AMD’s MI455X Helios production ramp is slowing as backplane reliability problems emerge in rack bring-up, adding strain to its first rack-scale AI system.
Up to 85% of the backplane must be retimed because of weak 200G SerDes performance, requiring more than 550 Broadcom Ethernet retimers per rack and complicating assembly and serviceability.
Internal software progress has also regressed because stable GPU clusters for development and CI remain scarce; AMD missed its Advancing AI 2026 target for open-source vLLM and SGLang nightly automated CI on MI455X.
The report says vLLM gating had been nearing 90% CUDA parity before cluster capacity was reassigned, with the MI455X software timeline now delayed to October 2026, though AMD is still trying to pull it back to August or September.
Despite those execution risks, the report argues AMD’s silicon and software trajectory has improved enough to give it a credible shot at taking AI accelerator share if Helios manufacturing and internal cluster capacity stabilize.