AI Agent UpdateOracle puts Google's TPUs, Nvidia's Vera Rubin and Cerebras on one cloud menu
At Oracle AI World, Oracle said Google's TPUs will run on its cloud from 2027, expanded its Nvidia Vera Rubin deal to multi-gigawatt scale, and added Cerebras and SambaNova — betting that AI customers want a menu, not a single chip.
The News
Oracle used its AI World conference in Las Vegas on October 27 to make the cloud's chip wars a little less tribal. Google's TPUs — the custom AI chips Google designs for its own data centers — will be available through Oracle Cloud Infrastructure (OCI) in calendar 2027, ramping over several years, with Oracle supplying the data-center power, cooling and networking. It is the first time TPUs will be offered on a hyperscale cloud outside Google itself, with the eighth-generation TPU 8t for training AI models and TPU 8i for running them.
In parallel, Oracle and Nvidia will expand OCI Superclusters — giant clusters of AI servers — with Vera Rubin NVL72 systems to multi-gigawatt scale, and Vera Rubin systems and Vera CPUs are now generally available on OCI. A gigawatt is a billion watts, roughly the output of a large power station, so 'multi-gigawatt' means Oracle is planning AI capacity on the scale of a small country's grid. Nvidia CEO Jensen Huang said on Nvidia's fiscal second-quarter 2027 earnings call in August that Vera Rubin systems can generate about $40 billion in revenue per gigawatt for Nvidia, though Oracle disclosed no deal value for this expansion.
Oracle is also diversifying beyond Nvidia: it will add Cerebras inference chips in 2027 and Intel-backed SambaNova systems (timing not specified in the cited report), and will skip some Nvidia networking chips in favour of its own in-house 'Acceleron' chips even inside Nvidia-based systems. Cerebras CEO Andrew Feldman called it 'a massive expansion of our customer base of tier-one customers.' Oracle shares closed up about 3.6% at $130.56 on October 27, and Cerebras closed up 10.1%.
The cloud wars used to be about whose chip you bought. Oracle just made it about whose menu you order from.
Why It Matters
The strategic signal is bigger than any single chip. For years, each cloud pushed its own silicon: Google had TPUs, Amazon had Trainium, Microsoft had Maia, and Nvidia sold to all of them. Oracle, which has no chip of its own to protect, is turning rival silicon into a menu — and the same morning it also announced OpenAI's ChatGPT and Codex embedded in Oracle business software (covered separately today). A customer that can pick TPU, Vera Rubin, Cerebras or SambaNova on one cloud is a customer less likely to leave.
For Nvidia, the read is two-sided. Multi-gigawatt Vera Rubin demand is an unambiguous win — it is the largest-scale commitment yet for its next-generation platform. But Oracle's diversification, including its own networking chips inside Nvidia systems, is the latest sign that inference — the business of running finished AI models — is becoming a multi-vendor market, not a Nvidia-only one. Nvidia keeps the training crown; the inference fight is now open.
For Google, putting TPUs on Oracle's cloud is a distribution breakthrough. TPUs were previously a Google-only advantage; selling them through Oracle turns a cost center into a product line and puts Google silicon in front of Oracle's enterprise customers, many of whom run Oracle databases and applications already.
Analyst's Take
Oracle's pitch is simple: you should not have to rebuild your company around one chipmaker's roadmap. That pitch lands because it is true — customers are tired of single-vendor lock-in, and the performance gaps between top chips are narrowing for many workloads. The missing numbers (no deal values, no chip counts, no launch quarter for TPUs) keep this from being a clean victory lap, but the direction is clear.
The test is execution: whether TPUs actually arrive on OCI in 2027 at competitive prices, and whether Oracle's multi-gigawatt buildout stays on schedule. If both hold, Oracle's 'menu' strategy becomes the template other clouds copy. Bullish, with the usual caveat that announcements are not deliveries.
The other winner here is the idea of competition itself. When four chip suppliers share one cloud's stage, each has to win on price, performance, or power for the next deal — and customers, not vendors, set the terms. That is a healthier market than a single-supplier world, even if it is a messier one to cover.
By The Numbers
FAQ
What is a TPU?
A Tensor Processing Unit is Google's custom chip for AI work — training models (TPU 8t) and running them (TPU 8i). Until now you could only use TPUs on Google's own cloud.
What does 'multi-gigawatt scale' mean?
A gigawatt is a billion watts, about what a large power station produces. Multi-gigawatt AI capacity means data centers drawing power on the scale of a small city's grid, just for AI servers.
Why would Oracle offer rivals' chips?
Oracle has no AI chip of its own to protect, so it can sell customers choice instead of loyalty. If a customer can get Nvidia, Google, Cerebras and SambaNova chips in one place, it has less reason to move to another cloud.
Does this hurt Nvidia?
Partly. The Vera Rubin expansion is a big Nvidia win, but Oracle using its own networking chips and adding Cerebras and SambaNova shows inference work is spreading beyond Nvidia. Training still leans Nvidia; inference is now contested.
Were any deal values announced?
No — no deal values, TPU quantities, customer orders or revenue figures were disclosed, and no launch quarter for TPUs beyond calendar 2027.
👁
Source: Reuters