Oracle used its Oracle AI World conference in Las Vegas on Tuesday to widen the menu of chips behind Oracle Cloud Infrastructure, expanding its supply deal with Nvidia to multi-gigawatt scale while adding rival processors from Google, Cerebras Systems and SambaNova Systems for the first time.
The Nvidia expansion is the centerpiece. Oracle and Nvidia said the cloud provider will build out to "multi-gigawatt scale" with Nvidia's newest Vera Rubin AI servers. Vera Rubin NVL72 systems and Vera CPUs are now generally available on Oracle Cloud Infrastructure, and Oracle also plans HGX Rubin NVL8 systems that put eight Rubin GPUs in an x86-based server for frontier model training and agentic workloads. The companies did not disclose the value of the deal, though Nvidia CEO Jensen Huang has previously said Vera Rubin systems can generate up to 40 billion dollars in revenue per gigawatt for the chipmaker.
Even within that Nvidia buildout, Oracle is keeping part of the system for itself. Reuters reported that some of the Nvidia-based systems will skip some of Nvidia's networking chips in favor of Oracle's own in-house Acceleron networking chips — a sign that Oracle wants more control over the fabric that ties its AI clusters together, and a smaller bill for the components where it believes its own silicon can do the job.
The bigger strategic shift is diversification. Oracle said it will begin offering Google's tensor processing unit chips for the first time, with availability expected to begin in 2027 and ramp over a multi-year period. That would make Oracle the first hyperscaler outside Google itself to offer TPUs, in both the TPU 8t configuration for training and the TPU 8i for inference. Cerebras will add high-speed AI inference to the cloud starting in 2027, and SambaNova's SambaRack SN50 inference systems will be deployed directly in Oracle data centers. Intel separately announced an expanded collaboration to support enterprise and agentic AI on the platform.
"Different AI workloads have different compute needs," said Mahesh Thiagarajan, president of Oracle Cloud Infrastructure. "That choice gives customers the flexibility to choose the right infrastructure for the performance, throughput and economics of each AI workload."
The new suppliers are aimed squarely at inference — the stage when a trained AI model answers requests from users. Nvidia dominates the market for training AI systems, but inference is more competitive and diverse, and both Cerebras and SambaNova build their chips specifically for it. "That's a massive expansion of our customer base of tier-one customers," Cerebras CEO Andrew Feldman told Reuters. "We're being deployed in other people's data centers at large scale."
For enterprise IT buyers, the practical effect is choice inside a single cloud contract. Customers who train on Nvidia hardware can weigh Google, Cerebras or SambaNova silicon for serving models, and Oracle can negotiate with several chip suppliers instead of one. Whether that choice translates into lower prices will depend on how quickly the 2027 deployments arrive — but Oracle has made clear it no longer intends to build its AI cloud on a single vendor's chips.