Three-quarters of new code at Google is now AI-generated, and the infrastructure is struggling to keep up.
Google’s engineers are running into capacity limits on internal AI coding tools because too many people inside the company are competing for the same finite pool of computing power.
CEO Sundar Pichai recently revealed that 75% of new code generated at Google is now produced by AI. That’s up from 50% just six months prior, which itself was a leap from the 25% figure reported in late 2024.
The internal AI tool driving much of this is Gemini. When you triple AI code generation inside a company with tens of thousands of engineers, the compute bill doesn’t just go up. The result is a classic supply-and-demand mismatch: Google aggressively promoted internal AI tool usage, engineers embraced it, and now the infrastructure is buckling under the load, creating bottlenecks that slow down the very productivity gains the tools were supposed to deliver.
Engineers defect to the competition #
Some engineering teams have reportedly sought exceptions to use Anthropic’s Claude for internal projects, despite company policy favoring homegrown tools. Some engineers have cited issues with performance and accessibility of Google’s internal AI offerings.
The situation also exposes a deeper philosophical divide within the company. Some teams are pushing for rapid deployment of AI models across every workflow. Others are pumping the brakes, arguing that code quality and reliability shouldn’t be sacrificed at the altar of AI adoption speed.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our