A supply shock does not reveal scarcity. It reveals which of your capabilities were leased. Every AI roadmap now terminates, several layers down, in someone else's policy, and the firms that survive substitution are the ones that kept their calibration assets portable.
In July 2023, China placed export licensing on gallium and germanium, materials for which it controls 98% of primary global gallium supply and a majority of refined germanium. Exports of both fell to zero the following month. Firms that held inventory had a cost problem. Firms that relied on continuous supply had an existence problem.
The distinction is the useful part, because every AI roadmap now has a dependency graph that terminates, several layers down, in someone else’s policy: a trade ministry, a pricing committee, a deprecation schedule. What a shock reveals is never scarcity. It reveals which of your capabilities were leased.
Restriction on owned inventory is cost; restriction on rented access is existence
A restricted component you own a stock of raises prices and forces planning. A restricted capability you rent disappears at the moment the contract does, and it takes every by-product with it: the data it produced, the workflows built around it, the expectations trained into your teams.
Most enterprise AI capability today is rented at exactly this level. The lease covers more than the model: it covers the judgement layer built into vendor platforms, the scoring, the routing, the summarisation quality your operations quietly depend on. Nobody prices the disappearance scenario, because on the day of purchase it looks like a service-availability question rather than what it is: an asset-location question.
Model deprecation is the rehearsal everyone gets
You do not need a trade war to run this experiment. Hosted model versions retire on published schedules, and every retirement runs the same drill: the model itself is replaceable, and usually by something better. What breaks is everything calibrated against the old behaviour.
The prompt library tuned over eighteen months. The evaluation suite whose thresholds were set empirically. The guardrails written against one model’s specific failure modes. The internal sense of what good output looks like. Firms discover during a forced migration that they accumulated a genuine asset and stored it in a form that only worked with one supplier.
That discovery has a price, and it is paid at the moment of least leverage. The behavioural regression suite exists precisely to make this migration a diff, not an excavation.
Know which layer you are accumulating value in
The instructive contrast: NASA’s Roman telescope will produce asteroid discoveries it was never funded for, because NASA owns the instrument, the survey cadence, and the data. Had it bought “dark energy measurements as a service”, the asteroids would have belonged to the vendor.
The argument here is not for building your own models but for knowing precisely which layer of your stack accumulates your value, and making sure that layer is one you control. Evaluations, prompt assets, routing logic, data schemas, and process knowledge can all live in supplier-neutral form. Value accumulated there compounds across every model you ever switch to. Value accumulated in the rented layer resets to zero at each substitution — the same trap, in a different layer, as the business case that destroys its own upside.
The dependency map
The exercise takes a fortnight and changes procurement behaviour permanently. For each AI-dependent capability, answer one question: is this accessed through an API someone else controls, or does it sit on data and interfaces we own?
If rented: assume it can be withdrawn or repriced within eighteen months, and invest in the layer around it — evaluations, prompts, routing, schemas — in a form that survives substitution.
If owned: stop optimising model choice and start extending retention, normalising schema, and exposing the system programmatically, because that is where by-products form.
If you cannot say which: that ambiguity is the finding. Commission the map before the next model contract renews, not after.
The seats you already pay for are the rented layer done right
For a Dubai firm inside the 295,000-company agentic push, this framing collapses into one practical rule. The Claude, ChatGPT, and Copilot subscriptions your firm already pays for are the rented layer, and renting there is correct: models improve on someone else’s capital.
The owned layer is what you wire those seats into: the process definitions, the evaluation habits, the logged decisions, the working agents whose design your team understands. Build value there and vendor substitution becomes a configuration change. Build value in the vendor’s proprietary walls and every renewal is a hostage negotiation.
Start the map where it pays back fastest. The five-step consulting method inventories which of your AI capabilities are rented, which are owned, and what must move between them before the next renewal. And the Claude certification program puts the evaluation and agent-building skills in your own team, which is the one layer no vendor can deprecate.
Frequently asked questions
What happens to AI assets when a hosted model is deprecated?+
The model itself is replaceable; there is usually a better one. What breaks is everything calibrated against the old behaviour: the prompt library tuned over eighteen months, evaluation thresholds set empirically, guardrails written against specific failure modes. Firms discover mid-migration that they accumulated a real asset and stored it in a form that only worked with one supplier.
Which AI layer should an enterprise own rather than rent?+
The calibration layer: evaluations, prompt assets, routing logic, data schemas, and process knowledge, kept in a form that survives supplier substitution. The model can stay rented. Value accumulated in the rented layer disappears with the contract; value accumulated in the owned layer compounds across every model you ever switch to.
How do you assess AI dependency risk?+
Commission a dependency map before the next model contract renews. For each capability, answer one question: is it accessed through an API someone else controls, or does it sit on data and interfaces you own? If the answer is unclear, that ambiguity is itself the finding.