AI Inference Hardware Specialization

AMD acquires Taalas to etch models into silicon for inference performance. SpaceX's $18.4B capex spike, mostly on AI clusters, points to orbital edge computing. Intel releases OpenVINO 2026.3 with expanded generative AI model support (SmolLM3, LFM2, FLUX.2), memory-saving disk offloading for MoE models, EAGLE-3 speculative decoding, and FP8 quantization. These moves signal a shift toward specialized, decentralized AI infrastructure beyond traditional data centers.