OpenAI plans to buy Google TPUs, reportedly easing its reliance on Nvidia GPUs
The Information reported on October 27, 2026, citing one person familiar with the matter, that OpenAI plans to buy Google's tensor processing units, or TPUs, which Google builds for AI workloads. TechRadar's report relays the same account. Reuters followed up, and OpenAI declined to comment while Google did not respond. The price, chip count, TPU generation and delivery schedule have not been disclosed, so the scale of the order is unknown.
The purchase would extend a relationship that began as rental. Reports from June 2025 said OpenAI had started renting TPUs, and the company then described the work as early testing with no plans to deploy them at scale. OpenAI is also hiring engineers to optimize its models for TPUs, and it is preparing its own Broadcom-built chip, Jalapeño, for deployment by the end of 2026. Nvidia remains central to the picture: Jensen Huang reportedly said OpenAI committed about $600 billion to Nvidia servers through 2030.
On Google's side, the company began selling TPU hardware directly to a small group of customers in April, letting them run the chips in their own data centers. Alphabet's CFO, Anat Ashkenazi, expects modest TPU sales revenue by the end of this year, with the bulk arriving in 2027. Oracle is also set to offer Google TPUs to cloud customers starting in 2027.
Why it matters
For companies building on AI infrastructure, the near-term effect is on inference costs. Google says its TPU 8i, designed for inference, delivers up to 80% better performance per dollar than the prior Ironwood generation, a figure from Google Cloud. Porting models and tooling to TPUs takes engineering time, which can cut into those savings, so hardware choice should be modeled as a cost line rather than assumed.
For vendors and funders, the report signals that a leading buyer is splitting compute spending across suppliers, which weakens single-vendor leverage and adds supply-risk questions to diligence. A purchase would also strengthen Google's case that TPUs can work as a standalone chip business, which may make the platform more attractive to other buyers. Founders selling to labs should expect more multi-chip stacks, and investors should ask which accelerators a company depends on and how portable its software is.
Key points
- 1The Information reported on October 27, 2026, citing one person familiar with the matter, that OpenAI plans to buy TPUs.
- 2Neither OpenAI nor Google has confirmed the deal, and OpenAI declined to comment to Reuters.
- 3The reports disclose no price, chip count, TPU generation or delivery schedule.
- 4OpenAI began renting Google TPUs in June 2025 and then said it had no plans to deploy them at scale.
- 5
