OpenAI reportedly plans to buy Google TPUs, but neither company has confirmed it and no terms are known

OpenAI wants to buy chips from a direct competitor, according to a report on October 27, 2026. If true, it is a notable shift for the company that has been among the largest buyers of Nvidia's GPUs. But the report is one source deep, neither company has confirmed it, and almost everything that would make it meaningful is missing. This piece lays out what is actually known.

What was reported

The Information reported that OpenAI plans to buy tensor processing units, or TPUs, from Google, citing a person familiar with the plan. The Next Web, which said it had not independently verified the report, followed on October 28 with the framing that the purchase would make Google a second major chip supplier for OpenAI alongside Nvidia. Reuters followed up as well. According to The Next Web, OpenAI declined to comment to Reuters and Google did not immediately respond.

Runtime Wire, citing the same original report, described it as "a plan, not a completed transaction" and noted that "The scale, timing and workloads remain unspecified." That is the accurate summary. We do not know how many chips, which generation, at what price, to be delivered when, or for training versus serving.

The unknowns that matter

Buying chips and renting them are very different arrangements. Renting TPUs through Google Cloud means paying by usage and leaving the hardware in Google's data centers. Buying implies OpenAI owning or controlling the hardware, possibly in its own or partner facilities. The reporting does not say which this is. Google's public documentation describes TPUs as application-specific integrated circuits built for the large matrix operations common in machine learning, and notes that code must be compiled with the XLA compiler, which has real software consequences for a lab built around other tooling. For a plain explanation of how the two chip types differ, see our explainer on TPUs and GPUs.

The software point is the one most coverage glosses over. OpenAI's models and tools were built largely on Nvidia hardware and its CUDA ecosystem. Running them efficiently on TPUs takes engineering. The Information, via the accounts we reviewed, said job postings show OpenAI hiring engineers to optimize frontier models for TPUs, and noted the work is still at an early stage. That is a hiring signal consistent with the plan, and also a signal that the plan is not near completion.

The history

We have seen this story before. In June 2025 Reuters reported that OpenAI had begun renting Google's TPUs to power products including ChatGPT. Soon afterward, according to The Next Web's account, OpenAI said it had "no active plans to use them in its products." That reversal is the reason we put this report in the thin category. A leak about OpenAI and Google chips has been walked back once already.

There is also the question of what Google would sell. Earlier reporting cited by the aggregators we reviewed said Google was not renting its most powerful TPUs to its rival. The current Google Cloud documentation lists TPU7x (Ironwood) as its newest generation. Whether any purchase would include that generation is not stated.

Why it would matter

If OpenAI buys TPUs at scale, three things change. Nvidia loses some of its leverage with its biggest customer. Google gains a validation of its chips from the company most associated with its main rival in chatbots. And OpenAI gets a second source of compute at a time when supply, not money, is the constraint on the labs. We reported on a related shift when Oracle and Nvidia said Google TPUs would sit in Oracle's cloud, which suggests TPUs are becoming an option for more buyers than Google's own customers.

None of that follows automatically from a plan. Large chip purchases are announced with numbers: gigawatts, dollars, delivery dates, as with the multi-gigawatt agreements we have covered elsewhere. A report with none of those details is closer to a leak of intent than a leak of a contract.

What would change our view

We would upgrade this to a confirmed story on any of four things: a statement from OpenAI or Google, a filing, a purchase figure from a named source, or a data center site tied to TPU deployment. Until then the useful discipline is to keep the claim at the size the evidence supports. One source, one report, two non-denials.

It is also fair to note what the report does not say. It does not say OpenAI is leaving Nvidia. It does not say Google agreed. And it does not say the chips would train the next frontier model rather than serve existing ones. Each of those is a larger claim than the one reported.

Why one source is not enough

The Information is a respected outlet for inside-the-industry reporting, and its reports on chip deals have often been right. But the story as relayed contains no named source, no document and no figure. In our standards that places it at reported, not verified. The Next Web said plainly that it had not independently verified the report, and Reuters's follow-up, as described in those accounts, added the OpenAI no comment and Google's non-response rather than new facts. Three outlets repeating one unconfirmed claim is still one claim. ## What to watch

Watch for confirmation from either company, for TPU related items in OpenAI's infrastructure announcements, and for any change in Nvidia's commentary about its largest customers. Also watch OpenAI's own Broadcom chip, which The Next Web says is expected in use by year end. A lab that builds its own chip, buys Nvidia GPUs and also buys TPUs is a lab telling you supply is the problem. That much is plausible whatever happens to this particular report.