AMD launches next-gen AI rack Helios with Samsung HBM4, challenging Nvidia

AMD, the world's second-largest AI accelerator company after Nvidia, has officially launched mass production of Helios, its next-generation AI infrastructure platform powered by Samsung Electronics' sixth-generation high-bandwidth memory, HBM4.

AMD said Helios outperforms Nvidia's AI superchip Vera Rubin and that demand from customers including OpenAI and Anthropic is already strong. The company signaled it will ramp up shipments in earnest from the fourth quarter, a development expected to further boost earnings at Samsung Electronics, which is supplying HBM4 to AMD on a priority basis.

AMD CEO Lisa Su took the stage Thursday (local time) at the annual "AMD Advancing AI 2026" event at Moscone West in San Francisco, California, where she unveiled the new AI accelerator Instinct MI455X and the Helios rack system built around it.

Su described the MI455X as "the engine of the Helios system and the highest-performing GPU in the industry," adding that each chip carries 432 gigabytes of HBM4.

Samsung Electronics' HBM4 12-layer product — the first in the industry to enter mass production and ship, in February — has a capacity of 24 to 36 gigabytes per unit, meaning the MI455X accommodates at least 12 HBM4 stacks. That is 50% more HBM than Nvidia's new GPU Rubin, which carries 288 gigabytes.

Helios is an AI server rack that integrates 72 MI455X GPUs, 18 units of the new Venice CPU, a power supply unit and a cooling system into a single large cabinet. The rack carries 31 terabytes of HBM4.

As AI computing grows increasingly complex, server racks have emerged as the core infrastructure underpinning AI data centers — pushing the rivalry between Nvidia and AMD beyond individual chips to the rack level.

Nvidia recently began mass production of its Vera Rubin rack, which pairs its Vera CPU with its Rubin GPU. AMD is now firing back with Helios, entering the supply race with big tech companies in its sights.

"Compared with the competing product — Vera Rubin — Helios delivers 15% higher computing performance, 50% more HBM4 capacity and 6% greater HBM4 bandwidth," Su said. "That means we have the capacity to handle longer context windows and the bandwidth needed to scale to thousands of racks."

She added that "Helios entered full mass production today," and that "shipments are scheduled to begin at the end of the third quarter, with volumes ramping up significantly from the fourth quarter."

At the event, Su brought executives from Anthropic, OpenAI and Meta onto the stage in succession to showcase an anti-Nvidia partnership coalition against the chipmaker's dominance. OpenAI plans to begin deploying Helios in the fourth quarter and expand its rollout in earnest next year, while Meta is currently testing the system ahead of a large-scale adoption.

The growing adoption of Helios among these companies is also good news for Samsung Electronics. Samsung Electronics signed an MOU with Su when she visited its Pyeongtaek campus in March and has since been supplying HBM4 to AMD's MI455X accelerator on a priority basis.

AMD has maintained a close partnership with Samsung Electronics, continuing to take Samsung's HBM supply even during the period when Samsung failed Nvidia's HBM3E qualification tests and was shut out of that supply chain.

Samsung Electronics recently began supplying HBM4 for Nvidia's Vera Rubin, entering a period of rapid growth. With AMD now pushing to expand sales of Helios — which requires 31 terabytes of HBM4 per rack — analysts expect the HBM supply shortage to deepen further.

AMD also announced that its next-generation MI500 GPU series is slated for 2027 and the MI600 series for 2028. Nvidia has its own roadmap, with Rubin Ultra in 2027 and Feynman in 2028. Samsung Electronics' supply of HBM4E and HBM5 is expected to take shape in line with both companies' schedules.