The week's apparent contradiction is the point. At the model layer, capability became cheaper and more portable: Moonshot's Kimi K3 arrived as a 2.8-trillion-parameter sparse model with aggressive API pricing, near-frontier benchmarks and a promised open-weights release, while sub-1-bit and ternary quantization pushed server-class checkpoints toward phones, vehicles and industrial endpoints. Neither is yet a kilowatt-hour measurement. Cheaper tokens redistribute inference among hyperscalers, neoclouds, sovereign clusters and enterprises; on the available evidence, they do not reduce aggregate electricity demand.
:quality(60))
:quality(60))