Google's AlphaEvolve Recovers 0.7% of Global Compute and Cracks a 55-Year-Old Math Problem

Google DeepMind’s AlphaEvolve is not a research demo. It’s running in production at Google right now, and the results are concrete: 0.7% of Google’s worldwide compute resources recovered continuously, a 23% speedup in Gemini’s training kernel, and — for the first time since 1969 — an algorithm that beats Strassen’s method for multiplying 4×4 complex matrices.

How it works. AlphaEvolve combines Gemini Flash and Gemini Pro in an evolutionary loop. Flash generates candidate solutions rapidly. Pro evaluates and refines the best candidates. Automated evaluators score each proposed algorithm against verifiable criteria. The system iterates — evolving solutions rather than retrieving them. This architecture is what separates AlphaEvolve from LLM-based coding agents that generate code in a single pass.

The infrastructure numbers. The data center scheduling heuristic AlphaEvolve discovered for Borg — Google’s cluster management system — has been in production for over a year. At any given moment, it’s freeing the equivalent of 0.7% of Google’s entire global compute footprint. That percentage sounds small. Applied at Google’s scale, it means millions of additional tasks completed on the same hardware every day. The Pallas kernel used in Gemini’s training architecture was sped up by 23%. FlashAttention operations saw a 32.5% improvement.

The math. AlphaEvolve applied to over 50 open problems across mathematical analysis, geometry, combinatorics, and number theory — improving the best-known solutions in 20% of cases. The headline result: matrix multiplication for 4×4 complex-valued matrices now requires 48 scalar multiplications instead of 49. Strassen’s 1969 algorithm had stood as the optimum in that configuration for 55 years. AlphaEvolve also pushed the kissing number lower bound in 11 dimensions from 592 to 593 and generated new theorems in complexity theory by discovering gadget structures too complex for human mathematicians to find manually.

The access question. AlphaEvolve is not publicly available. Academic researchers can apply for early access. Google frames the system as a general-purpose algorithm discovery engine applicable to drug discovery, materials science, and sustainability — but those applications are still in the research phase. What’s proven, right now, is infrastructure optimization and mathematical discovery.

The broader implication: AI systems are no longer just producing suggestions for humans to evaluate. AlphaEvolve autonomously generated production-grade infrastructure improvements that have been running at Google’s scale for over a year. The gap between “AI tool” and “AI colleague” is closing faster than most engineering teams realize.

Source: Google DeepMind Blog