AI chips diversify beyond GPUs for agents and edge

The article traces AI chip evolution from 2012, when Geoffrey Hinton's lab showed GPUs cut image recognition errors, a moment NVIDIA's CEO calls the 'big bang of AI.' That breakthrough led to deep learning and later generative AI, but training took hundreds of days even with thousands of GPUs. This demand for performance drove innovation beyond NVIDIA, yet NVIDIA's lead persists partly because its CUDA software ecosystem makes its chips easier to use. Now, with agentic AI, the article argues CPUs are better suited for workflow-driven tasks, and with physical AI, edge devices need on-device generative capabilities. This diversity means no single chip type dominates. However, designing advanced AI chips at 2nm is so complex that only a few design houses may handle it. That could limit which companies can produce data-center AI chips, potentially benefiting firms like Broadcom, Marvell, MediaTek, and Japan's Socionext, possibly aiding efforts to revive Japan's semiconductor industry.