AI Signal

The industry's center of gravity is RL post-training on verifiable tasks. Reasoning and agentic capability gains now come from post-training curricula more than from pretraining scale, so the binding constraint is shifting from raw compute to the supply of hard, checkable problems. Underneath that, the hardware base is diversifying slowly. Intel's 18A is credible again but still sits one node behind TSMC, so leading-edge AI silicon remains concentrated at TSMC. Enterprises are planning against a capability curve whose slope now depends on how labs manufacture training signal, and that input is less predictable than compute.