Vivodyne says the AI drug bottleneck is human tissue data, not model capability

Vivodyne, spun out of the University of Pennsylvania in 2021 after CEO Andrei Georgescu finished a bioengineering PhD there, opened last week what it calls the world's largest "human data center" just outside San Francisco ^[1]. The claim underneath the launch is narrower and more useful than the usual pitch: that AI drug discovery is constrained by the absence of causal data from living human tissue rather than by model capability ^[2].

Parts of that diagnosis do not depend on Vivodyne. Of drugs that work in animal testing and go on to clinical trials, 90% never win regulatory approval in humans ^[3], which leaves roughly one in ten ^[4]. AlphaFold won a Nobel and has still not produced a new drug ^[5]. Isomorphic Labs, founded to build on it, has moved its first trials from 2025 to the end of this year ^[6], and wrote in February that true drug discovery will require "highly accurate predictive models, across an expansive range of biochemical properties and interactions" ^[7]. A Nature Methods paper published last month found no clear data scaling laws when generative AI models are trained on existing cellular data ^[8]. That last finding is the empirical crux of Vivodyne's argument: if more of the same data does not buy more capability, the answer is different data rather than larger runs.

Georgescu's mechanistic version is that today's training data consists of static snapshots, so a model learns "this is cell state A" and "this is cell state B" but never that state B is the effect of inflaming state A ^[9]. Most available data comes from animal testing or studies of single cells and proteins, not living tissue ^[10]. His HIVE machines, modular robotic labs, grow 20 kinds of human tissue and then autonomously dose and monitor it ^[11], tracking hundreds of thousands of ongoing experiments in which diseased tissue is exposed to a stimulus ^[12], which Georgescu expects to yield reinforcement learning signal for models of human biology ^[13]. "Absent human testing, what are these [AI] models going to do?" he asks. "They're going to cure cancer in mice." ^[14]

The performance numbers are company-reported. Vivodyne says its liver cells reach 94% predictive accuracy against human toxicity trials, its airway tissue matches real human tissue 96% of the time, and its bone marrow achieved 100% concordance across 20 chemotherapy drugs ^[15]. Twenty is a small denominator: a single discordant result would take that figure to 95% ^[16]. Georgescu also says his team is already running twice the throughput of all US animal trials ^[17]. The company says it works with multiple major pharma companies but will not name them ^[18]. It has raised just under $80 million across two rounds led by Khosla Ventures ^[19]. Georgescu's framing is a crash test comparison: an automaker generally knows its car will pass NHTSA requirements before testing, while drugmakers enter trials that cost tens of millions without that confidence ^[20].

The rhetorical timing helps. Anthropic CEO Dario Amodei wrote over the weekend that claims AI will cure cancer have become more cliche than credible, and that "the thing that will work is actually curing cancer" ^[21], after floating the idea himself in earlier essays; Sam Altman has repeatedly cited curing cancer to justify OpenAI's AGI and compute buildout, and Demis Hassabis said last year AI could potentially cure all disease within a decade ^[22]. Meanwhile a handful of AI-designed drugs have entered human trials, one reaching Phase III ^[23].

Watch whether the concordance figures appear anywhere independently verifiable, whether any pharma partner is named, and whether Isomorphic starts trials by December ^[6][15][18].