Vivodyne opens human-tissue data center to train drug-discovery AI

Vivodyne opened a human-tissue data center near San Francisco that uses HIVE robots to generate human biological data for training drug-discovery AI, with each HIVE capable of testing 10,000 tissues at once and returning data in one to two weeks. The company, led by CEO Andrei Georgescu and chief scientific officer Dan Dongeun Huh, has raised $38 million in seed financing and $40 million in Series A, both led by Khosla Ventures, and claims to collaborate with a majority of the world's 10 largest pharmaceutical companies.

Vivodyne opens human-tissue data center to train drug-discovery AI Andrei Georgescu and Dan Huh's robotic lab says each HIVE can test 10,000 tissues, generating human biological data before clinical trials. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: TechCrunch https://techcrunch.com/2026/08/19/ai-isnt-close-to-curing-cancer-this-startup-says-it-knows-what-it-will-take/ Why it matters AI drug discovery is constrained by the experiments behind its training data. Vivodyne is spending its venture capital on producing causal human evidence before clinical trials. Vivodyne https://www.vivodyne.com/?ref=runtimewire opened what it calls a human data center near San Francisco, using HIVE robots to generate human-tissue data before drugs reach clinical trials. TechCrunch reported https://techcrunch.com/2026/08/19/ai-isnt-close-to-curing-cancer-this-startup-says-it-knows-what-it-will-take/?ref=runtimewire on August 19 that Vivodyne opened the facility last week. It is the latest step in CEO Andrei Georgescu and chief scientific officer Dan Dongeun Huh's effort to turn automated tissue experiments into training material for drug-discovery models. Inside the HIVE workflow Vivodyne's process begins with microfluidic chips called TissueDisks. Vivodyne says each disk can cultivate hundreds of self-assembling human tissues. Its HIVE robotic labs then grow, dose, monitor and analyze those tissues without manual handling. According to Vivodyne's product description https://www.vivodyne.com/?ref=runtimewire , each HIVE can test 10,000 tissues at once and return data in one to two weeks. Vivodyne says the platform measures visual changes, gene expression and proteins at single-cell resolution across more than 20 tissue types. The facility is designed to run the same workflow repeatedly: establish a tissue's initial condition, apply a drug or another stimulus, and record the resulting biological changes. HIVE experiments give Vivodyne control over the intervention and the measurement process, allowing Vivodyne to produce datasets that connect a defined action with a tissue response. That capacity places Vivodyne in an established organ-on-chip market alongside companies including Emulate https://emulatebio.com/products/?ref=runtimewire , CN Bio https://cn-bio.com/?ref=runtimewire , MIMETAS https://www.mimetas.com/?ref=runtimewire and Hesperos https://hesperosinc.com/technology/?ref=runtimewire . Emulate says its AVA system https://emulatebio.com/ava/?ref=runtimewire can handle as many as 96 Organ-Chip samples per run. Different tissue formats, measurements and experimental workflows make headline capacity figures difficult to compare. Vivodyne's stated distinction is the combination of tissue production, robotic testing and a data layer intended for model training. Public disclosures do not establish whether Vivodyne primarily sells experiments as a service, licenses datasets, supplies instruments or uses a combination of those models. Customer economics are also undisclosed. Vivodyne says it collaborates with a majority of the world's 10 largest pharmaceutical companies, a claim that appeared in its 2023 financing announcement https://www.vivodyne.com/blog/vivodyne-announces-38-million-seed-financing-led-by-khosla-ventures?ref=runtimewire . Vivodyne has not named those partners or published contract values. Venture capital paid for laboratory capacity Vivodyne announced $38 million in total seed financing on November 22, 2023 https://www.vivodyne.com/blog/vivodyne-announces-38-million-seed-financing-led-by-khosla-ventures?ref=runtimewire and a $40 million Series A on May 29, 2025 https://www.vivodyne.com/blog/vivodyne-to-replace-animal-testing-with-40-million-funding?ref=runtimewire . Khosla Ventures https://www.khoslaventures.com/?ref=runtimewire led both rounds. The Series A also included Lingotto Investment Management https://www.lingotto.com/who-we-are?ref=runtimewire , Helena Capital, Fortius Ventures, Kairos Ventures https://www.kairosventures.com/?ref=runtimewire , CS Ventures, Bison Ventures https://www.bison.vc/?ref=runtimewire and