Vivodyne is spending venture money on wet-lab throughput, not bigger models

Vivodyne opened a facility it calls a human data center near San Francisco last week, filled with robotic labs that grow human tissues, dose them and record what happens; TechCrunch reported the opening on August 19 ^[1]. The wager underneath the building is that the binding constraint on AI drug discovery is the rate at which causal human experiments can be produced, not the capacity of the models reading them ^[2].

The mechanics are more interesting than the label. The process starts with microfluidic chips the company calls TissueDisks, each of which Vivodyne says can cultivate hundreds of self-assembling human tissues ^[11]. HIVE robotic labs then grow, dose, monitor and analyse those tissues without manual handling ^[3]. Per Vivodyne's product description, each HIVE tests 10,000 tissues at once and returns data in one to two weeks, measuring visual change, gene expression and proteins at single-cell resolution across more than 20 tissue types ^[4]. The structural point is that the company owns both the intervention and the measurement, so every record ties a defined action to a tissue response ^[5].

That is the gap Georgescu says current biological training data has. He told TechCrunch that models learn "this is cell state A", "this is cell state B", but never that B followed from inflammation of A ^[12]. His blunter version: "Absent human testing, what are these [AI] models going to do? They're going to cure cancer in mice" ^[6]. The nearest thing to independent support is a 2026 Nature Methods study finding that adding pretraining data produced no clear scaling advantage for single-cell foundation models, with performance often plateauing as data volume rose ^[13].

The capital tells you how literally Vivodyne means it. The company announced $38 million in total seed financing on November 22, 2023 and a $40 million Series A on May 29, 2025, both led by Khosla Ventures ^[7]. The Series A paid for a planned 23,000-square-foot robotic facility in South San Francisco ^[14], which works out to roughly $1,700 per square foot if the round went to nothing else ^[24]. The research record puts total financing near $82 million across three rounds including an earlier $4 million raise, with no disclosed valuation ^[8]; a separate account describes just under $80 million across two Khosla-led rounds ^[9], so the accounting is not settled.

Vivodyne enters a market that already includes Emulate, CN Bio, MIMETAS and Hesperos ^[15]. Emulate says its AVA system handles as many as 96 Organ-Chip samples per run ^[16], about a hundredth of Vivodyne's stated per-HIVE figure ^[25], but tissue formats, measurements and workflows differ enough that headline capacity numbers do not compare ^[17]. Vivodyne's stated distinction is the bundle: tissue production, robotic testing, and a data layer built for model training ^[10].

What is not disclosed matters as much. Public filings do not establish whether Vivodyne sells experiments as a service, licenses datasets, sells instruments or mixes those, and customer economics are undisclosed ^[18]. Its claim to collaborate with a majority of the world's ten largest pharmaceutical companies dates from its 2023 financing announcement, with no partners named and no contract values published ^[19]. Georgescu's assertion that the team already runs twice as many studies as all US animal testing is a company statement ^[27], as are its reported concordance figures: 94% predictive accuracy on liver toxicity, 96% agreement for airway tissue, and 100% across 20 chemotherapy drugs in bone marrow ^[26].

Watch whether repeatable tissue experiments generate information buyers cannot get cheaply from public data or conventional preclinical work ^[20]. The backdrop is unforgiving: roughly 90% of drugs that look promising in animals and enter clinical trials never win approval ^[21], only one AI-discovered drug has reached Phase III ^[22], and Isomorphic Labs now expects first trials by year-end after originally targeting 2025 ^[23].