If the bubble bursts

Assume, for the length of this essay, that the bears are right. The valuations imply revenue that does not exist. The capital expenditure is circular — chip makers financing model labs financing data centres financing chip makers — and the GPUs depreciate faster than the business plans. Suppose it all goes the way bubbles go: down rounds, fire sales, consolidation, and one morning a frontier lab you have heard of files for protection from its creditors. This is not our forecast. But an enterprise strategy that only works if the optimists are right is not a strategy; it is a bet. So it is worth asking the question properly: if the AI bubble bursts, what exactly do you lose?

The answer depends almost entirely on one earlier decision — whether you rented your AI or built it. That single choice determines whether a crash in the AI industry is a catastrophe for your operations or a line item in someone else's newspaper.

A bursting bubble destroys valuations, not artefacts. The crash takes the companies. It leaves the infrastructure.

What bubbles leave behind

This has happened enough times to have a shape. The railway mania of the 1840s ruined a generation of British investors — and left behind the rails, which carried the next century of industry at prices the speculators would have wept at. The dot-com crash bankrupted the telecoms that had spent fortunes laying fibre — and the dark fibre they left in the ground made bandwidth nearly free for the decade in which the real internet economy was built, by companies that picked up the infrastructure for cents on the dollar. Economic historians call it the installation phase and the deployment phase: the bubble overpays to install capacity, the crash clears out the speculators, and the sober economy then runs on what they built. The frenzy is temporary. The infrastructure is not.

So take inventory of what this bubble has already installed, permanently, on the public side of the wall: open model weights that are downloaded onto tens of thousands of private clusters and can never be un-released; an open-source serving and training stack maintained by a global community rather than any single company; a generation of engineers who now know how to harness these systems; and more accelerated compute than the world has ever had, which a bust would dump onto the market at distressed prices. If every AI lab froze tomorrow, the capability that exists today — already sufficient to automate the bulk of enterprise routine work — would remain, in full, forever. The constraint on enterprise value was never the next model. It was, and remains, the organisation's capacity to absorb the current one.

The bankruptcy test

Run the thought experiment asset by asset. Your bespoke agentic layer: code, running on your infrastructure, wrapping your systems. A vendor's insolvency does not reach it; it keeps running the morning after. Your fine-tuned adapters: files in your artefact registry, trained on your own operational record. They keep working. The open weights on your metal: an MIT licence attached to an artefact already on your disks is not revocable by a bankruptcy court — there is no subscription to cancel, no server to switch off, no key to expire. Your inference bill: it goes down, because a bust floods the market with second-hand accelerators and idle data-centre capacity. For the builder, the bear case is mostly a procurement opportunity.

Now run the same test on the rented portfolio. The per-seat copilot: it dies with its vendor, or survives into the arms of an acquirer who reprices it. The agent platform hosted in a vendor's cloud: your workflows become an exhibit in someone else's restructuring, exportable in a format nobody else reads. The API-only strategy: capability that disappears on a deprecation schedule you do not control, from companies that — in this scenario — no longer exist. None of this is exotic. It is what happened to every category of rented software whose vendor failed, in every downturn, ever.

The renter is levered to their vendors' survival. The builder is exposed only to whether the thing works — and it already works.

The bubble is in the frontier, not the deployment

Be precise about where the speculative money actually sits. The hundreds of billions are being spent on the race to train the next model — the frontier. That is the wager that can fail to pay back. But enterprise value does not come from the next model; it comes from deploying the current one against processes that have not changed in twenty years: the reconciliations, the support queues, the document pipelines, the integrations. That deployment does not need the frontier to advance. It needs the weights to exist — and they exist irrevocably. The bubble and the enterprise opportunity are, to a first approximation, different markets that happen to share a headline.

Notice how this inverts the standard hesitation. Enterprises delay building because the model will be obsolete in six months. Follow both branches. In the bull case, models keep improving — and your architecture swaps the better one in through the router, one line of configuration. In the bear case, models stop improving — and the one you hold is as good as anyone will ever have, permanently. Building wins on both branches. Renting loses on both: in the boom you pay rising rents for someone else's compounding advantage, and in the bust the rent collector vanishes and takes the capability with them.

The honest exposure

There are things a bust would genuinely take, and pretending otherwise would make this comfort dishonest. Hosted frontier capability would degrade or consolidate — if your differentiation depends on the very best closed model being available by API forever, you are exposed. Subsidised inference would end — every venture-financed API price below cost is a teaser rate, and teaser rates end in one of two ways. And the ecosystem of thin wrappers — products that are a prompt and a margin on someone else's model — would vanish overnight, taking any workflow you built on them. The pattern is consistent: everything that dies in the bear case is something you were renting. Everything that survives is something somebody, somewhere, had the sense to build and hold.

Every subsidised API price is a teaser rate. Build as if the teaser ends.

So the bubble question, for an enterprise, is not a forecast to get right. It is a portfolio to construct so that the forecast does not matter. A built agentic layer, owned weights, adapters trained on your own memory, a router that treats every vendor as replaceable — that portfolio pays out if the optimists are right, and it pays out if the pessimists are right. When the same decision wins under both the descent and the bust, that is usually the market's way of telling you it was the right decision all along: build the thing that fits, own the layer, own the weights — and let the weather be someone else's problem.