Artificial intelligence
Training got the headlines. Inference gets the P&L.
The thesis
Frontier model quality is converging faster than distribution is. Two labs shipping near-identical capability at a two-month lag turns the model layer into a price-taker and pushes durable margin outward — to whoever owns the workflow the model is embedded in, and inward — to whoever owns the power and silicon it runs on.
The uncomfortable middle is the API reseller: a thin wrapper over a commodity endpoint, with no proprietary data loop and no switching cost. That layer reprices first.
Where margin accrues
Compute and power: constrained, contracted years ahead, and priced accordingly. The scarcity is physical, not algorithmic.
Workflow ownership: products where the model writes into a system of record accumulate correction data, and correction data is the only moat that compounds while model prices fall.
Evaluation and trust: as agents take actions rather than emit text, the buyer's question shifts from 'is it smart' to 'what happens when it is wrong'. Whoever answers that credibly sells to the enterprise.
The bear case
If inference cost falls faster than usage grows, the infrastructure buildout is overbuilt and depreciation lands on the wrong balance sheets. And if the frontier stops improving on the axes buyers actually pay for, the workflow layer discovers that its 'moat' was just a UI.
What we'd watch
- Inference price per token versus realised utilisation on new capacity
- Enterprise renewals on agent products at month 12, not month 3
- Power interconnect queues in the major datacenter corridors
- How many agent products still need a human in the approval loop
Pressure-test this note with the desk — the agent pulls live sources before it answers.
Ask the desk