A World Model of Life: Biohub's Alex Rives on Virtual Cells, Measurement, & the $1.8B Data Push

Full analysis

Alex Rives, who built Meta's ESM protein models and now runs science at the Chan Zuckerberg Initiative's Biohub, says a $1.8 billion effort with the Department of Energy, NIH, Google DeepMind, and Meta can produce a working "model of the cell" in five years. Not a cure for anything. A piece of infrastructure: a map of what cells do, big enough to train an AI on, the way the open internet trained the chatbots you use.

What's actually being decided here is where the next big AI data push goes, and who funds it. This is a bet that the bottleneck in AI-for-biology is not smarter models. It's that nobody can measure what a cell is doing in enough detail to feed a model. That bet is easy to walk back if it's wrong, because it's mostly research spend, not a product with customers. Nothing here ships to a business buyer this year. The deadline is self-imposed: Rives's own five-year clock, which started the week this was recorded.

The Skeptic

Five years for a "model of the cell" is the kind of timeline that gets you funded and never graded. Rives is honest about the catch, which is to his credit. The whole thing turns on verification: you only improve an AI by checking its guesses against reality, and in biology reality means a wet-lab experiment that takes days or weeks. In coding, the check takes milliseconds. That gap is not a detail. It's the reason this won't move at AI speed no matter how much money lands on it. He also concedes today's "virtual cell" models don't actually simulate anything. So what we have right now is a data-collection project and a hope that the models arrive once the data does.

The Researcher

The useful distinction Rives draws is between two targets. A full physics-based simulation of a cell, atom by atom, is far off. What he's actually promising in five years is narrower: a probabilistic "transition function." Feed it a cell's current state plus some input, and it predicts a spread of likely next states. That's a weather forecast for a cell, not a digital twin. His strongest evidence is real: protein models trained across all of evolution beat models trained on human data alone, and AlphaFold showed that physics-informed structure helps. The weak point is that proteins are a solved-ish shape problem. Whole-cell behavior across tissue is a different, messier beast, and he knows it.

The Compute Pragmatist

Notice where the $1.8 billion goes. Not GPUs. Lasers and microscopes. Biohub's headline advance is a standing-wave laser that lets an electron microscope see individual proteins working inside a living cell. The constraint here is sensors, not silicon. That flips the usual AI story. In language models, data was nearly free and compute was the bill. In biology, the data itself has to be manufactured, one expensive measurement at a time, and that production line can't be parallelized the way a training run can. Whoever owns the measurement machinery owns the data, and the data is the whole game. That's a slow, physical moat, not a cloud-rental one.

The Open-Source Advocate

Watch who's at the table and who isn't. Google DeepMind, including its drug arm Isomorphic Labs, and Meta are named partners. The output is "cell atlases," maps of hundreds of billions of cells. The open question nobody answered: who gets the data, and under what license? ESM, Rives's old work, was released openly and seeded a whole community. AlphaFold's structure database did the same for biology at large. If this atlas lands under open terms, every startup and university lab gets the raw material to train on and the field moves fast. If DeepMind and Isomorphic get first or exclusive access, it's a private moat funded partly by public money from DOE and NIH. That distinction decides whether this is a public good or a head start for two labs.

The Builder

For anyone building with AI tools today, there is nothing to integrate and nothing to buy. This is a five-year infrastructure bet with no API. The one portable lesson is the verification loop. Rives wants an AI that proposes an experiment, runs it, learns from the result, and repeats. That's the same loop that made coding agents good, where the model writes code and the test result comes back instantly. His problem is the test takes weeks, not milliseconds. If you build AI agents for anything where checking the answer is slow or expensive, that's your constraint too, and no model release fixes it. Fast, cheap, automatic verification is the thing worth engineering.

Where they split

The real disagreement is about what the five-year promise even means. The Researcher says the narrow version, a forecasting model for cell states, is plausible. The Skeptic says even that stalls on the wet-lab verification bottleneck, so "five years" is a funding horizon, not an engineering one. They can both be right: the data map gets built, the model that's supposed to sit on top lags behind.

The second split is who benefits. The Compute Pragmatist sees a measurement moat that favors whoever owns the lasers and microscopes, meaning Biohub and its best-funded partners. The Open-Source Advocate sees a fork in the road: public funding could produce an open dataset that lifts the whole field, or a private asset that hands DeepMind and Isomorphic a lead. The announcement doesn't say which, and that silence matters more than anything else in it.

What it hinges on

Three facts settle this, and none are in the press release. First, will the cell atlas be released openly, the way ESM and AlphaFold's database were, or held for the named partners? Second, can anyone close the verification loop fast enough to matter, or does biology's experiment-takes-weeks problem cap the whole thing? Third, does the narrow forecasting model show up before the money runs out, or does five years buy a beautiful dataset and no working model on top. The council leans skeptical on timeline, genuinely split on openness, and agrees the measurement advance is the one concrete, real thing here.

Prediction: Biohub's Virtual Biology Initiative will release its first large-scale cell atlas dataset under an open or openly-accessible license (comparable to the terms of the AlphaFold Protein Structure Database), rather than restricting it to the named consortium partners, by its first major data drop on or before 2027-10-21.

Confidence: Medium. CZI and Rives have a consistent open-release track record.

Why: The signal is in the people and the pattern. Rives built ESM and released its weights and data openly, which seeded an entire protein-AI community, and the Chan Zuckerberg Initiative has published its existing cell atlas work (the Human Cell Atlas lineage) openly for years. The initiative is also funded partly by the Department of Energy and NIH, public bodies whose grant terms push toward public data release. The mechanism that would push the other way, DeepMind and Isomorphic Labs wanting exclusive access to a competitive asset, is real, but the atlas is raw measurement data rather than a finished model, and CZI's whole stated mission is building shared infrastructure, so a closed atlas would contradict both its history and its funders' terms. The less likely outcome is full restriction to partners, which would invite public criticism of a philanthropy hoarding publicly-cofunded data.

Revisit by 2027-10-21: We're right if Biohub publicly releases a substantial portion of the initiative's cell-atlas data under an open or free-research license by this date. We're wrong if the first data drop is restricted to Google DeepMind, Isomorphic Labs, Meta, and other named partners, or if no data is released at all.