Oracle Cloud Adds Google TPUs and Rival AI Chips

Illustration: Fox Bulletin

Oracle Cloud will sell time on Google TPUs. The announcement came at Oracle AI World on 27 October, and it would make Oracle the first major cloud provider outside Google itself to offer the chips.

It is part of a wider opening. Alongside the TPUs, Oracle is bringing in accelerators from Cerebras, SambaNova and Positron, with AMD, Intel and OpenAI’s own Jalapeño chip expected to follow.

What Oracle Cloud is adding, and when

The Google hardware is the eighth generation, split by workload: the TPU 8t for training and the TPU 8i for inference. Capacity is due in calendar 2027 and will ramp over several years rather than appearing at once.

Google tensor processing unit hardware — Image: Google Cloud

Only one non-Nvidia option is actually bookable today. Positron AI’s Atlas inference chip is generally available on OCI now, and was the only one described that way on the day. Cerebras inference is slated for 2027, and SambaNova’s SN50 inference chips are planned with no date attached.

That gap between announced and available is the honest shape of the story. Four of the five non-Nvidia options named on the day are roadmap items, and the one customers can use immediately is an inference chip from the smallest vendor in the group.

Nvidia is not being displaced

Oracle was explicit that Nvidia remains its primary partner and the best general-purpose option. Nvidia’s Vera Rubin NVL72 systems and Vera CPUs are already generally available on OCI, which is more than can be said for any of the newcomers except Positron.

The framing matters. This is a cloud adding choice at the margins, not a cloud switching sides. Most of what customers can run on OCI tomorrow is still Nvidia silicon.

Why a cloud would want five vendors

Supply and price. An operator with one accelerator supplier takes whatever allocation and pricing that supplier offers, and allocation has been the binding constraint on cloud growth for two years. Several of these chips are also inference-only, which is the workload where cost per token decides whether a product is viable.

The same pressure is visible elsewhere in the week. Amazon intends to sell Trainium to outside buyers from late 2027, and SpaceX borrowed about $40bn to secure GPUs for delivery in the first half of the same year. Everyone is solving the same shortage from a different direction.

There is a second motive that Oracle has less reason to state. A cloud that can host a customer’s preferred accelerator keeps that customer, and OpenAI’s Jalapeño chip appearing on the list points at exactly that kind of arrangement.

What would make it real

Availability dates that hold. A roadmap with four vendors and one shipping product is a statement of intent, and intent in cloud hardware has a long history of slipping a year. The TPU ramp through 2027 is the item to watch, because it is the only one that puts a rival’s flagship training chip inside someone else’s cloud. CNBC covered the announcement from the event.

Written by Daniel Okafor

Daniel Okafor writes the business and legal explainers at Fox Bulletin, covering the paperwork small companies actually run into: company structures, insurance cover, employment rules and the state-by-state differences that catch owners out. The guides start from the assumption that nobody enjoys reading a statute.

View all posts →