The AI-Powered Biohub: Why Mark Zuckerberg & Priscilla Chan are Investing in Data, from Latent.Space
Summary
CZI is concentrating its next decade on Biohub because science—especially AI plus biology—proved the highest-impact part of its first 10 years. Chan says they tried education, community support, and science before concluding, “Oh my gosh, this is it.” Zuckerberg now calls Biohub the philanthropy’s main focus, while carefully framing its mission as helping scientists cure or prevent disease, not claiming the cures itself.
The institutional bottleneck is that transformative scientific tools can require 10 to 15 years and hundreds of millions of dollars, while conventional grants fund smaller, independent teams. CZI therefore operates institutes, co-locates biologists, engineers, and AI researchers, and connects Stanford, UCSF, and Berkeley. The wager is that new observation tools can unlock whole fields, as “the initial telescope” and microscope did before them.
Biohub’s central asset is a compounding data flywheel, not merely a collection of models. Its single-cell effort reached 125 million cells, with CZI responsible for roughly 25% and the wider ecosystem contributing 75%; a billion-cell project is now proceeding in months at a fraction of the earlier cost. AlphaFold’s reliance on 30 years of public data illustrates the strategic constraint: “These data sets aren’t going to get created by themselves.”
A useful virtual cell must connect molecules, proteins, cells, spatial structure, time, and eventually systems such as immunity. Biohub’s proposed edge is running “frontier biology and frontier AI in sync together,” designing instruments and experiments around the data a model needs. For investors, the implied stack runs from instrumentation and compute through datasets and models to wet-lab validation and clinical partners.
Wet labs remain the validation loop even if AI radically improves scientific throughput. Chan does not know whether validation capacity will become the bottleneck; biology cannot yet support the tens of thousands of cheap tests available to language models. The nearer-term payoff is models that generate and derisk hypotheses, letting grant-constrained scientists attempt more than “singles or doubles.”
The EvolutionaryScale team makes AI organizationally central to Biohub’s next phase. The ESM3 protein-model team joins Biohub’s existing researchers, with Alex leading the combined program—an AI researcher running the overall effort alongside leading biologists. Zuckerberg also commits to frontier models and large-scale biological compute, while Chan insists that models alone are not a satisfactory 10-year outcome.
The clinical destination is N-of-one medicine: predicting how an individual’s genetics and exposures alter cells, disease pathways, and treatment response. Chan’s sharpest example is depression, where patients may try a familiar antidepressant for months before learning whether it works: “Meanwhile, if it doesn’t work, it means the person’s suffering.” This is not a proposed “CZI app”; Biohub intends to build foundational tools that partners carry into clinical impact.
The mission’s clock may depend more on AI progress than biology, but AI cannot reason its way around missing empirical data. Zuckerberg says whether the goal takes 10, 20, or 40 years will probably track the pace of strong AI, conditional on continued frontier-biology investment. The practical bridge includes virtual immune-system work and engineered cells, but “cure and prevent all diseases” means making illness detectable early and manageable—not eliminating every infection. The separate question of immortality was left unresolved.
Deep dive
Not yet available upstream; scheduled sync will retry.