AMD has ramped up mass production of its Helios AI racks, which feature 72 Instinct MI455X accelerators per unit. Each accelerator is equipped with 432 GB of HBM4 memory. Production of these systems officially began in July 2026, with volume shipments expected to scale significantly in the fourth quarter of 2026.
The company plans to supply these systems to major clients, including Microsoft, OpenAI, and Anthropic. Due to demand for Helios exceeding initial 2027 forecasts, AMD intends to increase production capacity over the next three to five years.
Samsung, which began HBM4 mass production in February 2026, serves as a preferred supplier for AMD following a memorandum of understanding signed in March 2026. AMD CEO Lisa Su visited Seoul in early October 2026 to discuss securing additional supply.
Despite rising yields at manufacturers like Samsung and SK Hynix, supply constraints for HBM and advanced packaging remain a critical bottleneck. Industry capacity for HBM4 is reportedly fully committed through 2029, while Samsung aims to double its output in 2027.
