The classic AI hardware battle has centered on GPUs, accelerators, interconnects, and manufacturing scale—Nvidia’s CUDA moat, AMD’s MI-series push, custom ASICs from hyperscalers, and supply-chain fights. But as models evolve from large language models into multimodal, reasoning, agentic, and especially physical/spatial systems, pure chip performance is no longer sufficient.World models demand different compute profiles: heavy simulation, 3D understanding, real-time interaction, and tight coupling between model architecture and underlying hardware/software stacks. AMD explicitly frames the deal as giving it deeper insight into emerging workloads so it can design better hardware, software, and systems around them—and strengthen an open AI ecosystem spanning models, platforms, and compute.

Fei-Fei Li has argued that true progress requires foundation models grounded in the physical world beyond pure text/LLMs.