Google TPUs Coming to Oracle Cloud From 2027

Google’s Tensor Processing Units (TPUs), the AI chips it designs in-house, will be offered on Oracle Cloud Infrastructure (OCI) from 2027, the two companies said on 27 October.

Oracle’s press release calls it “the first time Google TPUs will be offered on a hyperscale cloud platform outside Google”. A Google Cloud blog post says: “we will deploy TPU capacity on OCI starting in the calendar year 2027.”

Neither company gave a capacity figure, a price or the names of any customers.

What the companies announced

The companies say no other large cloud provider has offered TPUs before. Under the deal, Google’s TPU systems will sit inside Oracle’s data centres. Oracle says “OCI will provide the power, cooling, and datacenter networking” to run them at scale.

Oracle says TPUs on OCI will be “available in calendar year 2027, ramping over a multi-year period to meet customer needs.”

The chips in question are Google’s eighth generation, which comes in two versions: TPU 8t for training models and TPU 8i for inference, the work of running them. Oracle’s release says TPU 8t delivers “nearly 3x the compute performance per pod over the previous generation”.

Google’s blog adds more figures, all its own claims: up to 12.6 PFLOPS of FP4 compute per chip, “over 97% goodput” (the share of time a cluster spends on useful work) on large training runs, and “up to 80% better performance per dollar for low-latency serving”.

Google’s blog says its TPU systems are built for direct liquid cooling at rack scale, and that Oracle’s newer data centres are designed for that kind of load. Oracle’s Acceleron networking will link the TPU clusters to the rest of OCI, including Oracle databases and storage. Inside each TPU cluster, Google’s own chip-to-chip links and optical switches carry the traffic. Google says those optical switches can route around faults without restarting jobs.

“Different AI workloads have different compute needs,” said Mahesh Thiagarajan, president of Oracle Cloud Infrastructure. Amin Vahdat, Google’s chief technologist for AI and infrastructure, said: “The next generation of AI will require a diverse range of compute.”

Why Google would put its chips in a rival cloud

Google and Oracle compete for cloud customers, so this is an unusual move. Google’s blog gives its reason plainly: “the global demand for AI compute exceeds any single cloud provider’s footprint.” In other words, Google wants its chips used more widely than its own data centres can manage, and Oracle has the sites, power and cooling to host them.

For Oracle, TPUs add a second kind of AI chip next to the NVIDIA GPUs it already sells. On the same day, it announced that NVIDIA’s Vera Rubin systems were now generally available on OCI. Offering both lets customers pick the chip that suits each job.

There is also a data angle. Many large companies keep their main records in Oracle databases. Google’s blog says running TPUs inside OCI lets them fine-tune and run models next to that data “without costly cross-cloud transfers”.

Software matters too. Google has been working to make TPUs easier to use with common AI tools, including by open-sourcing TorchTPU, its native PyTorch stack for TPUs. Its blog says teams can train and serve models on TPUs “using the tools you already rely on”, which should make it simpler to move work between Google Cloud and Oracle.

Still, 2027 is some way off and the deal has no stated size. Until Oracle or Google gives capacity numbers, it is hard to judge how much compute this adds. The deal fits Oracle’s push to sell AI to businesses, alongside opening its Fusion apps to ChatGPT and a $1 billion pledge to its partners.

What happens next

The companies say deployment starts in calendar year 2027 and ramps over several years. Pricing, regions and first customers have yet to be announced.