Where AI value actually accrues

The current debate about artificial intelligence is dominated by the question of whether it works. That debate is settled — for practical purposes, the answer is yes — and the more interesting question has quietly replaced it. Where in the stack does the value accrue, and where will it stay?

The four layers

Strip away the marketing and the modern AI economy resolves into four layers:

1. Compute. The silicon and the data centres housing it. NVIDIA today, plus an emerging tier of accelerators chasing them.
2. Foundation models. The general-purpose model labs — OpenAI, Anthropic, Google DeepMind, Meta, a small set of open-weights challengers.
3. Inference and middleware. The serving infrastructure, evaluation tooling, agent frameworks, and the routing layer that chooses which model handles which call.
4. Vertical applications. Software built on top of the above, sold to a specific industry or workflow — legal, biotech, code, support, design.

Every dollar spent on AI passes through some combination of these four layers. The question is how the dollar splits.

Where the durability actually sits

The reflexive answer is that the value accrues at the bottom — to NVIDIA and the foundation labs — because that is where the visible profit is concentrated today. This is true in a snapshot and likely wrong as a long-term thesis, for two reasons.

Compute capacity is fungible and reproducible. Every major sovereign and every major hyperscaler is now actively trying to break the present concentration. They will succeed, eventually, on some timescale. The economics may compress dramatically once demand-supply rebalances and once a credible second source ships at scale.

Foundation models commoditise on a predictable cadence. A state-of-the-art model from twelve months ago is open-source and free today. The frontier moves, but each prior generation collapses into a competitive substrate. The labs running near the frontier will continue to capture rent for a while; the labs running a step behind it may struggle to.

The layers most likely to retain pricing power over a decade are the application layer (where switching costs and data network effects compound) and the inference middleware layer (where the underlying model becomes a swappable component and the orchestrator captures the relationship with the buyer). This is the inverse of where attention is currently focused.

What this means for operators

The interesting question for an operator is not “what model do we use?” but “what is the smallest part of the stack we can afford to own, and what is the largest part we should treat as a substrate?”

The answer is almost always: own the data, the workflow, and the relationship with the customer. Treat the model as electricity. Companies that get this right will be model-agnostic by 2028, and they will not care which lab won the leaderboard that month.

What we look at

We back operators figuring out what AI is genuinely useful for in specific industries and workflows, and the supporting infrastructure that survives the next architecture shift. We are sceptical of companies whose entire value proposition rests on access to a single upstream model, and interested in the firms quietly building the data, evaluation, and integration layers that the next generation of applications will run on.

The interesting work for the next decade will not be at the foundation. It will be in figuring out what to do with one.