Oracle to Sell Google TPUs, Putting Google Chips Outside Google's Own Cloud

Oracle will offer Google's tensor processing units to its cloud customers, marking the first time Google's custom-built AI chips will be available outside of Google's own infrastructure. Alphabet announced the agreement on Tuesday, October 27, one day before its third-quarter earnings report.

Google has spent years marketing its chips, models and cloud services as a single package, a bundle Amazon and Microsoft have tried to match. The Oracle deal changes that approach, distributing Google's silicon to customers who run their workloads elsewhere. Thomas Kurian, who leads Google's cloud division and spent more than 20 years at Oracle before joining Google in late 2018, is driving the shift. Google opened its TPUs to outside data centers in the second quarter, before the Oracle announcement followed.

Google Cloud booked $25 billion in revenue in the second quarter, behind Amazon Web Services at $42 billion and Microsoft Azure at $29 billion, but it grew faster than both. Analysts expect Google Cloud revenue to rise about 90% from a year earlier in the third quarter, after an 82% increase in the second quarter. "Google was great at doing its own thing, but it was never able to serve customers," said Uday Cheruvu, a portfolio manager at Harding Loevner, which owns Alphabet shares. "That mindset has changed."

Estimates for TPU sales vary. TD Cowen models $9.4 billion in TPU sales for the second half of this year and $62.5 billion for 2027. Wells Fargo expects about $9 billion in third-quarter sales and $99 billion next year. CFO Anat Ashkenazi said on the July earnings call that only a small portion of TPU revenue from existing agreements will be recognized this year, with the "vast majority" arriving in 2027. She also said Google would use more outside capacity in the third quarter while it builds out its own infrastructure.

Selling TPUs outside Google carries tradeoffs. TD Cowen estimates a 20% operating margin on TPU sales, against an overall cloud margin of about 35% last quarter. Cheruvu does not expect Google's cloud operating margin to widen much beyond the mid-30s. Kurian, speaking at a Goldman Sachs conference in September, said Google's AI servers pay for themselves in under two years, and servers running its own silicon reach payback in roughly half that time. "It's because we can optimize that whole stack," Kurian said. Industry research firm SemiAnalysis estimates Google's latest Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server.

Anthropic, whose Claude models compete with Google's Gemini, agreed last October to use up to 1 million TPUs in a deal worth tens of billions of dollars. Mizuho says that agreement brings nearly a gigawatt of capacity online this year and describes the relationship as Google "arming its closest model competitor." OpenAI is also planning to buy TPUs from Google. This week, Broadcom, which works with Google on TPUs, detailed a $63 billion vehicle backed by Blackstone and three large banks that will buy TPU and networking gear and lease it to Anthropic for five years. Google declined to comment.