Hot Chips 2026, the annual academic-industry conference where chip architects present next-generation silicon designs, opened August 23 at Stanford and runs through August 25, with confirmed sessions from Nvidia, AMD, Intel, Google, Microsoft, Meta, OpenAI, IBM, Arm, and Fujitsu. Nvidia used its session to detail the next wave of its data-center accelerator platform: the Vera CPU, Rubin GPU, an updated NVLink Switch, an LPU-based interconnect architecture, Spectrum-X Ethernet networking, and BlueField-4 DPUs, positioned as a full rack-scale system rather than a single chip. AMD presented the architecture behind its Instinct MI400-series accelerators, its next answer to Nvidia in the AI training and inference accelerator market, while Intel walked through Diamond Rapids, its upcoming server CPU generation. For engineers building or buying AI infrastructure, Hot Chips is one of the few venues where vendors go past marketing slides into actual microarchitecture, interconnect topology, memory bandwidth tradeoffs, packaging choices, which is the level of detail that matters for capacity planning, cluster design, and understanding where the next jump in training and inference cost-efficiency is likely to come from. The presence of hyperscalers such as Google, Microsoft, Meta, and OpenAI presenting their own custom AI silicon alongside merchant vendors also underscores how much of the frontier AI compute stack is now split between off-the-shelf GPUs and internally designed accelerators, a split that increasingly shapes what is available to rent on public cloud versus what stays captive to a single company's own workloads.