AI isn’t close to curing cancer. This startup says it knows what it will take.

Everyone from Sam Altman to Demis Hassabis has promised that AI will cure cancer, or maybe all disease. Dario Amodei himself wrote over the weekend that such claims have become more cliche than credible, even though he's made versions of the same pitch before. Vivodyne, a biotech startup, thinks it knows why the promises keep outrunning the results: the AI models don't have the right data to work with.

Most of what trains today's drug-discovery AI comes from animal testing or studies of isolated cells and proteins, not living human tissue. Vivodyne's CEO Andrei Georgescu puts it bluntly: without human testing data, these models will cure cancer in mice, not people. His company built HIVE, modular robotic labs that can grow 20 types of human tissue and then autonomously dose and monitor it, generating what Georgescu calls causal biological data — the kind that shows not just what a cell looks like, but how it got there.

That distinction matters more than it sounds. Georgescu points to a study published in Nature Methods last month that found no clear scaling laws when generative models train on existing cellular data. The problem, he says, is that the training data is all static snapshots — this is cell state A, this is cell state B — with no record of the process that turned A into B. Without that causal thread, a model can't tell you what will happen if you inflame a healthy cell, only what an already-inflamed cell tends to look like.

Vivodyne claims its tissue systems track real human biology closely: liver cells matching human toxicity trial results 94% of the time, airway tissue matching real tissue behavior 96% of the time, and bone marrow hitting 100% concordance across tests of 20 chemotherapy drugs. Spun out of the University of Pennsylvania in 2021, the company has raised just under $80 million across two rounds led by Khosla Ventures, and last week opened what it calls the world's largest human data center near San Francisco. Georgescu says his team is already running experiments at twice the throughput of every animal trial happening in the US.

The pitch to pharma is essentially a confidence problem. Automakers know their cars will pass crash tests before submitting them; drugmakers rarely have that certainty walking into a clinical trial, where the vast majority of drugs fail to win FDA approval and roughly 90% of drugs that work in animals never clear human trials. Georgescu's bigger bet is that HIVE's hundreds of thousands of running experiments — diseased tissue exposed to stimuli, over and over — will eventually train AI models that actually grasp causality in human biology, which he argues is the only way to handle combination therapies that hit multiple disease pathways at once, since testing every combination experimentally simply isn't feasible.