Oracle to deploy 50,000 AMD GPUs by 2026 as competition heats up with Nvidia

Oracle

Oracle has partnered with chip manufacturer AMD to build an AI supercluster powered by 50,000 of AMD’s latest Instinct MI450 GPUs, in what marks the cloud provider’s biggest effort yet to challenge Nvidia’s grip on the AI compute market.

Set to go live in the third quarter of 2026, the new Oracle Cloud Infrastructure (OCI) supercluster will make Oracle, in what the company claims, the first hyperscaler to offer a publicly available AI cluster based entirely on AMD’s latest GPU lineup.

The company plans to expand capacity further in 2027 as demand for large-scale model training continues to rise, it announced on October 14, ahead of its flagship AI World event.

The announcement comes at a time when most hyperscalers, including Amazon, Microsoft, and Google, rely heavily on Nvidia hardware and software stacks for AI workloads. Over 90 percent of the world's heavy-duty GPUs are provided by Nvidia.

Oracle’s decision to double down on AMD-based infrastructure reflects growing industry interest in diversifying GPU supply and building open, interoperable systems that are less dependent on Nvidia’s proprietary CUDA ecosystem.

The upcoming AI supercluster will use AMD’s new “Helios” rack design, combining Instinct MI450 GPUs, next-generation EPYC CPUs, and Pensando networking chips.

Oracle said the system is designed for extreme scale and energy efficiency, which assumes importance at a time when training models are increasingly pushing existing AI clusters to their limits.

Mahesh Thiagarajan, executive vice president at Oracle Cloud Infrastructure, said the partnership is aimed at customers developing AI applications that require high-performance infrastructure at scale.

Each MI450 GPU will come with up to 432 GB of HBM4 memory and 20 TB/s of bandwidth, a specification aimed at workloads currently optimised for Nvidia’s powerful and widely used H100 and H200 GPUs.

AMD’s open-source ROCm software stack will also be available on OCI, offering developers an alternative to Nvidia’s CUDA framework.

The company has also been expanding its cloud footprint, with OCI already offering AMD’s MI300X and MI355X GPUs across its Zettascale superclusters.

With demand for compute power surging amid the AI boom, Oracle’s latest bet signals how major cloud providers are racing to secure alternative chip partnerships to keep up with Nvidia’s pace and, increasingly, to compete with it.

(This reporter was in Las Vegas at the invitation of Oracle)