Skip to content
News

October 28, 2026

The TPU Stopped Being Google's Private Weapon

Probir Sarkar
6 minute read

Google will put its TPUs on Oracle's cloud from 2027, the first time the chip is sold outside Google's own data centers. Analysts expect a thinner margin on those sales than on cloud, and the customers are Google's rivals.

The TPU Stopped Being Google's Private Weapon - BrowserStay blog card

Google spent thirteen years building a chip it would not sell. On Tuesday it put that chip on someone else's racks.

At Oracle AI World, Oracle and Google said Google's tensor processing units will be available through Oracle Cloud Infrastructure from calendar 2027, the first time TPUs are offered on a hyperscale cloud other than Google's. The same day, The Information reported that OpenAI plans to buy TPUs outright; CNBC said it confirmed that report, and OpenAI declined to comment.

One is a signed partnership, the other an unconfirmed purchase. Together they describe one decision: selling its best silicon to competitors is now worth more to Google than keeping it.

What Oracle actually got

Oracle will deploy TPU capacity in its data centres from 2027, with OCI supplying the power, cooling and networking. Oracle's release describes the eighth-generation parts, TPU 8t for training and TPU 8i for inference, and claims TPU 8t delivers nearly three times the compute per pod of the previous generation. Oracle's Karan Batta called it a "multi-year, multi-gigawatt partnership" and told Data Center Dynamics that "when Google launches a new TPU, we will have it on day one."

Neither disclosed terms, and Oracle declined to say how large the first deployment is. That "first" carries an asterisk: Google is separately building a TPU cloud platform with Blackstone in the US, so whichever starts first gets the label.

Until this year, the chip was not for sale

Google has worked on TPUs since 2013 and only began selling them externally in this year's second quarter. What changed is scarcity, not generosity. Oracle's infrastructure president, Mahesh Thiagarajan, told CNBC that supply remains a constraint and that customers benefit from several accelerators to choose from, while insisting this is not a pivot away from Nvidia, whose chips he called "the best, most fungible."

The margin Google is paying

TD Cowen estimates TPU sales carry about a 20 percent operating margin, against roughly 35 percent for Google Cloud when that estimate was made. Alphabet's third-quarter results, published on 28 October, complicate the arithmetic in Google's favour: cloud revenue more than doubled to $30.7 billion and segment operating margin rose to 36.2 percent from 23.7 percent a year earlier. TPU system sales were, in the company's words, "a meaningful contributor" to that growth, and Cloud chief executive Thomas Kurian said growth excluding TPU sales was "again incredibly strong."

Whatever selling silicon costs, it is not visible in the segment yet. The third-party bridging capacity Google has warned about came online only at the very end of the quarter, and cloud backlog grew to $561 billion from $514 billion three months earlier - so the demand being contracted is not in question, only the price of serving it.

Nobody agrees on the size. TD Cowen 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 in the third quarter alone and $99 billion next year. Alphabet still put no dollar figure on TPU revenue. Chief financial officer Anat Ashkenazi said a small share is recognised this year with "a more meaningful revenue ramp in 2027", and guided that TPU sales keep scaling in the fourth quarter while cloud margin faces near-term pressure from that bridging capacity.

The internal numbers argue for keeping them. Kurian said in September that Google's AI servers pay for themselves in under two years, and servers running its own silicon in roughly half that. SemiAnalysis estimates the current Ironwood TPU costs about 44 percent less to own and operate than an Nvidia GB200 server. Google is selling access to the thing it runs most cheaply.

The customers are the competition

Google's largest TPU customer competes directly with Gemini. The October 2025 Anthropic agreement covers up to one million TPUs and is worth tens of billions; Mizuho estimates it brings nearly a gigawatt online this year and calls the arrangement Google "arming its closest model competitor." This week Broadcom detailed a $63 billion vehicle, backed by Blackstone and three banks, that buys TPUs and networking gear and leases them to Anthropic for five years, about half of its 2027 needs according to William Blair.

OpenAI would be next: among Nvidia's largest GPU buyers, renting Google chips since June 2025, and putting its frontier models on Google Cloud last week.

None of this removes Nvidia. OCI will run GPUs, Google TPUs, OpenAI's Jalapeño accelerator and Cerebras hardware side by side: the bottleneck is total capacity, not brand loyalty.

What is still unconfirmed

  • The OpenAI purchase. Reported by The Information from one person familiar with the matter; CNBC says it confirmed the plan. OpenAI declined to comment; Google did not answer Reuters. No terms, volumes or delivery dates are public, so treat it as intent, not an order.
  • The scale of the Oracle deal. "Multi-gigawatt" is an executive's description, not a disclosed contract value or chip count.
  • Which generation lands on OCI. The announcements name TPU 8t and TPU 8i but not which parts Oracle will install.
  • Whether Oracle is genuinely first. The Blackstone platform could launch sooner.
  • The margin. The 20 percent figure is an analyst model, not a Google disclosure, and Alphabet's quarter shows the cloud segment margin without saying anything about the margin on TPU systems themselves.

Alphabet reported on 28 October and cloud margin held: $30.7 billion of revenue at 36.2 percent, TPU sales named but not quantified. The margin question is deferred rather than answered, because the capacity that dilutes it was barely live in the quarter. This post will be updated when Alphabet puts a number on TPU revenue, when either side discloses terms, and when the first TPUs switch on inside Oracle's data centres in 2027.

Sources

Tags:#AI#Business#Cloud#Technology
Share this post:X / TwitterLinkedInRedditWhatsApp

No comments yet

Related posts