AlphaEvolve Now Designs Google's TPU Silicon and Cuts Customer Costs

Google DeepMind says its Gemini-powered coding agent AlphaEvolve has proposed a circuit design so effective that engineers integrated it directly into the silicon of the company's next-generation TPUs — an AI system now shaping the chips that train future AI. The announcement marks AlphaEvolve's graduation from pilot testing to what Google describes as a core component of its infrastructure.

"AlphaEvolve began optimizing the lowest levels of hardware powering our AI stacks. It proposed a circuit design so counterintuitive yet efficient that it was integrated directly into the silicon of our next-generation TPUs. This is the latest example of TPU brains helping design next-generation TPU bodies," said Jeff Dean, Chief Scientist at Google DeepMind and Google Research.

The stakes are straightforward. If an evolutionary coding agent can outperform human engineers on problems once considered too low-level or too complex for automation, the economics of hardware design, cloud infrastructure, and enterprise optimization change. Google is now selling that capability to outside customers through Google Cloud.

What has AlphaEvolve changed inside Google?

The list of internal wins extends well beyond chip design:

• TPU design: AlphaEvolve now serves as a regular tool for optimizing the design of Google's next-generation TPUs.
• Cache replacement policies: it discovered more efficient policies in two days — work that previously required a concerted, human-intensive effort spanning months.
• Spanner: refinements to the database's Log-Structured Merge-tree compaction heuristics cut write amplification, the ratio of data written to storage versus the original request, by 20%.
• Compiler strategies: new optimization insights reduced the storage footprint of software by nearly 9%.

Each result landed in systems that serve billions of requests. A 20% reduction in write amplification across Spanner and a near-9% storage footprint cut translate into measurable infrastructure savings at Google's scale.

How are commercial customers using it?

Together with Google Cloud, Google is now offering AlphaEvolve to enterprises across industries. Five deployments came with numbers attached:

• Klarna (financial services): optimized one of its largest transformer models, doubling training speed while improving model quality.
• Substrate (semiconductor manufacturing): applied AlphaEvolve to its computational lithography framework and achieved a multi-fold increase in runtime speed, enabling significantly larger simulations of advanced semiconductors.
• FM Logistic (logistics): optimized routing problems including the Traveling Salesman Problem, finding a 10.4% improvement in routing efficiency over previous heavily optimized solutions — saving over 15,000 kilometers of distance travelled annually.
• WPP (advertising and marketing): refined AI model components navigating complex, high-dimensional campaign data, achieving 10% accuracy gains over competitive manual model optimizations.
• Schrödinger (computational material and life sciences): achieved roughly a 4x speedup in both Machine Learned Force Field (MLFF) training and inference.

The breadth matters as much as the numbers. The same system improved a payments company's model training, a logistics firm's routing, and a materials-science platform's molecular simulations. Google positions this as evidence that AlphaEvolve is becoming a general-purpose optimizer rather than a domain-specific tool.

Why does the Schrödinger speedup matter?

The 4x MLFF speedup carries direct commercial weight in drug discovery. Gabriel Marques, Technical Lead of Machine Learning at Schrödinger, said: "AlphaEvolve allows us to explore larger chemical spaces faster and more efficiently than ever before. Faster MLFF inference carries real business impact, shortening R&D cycles in drug discovery, catalyst design, and materials development, and enabling companies to screen molecular candidates in days rather than months."

Screening candidates in days rather than months compresses one of the longest and most expensive phases of pharmaceutical and materials R&D. For an industry where computational screening gates experimental work, a 4x inference speedup changes project timelines, not just benchmarks.

Who built it, and who backed it?

Google credits AlphaEvolve's core development to a team led by Matej Balog, Alexander Novikov, Ngân Vũ, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, and Pushmeet Kohli. The research grew out of a broader Google initiative focused on using AI for algorithm discovery.

A second wave of engineers — including Aja Huang, formerly of AlphaGo fame — joined to scale the system's impact. Separate teams built the AlphaEvolve UI and the API used to engage Google Cloud customers. The acknowledgment list spans leaders including Demis Hassabis, Jeff Dean, Pushmeet Kohli, and Sundar Pichai, and collaborators across Google DeepMind, Google Cloud, Google Labs, and Google Research — a signal of how broadly the company has mobilized around the system.

What comes next for self-optimizing algorithms?

Google frames the past year as proof that AlphaEvolve is rapidly becoming a versatile, general-purpose system. The company's stated thesis: the next breakthroughs will be driven by algorithms that can learn, evolve, and optimize themselves.

The roadmap is expansion. Google says it plans to extend these capabilities and bring the technology to an even broader set of external challenges. With Google Cloud already brokering deployments at Klarna, Substrate, FM Logistic, WPP, and Schrödinger, the next test is whether an agent that redesigns TPU circuits and cuts Spanner's write amplification by 20% can hold up on problems outside Google's own infrastructure.