AI
Google DeepMind's AlphaEvolve moves into production, optimizing TPUs and enterprise workflows
The Gemini-powered coding agent has graduated from testing to core infrastructure use, delivering efficiency gains across hardware design, finance, semiconductors, and drug discovery.
Google DeepMind Blog
Google DeepMind's AlphaEvolve, a Gemini-powered coding agent, has moved beyond pilot testing into active production, where it is now optimizing critical infrastructure and solving real-world problems across multiple industries.
Reshaping AI hardware infrastructure#
The system has become integral to Google's internal operations. AlphaEvolve assisted in designing the next generation of TPUs (Tensor Processing Units) by proposing a circuit design so efficient that it was integrated directly into the silicon, according to Jeff Dean, Chief Scientist at Google DeepMind. The system also identified improved cache replacement policies in just two days—work that previously took months of human effort.
On the database side, AlphaEvolve refined Google Spanner's compaction heuristics, reducing write amplification by 20% and cutting the storage footprint of compiled software by nearly 9%.
Commercial traction expands#
Working alongside Google Cloud, the technology is now rolling out to enterprises across sectors:
- Klarna (financial services) doubled the training speed of one of its largest transformer models while improving accuracy.
- Substrate (semiconductors) achieved multi-fold runtime improvements in computational lithography, enabling larger chip simulations.
- FM Logistic (logistics) cut routing distances by 10.4%—saving over 15,000 kilometers traveled annually.
- WPP (advertising) gained 10% accuracy improvements in AI model optimization for campaign data.
- Schrödinger (materials science) achieved a roughly 4x speedup in Machine Learned Force Fields, shortening drug discovery and catalyst design cycles from months to days.
According to Gabriel Marques, Technical Lead at Schrödinger, the ability to "explore larger chemical spaces faster" carries real R&D impact across molecular screening and materials development.
Self-optimizing systems as the frontier#
DeepMind frames AlphaEvolve as evidence that future AI breakthroughs will come from systems that can learn, evolve, and optimize themselves. The team plans to expand the platform's capabilities and apply it to a broader range of external challenges.
