If the AI Bubble Bursts, What Happens to Your Vendors?

Search interest in whether the AI bubble is bursting is climbing steeply, and almost everything written about it is addressed to somebody holding the stock.

Different question if you are a firm that has put these tools into your actual workflows. You are not asking whether the multiple is justified. You are asking whether the thing your associates now use every day will exist, at a price you can pay, in three years.

The arithmetic that drives the worry

Strip out the sentiment and the concern is mechanical. Four numbers.

Capital expenditure keeps climbing. Goldman Sachs revised its 2026 estimate up from 465 billion to 527 billion dollars.

Inference prices keep falling. Good for you as a buyer. It means each unit of the thing being sold generates less revenue than it did last year.

The infrastructure depreciates faster than the models assume. Useful-life assumptions on this hardware are contested, and if the real life is shorter than the assumption, the cost base is larger than the plan.

Some of the financing is circular. In September 2025 Nvidia announced a 100 billion dollar investment in OpenAI data centers, and OpenAI committed to buying Nvidia chips to build them. Money moving in a loop between a small number of companies inflates apparent demand, and anybody who watched 1999 recognises the shape.

Put those together. If revenue cannot grow fast enough, at falling prices, to service the debt and the depreciation on the buildout, then the buildout stops. Not gradually.

Why the sceptics of the sceptics are not stupid either

The counter-case is decent and deserves its paragraph.

Usage is real and growing, which was not true of most of 1999. The infrastructure is a general-purpose asset rather than dark fibre nobody wanted. And the bubble call has now been made annually for three years by serious people, which does not make it wrong but does say something about the difficulty of timing it.

There is also a version where the bubble bursts and the technology is completely fine. Railways were real. The railway bubble still ruined a lot of people. Both can happen, and for a buyer that scenario is actually the most likely one to plan for.

What actually happens to you

Three scenarios, in rough order of likelihood.

Prices go up. The most probable near-term outcome of a funding squeeze is not vendors disappearing. It is the end of subsidised pricing. A lot of current per-seat pricing is underwritten by investor money and does not cover the compute. If that stops, your renewal has a different number on it.

Your vendor gets acquired or folded. The tier most at risk is the thin wrapper. Companies whose product is an interface over somebody else's model, with no data moat and no distribution, are the ones that do not survive a funding winter. Several of those are sitting in advisory and accounting stacks right now.

Capability plateaus for a while. If capital expenditure stops, frontier progress slows. That is close to irrelevant if your use case is drafting, summarising and document handling, since today's models already do those well. It matters if you have bet on capabilities that do not exist yet.

What to do, and it is the same list either way

The useful test for any advice in an uncertain environment is whether it is correct under both outcomes. This list is.

Know what you would lose. Per tool: what data is in it, what would leave with it, what the workflow falls back to. If you cannot answer that in an afternoon, that is the finding, and it is the same inventory your examiner wants anyway.

Ask for the exit terms before you sign, not after. Export format, deletion timeline, who confirms it. This is already on the vendor question list for regulated firms and it is now also a solvency question.

Prefer tools where your data is portable. A tool holding your work product in a proprietary format has you by the throat in a renegotiation, and it is worthless to you in a wind-down.

Be wary of a price that cannot be real. If the per-seat price is obviously below the compute cost, somebody is paying the difference and it is not going to be them forever. That is not a reason to avoid the tool. It is a reason not to build a workflow that only works at that price.

Prefer the general tool to the bespoke build. MIT's own data supports this for reasons unrelated to the bubble, and it happens to also be the resilient choice. A workflow built on a general assistant survives a vendor failure by switching. A custom integration with a startup does not.

Do not prepay for years. Discounts on multi-year prepayment are attractive right now for a reason.

The honest disclosure

I run an AI consultancy. If the buildout unwinds, my pipeline suffers. You should read the above knowing that, and notice that none of it tells you to buy more.

What I would say is that the bubble question and the usefulness question are separate, and the conflation is doing damage in both directions. People who think the valuations are insane conclude the tools do not work. People who find the tools useful conclude the valuations must be fine. Neither follows.

The tools do what they do regardless of what Nvidia trades at. Whether your particular vendor is still there in three years is a procurement question with a procurement answer, and it is answerable now, at your desk, without a view on the macro.