AlphaEvolve Moves from Pilot to Core Google Infrastructure Across Five Industries

AlphaEvolve transitions from internal pilot to core Google production infrastructure

The system is now embedded in TPU hardware design, Google Spanner optimization, and compiler toolchains — no longer experimental but an operational engineering asset.

AlphaEvolve proposed a counterintuitive TPU circuit design integrated directly into next-gen silicon

The design passed engineering scrutiny and made it into production silicon. Jeff Dean: 'This is the latest example of TPU brains helping design next-generation TPU bodies.'

Google Spanner write amplification reduced by 20% via LSM-tree compaction heuristic refinements

Write amplification is a core efficiency metric in database storage engines. A 20% reduction at Spanner's scale translates to significant compute and storage savings across Google's global infrastructure.

Compiler optimizations cut software storage footprint by nearly 9%

AlphaEvolve identified optimization strategies that human compiler engineers had not surfaced, demonstrating value in low-level systems work beyond high-level ML tasks.

Cache replacement policy optimization completed in 2 days versus months of prior human effort

Illustrates the time-compression effect of automated algorithm discovery on engineering workflows that are typically bottlenecked by human trial-and-error.

Klarna doubled transformer model training speed while improving model quality

A 2x training speedup with no quality tradeoff is a significant commercial result for any company building ML models at scale.

Substrate achieved multi-fold runtime speedup in computational lithography

Computational lithography is one of the most compute-intensive workloads in chip manufacturing. Faster simulation expands what semiconductor designs are feasible to explore.

FM Logistic improved routing efficiency by 10.4%, saving 15,000+ km annually

Applied to the Traveling Salesman Problem — a classic NP-hard combinatorial optimization challenge. Km savings translate to fuel, cost, and emissions reductions at scale.

WPP achieved 10% accuracy improvement in ad campaign optimization over manual methods

In programmatic advertising, marginal accuracy gains compound across large media budgets, making a 10% improvement commercially meaningful.

Schrödinger achieved roughly 4x speedup in MLFF training and inference for drug discovery

MLFF models are used in molecular simulation for drug discovery, catalyst design, and materials science. A 4x inference speedup directly compresses R&D cycles in these domains.

Google Cloud expanding commercial AlphaEvolve access across five industry verticals

Commercial partnerships with Klarna, Substrate, FM Logistic, WPP, and Schrödinger span fintech/ML, semiconductors, logistics, advertising, and life sciences — indicating a broad multi-vertical go-to-market push.