Google DeepMind published the AlphaGenome Atlas on Monday. It is a one-petabyte table of predictions covering roughly nine billion single-letter changes to human DNA, which is every possible one. Each entry carries a single score, the AlphaGenome Variant Impact score, that combines two of DeepMind’s models into one estimate of how much damage a given change does. The dataset is more than 30 times the size of the AlphaFold Database.

The science press is treating it as a research milestone, and it is one. The business decision underneath it is the more interesting part: Google ran the model once, for everyone, and is putting the commercial version on its cloud.

A Table Instead Of A Model

AlphaGenome, released last year, predicts how a DNA change alters the molecular machinery around it, across both the 2 percent of the genome that codes for proteins and the 98 percent that regulates them. It is the field’s leading model and, in the words of one outside genomicist, “very slow and computationally intensive.” Running it on one variant at a time is a bottleneck for any lab that wants to scan a patient’s genome.

The Atlas removes the bottleneck by removing the model from the loop. DeepMind ran AlphaGenome across all nine billion variants in advance, plus more than 100 million short insertions and deletions, and stored the answers.

“When we started thinking about this project, it seemed impossible to do that computationally.”

That is Žiga Avsec, who leads DeepMind’s genomics work. The result is a lookup: a researcher with a variant of interest reads the row instead of running the model. The cost of asking the question has moved from the user’s compute budget to Google’s, and Google paid it once.

The AlphaFold Playbook, With A Cash Register

DeepMind did this before. The AlphaFold Database gave away predicted protein structures for free, and the model behind it became the standard for the field. The Atlas repeats the move with a difference in the terms. The website is free for non-commercial use , the API is available on GitHub for academic work, and the Atlas is wired into Google Antigravity as a skill for AI agents. Commercial use runs through Google Cloud’s Model Garden, with broader commercial availability “soon” and no price disclosed.

Those terms separate the science from the business. A frontier model is expensive to train and cheap to copy once its outputs are available at scale, as the US government said this week about six Chinese labs. A petabyte of precomputed answers is different. Its value sits in the hosting, the licensing, and the compute it took to build, and the commercial path for all three runs through Alphabet’s cloud. Google gives away the science and charges for the place where the science lives.

The Atlas is also an example of what AI does with a real record behind it. AlphaGenome was trained on decades of laboratory measurements of how genes are regulated, and the Atlas is that history extended to every variant nobody has measured yet. AI amplified decades of experiments into a table that no laboratory could have built by hand.

The authors are direct about the limits. The Atlas predicts molecular effects and is not validated for clinical use . A high impact score is evidence to be weighed, not a diagnosis. Carl de Boer of the University of British Columbia told IEEE Spectrum the score “has a clear use, but it also is probably going to be easily misinterpreted.”

Early users report results anyway. Teams at the Broad Institute and the University of Exeter say the Atlas surfaced a missed disease variant and found 22 percent more genetic associations in non-coding DNA in UK Biobank data. Those are research findings, and the clinical validation that turns them into products is years of work that Google is leaving to others.

The Atlas shows what a hyperscaler can do that a standalone lab cannot. It can spend the compute to precompute an entire field, give the result away to the people who set scientific standards, and collect from the companies that build on it through the cloud. Value in AI is moving from the model to the infrastructure that hosts what the model produced, and this is the cleanest example yet.

Alphabet is the beneficiary here, and it is not a company we have owned in preference to the other clouds. The same mechanism explains why we do own Microsoft and Oracle: the frontier labs’ science is trained and served on their capacity, and the lab pays the landlord whether or not the science pays off. The Atlas is Google running that model for itself, with its own science as the tenant.

For investors, the instruction is to own the compute layer where research becomes a product. Watch for Google Cloud’s pricing when commercial access opens, because the number it puts on a petabyte of answers will tell you what the hosting is worth.

Nine billion questions, answered once, stored on one company’s cloud. That is what the AI business looks like when it works.