Oracle Invites Nvidia's Rivals Into Its Own AI Cloud
By Daniel Rao | Las Vegas, 27 October 2026
Oracle has just done the one thing a loyal Nvidia customer is not supposed to do in public: it has invited Nvidia's rivals into its own cloud, given them a stage in Las Vegas, and then — in the same breath — promised Nvidia an even bigger deal. That is not a breakup. It is something more interesting, and more revealing about where the AI chip war is heading. On Tuesday, at its AI World event in Las Vegas, Oracle said it would expand its cloud AI chip work with Nvidia to what it called "multi-gigawatt scale" using Nvidia's newest Vera Rubin AI servers, while also beginning to offer chips from Google, Cerebras Systems and SambaNova Systems on Oracle Cloud Infrastructure for the first time. Nobody at Oracle put a price on the expanded Nvidia work. That silence matters, because it is the gap between a headline and a contract — and because, in this market, the absence of a number is itself a kind of information.
Read the announcements together rather than one by one and a pattern snaps into focus. Oracle is not walking away from Nvidia. It is building an escape hatch, a second shelf, and a third option, all while keeping the main supplier close enough that no customer has to worry about supply. For a cloud provider whose pitch has increasingly been "we will get you the chips, whatever it takes," that is a commercially sensible story. It is also, if you squint, a confession: even one of Nvidia's most committed cloud partners now thinks a single-supplier AI cloud is a risk it can no longer afford to present as a strategy.
A bigger Nvidia deal — with a catch
Start with the part Oracle wanted everyone to hear first. Oracle and Nvidia will expand to multi-gigawatt scale with Nvidia's newest Vera Rubin AI servers, according to Reuters, which reported the announcements from San Francisco on 27 October 2026. The value of the expansion was not disclosed. That is worth repeating plainly, because gigawatts sound like dollars and they are not. A gigawatt is a measure of power, not of revenue, and Oracle did not say how much it will pay, over how long, or for how many systems.
For scale, and only for scale, Reuters noted that Nvidia chief executive Jensen Huang has previously said Vera Rubin systems can generate up to $40 billion in revenue per gigawatt for the chipmaker. That is Huang's earlier framing of what the platform can be worth to Nvidia — not a price tag for this Oracle expansion, and it should not be quoted as one. We checked both reports used for this article for a deal value and found none. Anyone who tells you what this expansion is "worth" today is guessing, however confidently they say it.
What is concrete, according to MarketFlux's account of the AI World announcements, is availability. Vera Rubin NVL72 systems and Vera CPUs are now generally available on Oracle Cloud Infrastructure. Oracle also plans HGX Rubin NVL8 systems — eight Rubin GPUs in an x86 server — a configuration that matters because it fits the way many enterprises already buy and run servers, rather than demanding they rebuild everything around a single exotic rack design. The message to existing Oracle customers is deliberately boring in the best sense: the newest Nvidia hardware, in forms you can actually deploy, without a forklift upgrade to your whole worldview.
And then comes the catch, tucked inside the Nvidia systems themselves. Even the Nvidia-based systems will skip some Nvidia networking chips in favour of Oracle's own in-house "Acceleron" networking chips, Reuters reported. Networking is the unglamorous plumbing that lets thousands of GPUs act like one machine rather than thousands of expensive space heaters. Owning that plumbing is not a side project. It is leverage. We will come back to Acceleron, because it may be the most consequential sentence in the whole announcement — and the easiest to skim past.
The Google first: TPUs leave home
The genuinely historic piece of news is the Google one. Oracle will begin offering Google's tensor processing unit (TPU) chips for the first time, Reuters reported, and MarketFlux reported that the offering starts in 2027, making Oracle the first hyperscaler outside Google to offer them. Pause on that for a moment. Google built TPUs for itself, for its own search, its own models, its own cloud. Selling that silicon, in effect, into a rival cloud is a break with more than a decade of habit.
MarketFlux reported that the lineup includes TPU 8t for training and TPU 8i for inference — training being the enormously expensive process of building a model, inference being the everyday work of answering users once the model exists. That split matters, and Oracle knows it. A customer training a frontier model and a customer running a chatbot at scale do not need the same machine, the same price, or the same trade-off between speed and cost. Oracle's pitch is that it will stop forcing them to pretend otherwise.
