Oracle will rent Google’s TPUs from 2027. OpenAI buying them is still a report

Google and Oracle said on 27 October that Google’s tensor processing units will be offered on Oracle Cloud Infrastructure, and that OCI will be the first hyperscale cloud outside Google to rent them. Capacity is due to start in the calendar year 2027 and to build over the following years. The post, signed by Karan Batta of Oracle and Gurmeet Goindi of Google, does not say how many chips, which generation arrives first, or what a customer will pay. A date in 2027 is a plan. It is not a rack that is on.

The chip Google describes in the same post is the one it already runs. It says the eighth-generation TPU delivers up to 12.6 petaflops of FP4 compute per chip, and that the machines are built for liquid cooling at rack scale. Those are specifications of the current product. The post does not say that the 2027 OCI deployment is that generation, or the one after it. Oracle, the same day, also said it would expand Nvidia’s Vera Rubin systems to a multi-gigawatt scale. That is a different supply line. The two announcements share a cloud. They are not one order.

The OpenAI report should stay in its own sentence. The Information, and then Bloomberg and CNBC, said OpenAI plans to buy Google’s TPUs. Reuters said OpenAI declined to comment and Google did not respond. No price and no volume were in the report. OpenAI already rents Google’s chips through Google Cloud, and last week it said its own frontier models would be available there. A plan to buy the silicon, if it is real, is a step past renting it. Until one of the two companies says so, it is a report.

What changed on Tuesday is the landlord. For years a TPU was a Google Cloud product. From 2027, if the partnership holds, it will also be an Oracle product, in Oracle’s buildings, on a clock neither company has turned into a delivery schedule. Customers who wanted a second place to buy the chip have a name. They do not yet have a switch-on date finer than a year.