AMD takes on Nvidia with its Helios AI rack scale system

AMD chief executive Lisa Su walked through the first volume production details of the company's Helios rack scale AI system, according to TechCrunch, combining Instinct MI455X GPUs, sixth generation EPYC Venice processors, Pensando networking, and AMD's ROCm software into a single integrated platform priced between 5 million and 5.5 million dollars per rack. Each Helios rack targets up to 2.9 exaflops of FP4 compute with 31 terabytes of HBM4 memory, roughly 50 percent more total memory than Nvidia's competing Vera Rubin system offers, a specification AMD is emphasizing specifically because memory capacity has become an increasingly important bottleneck for the largest AI models. Engineering samples are shipping in the second half of 2026, with full mass production targeted for the second quarter of 2027, and AMD confirmed that OpenAI and Meta have together committed to a combined 12 gigawatts of AMD accelerator capacity, alongside Microsoft's previously announced deployment of Helios racks on Azure. The scale of the OpenAI and Meta commitments in particular signals that AMD's rack scale push is attracting genuine capacity commitments from the largest AI buyers rather than only smaller pilot deployments, a meaningful validation given how thoroughly Nvidia has dominated large scale AI infrastructure purchasing to date. The announcement continues a pattern this year of AMD steadily converting technical announcements into concrete customer commitments, gradually building a credible case that the AI infrastructure market may be moving toward genuine two vendor competition rather than remaining a near monopoly.