OCI president Mahesh Thiagarajan said in a statement: "Different AI workloads have different compute needs. That choice gives (customers) the flexibility to choose the right infrastructure for the performance, throughput and economics of each AI workload."
It is a corporate sentence, yes, but strip out the polish and it is also a direct challenge to the old cloud model, in which the customer took whatever chip the cloud had in bulk and shaped their workload around it. Thiagarajan is saying the shaping should run the other way. Whether Oracle can deliver that flexibility in practice — with real availability, real pricing, and real support for workloads that were built for somebody else's silicon — is a question no press release can answer. It will be answered in 2027, in procurement meetings, when engineers discover whether a TPU workload ports cleanly or whether "choice" turns out to mean three different toolchains, three different quirks, and three different support queues.
There is a second, quieter implication for Google. If TPUs succeed inside Oracle's cloud, Google is no longer just a cloud competitor to Oracle; it is also a chip supplier to Oracle. That is a strange, modern kind of rivalry — competing for the same enterprise customer on Monday and cashing a supplier cheque from the same rival on Tuesday. Technology has always produced these tangles, but AI has made them the norm rather than the exception. Everybody is everybody else's customer now, and everybody is everybody else's threat.
Inference is where the fight actually is
If training built Nvidia's empire, inference may decide who gets to keep a piece of it. Reuters was explicit on this point: Nvidia dominates training, while inference — when AI systems respond to users — is more competitive and more diverse. That is why the Cerebras and SambaNova announcements, which might look like footnotes next to the Nvidia and Google headlines, deserve more attention than they will probably get.
Oracle plans to offer chips from Intel-backed SambaNova Systems and Cerebras Systems, both making chips for inference, Reuters reported. MarketFlux added the timing and the shape: Cerebras high-speed inference comes to Oracle Cloud Infrastructure in 2027, while SambaNova's SambaRack SN50 inference chips are being deployed directly in Oracle's data centres. Note the difference in tense. One is a dated plan; the other reads as deployment under way. Neither report, as far as we could verify, named the first customers for these new chips, and we should not pretend otherwise. The customer list is the missing paragraph in this story, and until Oracle fills it in, the inference push is a capability announcement rather than a proven customer migration.
Still, the logic is easy to follow, and it is the logic reshaping the whole chip market. Training is a small number of gigantic, lumpy purchases by a small number of very rich companies. Inference is everybody else: every chatbot reply, every search answer, every agent action, every time a model does something useful for a normal user. It is high-volume, cost-sensitive, latency-obsessed work, and it rewards a chip that does one thing very fast and very cheaply rather than a chip that does everything brilliantly. That is a market where a specialist can win a workload even if it could never win the whole data centre.
Cerebras chief executive Andrew Feldman told Reuters: "That's a massive expansion of our customer base of tier-one customers. We're being deployed in other people's data centers at large scale."
Feldman is not a neutral observer, obviously — he is selling his own company's breakthrough — but the sentence captures what is at stake for the smaller chipmakers. A place inside Oracle's cloud is not just revenue. It is distribution, credibility, and a reference customer that every other enterprise buyer will ask about. For years, the hardest problem for a challenger chip company was not building a good chip; it was getting anyone with a large data centre to actually run it at scale. Oracle just offered three of them that chance at once.
Intel, separately, announced an expanded collaboration with Oracle for enterprise and agentic AI on Oracle Cloud Infrastructure, MarketFlux reported. Agentic AI — systems that do not just answer questions but carry out multi-step tasks — is exactly the kind of workload that hammers inference infrastructure, because an agent may call a model dozens of times to complete one job. If agents become the default way enterprises use AI, inference stops being the cheap afterthought of the AI stack and becomes the bill. Oracle is positioning for that world, and it is doing so with a bench, not a single star player.
Acceleron: the quietest, sharpest move
Now back to Acceleron, because this is where Oracle's strategy stops looking like simple diversification and starts looking like something bolder. Building your own networking chips to sit inside your biggest supplier's systems is a polite way of saying: we love your GPUs, we are less sure we need your everything else.
