THE AXIS HAS ROTATED

Build versus buy has stopped being the decisive question. When engineering capacity is the input the model now supplies, the axis rotates ninety degrees: what matters is whether you possess a substrate worth running a model against, and whether that substrate compounds faster than a competitor can copy it.

The case to build has never been stronger

A Wall Street firm built an internal drafting platform by pointing a frontier model at a decade of its own catalogued deals — down to provision level, including negotiation history rather than just executed documents. No software house. No thirty-person engineering team. Build cost was mostly people's time; compute was, in their word, nominal.

What made it possible was not the model. It was a corpus no vendor could replicate, plus the judgement to know what to point the model at. Crucially, that judgement sat inside the same people: legal authority and architectural fluency held together, not in two departments passing requirements across a wall.

Domain expertise is necessary but not sufficient. When build economics collapse, the fusion of domain judgement and technical architecture becomes the rare thing — not either competence standing alone.

The case to buy is the same case in reverse

The infrastructure is not what you buy. Knowledge graph, vector index, model calls, parsing pipeline — none of that is exotic. What takes the time is the domain layer: the entity taxonomy that survives contact with real cases, the conflict-tolerance rules that reflect how valuation uncertainty actually behaves, the extraction logic tuned to the document types that recur. That layer is months of work with practising specialists and it does not ship as an open-source package.

Build when you own the substrate. Buy when someone else does. The two cases are the same proposition stated from opposite ends of a corpus, and the model decides neither.

The artefact is not the moat

When the build is cheap, the build is cheap for everyone. A capability you can stand up in a quarter is by definition one your competitor can stand up in a quarter. What is being bet on instead is velocity: by the time the obvious thing is widely replicable, the team has moved to the problem that is not yet obvious. Defensibility lives in the rate at which the next thing ships, not in anything already shipped.

The substrate is the IP, and it is not promptable

The substrate is not the model — that is rented and common. Nor is it the stack; a vector index and a parsing pipeline are a weekend with a capable assistant. The defensible asset is the semantic layer between the documents and the reasoning, and it is built as a progression:

1. A domain-specific vocabulary hardens into a taxonomy comprehensive enough to survive real cases.
2. The taxonomy matures into an ontology expressive enough to drive retrieval that adapts to the shape of the question.
3. The ontology becomes the schema the graph actually runs on.

You can prompt a model into drafting an application. You cannot prompt a model into the data infrastructure that makes adaptive graph retrieval work at scale — multi-hop traversal over a typed graph with vector search living next to the edges rather than in a separate store.

Why the naive economics are already breaking

The dominant pattern treats the document as the unit of analysis: load the whole bundle into a long context window, or summarise everything first and stuff the summaries in. Both fail as volume grows. Loading two hundred thousand tokens to answer a question whose support lives in four thousand is not merely expensive — it is slower by an order of magnitude and it strips the system of any ability to decline, because once the bundle is in the window the model will produce an answer whether or not a basis for one is present. Summarisation is worse: each summary is a fresh generative artefact with its own capacity to distort, citations stop pointing at source pages, and the spend recurs every time a document is reclassified or the summarising model is upgraded.

What this means for operators

Your value is no longer the artefact you can produce, because the model can produce artefacts for anyone who asks well. Your value is your judgement about which problem to attack next, and the speed with which you attack it and move on. Treat every win as perishable and organise around motion rather than ownership.

Build what you own the ground under. Buy what you cannot prompt into being. Accept that the spend has moved rather than disappeared — and keep running.

Adapted from "The Axis Has Rotated" by Jonathan Aiken, first published in Nodes & Edges, 1 June 2026.