AMD launched Helios, its first full rack-scale AI infrastructure system, at its Advancing AI 2026 event, a complete, pre-integrated rack rather than a single chip, combining 72 of AMD's new Instinct MI455X GPUs with sixth-generation EPYC 'Venice' server CPUs, AMD's Pensando networking silicon, and its ROCm software stack into one open-architecture unit. The headline numbers are substantial: 2.9 exaflops of peak AI compute at MXFP4 precision across the rack, with each GPU carrying 432GB of HBM4 memory at 23.3TB per second of bandwidth, adding up to roughly 31TB of pooled high-bandwidth memory across the full system. What makes this significant beyond the spec sheet is the competitive positioning: this is AMD explicitly building an integrated rack-scale product to compete with Nvidia's NVL72-class systems, rather than selling individual GPUs for customers to integrate themselves, which has been Nvidia's structural advantage in the AI infrastructure market for the past several years. AMD says Helios is now in full production and will begin shipping to customers by the end of Q3 2026, and the company has also lined up partners like HPE and Schneider Electric to build reference designs and standardized rack form factors around it, which matters because data center operators need power, cooling, and physical rack specs to be predictable across vendors before they will commit to large deployments. For teams planning AI infrastructure procurement, Helios is the clearest sign yet that the 2026-2027 AI hardware market will have a genuine second rack-scale option beyond Nvidia, which historically has meant better pricing leverage and reduced single-vendor lock-in risk for large buyers, though software portability between the CUDA and ROCm ecosystems remains the practical hurdle for anyone considering a switch.