Networking inside an AI cluster is not decorative. When a model is spread across hundreds or thousands of GPUs, the speed at which those chips talk to each other can matter as much as the speed of any single chip. A bottleneck in the network turns a supercomputer into a traffic jam. By substituting its own Acceleron chips for some Nvidia networking in Nvidia-based systems, Oracle is claiming a piece of that critical path for itself — gaining engineering control, potentially cost control, and certainly negotiating leverage for the next round of talks with Nvidia.
We could not verify from these reports exactly which Nvidia networking components Acceleron replaces, in which configurations, or at what scale. That detail matters, and Oracle has not, in the reporting we relied on, spelled it out. A reader should hold two ideas at once: Acceleron could be a meaningful step towards Oracle controlling more of its own AI stack, or it could be a limited substitution in specific systems that leaves Nvidia's core networking position largely intact. The truth is probably somewhere in between, and it will emerge system by system, not in a single announcement. What is not in doubt is the direction of travel. Clouds that once bought whole stacks now want to own the connective tissue between the parts.
There is a precedent worth remembering, and it is one TechPulse USA readers will recognise. Oracle's cloud ambitions have never been modest, but they have not always been smooth either. Our earlier reporting on Oracle's $18 billion New Mexico AI data centre hitting the wall showed how brutally physical this boom can get: power, permits, and construction timelines do not care about keynote slides. This week's story is different — it is about supplier diversification, not a data-centre build — and it should not be conflated with that New Mexico project. But the two stories rhyme in one useful way: both are about Oracle trying to remove a single point of failure, whether that failure is a stalled site or a single chip supplier.
What this does — and does not — change
It is tempting, on a day like this, to write that the Nvidia era is over. It is not, and this announcement does not say that it is. The largest commitment in the package is still to Nvidia, at multi-gigawatt scale, on Nvidia's newest platform. The rival chips arrive in 2027, in Oracle's telling, not tomorrow. The market reaction reported by MarketFlux — Oracle shares up roughly 4–5% on the day, trading around $132 — should be treated with the caution that figure deserves: MarketFlux's read of the day's trading, not verified market data, and a single day's move is a mood, not a verdict.
What has changed, genuinely, is the shape of choice inside one major cloud. A customer walking into Oracle Cloud Infrastructure in 2027, if these plans land as described, will be able to pick Nvidia for training, Google TPUs for training or inference, and Cerebras or SambaNova for fast inference — and will do so on infrastructure partly stitched together by Oracle's own networking. That is a menu, not a monopoly. Menus create competition on price, on performance per watt, and on the unsexy things that actually decide enterprise deals: support, tooling, and whether the thing works on a Tuesday afternoon when something breaks.
The competitive ripples go beyond Oracle. Other clouds will be asked, loudly, why they do not offer the same breadth. Chipmakers that have waited years for a tier-one cloud shop window will point to Oracle and ask for their meeting. And Nvidia, which has built one of the great business moats in modern technology on the combination of chips, networking, and software, will have to decide how much of that bundle it is willing to unbundle before customers start unbundling it themselves. Nvidia's software ecosystem remains a formidable advantage — engineers know it, tools are built for it, and switching costs are real — but an advantage is not a guarantee, and Oracle just gave its customers a low-friction way to start learning alternatives.
For readers following the wider chip race, the context is worth keeping in view. Our earlier coverage of Nvidia's push to ship the AI factory itself captured the other side of this bargain: Nvidia wants to sell not just chips but whole systems. Oracle's answer, this week, is: fine, we will buy the factory — and we will also stock other people's machines in the same warehouse, and wire some of it ourselves. Similarly, our reporting on Cerebras's 100MW Gimlet deal showed that the inference challengers were already finding large-scale homes before Oracle came calling. AI World did not create this challenger wave. It gave it a much bigger stage.
Our judgment: diversification, or dependence-management?
So which is it — a genuine diversification, or dependence-management dressed up as diversification? Our judgment is that it is both, and that the second description is the more honest one. Oracle has diversified its shelf, but it has not diversified away its dependence. The multi-gigawatt Nvidia expansion is the anchor of the whole announcement; without it, the rival additions would read as a defection, and Oracle clearly did not want that reading. With it, they read as insurance. Insurance is valuable. Insurance is also, by definition, something you buy because you expect you might need it, not because you have already replaced the thing you are insuring.
