AMD will split its Instinct MI400-series into separate AI-focused (MI450X) and HPC-focused (MI430X) GPUs in H2 2026 to maximize performance per workload, but limited availability of UALink switches may hinder scalability.

AMD is set to launch its first rack-scale MI450X IF128 AI systems in 2026. The company hopes it will challenge Nvidia's VR200 NVL144 with up to 128 GPUs.

As power consumption of accelerators for supercomputers continues to grow in the coming years, AMD expects zetta scale-class datacenters to consume 500 megawatts, require nuclear reactor.