Oracle Corp. today announced plans to build new artificial intelligence clusters powered by rivals Nvidia Corp. and Advanced Micro Devices Inc. chips.

The first cluster, the OCI Zettascale10, will be hosted in the company’s OCI public cloud. It will enable customers to configure AI environments with up to 800,000 Nvidia graphics processing units. Separately, Oracle is building a 50,000-GPU cluster based on the Instinct MI450, AMD’s upcoming series of flagship AI accelerators.

Other key players in the AI market are also using GPUs from multiple suppliers to avoid overdependence on a single chipmaker. Oracle’s major rivals in the cloud market all offer a mix of Nvidia and AMD GPUs. OpenAI, which has commissioned $300 billion worth of AI infrastructure from the database maker, plans to deploy custom AI chips alongside the off-the-shelf silicon it currently uses.

The architecture that underpins the OCI Zettascale10 powers a data center Oracle is currently building for OpenAI in Abilene, Texas. The offering will support multigigawatt clusters with up to 800,000 Nvidia GPUs. Oracle estimates that OCI Zettascale10 will offer peak AI performance of 16 zettaflops, or 16 trillion billion computations per second.

The company plans to link together the GPUs in OCI Zettascale10 clusters using Nvidia’s Spectrum-X series of Ethernet network equipment. The product family is centered on two devices. The first is the BlueField-3 SuperNIC, a chip that connects GPU servers to a data center’s network and offloads certain computing tasks from their main processors. The chip is joined by a line of Ethernet switches called the Spectrum SN5000.

Oracle’s implementation of the network devices features a technology it calls Acceleron RoCE. Typically, moving data between GPUs requires sending it through the central processing units of the servers that host the GPUs. Acceleron RoCE skips that step to improve performance.

Oracle is currently taking orders for OCI Zettascale10. The offering will become available in the second half of 2026.

“Customers can build, train, and deploy their largest AI models into production using less power per unit of performance and achieving high reliability,” said Mahesh Thiagarajan, executive vice president of Oracle Cloud Infrastructure. “In addition, customers will have the freedom to operate across Oracle’s distributed cloud with strong data and AI sovereignty controls.”

Oracle plans to bring the OCI Zettascale10 online alongside an AI cluster equipped with 50,000 of AMD’s MI450 graphics cards. The GPUs will run in racks based on a new design, Helios, that the chipmaker detailed today.

A Helios rack can host 72 MI450 chips, each of which includes up to 432 gigabytes of HBM4 memory. HBM4 is a high-speed RAM variety that is not yet in mass production. AMD estimates that the technology will enable Helios to offer double the memory capacity and bandwidth of a system equipped with Nvidia’s upcoming Vera Rubin chips.

Helios racks also contain other components. They will incorporate AMD’s upcoming Venice server CPU series and Vulcano, a future addition to its Pensando line of data processing units. The company says that each Helios rack will be capable of providing up to 1.4 exaflops of performance when processing FP8 data.

The rack will use liquid cooling to remove the heat generated by its components. According to AMD, the system is based on a double-width design meant to make it easier for technicians to fix malfunctions. The company will enable hardware partners to extend the core Helios feature set to adapt it for their requirements. 

Oracle plans to install the first MI450-equipped Helios racks in its OCI data centers during the third quarter of 2026. The company will start bringing additional systems online in 2027.

Photo: Oracle