AlphaEvolve One Year On: From TPU Circuits to Company Customers
Google DeepMind's report on its algorithm-discovery agent lists results in TPU design, Spanner, genomics and power grids, plus company users such as Klarna and FM Logistic.
TL;DR
- A year after launch, Google DeepMind says AlphaEvolve now helps design TPUs, tunes Spanner and serves companies such as Klarna through Google Cloud [^1].
- The results are company-reported, and they share one requirement: each problem needs a program that can score candidate solutions automatically [^1] [^2].
Background
Google DeepMind introduced AlphaEvolve in May 2025 as an evolutionary coding agent [^2]. Gemini models propose programs that solve a given problem, automated evaluators run and score each one, and a database of programs decides which ones seed the next prompts [^2]. Gemini Flash supplies breadth of ideas, Gemini Pro depth [^2]. Because everything hinges on that score, DeepMind says the approach fits problems whose solution can be described as an algorithm and verified automatically [^2].
What happened
On May 7, the AlphaEvolve team published a report on what the agent has done since its launch [^1]. The largest part concerns Google's own systems. DeepMind says AlphaEvolve has moved from pilot testing to a core part of its infrastructure and is used regularly to optimize the design of the next TPU generation [^1]. Jeff Dean, chief scientist of Google DeepMind and Google Research, says it "proposed a circuit design so counterintuitive yet efficient that it was integrated directly into the silicon of our next-generation TPUs" [^1]. In the Spanner database, refined compaction heuristics for its log-structured merge tree cut write amplification, the ratio of data written to storage versus the original request, by 20 percent [^1]. AlphaEvolve also found more efficient cache replacement policies in two days, work that had taken a human-intensive effort over months, and suggested compiler strategies that shrank the storage footprint of software by nearly 9 percent [^1].
The report lists research results too. An improved version of DeepConsensus, a Google Research model that corrects DNA sequencing errors, made 30 percent fewer variant detection errors; scientists at PacBio use it to analyze genetic data [^1]. On the AC Optimal Power Flow problem from electricity grids, AlphaEvolve raised the share of feasible solutions found by a trained graph neural network from 14 to over 88 percent [^1]. Quantum circuits it suggested had ten times lower error than conventionally optimized baselines, which made complex molecular simulations possible on Google's Willow quantum processor [^1]. In mathematics, working with Terence Tao, it helped solve Erdős problems, and it improved lower bounds for the Traveling Salesman Problem and Ramsey numbers [^1].
Together with Google Cloud, DeepMind is now bringing AlphaEvolve to companies, and the report names five [^1]. Klarna doubled the training speed of one of its largest transformer models while improving its quality. FM Logistic improved routing efficiency by 10.4 percent over previous, heavily optimized solutions, saving over 15,000 kilometers of driving a year. Substrate got a multi-fold speedup in its computational lithography framework, WPP reported 10 percent accuracy gains over manual model optimization, and Schrödinger roughly a fourfold speedup in training and inference of machine-learned force fields [^1].
Why it matters
The report shows AlphaEvolve's reach growing, not its method changing. A year ago the headline wins were inside Google: a Borg scheduling heuristic that recovers on average 0.7 percent of Google's worldwide compute, a 23 percent faster matrix multiplication kernel in Gemini's architecture, and a Verilog rewrite of an arithmetic circuit for an upcoming TPU [^2]. Now it is a regular tool in TPU design, and companies outside Google bring their own problems [^1]. What links every case is the requirement from the original design: someone must write an evaluator that scores each candidate program objectively, which is why DeepMind pointed to fields where progress "can be clearly and systematically measured" [^2]. Routing distance, write amplification and error rates fit that description; goals without an automatic check fall outside it.
The figures also need a caveat. They come from Google and its partners, and some arrive without an exact number, such as Substrate's "multi-fold increase in runtime speed" [^1]. The report is a showcase, not an evaluation. For a team thinking about using AlphaEvolve, the more useful question is whether its own problem can be scored by a program at all.
Practical example
Say Jonas plans delivery routes for a regional grocery wholesaler. The routes come from a heuristic the team has tuned by hand for years, and a simulator already measures how many kilometers it produces on last year's orders. That simulator is the piece AlphaEvolve needs: an automated evaluator that runs each candidate program and scores it [^2].
Jonas's team gets access through Google Cloud, which now brings AlphaEvolve to companies, and hands over the routing code and the simulator [^1]. Gemini Flash proposes many variations of the heuristic, Gemini Pro adds deeper suggestions, and the evaluator scores each one. The best programs stay in a database and feed the next round of prompts [^2]. Round by round, the code evolves. FM Logistic, working on the same kind of problem, reported a 10.4 percent gain over solutions that were already heavily optimized [^1].
The catch sits in the scoring. If the simulator counts only kilometers, a program that ignores delivery time windows can win. So before the first run, Jonas adds the time windows and truck capacities to the evaluator, and a dispatcher reviews the winning heuristic before it goes live. Because the result is ordinary code, the team can read and debug it, an advantage DeepMind highlighted for its own Borg heuristic [^2].
Related gear
We recommend this book because it explains evolutionary algorithms, the loop of selection and variation that AlphaEvolve wraps around Gemini to improve its programs step by step.