Hacker News didn’t know what to make of it.
When the AlphaEvolve research paper dropped in May 2025, the thread exploded with over 1,000 points and 270 comments. One camp saw months of work dissolving into a weekend run. The other dismissed it as a toy for narrow, self-contained problems. That split hasn’t gone away. But with today’s general availability announcement, Google Cloud is betting that paying customers will settle the argument faster than any forum debate.
AlphaEvolve is now a full-fledged service on the Gemini Enterprise Agent Platform. What began as a DeepMind project that discovered faster matrix multiplication algorithms is now a line item on enterprise cloud bills. Users feed it a baseline “seed” algorithm and a scoring function; the system uses an ensemble of Gemini models—Flash for breath, Pro for depth—to propose mutations, evaluate them client-side, and evolve better candidates over generations. The code never leaves the user’s infrastructure. That privacy-by-design model was apparently a dealmaker for early adopters in finance and pharma.
The GA announcement came with an unusual level of customer proof. Here’s a snapshot:
| Customer | Domain | Reported Improvement |
|---|---|---|
| Klarna | ML training throughput | 2× speedup; explored ~6,000 candidates in 3 weeks with bit-exact reproducibility |
| JetBrains | IDE performance | Code completion latency down 15–20% |
| FM Logistic | Warehouse routing | 10.4% reduction on an already-optimized baseline; saves 15,000+ km of travel per year |
| Kinaxis | Supply chain forecasting | Accuracy up 22%, runtime down 90% |
| BASF | Supply chain digital twin | Planning and forecasting models improved by over 80% |
| Coolblue | Demand forecasting | 5%+ WMAPE reduction in only 200 iterations |
| Schrödinger | Molecular simulation | 4× speedup in ML force field training and inference |
| Oak Ridge NL | Scientific computing | Running on Frontier exascale system for GPU kernel optimization |
BASF’s case stands out—not for the numbers alone, but for what they imply. The company’s supply network spans 180 production sites and 5,000+ value chains. Deterministic models had failed repeatedly to capture the human decisions that keep things moving. According to Dr. Goetz Krabbe, VP for Global Supply Chain, AlphaEvolve finally made it possible to map the complex network and mirror the on-the-ground choices. “That’s where the real operational value lies,” he said in prepared remarks.
[SPONSORED]
COMFYUI WORKFLOW OPTIMIZATION
Reduce render times by 40% with our automated edge-silicon pipelines. Download Whitepaper.
FM Logistic saw similar bottom-line impact: a 10.4% cut in routing distance that Rodolphe Bey, Group CIO, called “particularly hard to come by” on an already highly optimized baseline.
Google tested it on itself first
Internally, Google says AlphaEvolve recovered 0.7% of global compute resources through smarter task scheduling, accelerated a Gemini training kernel by 23% (shaving 1% off overall training time), and reduced write amplification in Google Spanner’s LSM-tree compaction by 20%, cutting storage footprint by 9%. The team also used it to tweak arithmetic circuits in next-generation TPUs. These aren’t paper benchmarks—they’re production patches already merged into live systems.
Where the technical community stands
Fields Medalist Terence Tao gave a cautiously optimistic shout-out, saying constraint-based approaches like AlphaEvolve “make me a lot less pessimistic about AI use in mathematics and computer science.” On developer forums, though, the mood remains mixed. “AlphaEvolve didn’t have any measure of numerical stability, and lots of these ‘faster matrix multiplication’ algorithms have difficulty in this area,” reads a Julia forum critique. Another user countered: “It’s an evolutionary algorithm that uses an LLM for mutation—stop expecting it to write your entire application.” The truth is likely somewhere in between: the tool excels on well-defined, self-evaluable problems. For open-ended software engineering, it’s the wrong hammer.
Open-source alternatives have already sprung up. OpenEvolve, with over 6,200 GitHub stars, can run on a laptop and claims 2.8× speedups on Apple Silicon. CodeEvolve and ShinkaEvolve compete on benchmark scores. But as one Hacker News commenter noted, most of them assume you’ll throw a frontier model at every mutation—fine for experimentation, expensive for production. Google’s service bundles the models, the orchestration, and the enterprise governance layer. That’s a subscription many large companies will pay, even if hackers can roll their own.
The pricing question no one answers publicly
Google hasn’t published a per-API-call price for AlphaEvolve. Community estimates put a single optimization run somewhere between tens of thousands and hundreds of thousands of dollars in compute, depending on the evaluator cost. For a logistics company saving 15,000 km a year or a chip maker squeezing percentage points out of a tape-out, that’s a rounding error. For a startup, it’s a non-starter—which might explain why the customer list skews heavily toward enterprises with deep pockets and in-house HPC.
[SPONSORED]
AI INFRASTRUCTURE AUDIT
Is your tech stack bleeding resources? Let our engineers evaluate your architecture.
Whether AlphaEvolve becomes a core revenue engine for Google Cloud remains to be seen. In Q1 2026, Google Cloud reported $11.3 billion in revenue (up 35% YoY), largely driven by AI. The 7.5 billion dollar partner fund announced alongside the Gemini Enterprise Agent Platform signals serious commitment. AlphaEvolve is one piece of that broader push—a niche piece, perhaps, but one that turns a former research curiosity into a recurring billable service.
As InfoQ’s coverage put it, the JetBrains testimonial captured the real shift: “Engineers still own the benchmark, review, and release decision. The search space is what gets smaller.”