Google DeepMind, Meta and the US government just put their weight behind one problem: the data an AI needs to simulate a human cell. The project is the Virtual Biology Initiative, led by the nonprofit Biohub, and on October 7, 2026 it grew to $1.8 billion in funding, data, computation and new measurement technology.

What the Virtual Biology Initiative is

Biohub launched the initiative in April 2026. The target is a "virtual cell": a predictive AI model of biology. The way I think about it is ImageNet, but for living cells. Image models got good once there was a huge, shared, labeled dataset. Biology models do not have that yet. The bottleneck is data, and the initiative needs far more of it than exists today.

Who is paying

Here is the split from Biohub's release:

  • Biohub: a founding $500 million commitment. About $400 million goes to new technologies, and $100 million funds research outside Biohub.
  • US Department of Energy: more than $500 million over five years for lab measurement, modeling and computation, through the Genesis Mission.
  • NIH: coordinates datasets, repositories and knowledge bases built on more than $500 million in prior federal investment. This is existing data, not new money.
  • Google DeepMind, Isomorphic Labs and Meta: $300 million together. The release does not give a per company split.
  • NVIDIA adds compute, software and technical expertise, and Renaissance Philanthropy helps expand funding for data generation. No dollar amounts are given for either.

One small inconsistency: the headline says "nearly $2 billion" and the body says $1.8 billion. I use the $1.8 billion figure because that is the one the release spells out.

What the money buys

Mostly measurement. Two technologies stand out:

  • Cryo electron tomography, which resolves near atomic detail inside the cell.
  • Microscopy that can image millions to billions of cells in living tissue.

There is also money for engineering tools that build and perturb biology at the molecular, cellular, tissue and whole organism level. In plain terms: tools to change a cell on purpose and record what happens, which is exactly the kind of data a predictive model needs.

The best part: it is open

The partners describe it as an open resource for the research community, with shared standards, common identifiers and a single point of access. Biohub will work with NIH to standardize datasets for AI training, and it invites labs worldwide to join.

If you want to see what open cell data already looks like, open CELLxGENE. When I checked it on October 10, it listed 172.3 million cells across 2,240 datasets and 1,157 cell types. It is open to browse today.

The goal

NIH's Nicole Kleinstreuer put the target clearly: "universal cell models with sufficient biological complexity to predict how any cell responds to an intervention." In practice that means testing a treatment on a model before anyone runs it in a lab.

My take

This is the right bet. Model architectures are not what is holding back AI in biology: data is. Pushmeet Kohli of Google DeepMind said it in the release: "We will not solve this challenge without open, experimental biological data at an unprecedented scale." Two caveats. No timeline for a working virtual cell is given, and the open access details (formats, licenses, release dates) are not published yet. I will watch for the first data release before I call it a win.

Sources: Biohub: Virtual Biology Initiative expansion, Biohub: Virtual Biology Initiative launch and CELLxGENE