MBX Capital. The research record puts Vivodyne's reported financing at approximately $82 million across three financings, including an earlier $4 million round. Vivodyne has not disclosed a valuation. The Series A financed a planned 23,000-square-foot robotic facility in South San Francisco. That spending makes Vivodyne a capital-intensive data producer at a time when many AI drug-discovery companies emphasize proprietary models and compute. Vivodyne's defensibility will depend on whether repeatable biological experiments create information that customers cannot obtain cheaply from public datasets or conventional preclinical testing. A founder's bet on experimental data Georgescu reached that thesis through laboratory work. He studied bioengineering at Cornell University, completed his doctorate in Huh's organ-on-a-chip laboratory at the University of Pennsylvania, and worked on automated microfluidics, DNA-synthesis chips and lab-grown tissues sent into space with NASA https://pennovation.upenn.edu/news/accelerating-drug-discovery-vivodynes-andrei-georgescu-phd-great-entrepreneurs-podcast?ref=runtimewire . According to TechCrunch https://techcrunch.com/2026/08/19/ai-isnt-close-to-curing-cancer-this-startup-says-it-knows-what-it-will-take/?ref=runtimewire , Vivodyne spun out of the University of Pennsylvania in 2021 after Georgescu completed his doctorate there. His summary of the problem is blunt. "Absent human testing, what are these AI models going to do? They're going to cure cancer in mice," Georgescu told TechCrunch https://techcrunch.com/2026/08/19/ai-isnt-close-to-curing-cancer-this-startup-says-it-knows-what-it-will-take/?ref=runtimewire . Vivodyne is building experimental machinery that can produce new training records, with each intervention capturing how functioning human tissue changes over time. A 2026 Nature Methods study https://www.nature.com/articles/s41592-026-03120-y?ref=runtimewire supports a narrower point: adding more pretraining data did not produce a clear scaling advantage for single-cell foundation models. Performance often plateaued as the quantity of pretraining data increased. Vivodyne's proposed answer is to change the type of data available to models. Static observations can show that two cell states exist. Controlled experiments can record the stimulus, starting condition and subsequent response. The new facility gives Vivodyne more capacity to generate those interventional records across tissues, drugs and doses. Vivodyne eventually wants models to select subsequent experiments, creating a feedback loop in which software proposes an intervention, HIVE robots execute it, and the tissue response informs the next proposal. That resembles reinforcement learning conducted through physical experiments, although Vivodyne has not published evidence that such a loop can reliably discover or optimize a clinically successful medicine. Human tissue still is not a human trial Vivodyne says its liver tissues achieved 94% predictive accuracy against human toxicity results, its airway tissues matched real human behavior 96% of the time, and its bone marrow tests produced complete concordance across 20 chemotherapy drugs, according to TechCrunch's account https://techcrunch.com/2026/08/19/ai-isnt-close-to-curing-cancer-this-startup-says-it-knows-what-it-will-take/?ref=runtimewire . Those company-reported results do not demonstrate that medicines selected or optimized through HIVE will succeed in human efficacy trials. Even sophisticated tissue models simplify a body. Reproducibility becomes harder when a system combines living tissues, robotics, imaging and several molecular measurements. Regulators are creating room for such methods without granting blanket approval. The FDA's March 2026 draft guidance on new approach methodologies https://www.fda.gov/regulatory-information/search-fda-guidance-documents/general-considerations-use-new-approach-methodologies-drug-development?ref=runtimewire describes a validation framework for organoids, organ-on-chip systems and computational models. The recommendations are nonbinding, and developers must show that each method is reliable for its proposed use. Vivodyne can create value before proving that HIVE-selected drugs succeed in patients. Human-tissue evidence could help drugmakers abandon weak candidates earlier, compare drug combinations before a clinical trial or detect toxicity missed by an animal model. The larger ambition remains harder to measure: establishing that automated tissues provide enough fidelity and variation to train useful models of human biology. The new facility gives Georgescu and Huh the capacity to run that experiment at industrial scale.