Google's AlphaEvolve Rolls Out to All Google Cloud Customers, BASF Reports 80 Percent Accuracy Gain

Google Cloud has made AlphaEvolve available to all of its business customers. The tool, which instead of writing code from scratch learns to improve existing algorithms, has moved out of closed testing and onto the Gemini Enterprise Agent Platform as a generally available product.

AlphaEvolve is a technology out of Google DeepMind, built on Gemini models. Rather than generating software from scratch, the tool acts as an evolutionary collaborator: the user provides a baseline algorithm and a target to reach, and the system iteratively tests variants until it finds a version that beats the starting point. The result is human-readable code, not a black box.

From Private Testing to Product

Google began giving select partners access to AlphaEvolve in December 2025. After several months of testing under production conditions, the company decided on a full rollout as part of the Gemini Enterprise Agent Platform, its toolkit for building and running AI agents for business. Tom Beyer, Group AI Product Manager at Google Cloud, is among those overseeing the product's development.

Google points to microchip design, route optimization in logistics networks and accelerating medical research as target use cases. These are areas where even a small algorithmic improvement translates into real savings measured in millions of dollars or weeks of computing time.

The BASF Case

The most concrete use case described so far involves the German chemical group BASF. The company manages a network of 180 production plants worldwide, and a single product's bill of materials can run 30 levels deep. Fulfilling an order, from active raw materials to the finished product, can take up to two years, and planners make thousands of local decisions every day.

BASF had long tried to build a digital twin of its supply network using deterministic models. According to the company, every earlier approach failed because it couldn't capture the way people actually make decisions at the intersection of hundreds of variables.

We had several attempts at building a digital twin of our complex supply network using deterministic models, and all of them failed. With AlphaEvolve, we can not only map the complex network based on system data, but at the same time understand and replicate the human decisions that drive our daily operations - Dr. Goetz Krabbe, Vice President of Global Supply Chain, BASF

Trained on three years of historical data, the system discovered on its own rules that BASF had never explicitly documented, including patterns of production consolidation across plants, dynamic safety stock levels responding to demand seasonality, and coordination dependencies between different tiers of the network. The result was more than an 80 percent improvement in model accuracy over the baseline version.

What It Means for Companies in Poland

For Polish manufacturing and logistics companies that have spent years planning supply chains in Excel or in ERP systems without any machine learning components, the BASF example points to a different way of thinking about optimization. Instead of building a rigid mathematical model of the network, a company can let an AI agent search for rules on its own based on historical data.

The barrier to entry remains high, though. Deployment requires access to the Google Cloud platform, sufficiently large and well-organized sets of historical data, and a team that can define the optimization goal in a way the agent can understand. This is a tool for large organizations with extensive data infrastructure, not an off-the-shelf product for a small company.

Competition in Algorithm Optimization

AlphaEvolve fits into a broader trend in which major cloud providers are trying to sell not so much generative AI for writing new code, but agents specialized in improving existing systems. That's a different market segment from popular coding assistants, closer to the traditional optimization tools used in operations engineering, but enriched with the ability to learn from historical data rather than just from rules written by humans.

Google hasn't disclosed pricing for the service or how many customers have used it since the full rollout. The company says more case studies from other industries are coming in the next few months as additional Google Cloud customers begin testing the tool on their own optimization problems.

Sources: Google Cloud Blog (cloud.google.com), blog.google (blog.google), TechBuzz AI (techbuzz.ai)