The Moment Nobody Is Watching For
Most bubble narratives focus on the top: when does sentiment turn, when do valuations crack, when does the smart money start moving to the exits. That framing misses where AI's specific fragility actually lives.
The riskiest moment is not when investors get nervous. It is when the first major enterprise customer stands up and says, publicly, that the spending did not work.
That is the trigger. Not a rate decision. Not a model that fails a benchmark. One Fortune 500 CFO, in one earnings call, saying the AI investment did not produce the returns they projected.
Three Specific Scenarios
Analysts working the enterprise software space have identified three candidate triggers, each with its own mechanism.
The first: a major bank or insurer announces an ROI failure on an AI deployment. Financial services firms have been among the largest enterprise AI spenders, and they have internal audit cultures that will surface underperformance faster than most industries. A public announcement from a Citi or an Aetna carries a specific kind of credibility signal, because these are not companies that make that kind of admission without significant internal process.
The second: a Fortune 500 announces layoffs of its AI team. Not a restructuring, not a "refocus," but a reduction in headcount explicitly tied to AI initiatives that failed to produce. This would be the enterprise equivalent of a startup running out of runway. It tells every CFO across every sector that the ROI case they were sold was overstated.
The third is structural rather than anecdotal: Nvidia missing revenue guidance by more than 15% for two consecutive quarters. Nvidia's revenue is effectively a real-time demand index for AI infrastructure spend. Consecutive misses would mean the build slowed faster than the company's own projections, which would mean something large changed in the demand picture.
How the Feedback Loop Works
The specific danger is not a single announcement. It is what happens after.
One public failure triggers a predictable institutional response: CFOs across the sector demand ROI audits of their own AI spend. These audits surface more projects that are underperforming, because the honest answer to "is this producing measurable returns yet" is "not definitively" for a significant portion of enterprise AI deployments right now. Some of those projects get cut. Some of the cuts become public. More audits follow.
This is a negative feedback loop. Not a crash, exactly. A slow deflation, where each data point makes the next data point easier to surface. The correction does not need to be dramatic to be damaging. A 30% reduction in enterprise AI software renewal rates, spread across 18 months, produces a very different headline than a single day of market collapse, but the underlying economic effect is comparable.
The self-reinforcing quality is what makes timing so hard to call. The loop doesn't need a catastrophic event to start. It needs one credible admission.
The Cisco Parallel
In early 2000, Cisco Systems was the most valuable company in the world. Enterprise IT was spending at a pace that assumed the internet buildout would continue indefinitely. It did not.
When the first wave of enterprise buyers admitted they had overbought, the correction did not take years. Cisco's revenue collapsed in roughly 90 days. The company wrote down $2.2 billion in inventory. The infrastructure build that had driven its valuation was real, but the demand assumptions embedded in that build were not.
The structural parallel to AI is not perfect, but it is specific enough to be useful. The AI infrastructure build, like the late-1990s telecom build, is being driven by demand projections that assume current adoption rates continue and compound. The physical assets being built, data centres, power infrastructure, GPU supply chains, are long-lead investments. They cannot be quickly repurposed if the demand curve turns out to be flatter than projected.
What is different this time: cloud providers have locked in long-term enterprise contracts that slow the demand signal considerably. The flexibility that made the 2001 correction so fast, companies could cancel orders immediately, is partially absent here. The slowdown, if it comes, will move in slow motion compared to Cisco.
Why Enterprise AI Spending Is Different From Past Tech Cycles
Enterprise software spending in previous tech cycles collapsed when buyers could cancel quickly. Licenses, maintenance contracts, and hardware orders could be terminated or not renewed on annual cycles. The speed of the correction in 2001 was partly a function of that flexibility. Buyers realized they had overbought, and they stopped buying.
AI spending in 2025 and 2026 is structured differently in ways that create both risk and resilience. The resilience: hyperscaler cloud contracts are multi-year commitments with significant termination penalties. A company that signed a three-year Azure AI deal in 2024 is not walking away from that contract because a few AI projects underperformed. The spend is locked in.
The risk: locked-in spend without locked-in adoption creates a specific kind of pressure. IT departments that are paying for AI capacity they are not fully using have to justify that spend internally. When renewal comes up, they face a harder conversation. The correction may be slower than 2001, but the eventual adjustment may be larger precisely because buyers kept paying for longer than the returns warranted.
This also affects how the trigger works. Because enterprise contracts are multi-year, the signal that matters is not cancellations. It is non-renewals. By the time non-renewals become visible in earnings data, the decision to not renew was made six to twelve months earlier. The leading indicator of the correction will be internal budget allocation decisions that don't become public until the contract period ends.
The Two-Bubble Theory
Capital Economics has been developing an argument that helps clarify where we actually are. Their read: there were two AI bubbles, and one of them may have already deflated.
The first bubble was in stock valuations, concentrated in the 2022-2024 period, when AI enthusiasm drove speculative premium into any company that could credibly attach "AI" to its product. That bubble has partially corrected. The Nasdaq's AI-adjacent cohort has repriced. Some of that premium still exists, but the most egregious valuations have moderated.
The second bubble is different. It is the enterprise AI spend bubble: the actual dollars flowing into AI software licenses, AI-enhanced SaaS add-ons, AI consulting engagements, and AI infrastructure. This one deflates on a different timeline and with different mechanics. It deflates when procurement cycles turn, when renewal rates fall, when IT budgets get rewritten. That process is slower, less visible, and harder to time.
The warning indicators to watch, according to this framework, are practical: Salesforce and ServiceNow AI feature adoption rates, enterprise software renewal rates specifically for AI add-on SKUs, and data centre lease cancellation rates. These are the variables that reveal whether enterprise spending is translating into enterprise use, or sitting unused on a license manifest.
The Case That None of This Matters
The optimistic version of this story acknowledges the bubble and doesn't particularly care about it.
The argument: AI productivity gains are real. They are measurable in controlled settings. They will eventually show up in GDP data and in earnings, but that process takes three to five years, not one. The bubble popping does not mean the technology failed. It means the market priced in the gains before they arrived. The internet bubble of 2001 destroyed trillions in paper wealth and produced Amazon, Google, and the modern web as a byproduct.
On this read, the question is not whether a correction comes. It is whether the underlying capability is durable past the correction. Reasonable analysts disagree here, but the bears are not arguing that AI is useless. They are arguing that the current spend is priced for a speed of adoption that won't materialize on the projected timeline.
That is a narrower argument than "AI doesn't work." It is also, historically, the argument that has been correct about most major technology transitions.
The trigger exists. It has a shape. Watch for the admission, not the mood shift.