Here is the question: if you swapped your AI model tomorrow — different vendor, different family, no warning — which parts of your AI investment would survive unchanged?
Sit with the answer for a minute, because it sorts everything. The model will be swapped. Not might — will. Frontier models are leapfrogging each other on a cadence of months; prices move, licensing moves, your risk office's opinion moves. Renting model capability is the correct posture — it improves on someone else's R&D budget. But a rented thing is not an asset, and everything you build that only works with this model, this vendor's orchestration quirks, this quarter's prompt dialect, is a leasehold improvement on a building you don't own.
So run the test on a typical enterprise AI portfolio. The prompt library, lovingly tuned to one model's habits: mostly gone. The demos whose magic came from the model rather than from anything around it: gone, and they were theater anyway. The "AI platform" whose value proposition is proximity to one vendor: at risk on every contract cycle.
Now look at what passes.
Skills and memory files. Your ways of working — how a summary is structured here, what a review checks, which policy wins a conflict — written down where any agent finds them. That's institutional knowledge, finally in a form that executes. A different model reads the same files.
Gold sets and evals. The corpus of "here is the input, here is what right looks like," and the machinery that grades against it. This is the asset that gets more valuable at every model swap, because it's the instrument that tells you in an afternoon whether the new model is better for your work — while your competitors argue from anecdotes.
The ontology and the knowledge graph. What your business is — entities, relationships, policies with effective dates — made machine-navigable. Building it costs real effort exactly once. Every model that arrives afterward is smarter about your business on day one, because your business is written down.
Retrieval, done properly. Not the vector-store demo — the ingestion decisions, the chunking that matches how your documents are actually asked about, the measured pipeline. Ingestion work transfers; only the last-mile calls change.
Execution graphs and their contracts. The decomposition patterns, the specs, the signals and gates that turn agent work into repeatable delivery. The pattern is deterministic; only the intelligence inside each node is rented.
Controls and audit trails. Permission lines, hooks, human gates, evidence logs. Your regulator does not care which model produced the answer; they care that you can show how. That machinery is model-blind by construction.
The operating model. People who know how to own a skill, hold a gate, read a flow report. The hardest asset to build and the only one that appreciates on its own.
Notice what that list is. It isn't a wish list — it's an inventory of the layers around the model: context, verification, orchestration, controls, people. I've argued elsewhere that what enterprises are really buying from AI is business process capability at compressed speed — and the durable part of that purchase was never the model. It's the substrate underneath: the captured knowledge, the encoded judgment, the verification machinery. Companies did this analysis before, in the outsourcing era, and the ones who thrived were the ones who kept the process knowledge and rented the labor. Same play. New decade.
The test earns its keep in three rooms. In vendor meetings: ask what survives if their model relationship changes, and watch which pitches dissolve. In portfolio reviews: score each initiative by the fraction of its budget going to surviving assets versus rented magic — that ratio predicts which projects will still exist in two years. And in build decisions: when two designs compete, prefer the one that leaves more on the surviving side of the ledger, even when it demos worse this quarter.
None of this argues for one model less. Use the best model available; switch aggressively when a better one ships — that's the whole point of being free to. The argument is about where your own effort compounds. Model capability is a rising tide that lifts every boat in the harbor, including your competitors'. The substrate is the boat you own.
The model is rented. The factory is yours. Build accordingly.