Skip to content
Komposite News
Markets · Technology · Capital
Thursday, September 10, 2026Global EditionGot a tip?
AI Chips

Google's chips go to work in Oracle's cloud

Oracle will offer Google's tensor processing units in its cloud from 2027, the first large cloud outside Google's own to do so. The Information reported the same day that OpenAI plans to buy the chips; neither company confirmed it.

Dana Whitfield · House byline
October 29, 20263 min read
A blue circuit board carrying four chips under copper cold plates connected by grey coolant hosesZinskauf, CC BY-SA 4.0, via Wikimedia Commons
A Google Tensor Processing Unit v3 board with four chips under copper cold plates joined by grey coolant hoses, photographed on a display, in a file photograph.Photograph via Wikimedia Commons

Oracle will offer Google's tensor processing units, the custom AI chips Google designs for its own data centres, to customers of Oracle Cloud Infrastructure from 2027, the company said on Tuesday at its AI World conference in Las Vegas. Oracle says that makes it the first hyperscale cloud outside Google's own to offer the chips, The Register reported.

The chips will be selectable in Oracle's cloud from 2027, with capacity growing over several years, but Oracle did not say how many it will deploy. Oracle will house them in its own data centres and supply the power, cooling and networking, TNW reported. Google's eighth generation comes in two versions, the TPU 8t for training and the TPU 8i for inference; Google says the 8t delivers nearly three times the compute per pod of its predecessor. Karan Batta, a senior vice-president at Oracle's cloud unit, described a "multi-year, multi-gigawatt partnership", according to TNW's account of his remarks to DatacenterDynamics.

"Different AI workloads have different compute needs," Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, said in the announcement, as quoted by The Register. "Bringing Google TPUs to OCI gives customers ... the flexibility to choose the right infrastructure," he added. Oracle announced on the same day that Nvidia's Vera Rubin NVL72 systems were generally available in its cloud, and it has Nvidia commitments The Register put at "tens of billions" of dollars, so the TPUs arrive as an addition rather than a replacement.

The second report is less firm. The Information reported on Tuesday that OpenAI plans to buy Google's TPUs, citing one person familiar with the matter, and gave no price, volume or timing. OpenAI declined to comment to Reuters and Google did not immediately respond, TNW said; Komposite has not been able to confirm the report. OpenAI is one of the largest buyers of Nvidia's processors, and unveiled its first custom chip, designed with Broadcom, in June.

WhenWhoWhatStatus
June 2025OpenAIBegan renting Google's TPUsReported by Reuters; OpenAI then said it had no active plans to use them in its products
April 2026GoogleSaid it would sell TPU hardware directly to a small group of customers to run in their own data centresCompany statement
By September 2026Anthropic5GW of Google TPU capacity coming through BroadcomAnnounced; reported by Komposite on 18 September
27 October 2026OracleTPUs selectable in Oracle Cloud Infrastructure from 2027, in Oracle's own data centresAnnounced; volumes not disclosed
27 October 2026OpenAIPlans to buy Google's TPUsReported by The Information from one source; not confirmed
Sources: TNW, TNW, The Register and Komposite's earlier reporting. Timeline compiled by Komposite.

The pattern is a chip that used to stay at home. Google said in April that it would sell TPU hardware directly to a small group of customers who could run it in their own data centres, and Anat Ashkenazi, Alphabet's chief financial officer, has said she expects a small share of TPU sales revenue to be booked this year with most of it in 2027, according to TNW. Demand from customers including Anthropic, which has 5GW of Google TPU capacity coming through Broadcom, has been strong enough that Google's own researchers have had to wait for access, TNW reported.

It is also showing in Google's numbers. Alphabet's third-quarter results, published on Wednesday, show Google Cloud revenue up 103 per cent to $30.7bn, "powered by huge demand for our AI infrastructure and AI solutions", and cloud operating income of $11.1bn. By our arithmetic, that is an operating margin of 36 per cent, against 24 per cent a year earlier. The same results showed Alphabet spending more on data centres than it took in from operations, which is the subject of Komposite's lead story today.

Why it matters: Nvidia's position in AI computing rests partly on the fact that its chips can be rented from all the large clouds, while Google's could be rented mainly from Google. Oracle's deal and, if confirmed, an OpenAI purchase would make TPUs a second chip that can be bought or rented outside the company that designs it. Oracle's shares rose nearly 5 per cent on Tuesday, TNW reported, citing CNBC, but are down more than 30 per cent this year on investor concern about how it will pay for its data centres. The Register's Brandon Vigliarolo called much of the week's news "circular spending" among the largest AI companies.

OpenAI has been widening its suppliers for months: it was seeking thirty billion dollars of new funding this month, Komposite reported, and the chip start-up Tenstorrent more than doubled its valuation on the same demand for alternatives to Nvidia.

MORE ON THIS STORY