The AI Credit Signal: Eight Debt Indicators That Move Before AI Stocks Do

If the AI buildout starts to slow because funding gets harder, the bond market will usually show it before AI stock prices do. In September 2026, some of those debt signals moved at once.

This guide explains the AI credit signal: reading the buildout through what lenders charge rather than through equity prices. If you are exposed to the AI trade as an investor, a vendor, an employee or a buyer of compute, the payoff is practical: eight indicators to track, what normal looks like for each, and what counts as real deterioration rather than noise.

The mental model, and why the buildout became a credit story

For years the hyperscalers funded data-centre construction out of operating cash flow. They no longer do. Vanguard counted five hyperscalers issuing roughly $35bn of debt a year on average between 2020 and 2024, then $93bn in 2025, then about $132bn in the first seven months of 2026.

PIMCO supplies the arithmetic that forced the change: “Capex is now expected to absorb 94% of cash flow from hyperscaler operations in both years, versus just 40% in 2023, a sharp inflection that has fundamentally altered the funding equation.”

When capital spending absorbs almost all the cash you generate, the next dollar of construction comes from a lender. That makes credit pricing a useful early check on how willing lenders are to keep funding new projects.

There is another valid reading. Vanguard notes that “Railroads, electrification, and telecommunications, each of which changed how we lived and worked, were all substantially financed by debt. In that sense, the AI buildout arriving in the bond market is not a warning sign, but an indicator that the investment cycle is maturing.”

Key terms

Five basics first. Investment grade means debt rated comparatively safe; high-yield, or junk, means lower-rated debt that pays more to borrow. A basis point (bp) is one hundredth of a percentage point, so 100bp is 1 point. A spread is the extra yield a bond pays over a benchmark such as a government bond or a broad corporate index. Capex is capital spending on long-lived assets. Hyperscalers are the largest cloud and AI infrastructure buyers.

Three more terms carry the specific signal.

Powered shell. A developer builds the shell and the power, a hyperscaler signs a long lease, and the developer issues bonds repaid out of the rent. Penn Mutual Asset Management describes the structure investors “grew comfortable with.” What investors are really judging is the tenant’s credit, not the developer’s. The AI Daily Brief put that more sharply: “These deals have allowed hyperscalers to effectively rent their credit rating to smaller data center developers.”

Shadow borrowing. Off-balance-sheet leases, ring-fenced project vehicles and private credit. The BIS Quarterly Review names these structures directly: “These arrangements amount to “shadow borrowing”: obligations that are economically akin to debt but largely reside outside corporate balance sheets.” Separately, J.P. Morgan Asset Management counts roughly $1,400bn of disclosed data-centre lease obligations, about $1,100bn of it off balance sheet until the leases commence.

Excess bond premium. The part of corporate spreads not explained by expected defaults. Gilchrist and Zakrajsek showed in the American Economic Review that shocks to that non-default component lead declines in activity and asset prices, and the Federal Reserve publishes a monthly estimate plus a model-implied recession probability.

What the market did in September 2026

A single week does not prove a trend, but it can show what to watch. The third week of September 2026 put several indicators in motion at once.

A debut junk deal for a Meta-leased Georgia data centre priced at almost $2.28bn and 8.25%, and still drew about $10bn of orders, according to Bloomberg’s report. The facility is leased to a Meta subsidiary, which guarantees rent and operating expenses.

That same week, Yahoo Finance relayed an FT report that Oracle-linked data center loans totaling about $18 billion were trading at 89 to 91 cents on the dollar amid permitting and construction delays.

Practitioners read those deals differently. Investor Meltem Demirors posted that she was hearing banks stop compute lending and that a credit crunch was beginning, then added the part that matters more: “The long tail is drying up very quickly. Extreme divide between haves and have-nots.” Investor Jigar Shah pushed back: “Folks are freaking out about debt spreads widening for AI data center build-outs. No need to freak out.” His counter is that many sub-25-megawatt data centres can be built 50% cheaper on 120-day timelines.

Those are attributed opinions, not market facts. The deal data supports something narrower: funding is getting more expensive at the margin, but strong tenants can still raise money. As J.P. Morgan Asset Management put it, credit markets are “repricing, not rejecting, the AI buildout.” The AI Daily Brief’s summary of the week was that for signs the AI trade is rolling over, “this rather than stock prices is the corner of the market to keep an eye on.”

The eight indicators to watch

Use these as a checklist, broad market measures first. Each entry gives the baseline and what deterioration looks like.

