Every few years a technology arrives that resets the demand curve for computing. The internet did it. Mobile did it. AI is doing it now — and the most underappreciated part of the story isn’t the models, it’s the infrastructure required to train and run them.
The demand side is structural, not cyclical
Training frontier models is only the visible tip. The larger, more durable demand comes from inference — running these models billions of times a day inside products people actually use. Inference scales with adoption, and adoption is still early. That’s a multi-year tailwind, which is exactly where the Ovatek Lens tells me to start.
The companies that win won’t just sell chips. They’ll own a position in the value chain that’s painful to replace.
Where the structural edge lives
I’m most interested in the chokepoints: advanced accelerators, the networking that ties thousands of them together, the memory bandwidth that feeds them, and — increasingly — the power and cooling that make data centers physically possible. Each of these is a place where scale, IP, or manufacturing complexity creates a real moat.
The economics under the narrative
Here’s the part the headlines miss: the binding constraint on AI is shifting from chips to electricity. Cost per token falls with better silicon, but it’s capped by the cost and availability of power. That single insight links this thesis directly to my energy research — and it’s why I think the AI and energy trades are the same trade wearing two hats.
What I’m watching
Valuations are demanding, and demanding valuations punish disappointment. I want exposure to the structural buildout while respecting that the market has already priced in a lot of optimism. The setup I look for: a durable edge, improving unit economics, and a price that offers asymmetry on any short-term wobble.