On October 27, Oracle announced it will offer Google's AI chips, TPUs (Tensor Processing Units), through its cloud platform Oracle Cloud Infrastructure (OCI). Google began opening TPUs to outside data centers in the second quarter of this year, but this is the first time TPUs will be offered on a major cloud provider outside Google Cloud.
What happened?
According to Data Center Knowledge, Oracle will offer Google's eighth-generation TPU family: TPU 8t for training models and TPU 8i for running trained models. These chips are said to deliver up to twice the performance per watt of the previous generation, with TPU 8t giving up to 2.7 times and TPU 8i up to 80 percent better performance per dollar. The TPUs will be available on OCI in 2027.
In the same period Oracle also added inference chips from Cerebras, SambaNova (in a separate announcement), Positron AI and OpenAI's Broadcom-built chips to its lineup. The partnership with Nvidia is growing too: Vera Rubin NVL72 systems and Vera processors are now generally available on OCI.
According to CNBC, the move is part of Google's strategy to take its chip business beyond its own cloud. Anthropic has a deal worth tens of billions of dollars to use up to one million TPUs, and OpenAI is also reported to plan buying TPUs from Google. Analyst estimates span a wide range: for 2027, TD Cowen expects $62.5 billion and Wells Fargo about $99 billion in TPU sales.
Why does it matter?
Nvidia has dominated the AI chip market for years. Google starting to sell TPUs on other clouds offers big AI companies a serious alternative and could increase price competition. Cloud companies, too, are now bringing different chips under one roof instead of depending on a single maker. We told the story of the first processor that fit on a single chip in our Intel 4004 post.