AlphaEvolve Paid Back Its R&D Bill in Year One

While the AI discourse spent the last year arguing about whether agents can credibly book a flight, DeepMind quietly published a one-year retrospective on AlphaEvolve that should reset the conversation. The headline is not that an LLM-driven system discovered new algorithms. The headline is that the new algorithms made money.

The Strassen result is the loud one

In 1969, Volker Strassen showed you could multiply two 4×4 matrices using 49 scalar multiplications instead of the textbook 64. For fifty-six years nobody beat him. AlphaEvolve did, with 48. That is the kind of result that gets written into the next edition of every numerical-methods textbook.

But the textbook result is not the interesting part. The interesting part is the boring industrial deployment list DeepMind buried under it.

The boring industrial deployment list

DeepMind says AlphaEvolve recovered 0.7% of Google's worldwide compute — a number that sounds small until you remember the denominator. At Google scale, 0.7% is somewhere between a small data center and a medium one. That is the payback line. Everything after this is gravy.

• PacBio reduced DNA variant detection errors by 30% by letting AlphaEvolve rework the DeepConsensus model. Genomics throughput goes up, per-genome cost goes down.
• Klarna doubled training speed on one of its largest transformer models without sacrificing quality. Fintech ML pipelines run twice as fast on the same hardware.
• An AC Optimal Power Flow model went from finding feasible solutions 14% of the time to over 88%. The grid is one of those domains where a percentage point is measured in megawatts and people not freezing.
• FM Logistic shaved over 15,000 kilometers per year out of warehouse routing in a single Polish pilot — a 10.4% improvement over what was already a heavily optimized solver.

Why this matters if you run a 12-person business

I know. PacBio and Klarna and Google's worldwide compute footprint feel a thousand miles from a five-truck plumbing company in Spokane. They are. But the takeaway is structural, and the structure does apply.

The dominant AI narrative for 2026 is the autonomous agent — software that does multi-step work end-to-end. The reality is that almost nothing in that category is shipping at the level the marketing implies. Browser agents miss the dropdown. Calendar agents double-book. Customer service agents hallucinate refund policies.

Meanwhile, the actual frontier — the part where AI is unambiguously paying back its R&D — is narrow algorithmic discovery in domains where you can write down a fitness function and let a coding agent grind. That is a very different shape than "replace your accountant."

The owner-operator translation

For an SMB, this is permission to stop chasing the agent fantasy and start using AI in the place it's been quietly winning all year: as a specialist that grinds on a narrow problem you can measure. Pricing. Routing. Inventory turn. Ad creative variants. Schedule optimization on a route fleet.

You don't need AlphaEvolve. You need the AlphaEvolve mindset — pick the one thing where a 10% improvement compounds across every transaction, write down what "better" means, and let the model search. That is a far more defensible use of the next 12 months than waiting for an agent to credibly answer your phone.

"The agents will come. The optimizers are already here."

— the unspoken thesis of the AlphaEvolve year-one report

DeepMind built the most expensive engineering tool of 2025. In year one it paid for itself and started shipping into other people's businesses. That is the bar.