AMD

AMD's late-2026 coverage is led by data center AI hardware. At Hot Chips 2026 the company gave two talks on the Instinct MI400, moving the XCDs to TSMC N2 and the IODs to N3P, splitting Infinity Cache into a separate FCD, and carrying 12 HBM4 stacks. The Helios rack adds 72 lanes of UALoE alongside PCIe, with PCIe, UALink and UALoE switched at the protocol level. On September 28 AMD agreed to buy World Labs for about $8.2 billion in stock. That deal buys model researchers rather than another chip line: World Labs builds spatial intelligence models, and Fei-Fei Li becomes an executive vice president and chief scientist reporting to Lisa Su.

Software is moving toward shared or consolidated paths. AMD and Intel said they will fold PyTorch x86 optimization into oneDNN, with AMD integrating ZenDNN into that library and joining the Unified Acceleration Foundation as a steering member. The two are also working on AI Compute Extensions, an x86 instruction set aimed at matrix multiplication. On September 30 AMD released AMD Ross, an agentic assistant for embedded development that connects to Vivado, Vitis, ROCm and Ryzen AI Software. At ECS 2026 in August, AMD put a Unified Memory architecture shared by NPU, CPU and GPU at the center of its physical AI pitch.

Consumer graphics is less settled. Adrenalin 26.10.1 on October 20 brought FSR Upscaling 4.1 to RX 6000 series cards and to Ryzen AI Max 400 and 300 and Ryzen AI 9 400 and 300 integrated GPUs, well ahead of the January 2027 window AMD had given. Microsoft and AMD said FSR 4.1 is headed to Xbox Series X, Xbox Series S and the Xbox Ally X; the standard Xbox Ally does not support it. The RX 9050 reached Japan on July 31 at a street price around 55,000 yen, above the 54,000 yen cheapest RX 9060 XT 8GB, an inversion the review attributes to exchange rates and rising costs. A community project for running CUDA software on AMD Windows GPUs lists only the RX 9060 XT as verified.