Google TPUs on Oracle Cloud Infrastructure became official on October 27, announced at Oracle AI World in Las Vegas: the first time Google’s accelerators will be offered on a hyperscale cloud other than Google’s own. Coverage of the announcement has the rollout starting in 2027 and ramping to gigawatt scale over several years, with Oracle supplying the power, cooling and networking around them.

Oracle was quick to say this is not a break with Nvidia.

I believe that, and it still misses the point. The lineup Oracle laid out runs from Nvidia’s Vera Rubin superclusters and AMD’s Helios racks with MI455X to Cerebras wafer-scale systems, SambaNova inference boxes and Jalapeño, the chip OpenAI co-developed with Broadcom, with the TPU as the headline addition. Mahesh Thiagarajan, who runs OCI, said at the event that 2027 capacity sold out long ago and 2028 is already sold out as well, which is a strange moment to go shopping for more silicon vendors unless the customers are asking for them or the supply of Nvidia parts is the thing that runs out first. I lean toward the second reading. Oracle’s stock is reportedly down more than 30% this year, and being the neutral shop with every accelerator on the menu also pitches to investors who want to know where the data center money will come from. I wrote in September that the GPU was never Nvidia’s real moat, and a cloud that sells this many kinds of accelerators is a decent test of that claim.

The Jalapeño line surprised me most. It was only in early September that OpenAI put a benchmark number on Jalapeño against Blackwell, and here it is on a third party’s price list.

Cerebras is the detail I can’t resolve. The coverage says the first Cerebras deployments at Oracle are CS3 systems, landing in the first half of 2027, yet Cerebras launched the faster CS4 in August. Either Oracle is starting with the older generation or the reporting dropped a generation somewhere, and nothing I read says which.

Google’s side of the bargain is stranger than it looks. TPUs have been rentable in Google’s cloud for years, and Google has now agreed to let a direct cloud competitor run them in its own buildings. That only works if the software is easy to reach, which is where TorchTPU arriving a day after Amazon’s TorchNeuron starts to matter, and it fits a pattern of TPU capacity moving through financing arrangements like Broadcom lending Anthropic up to $42 billion to rent Broadcom-built chips.

Until the first racks go live in 2027, this is a slide with a very good logo wall. I’ll believe the gigawatts when someone posts the power contracts.

Sources

CNBC, Oracle expands cloud chip lineup and brings ChatGPT and Codex into its apps, October 27, 2026
PR Newswire via Yahoo Finance, Oracle brings Google TPUs to Oracle Cloud Infrastructure, October 27, 2026
Wccftech, Oracle Cloud Infrastructure brings Nvidia, AMD, Intel, Google, Cerebras, SambaNova and Broadcom under one roof, October 28, 2026
BiGGo Finance, Oracle AI World coverage: inference chip camp joins OCI, October 27, 2026
Yahoo Finance, Oracle, in effort to reduce reliance on Nvidia, adds AI chips from Google, Cerebras, and SambaNova, October 27, 2026