Google TPUs on Oracle Cloud: What the Deal Changes

Earlier this year Nvidia’s Jensen Huang told podcaster Dwarkesh Patel that Google’s TPU chips only grew because of one customer. “It’s 100% Anthropic,” he said. Business Insider quoted that line this week to make a point: it’s getting harder to say.

Two things happened on Tuesday, October 27. Oracle and Google announced that Google’s tensor processing units, the TPUs, will be rented out through Oracle Cloud Infrastructure. And The Information reported that OpenAI plans to buy TPUs from Google. They’re different kinds of news. One is a press release with two named executives. The other is a single report with one unnamed source.

Here’s the split, claim by claim:

Claim	Who says it	Status
Oracle Cloud will offer Google TPUs, in 2027	Oracle and Google, press release	Announced
Oracle is the first big cloud outside Google to do so	Oracle and Google	Their wording, with caveats below
OpenAI plans to buy Google TPUs	The Information, one unnamed source	Unconfirmed
TPU revenue of $67 billion in 2027	Mizuho analysts	Forecast

What did Oracle and Google actually announce?

Oracle’s press release says OCI will be “the first hyperscaler outside Google to offer Google TPUs.” Oracle supplies the power, cooling and networking in its own data centers. The chips arrive in calendar year 2027 and ramp “over a multi-year period.”

The release names Google’s eighth generation, which comes in two versions: TPU 8t for training and TPU 8i for inference. Google says the 8t delivers nearly three times the compute per pod of the previous generation. That’s Google’s own claim, and it gave no benchmark to check it against.

What Oracle didn’t say is how many chips it’s buying. The Next Web reports that Oracle declined to give a starting number. Oracle’s Karan Batta, a senior vice president at OCI, told DatacenterDynamics it’s a “multi-year, multi-gigawatt partnership” and “not an experiment.” Those are Oracle’s words, not a contract anyone has published.

Is Oracle really the first?

On paper, yes. In my read, the practice is less clear. The Next Web reports that Google is building its own TPU cloud, called Crux AI, with Blackstone (per DatacenterDynamics), and that the neocloud Fluidstack is already deploying TPUs for Anthropic. My take: Oracle’s “first hyperscaler” wording is accurate for the big-cloud club, and narrower than it sounds.

What’s the OpenAI part?

The Information reported on Tuesday that OpenAI plans to buy Google’s TPUs, citing a person familiar with the matter. No price, volume or timing came with it. The Next Web says it couldn’t independently verify the report, OpenAI declined to comment to Reuters, and Google didn’t respond right away.

There’s context that makes it plausible. Reuters reported in June 2025 that OpenAI had started renting Google’s chips, though OpenAI then said it had no active plans to use them in products. Business Insider says OpenAI currently has a job listing for an engineer to optimize its models for TPUs. And OpenAI unveiled its own Broadcom-designed chip, Jalapeño, in June, which CNBC reports will also run on OCI.

Plausible isn’t confirmed. My take: treat “OpenAI buys TPUs” as a report until one of the two companies says it.

Does this hurt Nvidia yet?

Not on the numbers. Business Insider points out that Nvidia booked $96.2 billion in revenue in its last quarter alone, and its share price stayed flat on Wednesday. Oracle’s own OCI president, Mahesh Thiagarajan, told CNBC that this isn’t a move away from Nvidia, whose chips he called “the best, most fungible.” Oracle also made Nvidia’s Vera Rubin systems generally available the same week.

The TPU numbers are mostly forecasts. Alphabet’s CFO, Anat Ashkenazi, said on the earnings call that Google expects a “meaningful revenue ramp in 2027” from TPU sales, per Business Insider. Mizuho projects TPU revenue of $28 billion for 2026 and $67 billion for 2027. Those are analyst estimates, and TD Cowen’s even higher figure for 2028 is a projection too.

The hardest data point is a company filing. In its IPO paperwork, Anthropic said it plans to spend $114.5 billion on Google compute between April 2026 and July 2033, as covered in our post on the Anthropic prospectus. The filing doesn’t say how much of that is TPUs, though Business Insider says it estimates $125 billion to $150 billion on TPU capacity from Google and other providers.

My read

Huang’s “100% Anthropic” line was probably true when he said it. Whether it’s true in 2027 depends on two things nobody has shown yet: how many TPUs Oracle actually deploys, and whether OpenAI’s reported purchase is a pilot or a real commitment.

What I find more interesting is the money side. Google’s AI spending is the subject of our look at Alphabet’s third-quarter free cash flow, and selling chips to other people’s clouds looks like one way to earn some of that back. It’s a bet that rivals will pay to use your hardware while competing with your products. Oracle gets a second chip supplier. Google gets a customer that isn’t only Anthropic.

For anyone who buys cloud AI, the near-term effect is zero. Nothing ships until 2027. Watch for Oracle naming a chip count, and for OpenAI or Google confirming or denying the purchase.

Sources

Note on sources

I read the Oracle release in full, and the Next Web and Business Insider articles apart from their last paragraphs. The Information’s report on OpenAI is paywalled and I did not read it; I’m relying on The Next Web’s and Business Insider’s descriptions of it. The DatacenterDynamics and CNBC interviews are quoted second-hand through The Next Web.