AI Bubble Risk: Five Indicators That Matter

The evidence does not support a clean market-wide verdict. AI can generate real earnings and useful infrastructure while particular stocks, private companies or data-center projects are priced for returns they will never deliver.

A stronger test tracks five indicators: earnings quality, valuations, capital-expenditure sustainability, financing conditions and infrastructure obsolescence. The warning becomes serious when weak cash returns, demanding prices and fragile funding appear together; one elevated measure is not enough.

1. Earnings quality: Are profits turning into cash?

Start with the relationship among revenue, net income and operating cash flow. Rising earnings are more credible when revenue is expanding and cash generation keeps pace; they are weaker when adjusted profit grows mainly through exclusions, accounting estimates or share-based compensation.

Fidelity’s February 2026 assessment identified earnings growth, earnings quality, valuations, capex sustainability and the interest-rate cycle as five directional signals. It found late-2025 S&P 500 earnings quality healthy and aggregate capex below free cash flow, although it also noted that AI monetization still lagged investment and that more private projects were using debt.

Score this indicator green when operating cash flow equals or exceeds net income and revenue supports the profit increase. Use amber when cash conversion falls below that level for a reporting period or adjusted earnings materially outpace GAAP earnings. Use red when weak cash conversion persists alongside falling revenue, repeated estimate cuts or restatements.

2. Valuations: How much future success is already priced in?

A high multiple is not, by itself, evidence of a bubble. The risk is the amount of growth embedded in the price relative to the company’s margins, competitive position and plausible duration of excess returns.

Historical figures provide reference points, not universal cutoffs. Fidelity calculated that the S&P 500 ended 2025 at about 22.3 times forward earnings, above its roughly 18.7 times 10-year average but below the 24.4 multiple it cited for July 1999. Those figures describe the broad index; they should not be applied mechanically to chipmakers, cloud platforms and software vendors with different economics.

Mark valuation green when earnings estimates rise fast enough to hold or reduce the forward multiple. Amber means the multiple remains above its own historical range while estimates flatten. Red requires a sharper disconnect: an extreme multiple combined with declining revenue growth, margins or forward estimates.

3. Capex sustainability: Can the buildout fund itself?

The useful measure is not spending alone but capex relative to operating cash flow, free cash flow and incremental revenue. A rational expansion can consume cash temporarily, but utilization, contracted demand or monetization must eventually improve.

Allianz Research’s March 2026 analysis estimated hyperscaler capex at approximately $575 billion for 2026, about 50% higher than its comparison period, while its monitor still classified bubble pressure as moderate. That combination captures the ambiguity: extraordinary spending raises the burden of proof without establishing that the entire buildout is uneconomic.

Individual filings show whether a company can carry that burden. Meta’s 2025 Form 10-K recorded $115.80 billion of operating cash flow, $72.22 billion of capex and $43.59 billion of free cash flow, then projected approximately $115 billion to $135 billion of capex for 2026. It also disclosed $103.77 billion of leases not yet commenced and $131.05 billion of contractual commitments, illustrating why current cash flow and future obligations must be assessed together.

Green means operating cash covers investment while utilization or related revenue improves. Amber applies when capex exceeds free cash flow without comparable gains for two reporting periods. Red combines that gap with rising commitments, cancellations, falling utilization or dependence on new financing.

4. Financing conditions: Who carries the downside?

The same project has different implications when funded from retained earnings, equity or leveraged debt. Track net debt, interest coverage, refinancing dates, lease guarantees, customer concentration and the quality of collateral behind data-center financing.

The Bank of England’s July 2026 Financial Stability Report found that AI companies’ use of public debt, private credit, leveraged finance and structured finance had accelerated in the first half of 2026. It judged immediate systemic risk contained by the still-modest stock of outstanding debt, but warned that growing external financing, complex structures and declining hyperscaler free cash flow could broaden the transmission of a future shock.

Score financing green when investment is funded mainly by operating cash and coverage remains stable. Amber means debt or long-term lease commitments rise as cash flow weakens. Red appears when refinancing difficulty, weak guarantees, maturity mismatches or forced asset sales occur together. This indicator distinguishes a sector valuation correction from a credit event capable of spreading beyond equity holders.

5. Obsolescence: Can equipment recover its cost before replacement?

Older AI servers remain operational but move to lower-value workloads as newer equipment takes over demanding tasks.

AI hardware can remain operational after losing access to the highest-value workloads. Faster accelerators or more efficient systems may push older capacity toward cheaper tasks before its accounting life ends, so physical durability and economic usefulness are not equivalent.

Monitor utilization by hardware generation, prices for older compute, impairment charges, changes in estimated useful lives and replacement frequency. Green means cash generated by the installed base covers depreciation and replacement. Amber means refresh cycles are shortening relative to accounting lives; red means impairments, idle capacity or falling rental prices show that assets cannot recover their cost.

Ricardo Caballero’s June 2026 MIT model explains why this distinction matters beyond individual machines: temporary overvaluation can finance productive capacity that survives a later price correction, but the transition fails if the correction arrives before enough sustainable capital is installed. The technology can therefore create lasting economic value even when investors who financed the peak earn poor returns.

What the combined signals say

Do not average the colors mechanically. Red valuation and obsolescence signals may imply severe losses for particular asset owners without threatening the financial system. Simultaneous deterioration in cash earnings, capex coverage and financing conditions is more consequential because it connects disappointed expectations to balance-sheet stress.

The five indicators separate outcomes that bubble headlines collapse into one question: AI can remain useful, infrastructure can become abundant and investors can still overpay. The strongest warning is a widening gap between cash returns and expectations, financed on increasingly fragile terms while the assets lose economic value faster than assumed.

Is AI a Bubble? Five Indicators Matter More Than the Headline