Biohub, Google, Meta and the US government put up US$1.8B to build the first AI 'virtual cell'

By Zendoric · October 9, 2026.

Biohub, the US Department of Energy (DOE), the National Institutes of Health (NIH), Google DeepMind, Isomorphic Labs and Meta announced on October 7 a coordinated US$1.8 billion investment to build the largest open repository of biological data prepared for training artificial intelligence models, according to the Biohub statement picked up by Reuters, MIT Technology Review, Bioengineer and News-Medical. The stated goal is to build a predictive model of the cell — a "virtual cell" — that would allow hypotheses to be tested on a screen before spending a single reagent in the lab.

The breakdown, detailed by Biohub and summarized by Unite.ai, mixes new commitments with money that already existed. Biohub is putting in US$500 million from the founding commitment it announced in April 2026 — part of a larger US$3 billion pledge from the Chan Zuckerberg Initiative: US$400 million for new measurement technologies (cryo-electron tomography and microscopy capable of recording from millions to trillions of cells in living tissue) and US$100 million for external research. The DOE is committing more than US$500 million over five years through its Genesis Mission, with exascale supercomputing, cryo-electron microscopy and autonomous labs across its national laboratory network; its Under Secretary for Science, Darío Gil, called it in the statement "a new standard for open science." The NIH contributes the coordination of more than US$500 million in prior federal investment — already spent on datasets and knowledge bases — which Biohub will standardize. And Google DeepMind, Isomorphic Labs and Meta are putting up US$300 million as founding partners of the so-called Virtual Biology Initiative. Put another way: of the US$1.8 billion in the headline, only a portion — the DOE's half-billion and the US$300 million from the companies — looks like genuinely incremental money; the rest repackages commitments and budgets that already existed. That does not make the announcement any less relevant, but it should not be read as a fresh US$1.8 billion check.

There is also a notable asymmetry in how access is shared out: data funded by the three private companies will be under a one-year embargo before being made public, while data generated with federal money is published without restriction, as Hoodline specifies. Google, Isomorphic Labs and Meta are, in practice, buying twelve months of exclusive advantage over a repository billed as "open science."

The scale the project is chasing is hard to grasp. The largest current single-cell datasets reach hundreds of millions of cells, close to a billion at best; Alex Rives, Biohub's head of science, told Reuters — as cited by Hoodline — that a genuinely useful predictive model will need "billions and eventually trillions" of cells. For reference, human tissues contain around 37 trillion cells, according to figures cited by Hoodline and Quartz. Biohub's plan is to compress into five years work that would normally take decades, according to Hoodline's coverage, with a first dataset ready in roughly a year, combining spatial transcriptomics — which maps molecular activity inside intact tissue — and cell screens that record how a cell responds to changes in its environment. Max Jaderberg, president of Isomorphic Labs, summed up the logic in the statement: "generating the data to solve predictive systems biology requires scaling beyond what any single organization can produce today." Pushmeet Kohli, vice president of AI for Science at Google DeepMind, made the same point: without "open, standardized data at an unprecedented scale" there is no way to solve the problem.

Biohub is not starting from scratch. It is the nonprofit lab founded in 2016 by Priscilla Chan and Mark Zuckerberg through the Chan Zuckerberg Initiative, and it has built up a decade of open projects such as Tabula Sapiens (a map of human cell types), OpenCell (an atlas of protein localization and interaction), Zebrahub (a zebrafish development atlas) and the community infrastructure CELLxGENE and CryoET Data Portal. Since 2023 it has also run a cluster of more than 1,000 Nvidia H100 GPUs dedicated to modeling cells, according to Hoodline. Joining as academic partners are the Allen Institute, the Broad Institute, Gladstone Institutes, the Wellcome Sanger Institute, the Human Cell Atlas and the Human Protein Atlas — institutions with prior experience coordinating consortia on the scale of the Human Genome Project — plus Nvidia as compute provider and Renaissance Philanthropy leading the search for additional funding.

The move is not an isolated one. Anthropic acquired the startup Coefficient Bio and set up a robotic wet lab in the Bay Area to test whether Claude can directly instruct automated equipment, as explainx.ai reports. The OpenAI Foundation launched a grant program of more than US$125 million for biological and medical datasets, according to MIT Technology Review (September 2026) — its first major award, US$40 million, went to the UNC Lineberger Comprehensive Cancer Center for cancer vaccine data, according to Hoodline, citing Venture Atlas. What Biohub claims sets it apart from those moves is not just the amount: it is the degree of coordination between the federal government, big tech and academia under a single program, with the promise that the data — barring the one-year corporate embargo — stays open.

This is probably the most concrete evidence yet of where our underlying argument about AI applied to biology is heading: if a model existed that could reliably simulate how a cell responds to a drug or a mutation, the cost and time of discovering a treatment would collapse, and the promise of eradicating diseases would stop being a slide in a presentation and become an engineering roadmap. But the announcement should not be confused with the result: what exists today is a funding coalition and a timetable — first dataset in a year, the bulk of the work compressed into five — not a working virtual cell. Recent AI history is full of five-year plans that end up taking ten, and compressing into five years work that, per the source, would normally take decades is a bet, not a done deal.

What is a fact is the repositioning of the bottleneck: it is no longer the algorithm or compute power — both relatively settled in generative AI — but clean, standardized biological data at a scale no single organization can generate on its own. That is the underlying reason government, philanthropy and three of the world's most powerful AI companies are converging on a single program: neither Google, nor Meta, nor Isomorphic Labs — Alphabet's drug discovery subsidiary — can buy access to 37 trillion cells on its own. And it is also why, despite the "open science" wrapping, the companies are reserving a year of exclusivity over what they fund: whoever trains on that data first gains months of advantage over anyone waiting for it to be released.

The pattern connects with something we were already seeing at Anthropic and OpenAI: competition between AI labs is shifting from model size to access to primary data and to the physical infrastructure — wet labs, microscopes, exascale supercomputing — that generates it. In the short term, none of this cures anything yet: it is infrastructure, money and institutional coordination, and it deserves the same skepticism we apply to any AI five-year plan. In the long term, it is exactly the kind of bet that, if it delivers even a fraction of what it promises, brings closer a medicine that simulates before it experiments — and with it, a handful of diseases we would finally stop fighting blind.