The AI models layer

This layer is the demand engine for the whole chain. Pretraining consumes the largest clusters; reasoning and multimodal models push context windows and compute per query higher; and inference serving — running the models for real users — is where cost per token, the single most important unit economic in AI, is set. When serving efficiency improves, the same hardware does more work; when it stalls, demand for accelerators, memory and power climbs. The agent stack and evaluation harnesses sit here too, shaping how much compute a given task actually needs. For investors, this layer matters less for its own listed names — many of the frontier labs are private — than for the direction it points the rest of the chain: the model layer decides whether the squeeze downstream tightens or eases. Read it as the throttle on demand for everything in the energy, chips and infrastructure layers below.