When the bubble pops, the only AI spend that survives is the kind that redesigned actual work

You've read the capex numbers and done the private arithmetic.

You don't say "bubble" in the exec meeting. Half the room holds the stock, the other half holds the strategy, and the word costs more than it clarifies. But on the drive home you have run the sum more than once: if this corrects — this year, next year — what happens to our programme? Which lines of your own AI budget would look embarrassing in the write-down year?

You suspect it's most of them. This piece is about finding out precisely, before anyone else does.

The numbers behind your arithmetic

First, the case that your private arithmetic is not paranoia. It is arithmetic.

Moody's analysts counted approximately $662 billion in future data-centre lease commitments across five companies — Amazon, Meta, Alphabet, Microsoft and Oracle — signed but not yet commenced¹. Under current accounting rules those commitments sit off balance sheet until the leases start. Moody's own comparison: that unrecorded pile equals 113% of the five companies' adjusted debt. Include the rest of their committed exposure and the undiscounted total reaches $969 billion.

The Bank for International Settlements — the central bank of central banks, an institution constitutionally incapable of hype — used its 2026 report to warn about an AI bubble burst, debt-funded capex and private-credit contagion in one sentence².

And the arithmetic has a demand side. MIT's researchers examined 300 enterprise AI deployments and found 95% delivered no measurable P&L impact³. Audit that number before you use it — the study leans on 52 interviews and defines "pilot" loosely, and its critics have a point. But even discounted heavily, it describes the same gap the BIS is pricing: an infrastructure build-out of historic scale, resting on enterprise returns that mostly haven't arrived.

So yes. The shape of a bubble. The debate about when is fully staffed without you.

The wrong layer

Here is the thing about that debate: it is conducted entirely at the market layer — valuations, capex, circular financing, who is holding whose equity. The market layer prices companies.

The market layer is not your layer. You run a company.

And at your layer, the history of bursts says something almost nobody quotes: a bubble bursting doesn't delete capability. It reprices it. The correction destroys spend that never became capability. It leaves actual redesigned work almost untouched — because redesigned work doesn't trade.

The dot-com crash is the clean experiment. Amazon fell 94% — from $107 to $6⁴ — and erased fourteen billion dollars of market value. Its revenue grew more than 140% from 1999 to 2002⁵, straight through the crater. The crash repriced the stock. It could not reprice the warehouse network, the one-click patent, the rewired retail operation — the work. Pets.com died in the same crash for the mirror-image reason: it was market story all the way down, with no redesigned work underneath to survive the repricing.

The companies that had genuinely rewired their operations for the web walked out of the wreckage owning the next two decades. The ones that had bought web presence — spend, wearing capability's clothes — wrote it all down.

That is the morning after. It doesn't ask what you spent. It asks what the spend became.

The correction has already started — inside

You don't have to wait for the market to see the sorting begin. It is running now, quietly, inside budgets.

S&P Global surveyed over 1,000 enterprises: 42% abandoned most of their AI initiatives in 2025, up from 17% a year earlier⁶. On average, 46% of projects died between proof of concept and adoption. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027 — escalating costs, unclear business value, inadequate risk controls⁷.

Read those numbers as a sequence and you see it: the private correction precedes the public one. CFOs are already doing to AI line items what the market has not yet done to AI stocks. The write-down year starts one budget cycle at a time.

Which puts you in the squeeze. Keep spending the way most programmes spend — tools first, adoption dashboards, workflow redesign left to emerge — and you are manufacturing the very line items the correction deletes. Freeze instead, and you hand the redesigned-workflow advantage to whichever competitor kept going through the winter — because the winter is precisely when capability gets cheap to build. Both roads are expensive. The wrong answer costs in both directions.

"Wait and see" is not a hedge. It's the sand pile with a pause button.

Three columns

So run the sorting yourself, before the market runs it for you. Every line of AI spend in your company lands in one of three columns.

Sand. Spend that evaporates in a correction because it never attached to work: licences bought ahead of need, pilots that ended in a demo, platform subscriptions "to be ready", the innovation-team showcase, adoption numbers with no workflow behind them. Sand can be enormous and green on every dashboard. Its tell: if it stopped tomorrow, no process would run differently.

Scaffold. Spend that enabled but isn't itself the building: training, enablement, data clean-up, integration work, the champions programme. Scaffold is honourable — nothing gets built without it — but scaffold left standing with no building inside it is just slow sand. Its tell: it points at capability that is supposed to arrive later. "Later" is the column's whole risk.

