Oracle is doubling down on Nvidia's next-generation AI hardware while simultaneously broadening its chip lineup to include rivals, a strategy that could reshape how the cloud provider negotiates with suppliers and serves its customers.
The enterprise software giant, which also runs Oracle Cloud Infrastructure (OCI), said it plans to deploy “multi-gigawatt” capacity using Nvidia's upcoming Vera Rubin systems. That's a huge scale-up—gigawatt-level data centers are rare and typically power entire cities. Nvidia CEO Jensen Huang has said Vera Rubin systems could generate up to $40 billion in revenue per gigawatt, underscoring the enormous financial stakes as AI infrastructure spending accelerates.
But Oracle isn't putting all its eggs in one basket. For the first time, it will offer Google's tensor processing units (TPUs), along with chips from Intel-backed startup SambaNova and AI hardware company Cerebras. These additions focus on inference—the process where trained AI models answer user queries—which is generally easier to shift between different chip types than training, the computationally heavy phase of building models.
This flexibility means customers can shop around for the best combination of speed and cost for their specific workloads. For Oracle, it also provides significant negotiating leverage with Nvidia, as it can point to alternatives when discussing pricing and supply terms.
Taking control of the stack
Oracle's move to use its own “Acceleron” networking in some Nvidia-based systems is another signal of its intent to control more of the technology stack. Networking is often where vendors add margin by bundling components, so by owning that layer, Oracle can capture more of the profit pool inside its cloud infrastructure.
This is a familiar pattern in the industry. As companies find different ways to fund AI buildouts, the battle for control over hardware and software components intensifies. Oracle's approach mirrors that of other hyperscalers, which have increasingly designed custom chips and networking gear to reduce dependence on dominant suppliers.
The company reiterated its plan to spend roughly $90–95 billion on capital expenditures in fiscal 2027, a massive outlay that reflects the scale of its AI ambitions. Notably, Oracle says customers are increasingly footing the bill for the chips, which helps offset the heavy upfront costs.
What it means for investors
For investors, Oracle's strategy is a double-edged sword. On one hand, the multi-gigawatt Vera Rubin buildout signals strong demand for AI infrastructure, which could drive revenue growth. On the other, the addition of multiple chip suppliers suggests that Oracle is preparing for a future where no single vendor dominates the AI hardware market.
“The more inference workloads can move between chip types, the harder it is for any one vendor to charge for the whole system,” said one industry analyst. “Oracle's choice to use its own networking is another clue: networking is where vendors often add margin. If Oracle controls that layer, Nvidia may capture less of the end-to-end profit pool.”
This shift in bargaining power could have broader implications for the AI supply chain. As Nvidia partners with other firms on chip design, and as companies like Tesla and Nvidia back Samsung's expansion, the competitive landscape is clearly evolving.
For everyday investors, the key takeaway is that Oracle is positioning itself to benefit from AI growth while hedging against over-reliance on any single chipmaker. That could mean more stable margins over time, but it also means the company's fortunes are tied to its ability to execute on this complex, multi-supplier strategy.
As Meta's similar $95B funding plans show, the AI infrastructure race is capital-intensive, and companies are exploring various ways to finance it. Oracle's approach—letting customers pay for chips—could be a smart way to manage that burden.
Ultimately, Oracle's success will depend on whether it can earn a bigger take from networking and software services even if inference hardware prices compress over time. That's a bet on the value of the full stack, not just the processor.
