AMD used Advancing AI 2026 to introduce the MI455X accelerator, a 72-GPU Helios rack and ROCm.ai software tools, framing the launch as a broader bid to narrow Nvidia’s lead in AI systems.
MI455X is positioned as AMD’s most advanced AI chip yet, with 2nm compute tiles, 432GB of HBM4 and 20 PFLOPS of FP8 performance, while Helios links 72 GPUs through 12 Broadcom Tomahawk 6 switches.
AMD also highlighted ecosystem momentum: Anthropic has publicly committed to deploy 2GW of AMD chips, Microsoft has reversed course to adopt MI455X Helios, and AMD is pairing with Cerebras on disaggregated inference.
The report says AMD’s software quality and upstream work in ROCm, vLLM and SGLang have improved sharply, but stable internal GPU clusters and CI capacity remain a bottleneck, delaying some parity goals to as late as October 2026.
Helios still faces execution risks in production—slow rack ramp, heavy use of retimers on up to 85% of the backplane, and reliability issues—underscoring that AMD’s challenge to Nvidia now depends on system delivery as much as silicon.