On the screen at AI World in Las Vegas on Oct. 27, four accelerators appeared side by side, and only one of them belonged to the company that has supplied nearly all of the industry’s AI computing.
Oracle said Google’s tensor processing unit will be available on its cloud next year and will scale up over the following years. That makes Oracle the first large data-center operator outside Google to offer the chip, and it puts Oracle in the position of selling a competitor’s silicon to customers who would otherwise buy Nvidia.
Oracle will supply the buildings, the power, the cooling and the networking. Google supplies the chip and the software that runs on it. The split matters because the constraints on AI capacity have moved from the processor to the site: a gigawatt of power, a cooling plant and a grid connection now take longer to arrange than a shipment of accelerators.
The menu does not stop there. Cerebras systems will be deployed in the first half of next year. SambaNova and Intel’s Diamond Rapids are scheduled to arrive in the same window. The three additions are aimed at inference, the stage where a trained model answers requests, and where cost per token rather than peak throughput decides who wins a contract.
Inference is where the money is starting to concentrate. Training runs are episodic and enormous; serving is continuous and priced per unit of output. A cloud operator that can offer four ways to serve a model can quote a customer three prices and let the workload pick the architecture.
The Nvidia relationship expanded on the same day rather than shrinking. Oracle said the Vera Rubin NVL72 clusters already running on its cloud will grow to multiple gigawatts, and that Nvidia’s Vera CPU is now commercially available on the platform. A gigawatt of AI capacity is roughly the output of a large nuclear reactor, which is why the figure is now used as a unit of account in procurement conversations.
So Oracle is not choosing sides. It is building a marketplace in which the largest supplier keeps the biggest allocation while smaller architectures take the workloads where they are cheaper.
The applications side of the announcement reaches further into Oracle’s actual business. The company said its ERP, procurement and finance applications and their data will be connected to ChatGPT, letting users complete actions inside approval flows from a chat window rather than by opening a separate console. Oracle said it already has 130,000 active ChatGPT users and more than 95,000 active Codex users internally.
That is an unusual concession from an enterprise software vendor. Oracle spent two decades selling the argument that its applications were the system of record and that the interface belonged to it. Handing the interaction layer to a model provider accepts that employees would rather type a request than learn a menu.
The financial picture around all of this is less comfortable. Capital spending guidance for fiscal 2027 was maintained at $90 billion to $95 billion, a figure that requires the cloud business to convert signed contracts into revenue on schedule. The stock remains 47% below its high for the year, set on June 1.
Investors have been asking a version of the same question since the spring: how much of the backlog is committed capacity at committed prices, and how much depends on customers exercising options they may not need. Adding third-party chips does not answer that question, but it does give Oracle a way to fill capacity it has already financed with workloads that do not require the most expensive silicon.
There is a strategic logic to the TPU arrangement that goes beyond cost. Google has been selling TPUs directly to selected customers this year and has agreed to place them in Oracle’s cloud in 2027. For Google, the deal monetizes a chip program without building the data centers; for Oracle, it fills space with hardware that a rival has already proven at scale.
The risk sits in support and accountability. When a customer’s training job fails on a rented GPU, there is one vendor to call. When it fails on a TPU inside someone else’s building, the diagnosis crosses a company boundary, and the enterprise contracts that govern uptime were not written with that in mind.
For customers, the practical effect is a second price in a market that has offered very little choice. Large buyers of AI capacity have spent two years accepting allocation, delivery schedules and prices set by a single supplier. A second source of custom silicon, available through a vendor with its own power contracts, changes the conversation at renewal.
The bet Oracle is making is that enterprises want a menu. The bet Nvidia is making, visible in its own multiyear commitments and financing arrangements, is that most customers will keep choosing the default because the software, the tooling and the engineers are already there.
Both can be right for a while. What the Las Vegas announcement shows is that the largest cloud operators no longer assume a single answer, and that the cost of building for one architecture is now high enough to justify the expense of supporting several.
