Oracle is expanding its Nvidia cloud AI chip arrangements to multi-gigawatt scale while adding Google, Cerebras and SambaNova offerings to Oracle Cloud Infrastructure.
Oracle said on Tuesday it is expanding its cloud AI chip arrangements with Nvidia to multi-gigawatt scale using Nvidia's newest Vera Rubin servers, while also adding chip offerings from Google, Cerebras Systems and SambaNova Systems to Oracle Cloud Infrastructure.
The companies did not disclose the value of the expanded Nvidia arrangement, though Nvidia chief executive Jensen Huang has previously said Vera Rubin systems can generate up to $40 billion in revenue per gigawatt for the chipmaker. Oracle said even the Nvidia-based systems will skip some of Nvidia's networking chips in favour of its own in-house Acceleron networking chips, and that it will offer Google's tensor processing units for the first time. A separate announcement from Cerebras and Oracle in Las Vegas on Tuesday said Cerebras fast inference will be made available on Oracle Cloud Infrastructure for AI labs, enterprises and developers building interactive applications.
The move matters because the two parts of AI computing are diverging: Nvidia dominates the market for training AI systems, while the inference market, when AI systems respond to requests, is more competitive and diverse. Cerebras chief executive Andrew Feldman said every step an agent makes requires inference, whether reasoning, checking its work or taking action, and that faster inference lets agents do more without making users wait. Oracle Cloud Infrastructure president Mahesh Thiagarajan said different AI workloads have different compute needs, and that choice gives customers flexibility on performance, throughput and economics.
The announcements themselves are the primary record here: the Cerebras-Oracle release and Oracle's statements as reported by Reuters. No dollar figure is given for the Nvidia expansion itself because none was disclosed, and the $40 billion-per-gigawatt figure is Huang's earlier statement about what Vera Rubin systems can generate for the chipmaker, not a price for this deal. Cerebras describes its Wafer-Scale Engine as integrating compute and memory on a single wafer-scale processor, a design claim the company itself makes for its hardware.
For customers, the practical effect is a wider menu on one cloud platform: Nvidia hardware for large-scale work, Google processors and specialist inference chips for responsive applications, and Oracle's own networking in place of some Nvidia parts. The test of the strategy will be availability and uptake on Oracle Cloud Infrastructure, which the companies have yet to detail in timing or pricing, and that will become clearer in Oracle's coming results and product updates.