Intelligence is the new infrastructure

When the marginal cost of cognition falls toward zero, every category gets re-founded. We invest at that fault line.

Every durable platform shift starts when something expensive becomes cheap. Bandwidth, then compute, then storage: each collapse in cost did not merely improve the incumbent businesses; it rewrote which businesses were possible at all. We are now living through the steepest of these curves yet. The marginal cost of intelligence is falling toward zero.

The first-order effect is obvious and already priced. Models get better, demos get slicker, every roadmap grows a copilot. The second-order effect is where the returns live. When cognition is abundant, the binding constraint moves elsewhere: to distribution, to trust, to the messy integration work that turns a capability into a governed, reliable, paid-for outcome inside a real organisation. The history of every prior cost collapse teaches the same lesson. The scarce resource is never the thing that just got cheap. Cheap bandwidth made attention scarce. Cheap compute made talent scarce. Cheap intelligence will make judgment, accountability, and proprietary context the things worth paying for.

We try to keep a cold head about this. Most of what is labelled AI today is a thin wrapper capturing a sliver of value that the underlying model could reclaim in its next release. That is not a business; it is a feature waiting to be absorbed. The companies that matter are the ones positioned to compound as the models improve rather than to be eaten by them. The right question to ask of any AI company is simple and unforgiving: when the next model arrives, does this company get stronger or weaker?

Where we underwrite

We back two layers. The substrate: the tooling, orchestration, evaluation, and infrastructure that make intelligence usable and safe at production scale. And the vertical applications: software that takes a specific, valuable workflow and re-founds it natively on AI, owning the outcome rather than renting a feature.

At the substrate layer, the unglamorous problems are the durable ones. Evaluation is the obvious example. As soon as a model is doing work that matters, someone has to prove it is doing that work correctly, repeatedly, and within policy. Observability, guardrails, the plumbing that routes a request to the right model at the right cost, the systems that capture human corrections and feed them back as training signal: these are the picks and shovels of an industry that is still, mostly, mining by hand. They are also defensible in a way the application layer often is not, because they sit between the model and the enterprise and accrue switching costs with every workflow they touch.

At the application layer, our filter is ownership of outcome. A company that merely surfaces a model's answer is exposed; a company that takes responsibility for the result, the underwriting decision, the resolved support ticket, the cleared claim, the booked freight, earns the right to a deeper relationship and a deeper margin. Owning the outcome means owning the data exhaust the outcome generates, and that exhaust is the raw material of the only moat that compounds in this market.

The winners will not be the companies that added AI. They will be the companies that could not have existed without it.

That distinction is our filter. A feature can be copied in a quarter. A company whose entire cost structure, data advantage, and product surface assume abundant intelligence is a different animal, and a far harder one to displace.

The new shape of a software company

Abundant intelligence does not just change what software does; it changes what a software company is. For thirty years the economics of software were set by the cost of writing it and the cost of selling it. Both are now in motion. When a small team can produce what once required a department, headcount stops being a proxy for ambition, and the centre of gravity shifts from how much you can build to how clearly you can decide what is worth building.

This has a sharp consequence for how value is captured. The classic per-seat model assumed a human sitting behind every license. When the software is doing the work rather than helping a person do it, pricing migrates toward the outcome: per resolved case, per transaction, per unit of work completed. That migration is not cosmetic. It re-rates the size of the prize, because the addressable market is no longer the software budget; it is the labour budget the software displaces. We spend real time underwriting whether a company is positioned to price into that larger pool or whether it will be trapped charging for seats while delivering work.

It also changes the failure mode. Software that advises can be wrong quietly; software that acts is wrong loudly and expensively. The companies we want to back treat reliability, auditability, and the graceful handling of their own mistakes as first-class product surface, not as a compliance afterthought. In a world where intelligence is cheap, the premium is on the systems trustworthy enough to be handed the keys.

What this means for how we invest

Practically, this thesis pushes us early and pushes us operator-led. The defensible AI companies are being founded right now by people who understand a specific domain deeply enough to know which workflow is worth re-founding and which is a trap. That knowledge is rarely found on a pitch deck; it is found in operators who have lived the problem. We would rather back a team that has felt the pain of a broken workflow for a decade than a team that discovered it last quarter through a model.

It also disciplines our patience. The substrate plays compound slowly and then suddenly, as standards settle and switching costs accumulate. We size positions to hold through the noisy middle, when a louder competitor raises a larger round on a thinner thesis. Our job is to be right about the fault line and to stay invested across it, not to win the quarter. When the marginal cost of cognition falls toward zero, the value does not disappear. It moves. We invest where it moves to.