CALIFORNIA: OpenAI has started using artificial intelligence chips from Google to power ChatGPT and other products, a person familiar with the matter told Reuters on Friday.
The company, known for being one of the biggest buyers of Nvidia’s graphics processing units (GPUs), traditionally relies on these chips to train and run its AI models. These models use GPUs not only for training but also for inference — the process where an AI makes predictions or decisions based on new data.
Earlier this month, Reuters exclusively reported that OpenAI planned to adopt Google Cloud services to meet its increasing computing demands. The move marks an unexpected partnership between two of the most well-known rivals in the AI industry.
For Google, the agreement comes as it begins offering broader access to its in-house tensor processing units (TPUs), which were previously reserved for internal projects. This strategy has helped Google attract major customers, including Apple and startups like Anthropic and Safe Superintelligence, both of which were founded by former OpenAI executives.
Renting Google’s TPUs marks the first time OpenAI is meaningfully using chips other than those from Nvidia. It also signals a gradual shift away from its heavy dependence on Microsoft’s data centres, where most of its computing needs were being met until now. According to a report by The Information, this could position TPUs as a lower-cost alternative to Nvidia’s widely used GPUs.
OpenAI reportedly hopes that using TPUs through Google Cloud will help bring down the high costs of inference.
However, Google is not providing OpenAI with access to its most powerful TPUs, The Information said, citing a Google Cloud employee.
Google declined to comment on the matter, while OpenAI did not respond to Reuters’ request for comment.
Adding OpenAI to its list of clients highlights how Google is leveraging its home-grown AI technologies, from hardware to software, to drive growth in its cloud business.
Read next: Design issues may push Apple’s foldable iPhone launch to late 2026