Google will distribute its tensor processing units outside its own infrastructure for the first time, announcing on Tuesday, one day before its third-quarter earnings report, that Oracle's cloud will offer TPUs to customers. According to CNBC, the move marks a structural change in how Google sells AI silicon and signals a direct challenge to Nvidia's dominance of the AI chip market.
The decision extends a distribution model that opened to outside data centers in the second quarter of this year. Google began developing in-house silicon in 2013 but only started selling TPUs externally in Q2. Oracle is now the first third-party channel.
Why distribution through partners changes the TPU market
Under Thomas Kurian, who joined Google in late 2018 after more than 20 years at Oracle, Google Cloud has pushed a full-stack proposition: chips, models and services sold as one package. That model worked inside Google's perimeter. It constrained reach. External distribution turns a vertically integrated asset into a protocol any cloud operator can plug into, a shift United States described as Google pushing its equipment outside its own walls to reach more customers.
Analysts see competitive consequences.
“They're a very strong competitor, and it's very possible that they could eventually take market share from Nvidia,” said Chris Ballard, managing director of Check Capital Management.
Estimates for the TPU business point to rapid growth. 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.
Margin tradeoffs of external TPU sales
Opening distribution imposes costs. TD Cowen estimates a 20% operating margin on TPU sales, versus an overall cloud margin of about 35% last quarter. Uday Cheruvu, a portfolio manager at Harding Loevner, which owns Alphabet shares, said he does not expect Google's cloud operating margin to widen much beyond the mid-30s.
The efficiency case offsets the margin compression. At a Goldman Sachs conference in September, Kurian 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 SemiAnalysis estimates Google's latest Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server. ICO Optics reported that the strategy relies increasingly on third-party partners to sell the custom chips.
“Google was great at doing its own thing, but it was never able to serve customers,” Cheruvu said. “That mindset has changed.”
Anthropic and OpenAI emerge as anchor TPU customers
Google's largest TPU customer is Anthropic, whose Claude models compete with Gemini. Anthropic 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 also plans to buy TPUs from Google, CNBC confirmed, after The Information first reported the ChatGPT maker's intentions on Tuesday.
The financing architecture behind these deals mirrors infrastructure-as-a-protocol. Broadcom, which co-builds TPUs with Google, detailed a $63 billion vehicle backed by Blackstone and three major banks that will buy TPUs and networking gear from Broadcom and lease them to Anthropic for five years. William Blair analysts estimate the plan covers about half of Anthropic's 2027 compute needs, implying more vehicles will follow.
In its IPO filing on Monday, Anthropic said it sources silicon from Google, Amazon, Nvidia and Advanced Micro Devices, letting it “select chips based on the best price-performance and efficiency characteristics for each workload.” On Google and Broadcom, Anthropic said the partnerships “provide us with access to compute capacity powered by Google TPUs across multiple generations of the TPU platform.”
Competitive position: third in revenue, first in growth
Google Cloud posted $25 billion in second-quarter revenue, behind Amazon Web Services at $42 billion and Microsoft Azure at $29 billion. Growth is the differentiator. Analysts expect Google Cloud revenue to rise about 90% year over year in the third quarter, after 82% growth in the second quarter, with TPUs contributing to the momentum.
“It's almost like the bigger the pie is getting, the faster they're accelerating into it,” said Ballard, whose firm held about $236 million in Alphabet stock as of June 30.
Google declined comment. BizToc noted that the full-stack pitch is one rivals Amazon and Microsoft have tried to replicate.
FAQ
When did Google start selling TPUs to outside customers?
Google began developing TPUs in 2013 and started selling them to outside customers in the second quarter of this year. The Oracle partnership, announced Tuesday, is the first time TPUs are available outside Google's own infrastructure.
How large is Google's TPU business expected to become?
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.
Who are Google's main TPU customers?
Anthropic agreed in October 2024 to use up to 1 million TPUs in a deal worth tens of billions of dollars. OpenAI also plans to buy TPUs from Google, CNBC confirmed on Tuesday.