OpenAI plans to buy Google’s AI chips, CNBC says, as Oracle prepares to offer TPUs in 2027
The Information reported the plan on Tuesday evening. CNBC said on Wednesday that it had confirmed a purchase of Google’s tensor processing units. Bloomberg and Reuters say the terms are not known, and that the companies declined or did not answer. What is signed is different: Oracle and Google say TPUs will run on Oracle Cloud Infrastructure from 2027, a first for a hyperscaler other than Google.
Deep dive5 minIn-depth
Maya Chen5 min read
- OpenAI
- TPU
- Oracle
- Nvidia

Photo: The InformationSource page
OpenAI plans to buy artificial-intelligence chips from Google, The Information reported on Tuesday, and CNBC said on Wednesday that it had confirmed a plan to buy Google’s tensor processing units. Ryan Barwick’s story at The Information, time-stamped 4:59 p.m. PDT, cited a person familiar with the matter. It said financial terms and other details of the planned purchase could not be learned, and that a deal for tensor processing units would be a win for Google’s chip business against Nvidia, which the story calls an OpenAI backer. CNBC’s account, by MacKenzie Sigalos and Jordan Novet, is one sentence on the purchase: OpenAI is planning to buy TPUs from Google, and CNBC confirmed that after The Information’s report. It does not add a price, a chip count, or a date.
Bloomberg’s story, updated at 5:14 a.m. EDT on Wednesday in the copy Yahoo carried, says OpenAI intends to buy AI computing chips from Alphabet’s Google, according to a person familiar with the matter who asked not to be identified because the discussions are private. Final terms are not yet known. OpenAI and Google declined to comment to Bloomberg. Reuters, in a story Investing.com shows as published at 9:23 p.m. and updated at 10:12 p.m. ET on Tuesday, said the report did not disclose financial terms. OpenAI declined to comment to Reuters. Google did not immediately respond to a Reuters request made outside business hours. Reuters added that OpenAI said last week it would make frontier models available through Google Cloud, and that Reuters reported in June 2025 that OpenAI had begun renting Google’s AI chips.
The dated offer is Oracle’s, and it starts in 2027
The announcement with a date is not the OpenAI purchase. Oracle and Google said on Tuesday, at Oracle AI World in Las Vegas, that Google’s tensor processing units will be available through Oracle Cloud Infrastructure. Oracle’s release says this is the first time Google TPUs will be offered on a hyperscale cloud outside Google. OCI will supply the power, the cooling, and the data-center networking. The chips on OCI are due in calendar year 2027, ramping over a multi-year period. No price is in the release.
The next generation of AI will require a diverse range of compute.
— Amin Vahdat, senior vice president and chief technologist, AI and infrastructure, Google, in Oracle’s announcement
Vahdat’s line continues that TPUs are a purpose-built option for demanding training and inference, and that putting them on OCI gives more organizations access. Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, said different AI workloads need different compute, and that the point of offering Google’s chips beside Oracle’s other accelerators is a choice of performance, throughput, and economics inside one cloud. The release describes an eighth-generation split: TPU 8t for training, which it says delivers nearly three times the compute performance per pod over the previous generation and cuts frontier model development cycles from months to weeks, and TPU 8i for inference, which it says is built for the iterative work of many specialized agents. It also says the chips can train and post-train dense models and mixture-of-experts models, and can run production inference for reasoning, agentic, and long-context work. Those performance lines are Google’s, printed in the joint release.
Selling the chip, not only the cloud around it
CNBC dates Google’s in-house silicon work to 2013, and says the company did not begin selling TPUs to outside customers until the second quarter of this year. The same story says Tuesday’s Oracle arrangement is the first time customers will be able to access TPUs outside Google’s infrastructure, and also that Google opened TPUs to outside data centers in the second quarter, before the Oracle announcement. Oracle’s release uses a narrower phrase: the first hyperscaler, other than Google, to offer the chips. CNBC says Thomas Kurian, who runs Google Cloud and spent more than 20 years at Oracle before joining Google in 2018, is using a partner cloud to get more of the chips into the market. Google declined to comment to CNBC.
On revenue timing, CNBC and Reuters both point to Anat Ashkenazi’s comments on the July earnings call. CNBC’s wording: only a small portion of TPU revenue from existing agreements will be recognized this year, with the vast majority in 2027. Reuters’ wording, in the Tuesday story: Alphabet expects to begin recognizing a small percentage of revenue from TPU sales agreements by the end of this year, with the vast majority in 2027. The bank models CNBC prints are not Google’s forecast, and they do not agree with each other. TD Cowen, CNBC says, models $9.4 billion of TPU sales in the second half of this year and $62.5 billion in 2027. Wells Fargo expects about $9 billion of third-quarter sales and $99 billion next year. SemiAnalysis, again via CNBC, estimates that Google’s latest Ironwood TPU costs about 44 percent less to own and operate than an Nvidia GB200 server.
Kurian, at a Goldman Sachs conference in September, told the audience Google’s AI servers pay for themselves in under two years, and that servers running Google’s own silicon reach payback in roughly half that time. “It’s because we can optimize that whole stack,” he said, in CNBC’s account. The same piece says Google’s biggest TPU customer is Anthropic, a competitor in models, which agreed last October to use up to 1 million TPUs in a deal worth tens of billions of dollars. Mizuho, CNBC writes, says that agreement brings nearly a gigawatt of capacity online this year. Anthropic’s IPO filing on Monday, CNBC notes, says the company uses silicon from Google, Amazon, Nvidia, and AMD, and that the Google and Broadcom partnerships provide compute powered by Google TPUs across multiple generations.
OpenAI’s own processor, and a print that is still ahead
Bloomberg says OpenAI has already announced plans to start using its own chips later this year. It names those processors Jalapeno, developed with Broadcom, and says they were created with help from Google veterans. Richard Ho, who runs OpenAI’s internal chip team, was previously involved in Google’s TPU work, Bloomberg reports. Reuters’ Tuesday story says OpenAI in June unveiled its first custom AI chip, designed with Broadcom, and said it planned to deploy that processor by the end of this year as part of a multi-generation program. Reuters also describes OpenAI as among the largest buyers of Nvidia’s graphics processors, used both to train models and to run them. A separate Reuters line in that story: Amazon, Microsoft, and Alphabet plan to spend as much as $600 billion combined this year on AI infrastructure.
Alphabet has scheduled its third-quarter call for 4:30 p.m. Eastern on Wednesday, a date it posted on 6 October. The chip stories above were published before that call. None of them says the OpenAI purchase has a signed price, or that Oracle’s 2027 TPU offering and OpenAI’s reported purchase are the same transaction. One is a cloud menu with a year on it. The other is a plan, confirmed by CNBC to the extent of the category of chip, with the commercial terms still unstated by every outlet that wrote it up.


