OpenAI plans to buy Google chips in move to expand computing options
Key Facts
OpenAI plans to buy tensor processing units developed by Google, The Information reported, citing a person familiar with the matter. The report framed the plan as part of OpenAI’s effort to diversify its chip sources, but did not specify the volume, timing or value of the purchases. The newspaper said it was unable to learn the financial terms or other details of the purchase. The scope of the deal and how it would be carried out therefore remain unclear. The report comes as OpenAI relies on Nvidia’s ecosystem for a large part of its computing infrastructure.
The background to OpenAI’s relationship with Nvidia shows the scale of that reliance and the commercial ties between the two companies. The Information described OpenAI as one of Nvidia’s leading customers and said Nvidia had invested tens of billions of dollars in the startup. The report added that Nvidia agreed in August to provide more than $100 billion in credit support for a massive data center project in Ohio, planned for OpenAI and dedicated exclusively to Nvidia’s chip ecosystem. The newspaper also quoted Nvidia Chief Executive Jensen Huang as saying OpenAI had committed to spending about $600 billion on Nvidia servers through 2030. These potential commitments place the reported plan to buy chips from another supplier against the backdrop of OpenAI’s continued large investments in Nvidia infrastructure.
A job posting by OpenAI suggests that adapting its models to run on Google’s units is still at an early stage. The Information said the company is seeking an engineer to lead efforts to optimize its advanced models for the units. The report therefore covers a purchase plan as well as engineering work needed to adapt the models to a different architecture. The job description said experience in training or inference would be acceptable. The company also sought a strong foundation in optimizing real workloads on tensor processing units, according to the text cited by the newspaper.
The reported plan fits with a broader infrastructure direction OpenAI has announced, without the company confirming that a purchase of Google units is part of it. OpenAI said in March that over the previous 15 months it had expanded its strategy beyond a limited number of core providers, aiming to meet global scaling and reliability requirements. In the same announcement, it said Nvidia remained the foundation of its infrastructure, with its training fleet and most of its inference stack running on Nvidia graphics processors. The company said its portfolio includes cloud services from Microsoft, Oracle, AWS, CoreWeave and Google Cloud, alongside chips from Nvidia, AMD and AWS Trainium, Cerebras, and a partnership with Broadcom to develop its own chips. It added that differing workload needs make flexibility across systems important, and that a single architecture cannot efficiently meet the needs of the entire spectrum of AI.
OpenAI has also previously announced a multiyear, multigenerational agreement with AMD to deploy 6 gigawatts of its graphics processors. The first gigawatt was scheduled to begin deployment in the second half of 2026, according to the company’s announcement in October 2025. The agreement shows that work with other chip suppliers was already underway before the report concerning Google’s units. In a separate announcement on infrastructure expansion, OpenAI said it had exceeded its initial target of securing 10 gigawatts of AI infrastructure in the United States by 2029. It added that more than 3 gigawatts had been added during the 90 days before its update issued on April 29, 2026.
The Google Cloud partnership announced on October 21 remains separate from the chip purchase plan reported by The Information. OpenAI and Google Cloud said the partnership would make OpenAI’s advanced models and Codex workflows available through Google Cloud’s model platform, while the announcement stressed that Microsoft remained OpenAI’s primary cloud partner. General availability was planned for the weeks following the announcement. In a later update, Google Cloud said OpenAI models had begun appearing in a private preview for selected enterprise customers through the model platform. Neither announcement linked the partnership to the reported purchase of Google chips.
In another move that could expand access to Google units beyond its own cloud, Google Cloud and Oracle said deployment of tensor processing unit capacity on Oracle Cloud Infrastructure would begin in calendar year 2027. That timeline provides a future window for the capacity to become available to Oracle customers, but does not specify whether OpenAI will use it. Google Cloud provided specifications for its eighth-generation units, saying they deliver up to 12.6 petaflops of native FP4 compute per chip. It added that their performance per watt is up to twice that of the previous generation. The Oracle and Google announcement did not specify the amount of capacity to be deployed or which customers would use it.