Overview
- AMD has begun volume production of its Helios 72‑GPU rack and says initial shipments will start late in Q3 2026 with broader rollouts in Q4 2026.
- The company publicly launched 6th‑Gen EPYC 'Venice' server CPUs built on TSMC 2nm and Instinct MI455X accelerators that AMD says offer large HBM4 capacities (up to 432 GB per GPU) and multi‑PFLOP AI compute.
- Major cloud and AI customers have made multi‑gigawatt commitments to AMD hardware, including multi‑gigawatt deals publicized for OpenAI, Meta, Microsoft, Oracle and Anthropic with staged deployments through 2027.
- Samsung is named as a primary supplier of HBM4 for the MI455X and Supermicro will ship validated Helios racks and H15 servers that integrate Venice CPUs and Instinct GPUs at data‑center scale.
- Key near‑term tests include independent performance benchmarks of AMD’s claims, the maturity of ROCm and software tooling versus NVIDIA’s CUDA, and the timing of staggered hyperscaler deployments that will determine revenue and market share shifts.