That is not a criticism, exactly. It is how mature buyers behave in a market dominated by one supplier. They do not storm out; they quietly build alternatives, learn the tooling, cultivate second sources, and make sure that the next negotiation starts from a position of choice rather than supplication. Oracle has done that work in public this week, which is itself a signal to Nvidia and to every other supplier: the next deal will be contested, workload by workload, and the winner will have to earn each one. If that sounds colossally less dramatic than a breakup, good — the chip war was never going to be won in a single keynote. It will be won, if it is won at all, in the slow accumulation of workloads that quietly stop defaulting to one supplier.
There is also a shoddy temptation to avoid here, and it applies to coverage of this story generally: do not mistake an announcement for an outcome. No deal value was disclosed. No deployment schedule was given beyond 2027 for the Google and Cerebras offerings. No customer names were announced for the new chips in these reports. Until those gaps close — with purchase orders, with named customers, with systems actually serving live traffic at scale — this is a strategy statement, not a market share shift. Strategies matter, but they are not results, and a publication that confuses the two will end up celebrating a roadmap as if it were a road.
Watch three things next. First, whether the 2027 dates hold — in AI infrastructure, a year is a long time, and a slip would tell you more than the original promise did. Second, whether Oracle names real customers for TPUs, Cerebras, or SambaNova, rather than speaking in generalities about choice and flexibility. Third, how far Acceleron actually goes inside Nvidia systems. If Oracle's own networking spreads, that is a structural shift; if it stays narrow, it is a negotiating tactic. All three are verifiable, all three are unanswered today, and all three are more informative than today's share-price wobble.
For now, the honest summary is this: Oracle has made its cloud harder to leave, easier to choose within, and slightly less comfortable for Nvidia — all at the same time. That is a neat trick, and a very Oracle trick. Whether it becomes a turning point depends on what happens after the Las Vegas stage lights go out, when engineers, not executives, decide which chips actually earn their workloads.
Daniel Rao covers enterprise AI and cloud infrastructure for TechPulse USA, with a focus on the chipmakers, cloud providers, and power constraints shaping the AI buildout. He is based in the US and writes about the business behind the models — who sells the silicon, who buys it, and what it costs when the keynote ends.
Oracle's AI World chip moves, answered
Is Oracle dropping Nvidia? No. The largest commitment in the package is still to Nvidia — a multi-gigawatt expansion on the Vera Rubin platform. The rival chips are additions, not a replacement.
When do Google's TPUs and Cerebras chips arrive on Oracle Cloud? In 2027, per MarketFlux's account of the AI World announcements — the Google TPU lineup (TPU 8t for training, TPU 8i for inference) and Cerebras high-speed inference are both dated 2027, not available at announcement.
What is Acceleron? Oracle's own in-house networking chips, which will replace some Nvidia networking chips even inside Nvidia-based systems. Exactly which components, in which configurations, and at what scale was not spelled out in the reports relied on.
How much is the expanded Nvidia deal worth? Not disclosed — no deal value was given for the expansion. The $40 billion-per-gigawatt figure is Nvidia CEO Jensen Huang's previous framing of what Vera Rubin systems can generate for Nvidia, not a price tag for this Oracle deal.
Who are the first customers for the new chips? None were named in these reports. Until Oracle names real customers, the inference push is a capability announcement rather than a proven customer migration.
Sources: Reuters (Stephen Nellis, San Francisco, 27 October 2026), reporting Oracle's expanded Nvidia deal, first-time Google TPU offering, Cerebras and SambaNova plans, Acceleron networking substitution, and executive statements; MarketFlux (27 October 2026, Oracle AI World, Las Vegas), reporting Vera Rubin NVL72 and Vera CPU general availability, planned HGX Rubin NVL8 systems, 2027 timing and TPU lineup, Cerebras and SambaNova deployment details, Intel's expanded Oracle collaboration, and the day's share-price movement. Where the two reports differ in emphasis, this article attributes each claim to its source. No deal value, customer names, or deployment schedule beyond 2027 for Google and Cerebras were disclosed in these reports.
Comments
Post a Comment