OpenAI plans to buy Google's TPU chips, report says
OpenAI, the company that makes ChatGPT, plans to buy Tensor Processing Units (TPUs) from Google, according to a report published on Tuesday. The report cited an unnamed source who knows about the plans. It did not say how much the chips would cost, how many OpenAI would buy, or when they would arrive. A second news organization later confirmed the reported plan.
Neither company has confirmed the deal. OpenAI declined to comment when asked about the report, and Google had not responded to a request for comment sent after normal business hours.
Buying the chips would build on an earlier agreement. In June 2025, OpenAI began renting Google's AI chips. Google designs TPUs specifically for artificial intelligence tasks. Companies have become interested in these chips because they may offer a cheaper option than Nvidia's leading AI processors.
Google began selling TPUs to outside customers this year. It started delivering TPU systems to customer data centers during the April-to-June quarter.
OpenAI is also one of Nvidia's biggest GPU customers. It uses Nvidia's graphics processors to train AI models and run them when people use them. OpenAI and Google compete in search and AI products, but their business relationship continues to grow.
Last week, OpenAI announced a partnership to make its most advanced models and its Codex coding tool available through Model Garden, Google's catalog of AI tools. Customers will pay for them through Google Cloud Marketplace. According to Google Cloud, three models—GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna—became available for private testing to selected business customers on October 21.
Thomas Kurian, head of Google Cloud, said the company wants to give customers “a broad selection of models, including those from OpenAI.” If OpenAI buys TPUs, Google would also become a hardware supplier to the AI company whose models it now sells to cloud customers.
OpenAI is also developing its own chips. On June 24, it introduced Jalapeño, its first AI accelerator designed in-house. The chip was developed with Broadcom, and OpenAI plans to begin using it before the end of 2026.
Broadcom CEO Hock Tan described the chip as “just the beginning of a multi-generation roadmap.” The report about Google's TPUs did not explain how a possible purchase would work alongside OpenAI's Jalapeño project.
Several important details remain unknown, including which generation of TPUs OpenAI might buy, how many chips it would order, and what it would use them for. The chips could help train AI models or run them after training. Anthropic, a competitor of OpenAI, already buys Google's TPUs, and its agreements have helped increase demand for the chips.
In April, Anthropic reached an agreement covering several gigawatts of future TPU computing capacity from Google and Broadcom. The first part of that capacity is expected to begin operating in 2027.
During Alphabet's July earnings call, Chief Financial Officer Anat Ashkenazi said revenue from existing TPU system deals would make up only a relatively small share of the company's revenue in 2026. She said, “the vast majority of the revenues from these agreements will be realized in 2027.”
Amazon, Microsoft and Alphabet together could spend as much as $600 billion on AI infrastructure in 2026. Investors are questioning when these huge investments will start paying off. Separately, Google DeepMind introduced Genie 4, a model that can simulate worlds, on October 27.
Alphabet is scheduled to publish its third-quarter financial results on Wednesday. Its conference call is set for 1:30 p.m. Pacific time. Investors will look for signs that spending on AI is helping Google Cloud grow. They will also listen for comments from company leaders about TPU orders and potential customers, including OpenAI.