From Capex to Credit: The AI Trade Faces Its Financing Test

For three years the AI trade was an equity story: demand for compute, record earnings at chipmakers, and hyperscalers that could fund their capital expenditure from operating cash flow. In the autumn of 2026 it became a credit story. The question investors are now asking is no longer whether AI demand is real, but who pays for the data centres, at what cost of capital, and over which horizon.

What Happened

The repricing arrived in two waves. In October, technology stocks lost about 7%, communication services about 15% and semiconductor stocks about 15%, while valuations of many AI chipmakers had already fallen more than 30% from their mid-year levels. The second wave came after the U.S. midterm elections on 3 November. On 5 November the S&P 500 fell 4.4%, its worst session since April 2025, and the Nasdaq slipped into correction territory. Some chipmakers lost up to a quarter of their value in the first week of November, and power companies tied to data-centre projects fell almost as sharply as the chips themselves.

The political trigger was specific. Democrats won control of both the House and the Senate, and investors moved quickly to price a slower pace of data-centre construction and tighter federal oversight of AI. Developers and utilities with the highest exposure to new capacity were sold first; parts of healthcare and clean energy rallied.

Why the Credit Market Matters More Than the Chip Market

The more telling signals came from fixed income. During the 5 November selloff, U.S. high-yield spreads widened to around 348 basis points, and to roughly 367 basis points for single-B borrowers. Individual data-centre bonds moved far more: notes issued by one data-centre project in late July with a 9.875% coupon traded at about 89 cents on the dollar on 5 November, lifting the yield to roughly 13%.

Even the strongest balance sheets are paying more. Bank of America estimates that ten-year hyperscaler spreads have widened by about 50 basis points in 2026, against roughly 20 basis points for comparable high-quality industrial issuers. Lenders are stress-testing data-centre cash flows and re-reading the clauses that allow large tenants to reduce or cancel capacity commitments, and some AI-related bond plans have been shelved.

The reason is supply. Outstanding hyperscaler debt had already doubled to around $460 billion between the end of 2024 and the summer of 2026, according to Moody's data reported by IFR, with lease liabilities tied to data centres growing faster still. Goldman Sachs Research expects global hyperscaler issuance to rise from an estimated $250 billion in 2026 to $420 billion in 2027. A build-out that was once self-funded is now competing for capital in a market where the ten-year Treasury yield touched roughly 5.3% at the end of September, its highest level in nearly two decades.

What Has Not Changed

It would be a mistake to read the selloff as a collapse in AI demand. Third-quarter U.S. GDP grew at an annualised rate of just under 3.5% in the first estimate, and the investment cycle behind it remains intact; several large spenders have said they intend to keep investing through 2027. What has changed is the price of the capital that funds the next leg, and the length of time investors are willing to wait for returns. When duration becomes expensive, the assets whose value depends most on distant cash flows are repriced first.

That is also why the repricing is uneven. The financing squeeze lands first on borrowers without diversified cash flows of their own: standalone data-centre developers, specialised cloud providers and the power and equipment suppliers whose order books depend on projects that still need funding.

The Concentration Problem

The episode also exposed how concentrated U.S. equity indices have become. In October the cap-weighted S&P 500 lost about 1%, while the average stock in the index gained about 5% and roughly three quarters of its members outperformed it. Healthcare, consumer staples and utilities each rose about 10% to 12%. An investor holding the index was, in practice, holding a leveraged view on a small number of AI-linked balance sheets.

For portfolio construction, that has two consequences. First, index-level volatility now understates single-theme risk: a portfolio can look diversified by count of holdings and still be dominated by one financing cycle. Second, correlations within the AI complex rise sharply in drawdowns. Chipmakers, power utilities, data-centre developers and their lenders moved together in November in a way that a backward-looking correlation matrix would not have predicted.

What We Monitor

Hyperscaler spreads relative to the investment-grade index. If the gap keeps widening while the broad index is stable, the market is pricing issuer-specific supply, not systemic stress.

New-issue concessions on high-yield data-centre debt. Discounts to par and amortisation features are the earliest signs that marginal projects are being priced out.

Breadth. A recovery led only by the largest names would leave concentration risk unresolved; a broadening recovery would suggest the repricing has done its work.

Implications for Systematic Investors

Narrative-driven markets are hard to time and easy to misjudge. A systematic process does not need to decide whether AI is a bubble; it needs to measure when the risk of holding equities rises above a defined budget. Value-at-Risk models respond to the jump in realised volatility within days, momentum signals register the break in trend, and mean-reversion signals help distinguish an orderly correction from a disorderly one. SPY3's design ranks a 99% Value-at-Risk veto above its return signals for exactly this reason: when tail risk rises, exposure is cut first and the reasons are evaluated afterwards.

The broader lesson is about the source of risk. The AI build-out is turning from a story about technology into a story about leverage and refinancing. That shift rewards investors who size positions to measured risk rather than to conviction, and who treat the cost of capital as a first-order input rather than a footnote.