Google will make its AI chips available through a major cloud provider outside its own platform for the first time. Under a partnership announced by Google and Oracle on October 27, Oracle Cloud customers will be able to use Google Tensor Processing Units (TPUs) starting next year. The deal broadens Google's route to market and gives it more access to customers in an AI chip market led by Nvidia.
Oracle Cloud adds a new route to Google TPUs
Until now, Google TPUs have been used mainly in the company's own data centers and products. Their availability on Oracle Cloud marks the first time the chips will be offered through a hyperscaler outside Google, giving customers another way to access them without relying solely on Google Cloud.
Nvidia CEO Jensen Huang played down the threat from rivals' in-house AI chips earlier this year, saying growth in Google's TPU business came “100%” from Anthropic. Recent customer and product developments, however, point to a broader potential demand base. OpenAI is hiring software engineers to optimize models for TPUs, and a report on the company said it plans to buy the chips from Google. Google, Nvidia and OpenAI did not respond to requests for comment; Google declined to comment.
Anthropic outlines major computing commitments
Anthropic said in an IPO filing this week that it plans to purchase $114.5 billion worth of computing resources from Google between April 2026 and July 2033. The filing does not specify how much of that spending would go toward TPUs, but estimates its spending on TPU computing capacity from Google and other suppliers at between $125 billion and $150 billion.
Anthropic also uses Amazon Trainium chips and AMD GPUs, alongside hardware from Nvidia and Google. Using several chip suppliers gives companies more options as they weigh availability, cost and computing needs, and could reduce reliance on any one vendor. The filing describes spending plans spanning several years, however, and does not allocate the full amount specifically to TPUs.
Analysts see room for TPU revenue growth
Mizuho forecast this month that Google's TPU business could generate $28 billion in revenue in 2026, rising to $67 billion in 2027 and as much as $100 billion in 2028. TD Cowen analyst John Blackledge estimated that TPUs could contribute $138 billion by 2028, equivalent to 34% of Google Cloud revenue that year. These are analyst forecasts, not guidance issued by Google.
Nvidia, by comparison, reported $96.2 billion in revenue in a single quarter last quarter, still far above current expectations for Google's TPU business. Nvidia shares were unchanged on Wednesday after the news, suggesting a muted market reaction. Mizuho analysts said consensus estimates may understate the potential of Google's TPU hardware business. Selling chips directly, rather than only renting computing capacity from its own data centers, could also ease Google's capital-spending burden.
Google continues to invest in AI capacity
Google developed TPUs more than a decade ago to improve products such as Search. The chips now also support the company's AI business. Google continues to invest in data centers and AI computing capacity, with analysts seeing no sign that it is slowing that spending. Bank of America analysts have said Google's spending is relatively efficient compared with peers, in part because its TPUs also serve its own AI projects.
Mizuho has also said the TPU business could provide some support for Google's core search business, particularly as AI agents may change how users find information. For Nvidia, the competitive implications will depend in part on whether Google can sell its chips at scale through external cloud platforms and whether customers continue to adopt multiple chip suppliers. Nvidia's revenue base remains substantially larger for now.