一句话总结
On October 27, 2026, Oracle said its OCI Supercluster will adopt Nvidia's next-generation Vera Rubin NVL72 at multi-gigawatt scale, become the first hyperscaler outside Google to offer Google TPU, and add Cerebras, SambaNova inference chips plus its own Accelon networking silicon — a multi-supplier reset of cloud compute procurement.
From exclusive tie-up to multi-supplier: why Oracle signed three chip vendors at once
According to Reuters as carried by The Economic Times, on October 27, 2026, Nvidia and Oracle jointly announced that Oracle's OCI Supercluster will deploy Nvidia's latest Vera Rubin NVL72 systems at multi-gigawatt scale, with no deal value disclosed. Nvidia's official newsroom entry that day confirmed the platform rollout. Crucially, Oracle did not narrow its other chip partnerships on the same day: it also unveiled plans to offer Google TPU, along with chips from Cerebras Systems and the Intel-backed SambaNova Systems.
Nvidia CEO Jensen Huang has previously said each gigawatt of Vera Rubin deployment could bring Nvidia up to $40 billion in revenue. That figure explains why Oracle keeps loading up on Nvidia — a multi-gigawatt buildout corresponds to tens of billions of dollars in system revenue. Yet by simultaneously opening three new supply lines, Oracle's goal is not to replace Nvidia but to preserve its core platform footprint while giving customers cross-architecture compute options.
TPU leaves Google Cloud for the first time: what OCI as an external hyperscaler signals
Per the joint Oracle-Google release (verified via IT Home), Oracle will become the first hyperscale cloud provider outside Google to offer Google TPU, with the service expected to launch in 2027 and scale over several years, supporting large-model training, inference and agentic workloads. Google's eighth-generation TPU comes in a training-oriented TPU 8t and an inference-oriented TPU 8i. FourWeekMBA, citing the joint press release, called this the first time TPU is offered on a hyperscale cloud platform outside Google.
TPU has long run closed-loop inside Google Cloud. Shipping it to OCI marks Google's move from self-use to external sales. For OCI, TPU offers customers a non-Nvidia training and inference option, appealing especially to model shops already familiar with the Google stack or looking to dilute Nvidia dependence on the training side. For Google, a hyperscale partner is the key channel to scale TPU external revenue — Alphabet said in July that the vast majority of TPU external-sales revenue will be recognized in 2027.
The second inference front: Cerebras and SambaNova fill Nvidia's layout gaps
Reuters reports that Oracle plans to offer chips from both Cerebras Systems and the Intel-backed SambaNova Systems, both focused on AI inference silicon. Cerebras CEO Andrew Feldman told Reuters this is a large-scale expansion of its tier-one customer base, with its products entering other companies' data centers at scale.
Nvidia's GPU dominates training, but inference workloads are more sensitive to cost, latency and throughput, with wider architectural variation. Adding inference-dedicated chips from Cerebras and SambaNova opens a second inference front alongside Nvidia's flagship systems, letting customers pick cheaper hardware for agent calls and batch inference. The pattern of training on Nvidia and routing inference by workload is exactly how leading cloud stacks are being layered.
Self-designed Accelon networking chips: the next step of cloud vertical integration
Per Reuters, even within Nvidia-based systems, Oracle will drop some Nvidia networking chips in favor of its own Accelon networking silicon. OCI President Mahesh Thiagarajan said different AI workloads demand different compute, and choice lets customers match performance, throughput and cost. As AI clusters scale, networking and interconnect become the bottleneck, and self-designed networking chips cut single-vendor dependence while tuning for a provider's own cluster topology.
Oracle targets roughly $90 billion to $95 billion in fiscal 2027 capital expenditure; last week's earnings call noted customers are increasingly covering chip procurement costs themselves, and Oracle will hold its investor day this week. At near-$100 billion capex levels, multi-supplier sourcing, cross-architecture options and in-house networking chips amount to a cloud vendor pursuing bargaining power, supply security and vertical integration at once.