Google lets Oracle sell its AI chips, and OpenAI is reportedly a buyer

Oracle will start selling access to Google’s custom AI chips, Alphabet said on Tuesday, according to CNBC. The chips are tensor processing units, or TPUs. CNBC describes the Oracle deal as the first time customers can reach them without going through Google’s infrastructure. Google spent years pitching its chips, models, and cloud as a single package. Now it is putting the chips in a rival’s catalog.

The timing is deliberate: Alphabet reports third-quarter earnings the next day. Thomas Kurian, who runs Google Cloud, worked at Oracle for over two decades before joining Google in late 2018, so the partner is a familiar one. Google Cloud revenue is forecast to climb about 90% year on year, after 82% in the second quarter.

Forecasts for the TPU business are big and far apart. TD Cowen puts TPU sales at $9.4 billion for the back half of this year, then $62.5 billion across 2027, while Wells Fargo models roughly $9 billion for the third quarter alone and $99 billion next year. Chief financial officer Anat Ashkenazi told investors in July that most of the revenue from signed TPU deals would be booked in 2027, with only a small slice this year.

Selling chips to other clouds costs Google something. TD Cowen estimates TPU sales carry a 20% operating margin, against about 35% for the cloud business overall last quarter. Harding Loevner portfolio manager Uday Cheruvu told CNBC he doesn’t see that overall margin rising much past the mid-30s. Google’s cloud still trails Amazon Web Services ($42 billion) and Microsoft Azure ($29 billion) in second-quarter revenue, at $25 billion, but it is growing fastest of the three.

Cost is the sales pitch. SemiAnalysis, an industry research firm, estimates Google’s newest Ironwood TPU runs about 44% cheaper to run and own than Nvidia’s GB200 server. Kurian said in September, at a Goldman Sachs event, that Google’s AI servers earn back their cost in under two years, and the ones built on its own chips do it in about half that.

Google’s largest TPU customer also competes with it. Anthropic signed up last October for up to a million TPUs, a contract CNBC values in the tens of billions, and Mizuho says it adds close to a gigawatt of capacity this year. Mizuho calls it Google “arming its closest model competitor.” OpenAI is lining up too: The Information reported Tuesday that it intends to buy TPUs, and CNBC then confirmed it.

Anthropic is getting help paying for all this. Broadcom, Google’s partner on the chips, laid out a $63 billion vehicle funded by Blackstone and three large banks. It will purchase TPUs and networking equipment from Broadcom, then rent that hardware to Anthropic over five years. William Blair analysts say that covers roughly half of what Anthropic needs in 2027 and expect further financing to follow. Anthropic’s filing for its IPO, submitted Monday, says it runs on chips from Google, Amazon, Nvidia, and AMD.

Nvidia still sets the price that every other chip gets measured against. A cheaper alternative on Oracle, with Anthropic and OpenAI as references, hands buyers a stronger hand in their next negotiation. Ashkenazi has said the money mostly lands in 2027, so the $62.5 billion and $99 billion forecasts won’t be tested for another year.

Reported by CNBC (MacKenzie Sigalos and Jordan Novet) on 28 October 2026.