Stanford ran a virtual biotech of 37,000 agents
A Stanford team ran 37,075 AI agents as a "virtual biotech" for a week, and the therapeutic strategy it proposed for lung cancer is one Merck and Daiichi Sankyo independently took to an FDA breakthrough designation last year — but no molecule came out of it, and nothing was tested. The paper, published in Science on 17 September by Harrison Zhang, Peter Eckmann, Jiacheng Miao, Andrew Mahon and senior author James Zou, points an organisation of more than 37,000 agents at 55,984 annotated clinical trials, each agent assigned to work one later-stage trial under a chief-scientist agent, finishing in under a week. It derives a rule from the pile: drugs aimed at cell-type-specific, switch-like genes were 48 percent more likely to reach market with 32 percent fewer adverse events. It then asks the system to design a therapy with only information available before January 2025. The answer was a B7-H3-targeting antibody-drug conjugate — protein CD276 — and the proof offered is that ifinatamab deruxtecan, the B7-H3 ADC that Merck and Daiichi Sankyo won breakthrough therapy designation for in August 2025, is the same target and the same modality.
Read that claim precisely, because most of the coverage did not. The paper's own abstract says the system "propose[s] a therapeutic strategy in lung cancer." There is no candidate molecule, no potency figure, no in vitro or in vivo experiment run by the authors, and no clinical data. Nature's news write-up — the source the Chinese aggregators translated — quotes outside scientists saying the system "has not been vetted in the crucible of real-world drug discovery, and its predictions were not validated through experiments, let alone clinical trials." The third result in the paper, an analysis of a terminated ulcerative colitis trial and why it failed, is the one most write-ups dropped, and it is arguably the more useful demonstration: retrospective mechanism work is where an agent that has read 55,984 trials should be strongest.
What is genuinely new is the organisational claim, not the biology. Trials are messy, multi-document, and full of design choices that are never written down as rules; reading tens of thousands of them end-to-end and producing one testable hypothesis is a real capability, and the target it landed on is not a random one — B7-H3 is a live clinical bet. But a target match is not a discovery, and the field has been here before. We looked at the failure mode in More molecules, same failures: the crack in AI drug design — generating more candidate structures has not moved clinical success rates, because the bottleneck is validation, not ideation. This paper sits on the ideation side of that line. The agents cost almost nothing to run; the trial that would settle whether their rule is real costs hundreds of millions.
Nscale's IPO filing says its largest customer was 73 percent of its 2025 revenue, and the Financial Times reports that customer is ByteDance — which used the company's Norway data centre to reach Nvidia chips it could not buy at home. The numbers in the filing are hard: revenue of $33 million in 2025, up from $19.1 million, with a single unnamed customer at 73 percent, down to 52 percent in the first half of 2026 as revenue jumped to $140.6 million. The FT, citing people familiar with the matter, identifies the customer as ByteDance and describes a mechanism that is legal but awkward — US export controls restrict chip sales into China, not remote access to chips sitting in a non-EU data centre in Norway. ByteDance is not named anywhere in the filing, so the attribution belongs to the FT, not to the document.
The concentration is the story and it is already unwinding on paper. Nscale told the SEC that the 73 percent customer will be under 20 percent of 2026 revenue, and the same filing discloses $43.8 billion of Microsoft payments through 2033 and $44.6 billion from Anthropic — about 85 percent of a $103.4 billion contract book against $33 million of actual 2025 revenue. We covered that gap in September — Microsoft and Anthropic hold 85% of Nscale's $103.4 billion in AI contracts — and the ByteDance disclosure sharpens it rather than changing it: this is a company that has already swapped a Chinese customer it cannot advertise for two American ones it cannot yet bill.
The filing names its own regulatory exposure, which is the part worth watching. Nscale flags a Commerce Department rule that could force cloud providers to verify the identity and beneficial ownership of foreign customers, a Remote Access Security Act that would let the government restrict cloud access to export-controlled chips, and the BIS Affiliates Rule becoming operative again on 10 November 2026. It also notes that Norway is not an EU member, so European measures may not apply there on the same terms. The company went public on the strength of American contracts; the Norwegian chapter of its revenue history is exactly the kind of thing a rule change can close.
What to watch: whether Nscale's IPO pricing forces it to disclose the 73 percent customer by name, and whether any US agency acts on cloud access to export-controlled chips before 10 November.
If an AI system proposes the same drug target a pharma company already bet on, does that count as discovery — or as very fast literature review? Tell us in the comments.
Sources: Science — The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development · Nature · Stanford Medicine · Merck · SEC — Nscale S-1 · Financial Times · TechCrunch · DataCenterDynamics
