Is AI a Bubble? Watch Cash Flow, Capex and Credit

AI investment is not one uniform bubble. The evidence supports a real technology buildout with uneven financial risk: some companies can fund expansion from current cash generation, while others require distant monetization, generous valuations or continued access to credit.

The useful question is which infrastructure, model and application companies are priced and financed as though execution cannot fail. A five-part worksheet—cash flow, capex payback, monetization, valuation and credit dependence—makes that risk measurable without pretending one verdict fits the entire AI stack.

Score companies, not the AI narrative

Create one row per company and score each of the five indicators from zero to two. Zero means current operating evidence supports the investment case; one marks mixed or deteriorating evidence; two means the case relies heavily on outcomes that have not appeared in cash flow or customer economics.

0–3 points: current fundamentals broadly support the investment cycle.
4–6 points: real demand exists, but execution, financing or valuation risk is material.
7–10 points: the investment case depends heavily on future cash flows, refinancing or sustained high valuation multiples.

These ranges are an editorial screening heuristic, not a validated trading model. Compare a company with peers in the same layer before comparing it with the rest of the stack: an early-stage model developer will naturally have different margins and funding needs from a mature cloud provider.

An arXiv diagnostic preprint supports this segmented approach. It finds genuine revenue, adoption and productivity evidence alongside localized fragility, including layers where capital expenditure has grown faster than observed monetization.

1. Start with cash flow and earnings quality

Score zero when operating and free cash flow are positive, recurring and sufficient to fund ordinary operations plus most AI investment. Score one when the business generates cash but capex is absorbing a rising share of it. Score two when operating losses, weak cash conversion or persistent working-capital demands leave the company reliant on outside funding.

Then inspect what produced the revenue. Repeat usage, renewals and cash collection provide stronger evidence than nonbinding capacity reservations, promotional credits or transactions among strategic partners. Stock-based compensation is noncash, but persistent dilution still matters because it can transfer a substantial share of operating gains away from existing investors.

Fidelity’s five-factor review found that aggregate technology capex remained below free cash flow in January 2026, unlike the prolonged cash-flow deficit before the dot-com crash. Fidelity nevertheless identifies shrinking free cash flow, deteriorating leverage and wider credit spreads as warning signs, so the aggregate result does not protect a weak individual issuer.

2. Estimate capex payback for each layer

Large capex is not automatically wasteful. The relevant question is whether incremental cash gross profit can repay the investment before equipment becomes obsolete, customer demand changes or financing must be renewed. A practical estimate divides incremental capex by the annual incremental cash gross profit attributable to the added capacity; use a range when attribution is uncertain.

For infrastructure companies, examine utilization, contracted capacity, power availability and the duration and credit quality of customer commitments. Score two when credible payback extends beyond the likely economic life of the assets, or when projected utilization depends on customers that have not made enforceable commitments.

For model developers, include training investment, inference cost per unit of paid usage and the cost of serving free demand. For application companies, replace physical capex with the broader product investment needed to generate revenue: model fees, engineering, customer acquisition and implementation. Improving gross margin and retention as paid usage grows are stronger signals than user growth alone.

Allianz’s capex-cycle analysis put hyperscalers’ planned 2026 capital expenditure at about $575 billion, approximately 50% above the previous year, while its risk monitor still classified overall bubble pressure as moderate. The combination matters: the spending scale raises the required future return, but it does not establish that every project or company is speculative.

3. Match valuation to visible monetization

Valuation earns a high-risk score when the current price can be justified only by an unusually large market share, exceptional margins and little future dilution. Compare enterprise value with revenue, gross profit and free cash flow, using peers at a similar development stage and the company’s own valuation history. Avoid comparing an asset-heavy data-center operator directly with a high-margin software vendor on revenue multiples alone.

Monetization evidence also changes by layer. Infrastructure requires paid utilization and acceptable returns on installed assets. Model providers require durable paid volume after inference costs. Applications require renewals, expansion and demonstrated willingness to pay rather than trials or engagement that produces no economic return.

Stress the valuation by reducing expected revenue growth, gross margin and the terminal multiple together. Score two if a plausible slowdown removes most estimated equity value or pushes capex payback beyond the relevant assets’ useful economic lives. That result does not prove fraud or technological failure; it shows that the market price allows little room for normal execution error.

4. Use credit as an independent check

Equity investors can tolerate distant returns; creditors still require fixed payments. Track net debt, interest coverage, maturity schedules and the spread over comparable government or corporate bonds. Rising debt-funded capex is not conclusive on its own, but wider spreads combined with weaker free cash flow indicate that lenders are charging more for risks equity prices may not reflect.

Include obligations that sit outside the headline debt figure, such as equipment leases, purchase commitments, minimum cloud contracts and guarantees. For private companies without publicly traded bonds, examine financing terms, collateral and whether new funding expands proven demand or merely covers continuing operating cash burn.

In a May 2026 speech, Federal Reserve Governor Lisa Cook’s assessment said hyperscalers had recently used investment-grade bond markets for AI capex, while smaller data-center developers were raising private and asset-backed debt. She described many large investors as strong borrowers but warned that sustained debt issuance for an emerging technology could eventually create a financial-stability concern.

Read the five indicators as a cluster

No single metric settles the bubble question. An infrastructure provider can have contracted demand but a payback period vulnerable to obsolescence; a model company can grow paid usage while remaining dependent on external capital; an application can produce attractive cash returns despite using commoditized models.

The strongest bubble warning is therefore a cluster: capex rising faster than cash gross profit, payback extending beyond a credible asset life, weak paid retention, valuation requiring near-flawless growth and credit becoming more expensive. Strong cash conversion, improving unit economics, defensible payback and mostly self-funded investment point instead to a productive buildout, even if individual prices remain vulnerable.

Update the worksheet when earnings, financing terms or customer commitments materially change. Its purpose is not to attach a permanent label to AI, but to identify where financial expectations have moved further than operating evidence.

Is AI a Bubble? Watch Cash Flow, Capex Payback and Credit—not Hype