Blog · Delta V

"The number that made me check the math"

"Two thousand likes in a day is normal for a Google AI announcement. What stopped me was the unit: "one petabyte". On September 8, Google DeepMind introduced AlphaGenome Atlas — precomputed predictions for the molecular effect of "every possible single-letter change in the human genome", roughly 9 billion variants, served through a zero-code web portal that any biologist can query from a browser."

"I did the arithmetic before accepting the headline. The human genome is about 3.1 billion letter positions; at each position, one DNA letter can mutate into 3 others. That is 3 × 3.1 billion ≈ "9.3 billion possible single-nucleotide variants" — \"9 billion\" is the honest rounding. DeepMind ran its AlphaGenome model against all of them and saved the outputs. One petabyte of stored inference — more than 30 times the size of the AlphaFold Database, which itself holds 200+ million protein-structure predictions."

"That last comparison is the real story. AlphaFold DB changed biology by turning a model into an artifact. AlphaGenome Atlas is the same move applied to the other 98% of the genome — and it is a lesson in the economics of inference that has nothing to do with biology."

Fig. 1 — Inference as an artifact: pay the compute once, freeze the answers, serve queries from dumb storage. Per-query metering dies; the publisher of the artifact owns access. Sources: Google DeepMind announcement (Sep 8, 2026); Forbes analysis (Sep 10, 2026).

"What the Atlas actually is"

"The 2025 AlphaGenome model predicts how a genetic variant affects molecular processes — how much of a protein gets made, whether a splice site breaks, whether a regulatory switch turns on. Useful, but you had to run the model yourself, variant by variant."

"The Atlas inverts that. DeepMind's researchers pre-ran the model across the entire space of single-letter changes and published the results as a queryable catalogue:"

Coverage" — all ~9 billion single-nucleotide variants, plus, per the accompanying paper, a catalogue of short insertions and deletions observed in human populations. The 2% of DNA that codes for proteins and the 98% that regulates it are both included."
The AVI score" — each variant gets an AlphaGenome Variant Impact score: one readable number for how disruptive a change could be, decomposed into interpretable contributions across chromatin accessibility, splicing, conservation, and other categories."
Mechanistic context" — a compendium of more than 2,500 recurrent DNA motifs, so a high score comes with a hypothesis about "why"."
Access" — a free web portal requiring zero coding (the AlphaFold-DB playbook, deliberately repeated), plus availability through the AlphaGenome API and as a skill in Google's Antigravity agentic platform."

"The headline validations came from UK Biobank whole genomes (University of Exeter) and the Broad Institute. Nature's coverage frames it as a successor to AlphaFold DB in ambition: a map of terrain scientists previously had to walk on foot."

"Why this matters beyond genomics: inference as an artifact"

"Here is the pattern worth stealing. A frontier model is expensive to run — every query pays the electricity, the GPUs, the queue. DeepMind did something economically clever: "they paid the compute once, up front, and froze the answers". One petabyte of static predictions now serves millions of queries from dumb storage. No researcher querying the Atlas needs access to the model, the weights, or a GPU. The intelligence was amortized into an artifact."

"Call it precomputed inference, and notice its properties:"

The marginal cost of a query is near zero." Serving a petabyte from object storage costs orders of magnitude less than running the model per request."
The artifact outlives the pricing decision." AlphaFold DB has been free for years; the predictions published today remain queryable regardless of what any API costs tomorrow."
The moat moves from the model to the dataset." Nobody else can regenerate this catalogue without the same compute budget — but once it exists, the "publisher" of the artifact controls access to it."

"That last property is where the sovereignty question enters."

"The honest math"

"Three flags before anyone treats this as a finished map of human biology."

1. Prediction is not measurement." Every entry in the Atlas is a model output. Press coverage flagged the accuracy gap: AlphaFold's protein structures were validated against decades of experimental crystallography, while variant-effect prediction has no equivalent ground truth at scale. For the variants that matter, laboratory validation is still mandatory. The Atlas prioritizes hypotheses; it does not confirm them."

2. Most of the 9 billion variants will matter very little." The distribution of variant impact is heavily skewed toward \"no detectable effect.\" A high AVI score is a lead, not a diagnosis. The hard problem in clinical genetics — interpreting variants of uncertain significance for an actual patient — is a deployment, regulation, and liability problem that no petabyte closes by itself."

3. The free portal is free like a demo is free." Access today is non-commercial only. Commercial use lands \"soon\" on Google Cloud — with "no published pricing". The pattern deserves clear eyes: the academic halo is funded by the commercial meter that comes later. If your biotech pipeline depends on the commercial tier, you are building on infrastructure whose invoice Google has not written yet. The predictions are open for science; the platform is the product."

"The sovereignty clause"

"Our house thesis is that whoever pays the electricity bill decides who runs models. The Atlas adds a second clause: "whoever precomputes owns the answers.

"DeepMind just demonstrated the sovereignty pattern at petabyte scale — compute once, own the artifact, serve forever — while simultaneously demonstrating the centralizing pattern: the artifact lives on Google's cloud, behind Google's portal, on terms Google will set for commercial use. Both things are true, and the tension between them "is" the lesson."

"The same economics work at client scale, pointed the other way. Any organization doing repeated expensive inference against a stable domain — document review, threat triage, signal classification — can batch the hard computation, store the results as an artifact it owns, and serve queries locally from cheap hardware. Your precomputed layer does not rate-limit, does not change terms of service, and does not report back. We design exactly this shape of pipeline in the Sovereign AI Agent Forge: turn frontier-model spend into an asset on your own meter, and reserve live inference for what genuinely needs to be live."

"What to watch"

The commercial price list" — when Google Cloud pricing lands, it becomes the reference point for what \"access to precomputed intelligence\" costs. That number will echo well beyond genomics."

Independent validation studies" — how do AVI predictions hold up in prospective lab studies? The gap between prediction and measurement is where the next papers live."

The precompute-and-publish pattern" — whether other labs (and other fields) follow with precomputed catalogues of their own, or keep the pattern and drop the publishing. The compute race is now also a cataloguing race."

Cooperation over defection, at petabyte scale: nine billion predictions published free for the people trying to understand disease is a genuinely generous act of science. The commercial meter arriving later is the defection test — watch what Google charges.

Sources:" Google AI — AlphaGenome Atlas announcement (X, 2026-09-08) · Google DeepMind — \"AlphaGenome Atlas: Molecular predictions for 9 Billion human DNA variants\" (2026-09-08) · Nature news — \"DeepMind's new genome 'atlas' charts effects of all 9 billion human gene mutations\" (2026-09-08) · The Verge (2026-09-08) · Scientific American (2026-09-08) · The Next Web (2026-09-08)"