The dominant question in AI investing today is who sells the picks and shovels? It is the wrong question. The right question is who quietly uses AI to permanently re-rate their own unit economics?
Producer-side AI is loud. The pitch deck has 'AI-powered' in the headline. The moat is unclear. The user is hypothetical. The valuation is priced for a market that has not yet decided whether the product is necessary.
Beneficiary-side AI is quiet. It looks like an operator quietly compressing their cost-to-serve by 30%, expanding their addressable market into segments that were previously uneconomic, or breaking a competitor's moat by automating a workflow the competitor still pays humans to do.
The test
Before we own any name on an AI thesis, we run a single check: if you stripped 'AI' out of the company's marketing tomorrow, would the unit economics still get materially better over the next 24 months? If the answer is yes, it is a beneficiary. If the answer is 'we don't know yet,' it is a producer story we are not yet willing to underwrite.
This test eliminates most of the names that dominate retail AI portfolios. It also eliminates a non-trivial portion of the sell-side AI universe. That is the point. The beneficiary set is structurally narrower and structurally more durable.
Why we can see this clearly
We are running our own fund on AI. We feel the cost curve. We feel the latency floor. We feel the difference between a workflow where the model works and one where it doesn't. When a company talks about agentic workloads on an earnings call, we have a calibrated sense of what is real, what is marketing, and what breaks at scale.
That operator-grade tape reading is the asymmetry. Most analysts read about agents. We run them. The difference shows up in which names we are willing to size into.