Google will sell its TPU chips through Oracle's cloud, and OpenAI reportedly plans to buy them too

Oracle said on October 27 at its AI World event that it will make Google's TPU accelerators available through Oracle Cloud Infrastructure. SiliconANGLE describes it as the first time Google is supplying the chips to another hyperscaler. A separate report from The Information, relayed by Reuters and Yahoo Finance, says OpenAI plans to buy Google's AI chips as well. That second report cites a person familiar with the matter, and financial terms were not disclosed.

The Oracle side has more detail. According to Wccftech's summary of Oracle's announcement, deployments start in 2027 and ramp to gigawatt scale over several years. Oracle says the TPUs will link to each other using Google's ICI and optical circuit switching technology, and connect to the rest of the data center through Oracle's own Acceleron networking. Oracle and Google Cloud executives wrote that Acceleron's SmartNIC provides hardware-level tenant isolation and moves data between Oracle databases, storage and TPU clusters with high throughput and low latency.

TPUs are not new. They began as chips Google used internally, and outside customers have rented them through Google Cloud since 2018. What changes here is where they sit. Oracle also announced it is expanding its other chip suppliers: multi-gigawatt Nvidia Vera Rubin deployments, AMD's 6th Gen EPYC and Helios racks, Intel's upcoming Xeon chips, and Cerebras systems for fast inference, with first Cerebras deployments planned for 2027. CNBC, cited by SiliconANGLE, also reported Oracle intends to deploy OpenAI's Jalapeño inference processor.

The timing is notable. Alphabet reports third-quarter earnings on Wednesday, and Anthropic's IPO filing laid out large TPU commitments, including a plan to spend $114.5 billion with Google Cloud. Taken together, the reports suggest demand for non-Nvidia accelerators is spreading from Google's own cloud to rival clouds and to OpenAI, a major Nvidia customer. OpenAI has not confirmed the purchase in the material we reviewed.

For developers and AI builders, the practical point is choice. If TPUs become available on Oracle's cloud, teams that train or serve models on JAX or other TPU-friendly stacks could shop across providers instead of being tied to Google Cloud. Google also claimed in the Oracle material that its latest TPU generation delivers up to 80% better performance per dollar for low-latency serving, a vendor claim worth testing on your own workload.

What to watch next: whether OpenAI or Google confirms the chip purchase, what Alphabet says on its earnings call, and when Oracle publishes TPU instance types and pricing. Since deployments are slated for 2027, nothing here is something you can rent today, so treat it as a roadmap signal rather than a product you can try this week.

Building with AI? See this week's hot AI repos and how to use each one.