Oracle Will Offer Google AI Chips as Cloud Hardware Choice Widens

Oracle plans to offer Google’s Tensor Processing Units through Oracle Cloud Infrastructure from 2027, giving AI customers another way to buy computing power without using Google’s own cloud. The companies announced the deal on 27 October, saying OCI would become the first hyperscale platform outside Google to offer the chips. It is a notable shift in an AI market where cloud providers compete fiercely for customers but increasingly share infrastructure when demand outstrips a single supplier.

The joint announcement says Oracle will supply the data-centre power, cooling and networking around the TPUs. The chips are not available on OCI yet. The companies expect availability in calendar 2027, followed by a multi-year ramp, but have not disclosed pricing or the size of the initial deployment.

Google’s TPUs were designed to accelerate AI training and inference, the stage at which a deployed model answers requests. Allowing another cloud to offer them can increase the number of customers using Google’s silicon even when those customers prefer Oracle’s infrastructure or already run important workloads there. Oracle, in turn, gains a differentiated chip option for customers who want to compare cost, performance and software compatibility across architectures.

The arrangement does not mean Google is abandoning its own cloud business. Nor does it establish that TPUs will be cheaper or faster for every workload. The practical advantage will depend on the model, the tools developers use, the networking around the chips and the terms Oracle offers when capacity becomes available.

Oracle is also widening its hardware menu beyond its long-standing Nvidia relationship. In a separate infrastructure update, the company described plans involving Nvidia, AMD, Intel, Cerebras and SambaNova. Some systems are already available, while others remain planned for 2027. Grouping them under one cloud gives customers more choice, but it also leaves Oracle with the difficult task of integrating different chips, networks and software stacks.

AI buyers are increasingly focused on the cost of running models at scale, not only on the speed of the chip that trained them. A useful accelerator must be available in sufficient quantities and supported by software that lets teams move from experiments to production. Oracle’s pitch is that customers should be able to match the hardware to each workload rather than fit every job into one supplier’s system.

TechBooky recently covered how AMD’s data-centre growth reflects demand for alternatives in AI computing. Oracle’s agreement with Google takes that diversification further. The outcome to watch is not the announcement alone, but whether OCI can make TPUs widely available in 2027 and demonstrate real cost and performance gains for customers.

For now, this is a planned commercial expansion with a clear start window, not a completed global rollout. Its larger significance is that even the companies building their own AI chips increasingly see value in selling them beyond the walls of their own cloud.