Google said Oracle ($ORCL) will offer its custom tensor processing units, the first time customers can use the chips outside Google's own infrastructure.
The announcement landed a day before Alphabet's third-quarter earnings, on Oct. 28.
Google has built these chips since 2013 and began opening them to outside data centers in the second quarter.
Oracle paired the move with chips from Cerebras and SambaNova, part of a push to reduce dependence on Nvidia and avoid being locked into a single supplier.
Nvidia has said roughly half its revenue comes from hyperscalers, the largest cloud providers, which is why selling into their data centers is a direct challenge.
The numbers analysts are putting on TPUs
TD Cowen models $9.4B of TPU sales in the second half of this year and $62.5B in 2027.
Wells Fargo is higher, projecting roughly $9B in the third quarter alone and $99B next year.
Google's finance chief, Anat Ashkenazi, said in July that most TPU revenue from existing deals lands in 2027.
Demand is already contracted. Anthropic, Google's biggest TPU customer, agreed last October to use up to 1M chips, and OpenAI plans to buy them too.
Broadcom detailed a $63B financing vehicle backed by Blackstone and three banks that will buy TPUs and networking gear, then lease them to Anthropic for five years.
The cost pitch is specific. Research firm SemiAnalysis estimates Google's latest Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server.
The margin trade-off
Selling chips earns less than renting them. TD Cowen estimates a 20% operating margin on external TPU sales, against the roughly 35% cloud operating margin Google posted last quarter.
Google Cloud brought in $25B of revenue in the second quarter, behind Amazon Web Services at $42B and Microsoft Azure at $29B, though it's growing fastest of the three.
Quarterly Cloud Revenue by Provider (Q2 2026)
Q2 2026 quarterly cloud revenue by provider.
Pie chart. AWS: Revenue ($B) $42. Azure: Revenue ($B) $29. Google Cloud: Revenue ($B) $25.
Analysts expect Google Cloud revenue to grow about 90% from a year earlier, after 82% growth in the second quarter.
Oracle is also bringing Nvidia's Vera Rubin NVL72 superclusters and Vera central processing units onto its cloud, so the Nvidia spending continues alongside the new suppliers.
Amazon, Microsoft, Meta, OpenAI, and Anthropic all have custom AI chips built or in development, and Amazon's AI chief said in June the company was in early talks to sell its own silicon to outside data centers.
Whether lower-margin chip sales grow fast enough to offset the revenue Google gives up per dollar is the open question.
