Inference, not training, sets the economics

Most of the capital, and most of the headlines, still attach to training: the size of the next model, the cluster that trained it, the cost of the run. That number is large and visible. It is also a one-time cost, and it is not where the economics of the build-out will be decided.

Inference is the recurring cost. Every query, every agent step, every generated token is paid for again and again, and demand for it compounds with adoption rather than with the release schedule. A business that serves a model pays for inference the way a utility pays for fuel: continuously, at scale, with little patience for waste.

That changes what matters. The binding constraints become memory bandwidth, interconnect and power per useful token, not peak training throughput. Hardware that is efficient at high-utilisation serving earns its keep; hardware built only for the training headline does not. The same logic favours the parts of the stack whose revenue recurs over the parts whose revenue arrives once.

We position the firm's own book accordingly: weighted toward the recurring, inference-driven layer, and sceptical of valuations that rest on the training cycle alone. This is how we read it today. It is informational, not advice, and we will revise it if the evidence does.