Oracle wants to be the cloud where you can rent any AI chip you like. That’s a notable shift for a company whose AI buildout has been tied so closely to Nvidia hardware, and it says a lot about where the economics of AI infrastructure are heading.
Oracle announced on Tuesday that it is expanding its Nvidia relationship to “multi-gigawatt scale” with the new Vera Rubin servers. Neither company gave a dollar figure. But Jensen Huang has said Vera Rubin systems can bring Nvidia up to $40 billion in revenue per gigawatt, so even a couple of gigawatts implies a very large check.
The more interesting part is what Oracle is adding around Nvidia. It will offer Google’s TPUs for the first time, plus inference chips from Cerebras and Intel-backed SambaNova. Even the Nvidia systems will skip some of Nvidia’s networking silicon in favor of Oracle’s in-house Acceleron networking chips.
Why inference? Nvidia dominates training, where its software ecosystem is a deep moat. Inference, where a model answers a user request, is a more open fight. Workloads are repetitive, cost per token matters a lot, and specialized chips can win on speed or price. Oracle Cloud Infrastructure president Mahesh Thiagarajan framed it plainly: different workloads have different compute needs, and customers should pick the right mix of performance, throughput and economics.
This puts Oracle in line with the hyperscalers. AWS pushes Trainium and Inferentia, Google pushes its own TPUs, and Microsoft has Maia. Oracle doesn’t have a large chip program of its own, so it is playing aggregator instead, and that could appeal to AI labs and startups trying to avoid a single supplier and a single price list.
For Cerebras, the win is bigger than it looks. CEO Andrew Feldman told Reuters the deal is “a massive expansion of our customer base of tier-one customers,” and that Cerebras is being deployed in other people’s data centers at large scale. Until now, much of its business has run through its own facilities. Oracle’s footprint changes that.
The open question is money. Oracle didn’t say how it will pay for all this hardware. On last month’s earnings call it said customers are increasingly footing the bill for chips, and it held its fiscal 2027 capital spending target at roughly $90 billion to $95 billion. That’s a huge number, and investors will want details at this week’s investor day.
So does it matter? Yes, with a caveat. Nvidia still wins the biggest share of the spend, and the Vera Rubin commitment proves it. But the multi-vendor menu is a real signal that inference is where Nvidia’s grip loosens first, and Oracle is positioning to profit from that.
