Google Routes AI Chips Through Oracle’s Cloud

Google and Oracle announced Tuesday that Oracle cloud customers can use Google’s Tensor Processing Units starting in 2027. The deal gives Google a new channel for selling its AI chips as it tries to grow the business and compete with Nvidia. The companies that need computing power make up the market; those that own and sell the hardware set the terms.

A new route to the chips

For the first time, Google will offer its TPUs on another hyperscaler’s platform. Google began work on its in-house silicon in 2013, then started selling TPUs to outside customers in the second quarter of 2026. Oracle’s cloud will now bring the chips to its customers, while Google looks for growth beyond its own data centers.

The deal fits a wider corporate push to find alternatives to Nvidia. Analysts forecast rapid growth for Google’s TPU business as companies adopt multichip strategies. But Nvidia’s chip business still exceeds Google’s, and Google remains a major customer for Nvidia GPUs. The contest is between companies with the money and infrastructure to build, buy and distribute the hardware.

OpenAI listed a software engineer job to optimize its models for TPUs. The Information reported Tuesday that OpenAI planned to buy TPUs from Google. Those developments point to possible demand, but the article describes corporate plans and market forecasts—not community-controlled computing or a mutual aid response. That distinction matters.

Billions behind the hardware

Anthropic’s IPO filing said the company planned to spend $114.5 billion on Google compute from April 2026 through July 2033. The filing didn’t say how much would go to TPUs. It estimated $125 billion to $150 billion for TPU capacity from Google and other providers.

Anthropic also said it used Amazon’s Trainium chips and AMD GPUs, along with chips from Nvidia and Google. CNBC reported Anthropic held $112.4 billion in cash and cash equivalents and $19.3 billion in short-term investments as of June 30, 2026. Those sums topped the cash and short-term investments of every non-bank U.S. company in CNBC and S&P Capital IQ data except Berkshire Hathaway and Alphabet. Here, enormous corporate reserves meet enormous spending plans.

Mizuho projected TPU revenue of $28 billion in 2026 and $67 billion in 2027, with the possibility of $100 billion in 2028. TD Cowen’s John Blackledge projected $138 billion in 2028, equal to 34% of Google Cloud’s projected revenue that year. Nvidia reported $96.2 billion in revenue in the previous quarter. In a note published Oct. 18, Mizuho analysts wrote: “We believe Google’s TPU hardware business is emerging as a materially larger opportunity than what is currently reflected in consensus estimates.”

The companies set the terms

Mizuho analysts said outright TPU sales, rather than rentals from Google data centers, could ease Google’s capital spending. Bank of America analysts called Google’s spending “relatively efficient” compared with peers, noting that TPUs also power Google’s own AI efforts. Google declined to comment; Nvidia and OpenAI didn’t respond to requests for comment.

Google developed the chips to support Search and other products. Mizuho said the TPU business could also help buffer Google’s core search business as AI agents challenge it. At the eighth annual Technology Executive Council Summit in New York, speakers discussed AI risks and costs, regulation, agents, worker use of AI and the technology’s effects on society. Cisco President & Chief Product Officer Jeetu Patel said companies should think of rogue AI agents like teenagers and discuss how to use them safely.

The summit’s listed speakers included executives from Forum AI, Publicis Sapient, Calix, Brooklyn Sports & Entertainment and Levi Strauss. The account covers forecasts, corporate spending and executive discussion; it describes no worker-led or community-run alternative to the companies building and selling the infrastructure.