AMD ramps up production of HBM4-based Helios AI systems

AMD is putting its foot on the gas for its next generation of AI hardware. The company is increasing production of systems built around HBM4, the newest generation of high-bandwidth memory. Inside the Helios ramp Each Helios rack packs 72 Instinct MI455X accelerators. Every one of those chips carries 432 GB of HBM4 memory. The math works out to 31,104 GB of HBM4 in a single rack. For AI workloads, memory bandwidth matters enormously. Large models constantly shuttle huge amounts of data, and a fast chip starved of memory bandwidth mostly sits around waiting. Mass production of Helios began in July 2026. Initial shipments were scheduled for the end of Q3 2026, with significant volume increases expected in Q4 2026. Volume shipments are expected to serve major customers including OpenAI, Anthropic, and Microsoft. AMD has also found that demand outran its own projections. Interest in Helios has far exceeded the company’s initial forecasts for 2027, prompting AMD to plan substantial production increases over the next three to five years. Samsung takes the priority supplier seat Samsung began HBM4 mass production in February 2026. A month later, in March 2026, the two companies signed a m...