Vivodyne opened what it calls the world's largest human data center outside San Francisco last week, staking a claim that AI drug discovery is stuck on a data problem no frontier model can solve alone. The Khosla Ventures-backed startup has raised just under $80 million across two rounds to build HIVE, a modular robotic lab that grows 20 kinds of human tissue, doses them autonomously, and monitors the results. CEO Andrei Georgescu says the facility already runs at twice the throughput of all animal trials being held in the United States.
The argument cuts against a favorite talking point of frontier lab CEOs. Sam Altman has repeatedly cited curing cancer as a justification for OpenAI's push toward AGI. Demis Hassabis said last year that AI could cure all disease within a decade. Even Anthropic's Dario Amodei — who has floated similar claims in prior essays — wrote over the weekend that the pitch has become more cliche than credible.
Georgescu's framing is blunter. He argues that today's biological AI models are trained almost entirely on static snapshots of cells and proteins, which is why the industry keeps producing candidates that work in mice and fail in humans. Roughly 90% of drugs that clear animal testing never receive FDA approval, and each failed clinical trial burns tens of millions of dollars.
Key facts
- 01Vivodyne has raised just under $80M across two rounds led by Khosla Ventures since spinning out of the University of Pennsylvania in 2021.
- 02The company's HIVE robotic labs grow 20 kinds of human tissue and run at twice the throughput of all US animal trials combined.
- 03Vivodyne reports 94% predictive accuracy on liver toxicity, 96% match on airway tissue, and 100% concordance across 20 chemotherapy drugs on bone marrow.
- 0490% of drugs effective in animal testing fail to win FDA approval, with each clinical trial costing tens of millions of dollars.
- 05Isomorphic Labs, spun from AlphaFold, has pushed its first trials from 2025 to end of 2026 — no AI-designed drug has yet reached market.
Vivodyne, spun out of the University of Pennsylvania in 2021, says its lab-grown tissue closely tracks the behavior of real human organs. The company reports 94% predictive accuracy for liver cells versus human toxicity trials, a 96% match for airway tissue behavior, and 100% concordance for bone marrow across tests of 20 different chemotherapy drugs. Georgescu compares the goal to automotive crash testing: an automaker is confident its car will pass NHTSA requirements before it hits the wall, while drugmakers rarely walk into an FDA trial with anything close to that certainty.
The near-term business is contract work with pharma. Vivodyne won't name partners publicly but says it is working with multiple major drugmakers to pre-screen candidates before the clinical stage. The pitch is straightforward — if the tissue panel says a molecule will fail in humans, kill it before spending nine figures to prove it.
The longer bet is on training data. Georgescu points to a study published last month in Nature Methods that found no clear data-scaling laws when training generative AI models on existing cellular data. His read is that the field has been feeding models the wrong kind of information — before-and-after snapshots without the causal chain that produced the change.
HIVE is designed to close that gap by running hundreds of thousands of ongoing experiments where diseased tissue is exposed to a stimulus and the trajectory is captured over time. Georgescu expects that to enable the kind of reinforcement learning that has been missing from biology-focused foundation models. The comparison lab to watch is Isomorphic Labs, the DeepMind spinout built on AlphaFold, which wrote in February that real drug discovery will require highly accurate predictive models across an expansive range of biochemical properties. Isomorphic's first human trials, originally planned for 2025, have slipped to the end of this year.
The stakes rise sharply once the industry moves past single-target drugs. Combination therapies — the likely shape of next-generation cancer and immunology treatments — expand the search space beyond what wet-lab experimentation can cover.
There are reasons for skepticism. Vivodyne's accuracy numbers are self-reported and drawn from internal benchmarks rather than peer-reviewed head-to-heads against animal models. Organ-on-chip and lab-grown tissue platforms have been pitched as animal-testing replacements for more than a decade without dislodging the incumbent workflow, and regulators still require animal data for most submissions. Georgescu's causal-data thesis is also unproven at model-training scale — no one has yet shown that a foundation model trained on HIVE-style trajectory data outperforms one trained on the static datasets he critiques.
The bigger point Vivodyne is making lands regardless of whether its own tissue platform wins. The frontier labs have spent three years promising that scaling compute and parameter counts will bend the curve on medicine, and the concrete results — one AlphaFold-derived candidate in Phase III, Isomorphic's trials sliding by a year — have not matched the rhetoric. If Georgescu is right that the bottleneck is causal human-biology data rather than model architecture, then the AI drug-discovery race is going to be won in the wet lab, not the data center. That would reshape where the money flows in the next round of biotech-AI funding, and it would put Anthropic, OpenAI, and Google DeepMind in the awkward position of needing to buy or partner their way into physical infrastructure they have so far treated as someone else's problem.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