Asset. A named workflow that runs measurably differently because of the spend — claims triage cut from ninety minutes to nine, the month-end close that lost a week, with an owner who can say so in numbers. Asset survives any correction, because a repriced model or a cheaper vendor doesn't un-redesign your work. Amazon's warehouses did not care what the stock did. Its tell: it would keep paying even if the AI market halved.

If you have read this far and mapped your own budget in your head, you already know your problem: honestly sorted, the asset column is thin. That thinness — not the Nasdaq — is your exposure to the burst.

The Morning-After Ledger

The sorting takes one afternoon. Pull the AI budget — every line over a threshold you pick — and run each line through nine prompts. Three per column, in order.

Sand checks: 1 · The stop test. If this line stopped tomorrow, which process would run differently? No answer — sand. 2 · The demo test. Has this produced anything beyond a demonstration, a pilot report, or a dashboard? No — sand. 3 · The reason test. Was this bought because of a named workflow, or because of the year it was purchased? Be honest. The second answer — sand.

Scaffold checks: 4 · The pointer test. Does this line exist to enable a specific named build — or "capability in general"? Named build — scaffold. In general — sand. 5 · The date test. Does the thing it enables have a delivery date and an owner? Yes — scaffold, hold. No — reclassify sand at next budget. 6 · The orphan test. Is what it was scaffolding still funded and alive? If the building died, the scaffold is standing sand.

Asset checks: 7 · The workflow test. Name the end-to-end process that runs differently. Name its owner. Both names or it isn't an asset. 8 · The number test. State the cycle-time, cost or quality change, in units. "Feels faster" is scaffold at best. 9 · The halving test. If AI prices doubled or the vendor vanished, would the redesigned process retain most of its gain? Yes — asset. No — scaffold wearing asset's badge.

Then compute one number: the sand ratio — sand dollars over total AI dollars. The verdict comes in three words:

• Sand ratio under a third: compounding. Your programme survives a burst and exits it ahead. Protect the asset column's owners above all.
• A third to two-thirds: covered — half-built. You have real assets and real exposure. The move is not more spend; it is converting named scaffold to named asset before the cycle turns.
• Over two-thirds: exposed. Most of your programme is repriced to zero in a correction, whatever the dashboards say. You now know before the market does — which is the entire advantage this afternoon buys you.

One boundary, honestly: the ledger cannot time the burst. Nothing can — the BIS can't and Moody's won't. It can only guarantee that whenever the morning after arrives, your budget is already sorted into what it deletes and what it can't touch.

The winter advantage

Here is the ending most bubble commentary refuses to write: for a company whose spend is mostly asset, a burst is good news. Capability gets cheaper. Talent gets available. Competitors whose programmes were sand spend two years explaining write-downs while your redesigned workflows keep compounding — exactly as the web kept compounding after 2001, under new management.

Sort the budget before the market sorts it. The strategy conversation follows the sand ratio, not the other way around.

By the end of one afternoon you will have every AI line in a column, one ratio, and a word — sand-proof answers for the meeting where somebody finally says "bubble" out loud.

Start with the stop test. Run it on your three biggest AI lines before Friday. If none of the three passes, you don't need the other six prompts to know which column your programme lives in.

References

Footnotes

1. Yahoo Finance. (2026). Moody's flags $662 billion risk at the heart of the data-center buildout by just 5 companies. https://finance.yahoo.com/news/moody-flags-662-billion-risk-081000616.html ↩

2. Fortune. (2026, June 29). The central bank of central banks just released its flagship annual report — and it sees a $1 trillion AI investment boom headed for a reckoning. https://fortune.com/2026/06/29/bis-central-bank-warning-hyperscaler-data-center-1-trillion-gamble-recession/ ↩

3. Legal.io. (2025). MIT report finds 95% of AI pilots fail to deliver ROI, exposing "GenAI Divide". https://www.legal.io/blog/5719519/MIT-Report-Finds-95-of-AI-Pilots-Fail-to-Deliver-ROI-Exposing-GenAI-Divide ↩

4. Shortform. (n.d.). Amazon in 2001: How Jeff Bezos rose above bad share prices. https://www.shortform.com/blog/amazon-2001/ ↩

5. Amazon.com, Inc. (2003). Form 10-K for fiscal year 2002. U.S. Securities and Exchange Commission, EDGAR. https://www.sec.gov/Archives/edgar/data/1018724/000095014903000355/v87419ore10vk.htm ↩

6. CIO Dive. (2025, March 14). AI project failure rates are on the rise: Report. https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/ ↩

7. MarTech. (2025, June 25). Gartner: 40% of agentic AI projects will fail, making humans indispensable. https://martech.org/gartner-40-of-agentic-ai-projects-will-fail-making-humans-indispensable/ ↩