The inference crisis

NVIDIA's GTC 2026 showcased 100x performance gains for vision AI. Yann LeCun's AMI Labs raised $1.03 billion for "world models." Google's Turing Award winner published on the hardware crisis: the bottleneck in AI has shifted from training to inference — from learning to thinking. The machines can learn; they can't think fast enough. Every platform is shipping smaller, faster models (GPT-5.4 mini/nano, Gemini 3.1 Flash-Lite) because the economics of intelligence now favour speed and cost over raw ...

Neither processes the philosophical inversion: intelligence is now cheap to create and expensive to run. The scarcity has moved from the spark to the metabolism.