OpenAI plans to buy Google TPUs as Oracle starts selling them

For ten years the TPU was the one advantage Google refused to sell outside its own walls. This week it sold the distribution, and within a day the buyer on the other end was Nvidia's largest customer.

Oracle will offer Google's tensor processing units through Oracle Cloud Infrastructure starting in calendar 2027, the first time TPUs will run on a hyperscale cloud other than Google's own, Oracle announced at its AI World conference. Hours later The Information reported that OpenAI plans to buy TPUs directly, citing a person familiar with the plan. Bloomberg reported the same thing and CNBC said it had independently confirmed it. Neither OpenAI nor Google would comment to Bloomberg. Price, volume, chip generation and delivery dates are all undisclosed.

That leaves one hard artefact to go on: OpenAI has posted a job for an engineer to lead optimisation of its frontier models on TPUs, across model code, kernels and hardware execution. A job listing is not a purchase order. It is, however, what a company writes when the porting work has already started.

Why a TPU is suddenly something you can buy

The Oracle deal is the part that is signed, and it is the reason an OpenAI purchase is even a sentence someone can write. Google only began letting TPUs run in outside data centres in the second quarter of this year; Fluidstack is one of the operators, and Crux AI — the Blackstone and Alphabet venture — borrowed $22 billion against the chips themselves, an arrangement we walked through in What is a TPU?. OCI supplies the power, cooling and datacentre networking, and the TPUs talk to each other over Google's interconnect before handing off to Oracle's Acceleron fabric.

The silicon on offer is Google's eighth generation, split into two chips for the first time: TPU 8t for training, which Google says delivers nearly three times the compute per pod of the previous generation, and TPU 8i for inference and agent workloads. Per Google Cloud's technical post on the partnership, the eighth-generation part reaches up to 12.6 petaflops of native FP4 compute per chip — more than 100 times the 92 teraflops of the original TPU v1 — at up to twice the performance per watt of the last generation.

The claim worth pinning to the wall is Google's, that workloads can move between accelerators "without rewriting model code". If that is true at frontier scale, the CUDA moat is a switching cost rather than a wall, and every large buyer should be running a second-silicon pilot. If it is true only for JAX-shaped models trained by teams who already know XLA, then the OpenAI job posting is the honest signal: they are hiring the person who will find out.

Karan Batta, a senior vice president at OCI, told Data Center Dynamics the arrangement is a "multi-year, multi-gigawatt partnership" and that when Google launches a new TPU, Oracle intends to have it on day one. Gigawatts, not racks: Oracle is not treating this as a pilot SKU.

The business Google just opened

Sell-side estimates for the TPU business have been climbing all year. TD Cowen models $9.4 billion of TPU sales in the second half of this year and $62.5 billion for 2027; Wells Fargo expects roughly $9 billion in the third quarter alone and $99 billion next year, both per CNBC. That second number is what turns Google from an also-ran in accelerators into a second source at the scale Nvidia operates. Analysts also expect Google Cloud revenue to grow about 90% year on year this quarter.

Which raises the obvious question: why sell the family silver to a company you are fighting in search, in models and in front of European regulators? Three reasons, in ascending order of honesty.

Capacity monetisation is the first. Google's own blog makes the argument plainly: global demand for AI compute exceeds any single cloud provider's footprint. A TPU Google cannot house is a TPU not earning, and the wafer commitments to TSMC and Broadcom do not shrink when Gemini's training run finishes.

The second is that TPU revenue does not depend on Gemini winning. Google Cloud has spent two years selling a bundle — chips, models and cloud in one place — and the bundle only pays if the customer wants the model. Selling the chip alone reaches buyers who will never use Gemini, including the one that makes ChatGPT.

The third is strategic and less comfortable. Every external gigawatt of TPU capacity creates a constituency that needs Google's roadmap to succeed: Broadcom co-designs the chips, Oracle is on the hook for the datacentre build, Crux's lenders hold the silicon as collateral, and OpenAI would own a fleet. That is a lot of balance sheets aligned behind a Google product line — useful to have in a year when regulators are deciding how much of Google's search data must be shared with competitors.

Who wins, who loses

Broadcom collects either way. It co-designs the TPU, so every unit Oracle, Fluidstack, Crux or OpenAI takes is Broadcom content, and merchant TPU sales are a better business than captive ones.

Nvidia does not lose its lead this week, and anyone calling this the end of the GPU era is skipping a step. What it loses, at the margin, is the assumption that there is no alternative at the frontier — the assumption that lets it price at $10 per index task while a competitor sits at $1.27. A buyer with a working TPU path negotiates differently at renewal, even if it never moves the majority of its training.

AMD has the worst week of the three relative to expectations. Its Helios racks went live at Meta and it bought $1 billion of Meta stock, as we covered this morning — genuinely good news. But the pitch AMD has been selling hyperscalers is "we are your second source". Google just offered a better version of that pitch, with a decade of compiler work behind it and a cloud partner to run it.

Oracle wins the most on positioning. The same day it announced TPUs, it made Nvidia's Vera Rubin NVL72 systems generally available, and OpenAI's Broadcom-designed Jalapeño chip and Cerebras systems are headed to OCI too. That is a chip-neutral cloud, and it is a sharper differentiator than AWS's Trainium or Microsoft's Maia, both of which remain captive to their owners.

What the skeptics get right

"Plans to buy" is the weakest verb in technology reporting, and it is doing all the work here. There is no contract, no disclosed volume, no date, and two companies declining to comment. OpenAI has rented TPU capacity since June 2025 — a fact that has so far changed nothing about its Nvidia orders — and the base rate for reported procurement plans turning into fleet-scale deployments is not good.

The porting argument is also stronger than the press release suggests. Training a mixture-of-experts model on TPU means different sharding, different collective communication, and a compiler team that knows XLA's failure modes. Google's own framing — that this is smooth if your stack was co-designed for it — is a claim about Google's stack, not about everyone's.

And Google has an incentive to ration. TPUs are the scarce input to Gemini, and every chip sold to OpenAI is one not training the model Google needs to beat OpenAI. A company that says supply is short through 2030 cannot sell to a rival without a price that reflects the favour.

What to watch

Alphabet reports third-quarter results on Wednesday: the first question is whether anyone puts an external TPU revenue number on the record rather than leaving it to TD Cowen. After that, three tells. Whether Oracle or OpenAI disclose a volume when the 2027 capacity is allocated. Whether Nvidia's investment in OpenAI closes on unchanged terms — an investor whose customer is buying the competitor's silicon has a different conversation. And whether AWS responds by letting Trainium out of its own data centres, which is the moment the merchant-silicon model stops being Google's experiment and becomes how AI compute is sold.

Would you run frontier training on a chip your competitor designs? Write to us via our contact page.