Three billion reference bases, each with three alternative letters, is how you arrive at 9 billion [9][1], and at an average of about 27,000 predictions per variant across hundreds of human and mouse cell and tissue types [10], the catalogue carries roughly 243 trillion individual predictions [2]. DeepMind paid that compute bill once and has not said what it came to. The marginal cost of the next lookup, to the researcher doing the looking, is a page load [6]. The comparison DeepMind offers is the old workflow of running a model one variant at a time or testing in a laboratory [3]; the comparison a large buyer will actually run is against its own hardware bill for scoring variants in house, and that is the number the unpublished licence terms will be measured against [6].
Exhaustiveness stops at substitutions. The same release scores more than 100 million insertions and deletions, but those are variants observed in population databases including the UK Biobank and the NIH's All of Us [11], which is about 1.1% of the substitution count [4] and a different sort of object: what has been seen in people, rather than what is chemically possible. Chase an indel and you are back to asking whether anyone has recorded it.
The internal licence matters most here. If Alphabet's drug-discovery arm has to take a commercial licence to use a dataset produced by its sibling [7], Atlas is being run as a priced product line rather than a shared corporate asset, and the free academic tier is what establishes it as the reference everyone benchmarks against before the paid tier has a number attached. That is the thesis, and the counter-thesis sits inside the same missing figure: a nominal licence written for accounting hygiene and a seven-figure one imply opposite answers about where value accrues, and nothing announced distinguishes them [7]. How it plays out depends on the pricing. If the licence is set to displace in-house variant pipelines, Google books the difference. If it functions instead as a funnel, the Cloud compute around the data ends up earning more than the data itself. And if it is priced above what pharma already pays to run its own models, Atlas settles into being an academic public good with a toll booth nobody walks through.
Other than published terms, the thing that would settle it is validation. The account of how Atlas was built is a bioRxiv preprint [8], and the claim being made for the exercise is Kohli's, that the Human Genome Project bought the book in 2003 without learning to read it [5]. A reading aid gets judged on whether the reading changed, which shows up as confirmed variant-to-disease links, not as a prediction count.
