Oracle is doubling down on Nvidia and hedging at the same time. On Tuesday, the cloud company said it will expand to “multi-gigawatt scale” with Nvidia’s newest Vera Rubin AI servers, while also offering chips from Nvidia rivals for the first time. It’s a big order of confidence in Nvidia’s hardware, wrapped in a clear message: Oracle doesn’t want to be locked into one chipmaker.

The companies did not disclose what the expanded deal is worth. But Nvidia CEO Jensen Huang has previously said Vera Rubin systems can generate up to $40 billion in revenue per gigawatt for the chipmaker, which gives a sense of the scale involved when a cloud provider talks in gigawatts rather than server racks.

Oracle also said it will begin offering Google’s tensor processing unit (TPU) chips for the first time. That’s a notable shift. Google has historically kept its TPUs mostly for its own cloud, and Oracle putting them on its menu turns a rival’s in-house silicon into a product Oracle customers can rent.

What Oracle announced

The core of the announcement is the Vera Rubin expansion. Vera Rubin is Nvidia’s newest generation of AI servers, the successor to the Blackwell systems that powered the last wave of data center buildouts. Oracle plans to deploy them at multi-gigawatt scale, a measure of power consumption that has become the industry’s shorthand for how serious a buildout really is.

Alongside that, Oracle plans to offer chips from Intel-backed SambaNova Systems and Cerebras Systems. Both companies build chips aimed at inference, the stage where a trained AI model answers questions and serves users, rather than the training stage where models are built in the first place.

“Different AI workloads have different compute needs,” Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, said in a statement. “That choice gives customers the flexibility to choose the right infrastructure for the performance, throughput and economics of each AI workload.”

For Cerebras, the deal is a breakthrough. “That’s a massive expansion of our customer base of tier-one customers,” CEO Andrew Feldman said in an interview with Reuters. “We’re being deployed in other people’s data centers at large scale.”

Why Oracle is adding rivals’ chips now

Nvidia still dominates AI training, the compute-heavy work of building frontier models. Nobody in the industry disputes that. Inference is a different story. Once a model is trained, running it for millions of users is a cost-sensitive job, and buyers care a lot more about price per answer than raw peak performance.

That’s where the opening sits for Google, Cerebras, and SambaNova. Their chips are designed to serve models fast and cheaply, and cloud customers running chatbots, search features, and agents at scale feel every cent of that difference. Oracle is betting that some of its customers will pick a cheaper inference chip for day-to-day serving while still renting Nvidia hardware for training.

Timing matters too. Oracle’s cloud business has grown fast on the back of AI demand, and it needs capacity from wherever it can get it. Chip supply is tight, power is tighter, and a cloud provider that can only offer one vendor’s hardware is at the mercy of that vendor’s roadmap and delivery schedule. Offering four chip families instead of one gives Oracle more ways to fill its data centers.

The networking detail that matters

Buried in the announcement is a detail that could matter more than the headline. Oracle said even its Nvidia-based systems will skip some of Nvidia’s networking chips in favor of its own in-house “Acceleron” networking chips.

Networking is the connective tissue of an AI data center. It moves data between thousands of chips working on the same model, and it is also a high-margin business for Nvidia, which sells its networking gear as part of a tightly bundled system. By swapping in its own silicon for part of that layer, Oracle keeps more of the system design, and more of the profit, for itself.

It also weakens one of Nvidia’s strongest holds on its customers. When the chips, the networking, and the software all come from one vendor, switching gets expensive. Oracle is quietly loosening that grip, one component at a time, while still buying Nvidia’s flagship servers in bulk.

What this means for the AI chip race

The announcement fits a pattern running through the whole industry this year. Cloud giants are still buying Nvidia by the gigawatt, but they are also building or buying alternatives wherever they can. Google has its TPUs, Amazon has Trainium, Microsoft has Maia, and Meta is buying into the ecosystem too. Now Oracle, which doesn’t design its own AI training chips, is assembling the same kind of multi-vendor menu from other companies’ silicon.

For customers, that should mean real choice and, eventually, better prices. Inference costs have been falling fast, and competition between chipmakers is one reason why. A company running a customer-service chatbot doesn’t need the most powerful chip on the market. It needs the cheapest chip that answers fast enough, and Oracle can now offer it several options for that job.

For Nvidia, the message is mixed but not bad. Vera Rubin demand is clearly strong if Oracle is committing at multi-gigawatt scale. But the Acceleron swap and the rival inference chips show that even Nvidia’s biggest customers are now negotiating harder and designing around the edges of its bundle. Nvidia will still sell a lot of silicon. It just won’t sell every part of every system.

Watch what happens next on two fronts. First, whether other cloud providers follow Oracle in offering Google TPUs to outside customers. Second, whether Oracle’s customers actually shift inference work to the new chips, or keep defaulting to Nvidia out of habit. The hardware is now on the menu. What customers order will decide how much this really changes the market.

Frequently asked questions

What did Oracle announce about Nvidia Vera Rubin chips?

Oracle said on October 27, 2026, that it will expand to multi-gigawatt scale with Nvidia’s newest Vera Rubin AI servers, though the companies did not disclose the value of the deal.

Why is Oracle offering Google TPUs and Cerebras chips?

Oracle says different AI workloads have different compute needs, so customers want a choice of hardware. Google TPUs, Cerebras, and SambaNova chips target inference work, where AI models answer user requests, a market where Nvidia faces more competition.

What are Oracle’s Acceleron networking chips?

Acceleron is Oracle’s in-house networking chip. Oracle said even its Nvidia-based systems will skip some of Nvidia’s networking chips in favor of Acceleron, keeping more of the system design under Oracle’s control.

Does this mean Oracle is moving away from Nvidia?

No. The Vera Rubin expansion deepens the Nvidia relationship, but Oracle is no longer betting its whole AI cloud on one supplier. It is adding rival chips alongside Nvidia, not instead of it.

Sources: Reuters, The Economic Times, Finimize