Inside Oracle’s push to end single-vendor GPU dependence

Oracle

For the past decade, the market for large-scale AI has largely centered around Nvidia’s hardware and software ecosystem, resulting in an industry where access to GPU capacity, networking choices and system design patterns shaped by a single supplier. Oracle is now trying to break that pattern.

On October 15, the US tech giant announced plans to deploy 50,000 AMD Instinct MI450 GPUs in a new supercluster set to go live in 2026 with capacity expansion planned for 2027.

The aim is not to replace Nvidia, said Oracle, but to shift AI infrastructure from a single-vendor to a workload-driven model. “We strongly believe in the power of ‘let the market decide the winner,’ and which means you need to invite more players to the playing ground,” Sudha Raghavan, who leads cloud engineering for Oracle, told Moneycontrol on the sidelines of the company’s flagship AI World 2025 event.

Why Oracle wants to Widen the AI Market?

With demand for foundation model training continuing to accelerate, cloud providers have been forced to rethink on how they can secure compute power, and how much control they should cede to the vendor that dominates it. Instead of treating non-NVIDIA accelerators as fallback options, Oracle has begun building superclusters with AMD’s Instinct platform at the same scale as its Nvidia deployments.

In Raghavan’s view, constraints such as cloud infrastructure design and procurement of entire GPU clusters need to be dealt with.

NVIDIA has been the sole player, which means everything from supplier contracts to switch manufacturer is picked by the company that manufactures over 90 percent of the world’s GPUs.

Also Read: The trillion-dollar merry-go-round: Inside AI's circular economy

Oracle is not suggesting that customers walk away from Nvidia. Its largest training workloads still run on Nvidia’s H100 and H200 superclusters, and those fleets continue to expand.

Cloud Moving With the Workload

Oracle, which has gone from building databases to becoming a hyperscaler and infra player, said it has designed Oracle Cloud Infrastructure (OCI) so that the same cloud can run in Oracle regions, inside enterprise data centres and within other hyperscale clouds.

“We can also put our cloud inside a customer's firewall,” Chris Chelliah, Oracle’s senior vice president for technology and customer strategy told Moneycontrol.

“You can buy $100 worth of cloud credits, picking an arbitrary number, and you can use 20 of it in Azure, 10 of it in Google, and 70 of it on Oracle. You're drawing from the same set of one contract, one set of terms.”

Analysts argue that most enterprises want the option to shift workloads between Nvidia and AMD for cost reasons, availability reasons or model-stage reasons.

Also Read: Infosys, Cognizant, Accenture, LTIMindtree invest in Oracle’s $1.5 bn AI Data Platform

AMD a Strategic Counterbalance

The timing of Oracle’s AMD expansion is not coincidental. AMD’s Instinct MI300 and MI450 series now target the same high-bandwidth, large-memory workloads that form the backbone of LLM training and inference. With Nvidia capacity stretched across both hyperscalers and model labs, any credible alternative assumes significance.

OpenAI Vouches

OpenAI’s infrastructure team too has highlighted this shift. Peter Hoeschele, who leads infrastructure at OpenAI, said the lab worked with Oracle when it needed to expand on AMD systems.

“When AMD and Lisa Su were saying they really want to push the next generation of AMD, we went to Oracle, and Oracle said, we will do anything to make this work with you,” he said during an AI World session.

Shift from Vendor Logic to Workload Logic

Oracle is betting that as AI workloads diversify, choice will become a form of resilience against pricing swings, supply shortages and architectural lock-in. This view reflects Raghavan’s description of GPU selection. “This is not an if-else question, it’s a matrix question.”

Nvidia remains central to the AI market and Oracle is not trying to refute that, however, it is trying to ensure that Nvidia does not remain the only plan.