Google Lets Oracle Sell TPUs, Cracking Nvidia's Cloud Grip

Google's custom AI silicon is stepping outside Google's own data centers. On Tuesday, Oracle and Google announced that Oracle Cloud Infrastructure (OCI) will offer Google's Tensor Processing Units, making OCI the first hyperscale cloud outside Google to sell TPUs. According to Oracle's announcement, TPU capacity will land on OCI starting in calendar year 2027.

The timing matters: Alphabet reports third-quarter earnings on Wednesday, a day after the deal was revealed at Oracle AI World in Las Vegas.

What happened

Oracle will supply the power, cooling and datacenter networking to deploy Google's accelerators at scale, while Google supplies the chips and the compiler stack. The Google Cloud blog post frames it around economics rather than raw throughput, citing eighth-generation TPUs at up to 12.6 PFLOPS of native FP4 compute per chip and roughly 2x the performance per watt of the prior generation.

Google only began selling TPUs to outside customers in the second quarter of this year, after building them internally since 2013. CNBC reports that estimates for the TPU business now range from $9 billion in third-quarter sales (Wells Fargo) to $62.5 billion for 2027 (TD Cowen). Oracle president Mahesh Thiagarajan pitched the deal as workload matching: customers want multiple leading AI compute architectures inside one cloud.

Why it matters

Demand is already there from competitors. Anthropic agreed to use up to one million TPUs, and CNBC confirmed OpenAI also plans to buy Google silicon after The Information reported it on Tuesday. Broadcom, which co-designs TPUs with Google, detailed a $63 billion financing vehicle backed by Blackstone and three banks that will buy TPU and networking gear and lease it to Anthropic for five years.

The economics are not free for Google. TD Cowen models a roughly 20% operating margin on external TPU sales, against about 35% for Google Cloud overall. But SemiAnalysis estimates an Ironwood TPU costs about 44% less to own and operate than an Nvidia GB200 server, and cloud boss Thomas Kurian has said Google's AI servers pay for themselves in under two years, or roughly half that on its own silicon.

Builder impact

For teams deploying AI agents, the accelerator market is turning from a default into a decision. A second hyperscaler selling non-Nvidia capacity puts real pressure on inference pricing, and it gives buyers leverage in multi-year compute negotiations.

The catch is software, not silicon. TPUs are co-designed from compiler up, so portability depends on JAX, XLA and framework support rather than a simple lift-and-shift from CUDA. Builders should treat accelerator choice as a portability question: keep training and serving code abstract enough to move, and price out a TPU path before locking a long Nvidia commitment. The chip war's next front is the toolchain, and that is where lock-in will actually live.