Oracle will bring Google's TPUs to its cloud as AI stops depending on a single chipmaker

29 Oct 2026 4 min read 610 views Technology
Technology

On October 27, 2026, at its AI World conference in Las Vegas, Oracle announced that it will offer Google's eighth-generation TPUs on Oracle Cloud Infrastructure (OCI) starting in 2027. It is the first time a hyperscaler other than Google Cloud will make these chips available to its customers, and it comes alongside deals with Cerebras, SambaNova, Nvidia and OpenAI itself.

The underlying message is clear: the AI cloud is no longer synonymous with Nvidia GPUs. In a single day, Oracle presented a lineup from five different silicon vendors, and the next day it emerged that OpenAI is considering buying Google TPUs, something the company had publicly ruled out only a year earlier.

This week's context

According to Oracle's announcement and coverage by CNBC and The Register, Google will bring two variants of its eighth-generation TPU to OCI: TPU 8t for training and TPU 8i for inference. Oracle's Acceleron network will connect those clusters to OCI's databases and storage. Futurum sums up the event as Oracle turning its "Supercluster" into a multi-silicon platform: besides Google, it includes Nvidia Vera Rubin at multi-gigawatt scale, Cerebras (with deployment planned for the first half of 2027), SambaNova SN50 for agentic inference, and Intel Diamond Rapids. Oracle and OpenAI also announced ChatGPT and Codex integrations in the Fusion and NetSuite applications. Oracle shares rose nearly 5% and Cerebras shares around 8%.

What has changed

  • TPUs leave Google Cloud: until now, the only way to use TPUs was to rent them in Google's cloud. CNBC describes the move as a strategy shift for Google, which is starting to lean on partners to sell more AI chips.
  • A multi-chip catalog: Oracle will offer accelerators from Google, Nvidia, Cerebras and SambaNova, plus Intel CPUs, in the same cloud, so each workload can pick the silicon that fits best.
  • OpenAI's own chip reaches third parties: CNBC reports that OpenAI's Jalapeño inference chip will also run on OCI, although Techzine says it has no date yet.
  • OpenAI diversifies its compute: The Information and CNBC reported that OpenAI plans to buy Google TPUs. Other reports say it will not do so for now, so it should be treated as an unconfirmed plan.
  • More plans than services: TPUs will not reach OCI until 2027, and no pricing, regions or capacity have been announced.

Impact for development and product teams

For anyone training or serving models, competition between chips is good news: more supply usually means better prices and shorter waits for capacity. But it also adds complexity. Each accelerator family has its own software ecosystem (CUDA on Nvidia, JAX and XLA on TPUs, proprietary SDKs on Cerebras and SambaNova), and code tightly coupled to one of them does not move to another for free. Organizations that already build on portable frameworks or consume models through APIs will be the quickest to benefit from this diversity. For everyone else, the announcement is a reminder that depending on a single hardware vendor is a business risk, not just a technical one.

Practical recommendations

  1. Review how much of your training and inference code depends directly on CUDA and where a portable layer such as PyTorch/XLA, JAX or OpenXLA could be used.
  2. Separate training and inference workloads when planning capacity: the new specialized chips (TPU 8i, SambaNova, Cerebras) mainly target inference.
  3. Ask cloud providers for prices per task or per token served, not just per accelerator hour, so you can compare different architectures.
  4. Do not reserve long-term capacity based on announcements: wait for real pricing, regions and availability in 2027.
  5. If you use OpenAI models inside Oracle applications, keep track of the terms of the new ChatGPT and Codex integrations in Fusion and NetSuite.

What to watch next

  • Concrete pricing, regions and dates for Google TPUs on OCI.
  • Whether other cloud providers follow Oracle and offer TPUs.
  • Confirmation (or not) of OpenAI's TPU purchase and its effect on the relationship with Nvidia.
  • When OpenAI's Jalapeño chip becomes available to outside customers.

Conclusion: Oracle's announcement marks the shift from an AI cloud built almost entirely on Nvidia to one where several chipmakers coexist. For technical teams, the competitive edge will come from writing software that can move between chips and from negotiating compute as what it is becoming: a market with real alternatives.

Sources and documentation

Share