursion by name. This remains the purest "generate proprietary biological data at scale" bet among the public names, still waiting for a clinical or commercial event to break the "show-me" discount the stock trades at.
  • Schrödinger (SDGR): $18.46, up about 3.7% on the week, holding onto the big jump that followed its Aug 5 beat-and-raise. 52-week range $10.95–$23.02; market cap ~$1.4B. Also no material news this week. Notably, its "Bunsen" AI co-scientist product, and its new at-scale deployment deal with Bristol Myers Squibb, still drew zero podcast discussion, a mindshare gap that has now persisted for three straight weeks even as the stock has re-rated.
  • Eli Lilly (LLY): $1,209.85, up about 1.4% on the week and pressing against its 52-week high of $1,249.45; market cap ~$1,139B. Driven this week by the UK pill approval and continued post-earnings analyst enthusiasm rather than anything AI-specific.

One frame worth keeping in mind, courtesy of the Chai conversation: privately held AI-design companies are now commanding capital that dwarfs the public pure-plays. Chai's fresh $400 million raise, on ~30 employees, lands against Schrödinger's ~$1.4B market cap and Recursion's ~$1.7B. The most aggressive money in this theme is increasingly going to private, partnership-model "software factory" companies, not the two names public investors can actually buy.

Read-throughs

  • The "picks-and-shovels" thesis just got a sharper edge. Chai is the cleanest example yet of a company betting it can win by being the neutral toolmaker everyone rents, rather than a drugmaker. If that model works, it is bullish for the idea of AI drug discovery but bearish for the notion that any single AI-native company will capture the value by owning drugs. It also raises the competitive bar for Schrödinger, whose software business competes for the same "sell the tools, not the drugs" budget, while Schrödinger, unlike Chai, also runs its own drug pipeline.
  • Compute scarcity is a real, under-appreciated constraint on the whole theme. When a well-funded, buzzy startup says it is "fighting over the scraps" for GPUs behind the hyperscalers, that is a reminder that access to compute, and to biology-appropriate compute, not just chatbot chips, is itself a moat. It favors players who can strike compute-layer deals (big pharma, the largest labs) and squeezes everyone else.
  • AI's first real money in pharma may be commercial, not scientific. The Lilly–Veeva "agentic commercial" deal and the Fidelity manager's comments point the same way: the nearest-term, most bankable AI wins are in running trials, finding patients, and selling drugs, the parts with abundant data, not in the still-hard science of inventing them.
  • Federated learning is the structural story to watch. If rival drugmakers actually start pooling the lessons from their private data, it would blunt the data-moat advantage of the biggest players and reshape who benefits from AI in biology. It has been discussed for years without happening; this week's framing that it is "reaching an inflection point" is the item to verify over coming months.

What changed vs last week

  • The spine flipped from earnings to the podcast conversation. Last week was dominated by the Aug 5 prints (Schrödinger's beat-and-raise, Recursion's soft quarter, Lilly's blockbuster). This week the public pure-plays went silent, and the center of gravity moved to the private/expert discussion, above all Chai Discovery's marquee appearance and its new $400M raise (a fresh data point; last week we only knew Chai's four pharma partners, not this round).
  • The "data, not the model" theme hardened again, now with a named solution. For several weeks the sector agreed the moat is proprietary data. This week it named the mechanism it's converging on to share data without giving it away: federated learning (vendor Aferis cited). That's new and concrete.
  • The AlphaFold reality-check got its clearest articulation. Prior weeks featured skeptics saying "the clinic is the real bottleneck." AlQuraishi added the missing why on the discovery side: AlphaFold is weakest exactly where medicine needs it most (drug-bound and mutated protein shapes), and its power came from evolutionary signal, not dataset size.
  • Chai got real, deep coverage, a first. In prior weeks Chai appeared only as a passing VC portfolio mention or a partner name. This week it was the single most substantive episode.
  • Still no podcast coverage of RXRX, SDGR (or Bunsen), or the Lilly AI angle by name. The three tickers this newsletter tracks continue to be discussed almost entirely through themes and read-throughs rather than head-on, a persistent gap.
  • Discovery Loop and the Isomorphic/DeepMind reshuffle rolled forward as continuing threads rather than fresh news.