OpenAI plans to buy Google’s AI chips: the ChatGPT maker moves from renting TPUs to owning them

OpenAI is preparing to buy AI chips from Google, according to a report published on Tuesday — a move that would turn one of the industry’s most watched rivalries into a supplier relationship, and loosen Nvidia’s grip on the ChatGPT maker’s infrastructure.

The Information reported the planned purchase, citing a person familiar with the matter. No price, volume, or delivery timetable was disclosed, and neither company confirmed the report: OpenAI declined to comment, and Google did not immediately respond to a request for comment. That makes this a reported plan, not a done deal — but the surrounding evidence suggests it is more than a rumour.

From renting to owning#

The backstory starts in June 2025, when Reuters reported that OpenAI had begun renting Google’s AI chips through Google Cloud to run ChatGPT and other products. It was OpenAI’s first meaningful use of silicon from outside Nvidia’s lineup, and the motivation was cost: running a chatbot at ChatGPT’s scale means paying for inference — the compute consumed every time a user sends a prompt — on millions of queries a day.

At the time, Google was not renting its most powerful TPUs to OpenAI. A purchase, rather than another rental agreement, would signal a deeper commitment: hardware you own gets built into your long-term capacity planning, while rented capacity is a stopgap for demand spikes. The distinction matters because training and serving frontier models now runs on multi-year infrastructure bets, not quarterly purchases.

Why inference cost is the pressure point#

Google’s Tensor Processing Units, or TPUs, were designed specifically for AI workloads, and they have gained traction as an alternative to Nvidia’s market-leading graphics processors as buyers look for cheaper options and for silicon tuned to running models rather than training them. Google began selling TPUs directly to some customers this year, rather than only offering them through its cloud — a shift that makes a direct purchase by OpenAI plausible in a way it would not have been a few years ago.

The clearest public corroboration is a job listing. OpenAI is hiring a Software Engineer, TPU Compilers, based in San Francisco, whose role is described as leading the optimisation of the company’s frontier models on TPUs — covering performance across the full forward pass, from model code through kernels to execution on the hardware, with quantization called out as one lever for inference efficiency. Companies do not usually hire compiler engineers for hardware they only intend to rent briefly.

The hire would also land in an organisation with Google DNA. Trade-press reporting notes that OpenAI’s hardware division is run by Richard Ho, who was previously a senior director of engineering at Google and led development of its Cloud TPUs. None of this confirms a purchase on its own, but together the listing and the leadership fit a company preparing to run a meaningful share of its workload on Google silicon.

What this does — and doesn’t — change for Nvidia#

None of this dethrones Nvidia overnight. OpenAI remains one of the largest buyers of Nvidia’s GPUs, which it uses both to train models and to run inference, and porting large models to a new chip architecture is slow, specialised work — which is exactly why the compiler role exists. The realistic reading is diversification, not replacement: a second supplier for leverage on price, insurance against shortages, and a hedge against any single vendor’s roadmap.

For Google, the prize is validation. Selling TPUs to the company behind ChatGPT would mark the moment its custom silicon stops being an internal advantage and becomes merchant hardware that frontier labs plan around. It also feeds the cloud business: Alphabet reports quarterly results on Wednesday, with investors watching for signs that its heavy AI investment keeps driving growth in Google Cloud.

What to watch#

• Confirmation, or a denial: whether either company comments on the record, and whether the report’s single-source account gains a second confirmation.
• The shape of the deal: which TPU generation, how many chips, and over what timeframe — the difference between a pilot order and a fleet-scale commitment.
• Training or inference only: whether the chips will serve ChatGPT traffic, train future models, or both. Inference first is the likelier path, given the rental history.
• Alphabet’s earnings call: any mention of TPU momentum or large new customers, even without naming OpenAI.
• Whether rivals follow: other frontier labs face the same cost and supply pressures, and a public OpenAI purchase would make TPU adoption easier to justify internally.

The story to watch is not whether OpenAI keeps buying Nvidia chips — it will — but whether a second silicon supplier becomes a permanent part of its stack. If it does, the AI chip market stops being a one-supplier story, and the rivalry between the two companies that define consumer AI gets a lot more complicated.