Anthropic Is the World's Fourth-Biggest Drugmaker

Last week I dug into Claude Science and came out with a simple conclusion, that in drug discovery, the scarce thing is the data, not the model. Anthropic can wrap Claude around any lab but will still hit the same wall in every drug program because no public record exists for whether a molecule works in a person. That was one company. Looking back on the week, I wonder if a new trend is forming, that these frontier model companies are already valued like drugmakers, but not one of them has made a drug.

Let’s look at Anthropic again. After its February Series G, it ranks fourth among the world's largest drugmakers, ahead of Merck, Novartis, AstraZeneca, and Pfizer. There are now reports of it going public at almost a trillion-dollar valuation.

Now watch what a company that size pays to get into the drug business. When a frontier lab (OpenAI, Anthropic, Google, and the like) buys in, it pays in a currency that only comes due if the science works.

Anthropic bought Coefficient Bio in April, a stealth team of fewer than ten, mostly ex-Genentech computational biologists, no disclosed pipeline, for about $400 million in stock. Against Anthropic's $380 billion valuation, that's a slice so small it barely registers, pocket change for a company that size.

Isomorphic Labs’ deals with drugmakers run past $4 billion in headline value on $82.5 million of upfront cash, near 50 to 1, and the milestones are written so the partner loses almost nothing if the molecules fail. Both ends of the market hand over enormous nominal sums that cost nothing unless the drug clears. To me, these aren't really prices; they're bets. The money doesn't cost anyone anything unless the drug actually works.

Nobody can price the wager because drug value does NOT sit where AI is strong. Clinical development is 50 to 70% of the cost of one approved medicine. Around 90% of drugs that enter human trials fail, and they fail on efficacy and toxicity, the two things a model cannot know until the molecule is tested inside a body. AI floods the cheap, crowded front of the pipeline. The cull happens downstream, where there is no model, only years, patients, and money.

To be fair, this is changing as we’re also seeing some AI drug companies push into the tougher stages of the pipeline, and with some success. Isomorphic raised $2.1 billion in May to carry its own molecules into the clinic, aiming at first trials by year-end. Insilico's rentosertib reached Phase IIa with positive results in Nature Medicine, and has now started Phase III. Validation itself is starting to admit AI, with Absentia's in-silico liver model entering the FDA's ISTAND program in June. If these companies can prove they can use AI beyond generating candidates, that's something big pharma should really worry about.

Fifteen to twenty AI-designed programs are expected to reach Phase III this year. If even one clearly works, the whole story changes overnight. AI isn’t just creating value by modeling endless new drug candidates; it’s proving that these candidates can make it all the way to the market. Until that happens, being “richer” than Merck proves nothing.

So, what I'm watching isn’t the market caps or the sky-high valuations. It’s where AI is creating value in the drug pipeline. Is it going back toward designing molecules, which is the easy part and where last week’s data problem still sits, or forward into proving they work in people, which is the part that actually makes a drug?

Xaira pulled back the curtain on X-Cell, a virtual-cell model built to simulate what happens inside a cell before anyone runs the experiment. The stealthy biotech, $1.3 billion raised, barely a word said publicly, trained the 4.9-billion-parameter model on 25.6 million perturbed single-cell transcriptomes. Virtual cells are the frontier everyone's chasing, because a model that predicts perturbation response could shrink the lab loop. Whether X-Cell actually does that, or is just the biggest one yet, isn't shown. [Link]

Recursion's co-founder filed to sell shares twice in one week, while short-sellers held roughly 35% of the float. Chris Gibson logged Form 144s on July 2 and July 7; short interest near 35% means a lot of money is betting the stock falls. Insider sales are often pre-scheduled and mean nothing, but stacked against that short position, they're the skeptic's data point in a week full of bullish AI-biotech deals. Planned selling or a confidence crack? The filing won't tell you. [Link]

Oxford researchers built a generative model that produces protein motion, not just a static shape. Spectral Diffusion represents a protein's dynamics in the frequency domain, then generates full conformational trajectories as temporally coherent sequences, where AlphaFold hands you one frozen snapshot. Proteins do their jobs by moving, so modeling the movement is the real prize. It's a preprint on arXiv, unreviewed and unbenchmarked against wet-lab dynamics, so treat it as a promising direction, not a solved problem. [Link]