Oracle Expands Nvidia AI Chip Deal to Multi-Gigawatt Scale, Adds Google and Cerebras

By Grace Kim

Oracle will expand its Nvidia AI server deployments to "multi-gigawatt scale" while adding chips from Nvidia's rivals Google, SambaNova Systems and Cerebras Systems, the company announced Tuesday from San Francisco.

Neither Oracle nor Nvidia disclosed the deal's value. But Nvidia CEO Jensen Huang has previously said the newest "Vera Rubin" AI server generation — the systems covered by the expansion — can generate up to $40 billion in revenue per gigawatt for the chipmaker. That figure frames the stakes. Oracle did not say how many gigawatts the agreement covers.

What does this change for Oracle's cloud customers?

Three things, all announced at once.

• More Nvidia capacity. Oracle Cloud Infrastructure will scale up deployments of Nvidia's Vera Rubin servers, the newest generation of the chipmaker's AI systems.
• First-time Google TPUs. Oracle will begin offering Google's tensor processing units (TPUs), Alphabet's in-house AI accelerators, for the first time.
• New inference silicon. Chips from Intel-backed SambaNova Systems and Cerebras Systems join the lineup. Both target inference — the phase where AI systems respond to requests from human users, as distinct from training.

Even the Nvidia-based systems carry a tweak. Oracle said those deployments will skip some of Nvidia's networking chips in favor of Oracle's own in-house "Acceleron" networking silicon.

"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."

Why the inference market matters

Nvidia dominates the market for training AI systems, the compute-heavy process of building models. Inference is different. It's more competitive and more diverse, with a wider field of viable suppliers — which is exactly where Cerebras and SambaNova operate.

Cerebras CEO Andrew Feldman cast the Oracle deal as a major commercial breakthrough for his company. "That's a massive expansion of our customer base of tier-one customers," Feldman told Reuters in an interview. "We're being deployed in other people's data centers at large scale."

The arrangement shifts Cerebras from selling its own systems toward operating inside a hyperscaler's infrastructure — a distribution channel the startup has not had at this scale before.

How will Oracle pay for the chips?

That question went unanswered Tuesday. Oracle, which plans to hold an investor day this week, did not disclose how it intends to finance the chip purchases.

There is a partial answer from last month. On an earnings call, Oracle said customers are increasingly footing the bill for chips themselves. The company also maintained its capital spending target of about $90 billion to $95 billion for the 2027 fiscal year.

That capex figure now sits alongside a multi-gigawatt Nvidia commitment and a broadened supplier roster. Whether the $90-95 billion range absorbs the new deals, or the deals rely on customer-funded financing, is the open item heading into the investor day.

The strategic direction is clear regardless. Oracle is hedging its AI infrastructure bet across at least four silicon vendors — Nvidia, Google, SambaNova and Cerebras — while pushing its own Accelerator networking into systems that would otherwise carry Nvidia's. For customers, that means more compute options on a single cloud. For Nvidia, it means a bigger deployment with a partner that is no longer exclusively theirs.