4 AI Bubble Indicators Economists Actually Track

Chart-style illustration of AI investment bubble risk indicators including Fed survey data, bank loan exposure, and hyperscaler depreciation schedules

An AI bubble indicator is a specific, measurable signal — a survey number, an accounting change, a loan balance — that economists and regulators track to judge whether AI spending is outrunning AI revenue, rather than a vibe about whether the technology itself is overhyped. Several of these indicators moved sharply in 2026, and unlike most bubble talk, they come from the companies' own regulatory filings and from central banks whose job is specifically to watch for exactly this kind of risk. This explainer walks through four of them, in the exact language the source documents use.

This explainer is based on official Federal Reserve publications, IMF notes, and company SEC filings verified as of 14 September 2026 — the same documentary standard we used to unpack how on-device AI actually works.

Indicator 1: How many people at the Fed itself are worried

The Federal Reserve doesn't just publish opinions about AI risk — it surveys market participants directly and reports the raw numbers. The Federal Reserve's May 2026 Financial Stability Report is built around exactly this kind of survey data on what respondents see as the biggest threats to the financial system.

The scale of the shift is the indicator itself: coverage of that survey data reports that the share of market contacts naming AI as a possible shock jumped from roughly 9% a year earlier to about 50% in the spring 2026 survey — more than a fivefold increase in twelve months. That's not a prediction of a crash; it's a measurement of how quickly professional risk-watchers changed their minds.

Indicator 2: Whether the loans behind AI are showing stress yet

The Federal Reserve Bank of Chicago publishes its own analysis specifically on how exposed the banking system is to AI investment, treating it as a distinct, trackable risk category rather than a general worry. Its research states plainly that large bank commercial-and-industrial loan commitments to the software industry — the sector doing much of the AI infrastructure borrowing — grew from $150 billion in early 2022 to $191 billion by late 2025.

Critically, the Chicago Fed's own data shows delinquencies on those loans remained low as of the third quarter of 2025 — under 50 basis points of Tier 1 capital at each large bank. That's the indicator to actually watch going forward: rising loan balances are not themselves alarming, but rising loan balances combined with rising delinquency rates would be.

Indicator 3: A hardware accounting choice that four companies made differently

This is the indicator that's hardest to spin, because it comes directly from the companies' own SEC filings, in their own words, and because two of them moved in opposite directions on the same underlying technology.

Company	Useful-life change	Company's own stated reason (SEC filing)
Amazon	Extended servers 4→5→6 years (2022–2024), then reversed a subset back to 5 years effective January 2025	"The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning"
Meta	Extended to 5.5 years, effective fiscal year 2025	Described as the result of a completed useful-life assessment
Microsoft	Currently depreciates servers and network equipment over a 2-to-6-year range	Per Microsoft's FY2026 10-K property and equipment disclosure

Extending an asset's useful life lowers a company's reported depreciation expense and raises reported operating income in the near term — it's a legitimate accounting estimate, not automatically a red flag. What makes this indicator worth tracking is the divergence: Amazon's own 10-Q explicitly ties its 2025 reversal to AI hardware moving faster than expected, while other hyperscalers extended their schedules over the same period. When companies facing the same GPU replacement cycle draw opposite conclusions about how long that hardware will last, the disagreement itself is the signal — it means at least one side's depreciation expense doesn't match the economic reality of the hardware.

Indicator 4: How the IMF is framing the transition, officially

The IMF's own 2026 scenario-planning note doesn't predict a crash — a similarly conditional, data-grounded approach to the one we took in explaining how dating app algorithms actually work. It frames AI as what it calls a "macro-critical transition" whose outcome depends on diffusion speed, institutional readiness, financial stability, and global coordination — deliberately conditional language from an institution that avoids single-outcome predictions. The same note points out that inference prices for certain frontier models have dropped by over 99%, which cuts both ways as an indicator: falling prices could reflect healthy competition making AI cheaper to deploy, or they could mean AI companies can't charge enough to justify the infrastructure being built to serve them.

What these four indicators don't tell you

None of them, individually or combined, tells you when or whether a correction happens — they're diagnostic tools, not forecasts, and the sources themselves are explicit about that limitation.
The Fed's own survey measures sentiment among market contacts, not a Fed conclusion that a bubble exists — New York Fed President John Williams stated publicly in August 2026 that he does not see the situation as a bubble, even as his own institution tracks these risk indicators.
Depreciation schedule choices are estimates within GAAP's normal discretion; a company extending useful life is not committing fraud, even when critics argue the assumption doesn't match GPU replacement cycles.

How to actually use these four indicators

Rising survey percentage without rising loan delinquencies: sentiment is shifting faster than actual credit stress — a warning sign worth tracking, not yet a crisis signal.
Loan delinquencies start climbing alongside the survey numbers: this is the combination Chicago Fed's own framing suggests would matter — stress moving from sentiment into actual bank balance sheets.
More hyperscalers start shortening useful-life estimates, following Amazon's move rather than the earlier extensions: that would suggest even hardware buyers with the best internal data on GPU performance are conceding shorter real-world lifespans.
Inference pricing keeps falling while infrastructure spending keeps rising: the IMF's own note frames this exact combination as the crux of whether AI's economics work — watch whether falling prices are being offset by rising usage volume, not just headline capex figures.

Frequently asked questions

Has the Federal Reserve officially called AI spending a bubble?

No. Its own Financial Stability Report presents survey data showing more market participants view AI as a possible shock, but Fed officials, including New York Fed President John Williams, have publicly stated they don't currently characterize the situation as a bubble.

Why does it matter that Amazon reversed its depreciation schedule while other companies extended theirs?

Because all four companies are buying largely the same category of Nvidia-based hardware under the same technological conditions. When one concludes the equipment wears out faster while others conclude it lasts longer, at least one assumption likely doesn't reflect economic reality — and that gap eventually has to be reconciled with future write-downs or continued extensions.

Is falling inference pricing a good sign or a bad sign for AI's economics?

The IMF's own note treats it as ambiguous rather than clearly positive or negative — it depends on whether falling per-query prices are being offset by rising query volume, or whether providers are pricing below what's needed to justify their infrastructure spending.

Last verified: 14 September 2026. This explainer draws on Federal Reserve, IMF, and SEC filing data current as of publication; these figures update with each new quarterly filing and Fed report.