1. The broad high-yield bond spread. The ICE BofA US High Yield Index option-adjusted spread measures the extra yield riskier corporate bonds pay over safer ones. It is free on FRED, the St. Louis Fed’s data site, as series BAMLH0A0HYM2. Baseline: a market weather report, not an AI-only number. Deterioration: it widens for weeks while AI-linked debt widens more than the market as a whole.
2. Hyperscaler spread versus the broad investment-grade index. Baseline from Penn Mutual: roughly 20 to 25 basis points in early 2025, about 35bp by July 2026, having reached nearly 50bp in the first half. Deterioration: the gap stays above 50bp or keeps widening while issuance continues.
3. High-yield data-centre spread versus the high-yield index. Baseline: about 220bp at the December 2025 peak, a sharp compression in the first half of 2026, then re-widening from June. Deterioration: a return toward 220bp while index spreads sit flat.
4. Data-centre bond spread versus the tenant’s senior spread. BlackRock put the gap at “over 100bps” for both investment-grade and high-yield data-centre bonds in its 12 August note. Deterioration: past roughly 150bp, a sign investors are assigning more risk to the lease or the construction timeline.
5. New-issue concessions and order-book coverage. A new-issue concession is the extra yield a borrower offers to get a fresh deal sold; order-book coverage is demand, or dollars of orders per dollar of bonds offered. Goldman Sachs recorded concessions moving from “two, three basis points” to as much as 20 basis points on a very large hyperscaler deal. Baseline: healthy deals clear with modest extra yield and several times as much demand as supply. Deterioration: books below 2x, or pulled deals - the strongest single signal.
6. Migration into private and off-balance-sheet structures. Some financing is shifting into leases, project vehicles and private deals that are harder to track than public bonds. Part of it is visible in company 10-Qs, the quarterly reports US public companies file with the SEC. Baseline: normal in project finance. Deterioration: more borrowing moves into private structures while disclosure thins.
7. Rating actions and the fallen-angel line. A fallen angel is an issuer downgraded from investment grade to junk; index-tracking funds can be forced sellers when a bond leaves the index. Baseline: isolated outlook changes are noise. Deterioration: a major AI-linked borrower is cut below investment grade, or several issuers go on negative outlook at once. The AI Daily Brief reported Oracle as already at risk of being downgraded.
8. Real-economy cross-checks. JLL’s mid-year EMEA report puts 74% of under-construction capacity as pre-let, meaning leased before completion, with sub-megawatt rents pushing through EUR177 while wholesale space eases from 2025 peaks. Deterioration: pre-lets falling and smaller formats softening too, which would weaken the argument that small, fast builds offset stress in large campuses.

The reading rule matters as much as the list: any single indicator can be noise; the signal is three or more moving the same way in the same quarter.

Two things would make September look like noise instead of a turn: AI-linked issuance clearing at normal concessions with order books staying strong, and project-level spreads stopping their widening relative to the broad high-yield market. Credit can also widen for reasons unrelated to AI, such as a general risk-off move, and equity prices can move first when a shock is regulatory or competitive rather than financing-led.

Common misconceptions

“This is a bubble call.” It is not. The same evidence base holds the bull case.

“It tells you when.” It claims directional priority, not calendar precision. Spreads widened, compressed and re-widened inside 2026 alone. A widening can run for months without a bust.

“There is one number to watch.” Every total here is definitional: some count hyperscaler corporates only, others add utilities, chipmakers or private vehicles. Do not average them.

“Widening spreads mean AI is unprofitable.” Revenue growth and credit stress can coexist. The open question is whether cash flow can carry the capital spending.

“Repricing” and “crunch” are the same thing. They are not. The Richmond Fed traces the term to 1966, when two Salomon Brothers economists coined it to describe how that year’s tight-credit episode differed from the 1950s.

Done means

You have this working when:

• You can name the eight indicators and where each one is published.
• Every spread, coupon or yield you repeat carries a date and a source.
• You can state the difference between a repricing and a crunch in one sentence.
• You have picked two indicators to check monthly, not all eight daily.
• You treat a practitioner post as a claim to verify, not as a data point.

What this article does NOT cover

It does not forecast AI stocks or recommend a trade, and it does not reconcile the divergent totals for AI-related debt, because no definition-compatible total exists. It does not name a credit crunch, assert that any hyperscaler is distressed, or claim a bust is imminent. Re-verify any figure here before repeating it: spreads, coupons and order-book cover ratios are dated observations, not standing facts.

Related guides

• Pacing the Frontier - the policy position this financing pays for
• When Labs Control the Model and Harness - enterprise exposure if compute financing tightens
• AI Operator Shift - why operator judgement is the bottleneck
• Loop Engineering - the other half of the AI cost story

Research basis: shared/abs-research-briefs/aidb/ai-credit-signal/ai-credit-signal-research-2026-09-21.md (verified 2026-09-21).