STRATA
Operationalizing the six-layer AI stack into signals you can actually track.
Every AI market map I have read is a picture of where value sits today. Almost none of them tell you what to watch to know when it moves. That is the gap STRATA is trying to close.
The structural shift
Value in the AI stack is not migrating upward in a straight line, the way it did in cloud. It is oscillating. Margin pools open at one layer, get compressed by the layer below commoditizing, and reopen somewhere else eighteen months later. Anyone holding a static thesis is holding a photograph of a moving object.
The useful question is therefore not which layer wins. It is what would have to be true for value to leave the layer it is in now, and what observable signal fires first when that starts happening.
The six layers
S · Substrate. Silicon and the power to run it. The physical floor.
T · Training. Foundation models and the labs that make them.
R · Retrieval. The data, memory, and context that make a general model useful on a specific problem.
A · Agency. Orchestration, tool use, and the machinery that turns a model from an answerer into a doer.
T · Trust. Evaluation, audit trails, permissioning, and the regulatory scaffolding that lets a serious institution deploy any of the above.
A · Adoption. Interface and distribution. Who actually holds the customer.
The acronym is not decoration. The order is the dependency chain: no layer can capture durable margin faster than the layer beneath it can supply it.
The wedge
Most capital is concentrated at T (Training) and A (Adoption), because those are the two layers with legible narratives, the frontier lab and the consumer app. The two least-crowded layers are Retrieval and Trust, and they are the two that get structurally more valuable as the others commoditize. Retrieval is where a general model becomes a specific one. Trust is where a demo becomes a deployment.
That is the wedge: buy the layers that only matter once the exciting ones are boring.
The signals
A thesis you cannot falsify is a mood. Each layer gets one primary signal, and I am wrong when it moves against me.
| Layer | Signal I track | What it tells me |
|---|---|---|
| Substrate | Cost per useful token, not per FLOP | Whether the floor is still dropping |
| Training | Gap between frontier and best open weights | Whether model quality is commoditizing |
| Retrieval | Share of enterprise spend on context vs. inference | Whether specificity is becoming the product |
| Agency | Fraction of agent runs completed without a human touch | Whether autonomy is real or theatre |
| Trust | Time from pilot to production in a regulated buyer | Whether the deployment bottleneck is loosening |
| Adoption | Where the default interface lives | Who owns the customer relationship |
My position
I am long Retrieval and Trust, sceptical of durable margin at Training, and I think Adoption is the layer most likely to surprise everyone, because the interface that wins probably does not look like a chat box.
The prediction
Within the next two years, the binding constraint on enterprise AI stops being model capability and becomes the ability to prove what an autonomous system did and that it was allowed to. The first signal that this is happening is not a model release. It is the Trust row above: pilot-to-production time in a regulated buyer starting to fall.
I will update this page when it does, or when it doesn’t.