Hyperscaler AI capex: $266 billion in bonds

Hyperscaler AI capex has shifted markets: from technology company balance sheets to the bond market, and from there to the macro level.

1847: When industrial capital shifts from profits to credit

In October 1847, the Bank of England suspended the Bank Charter Act of 1844 to keep the credit system functioning.

The engine of the crisis was a boom in railway investment. British rail lines were absorbing capital at an accelerating rate, and financing shifted from retained earnings to capital calls and bank deposits. When rates rose, debt service broke the upstream chain with creditors.

The mechanism sat in three words: capex, leverage, maturity. The central bank intervened at the end, when risk had already passed from companies to the banking system.

This sequence has a recognizable signature.

It repeats in Frankfurt in 1992, with the Bundesbank defending the Deutsche Mark while post-reunification credit tightened. The sector changes, the century changes. The structure remains identical.

$266 billion: the number that defines 2026

Investment-grade US dollar bond issuances tied to AI capital expenditure reached $266 billion since the start of 2026, according to a JPMorgan strategy report from September 9, 2026 covered on September 11^[1]. The total includes $182 billion from hyperscalers, $42 billion from data centers, and $42 billion from semiconductor companies.

Historical comparison matters more than absolute levels.

All of 2025 closed at $139 billion. 2024 was worth $29 billion. Nine times in two years, with hyperscalers rising from $17 billion in 2024 to $93 billion in 2025 and $182 billion this year.

A slope like this on a single balance sheet item has few precedents in postwar American corporate markets. J.P. Morgan Asset Management documents the same debt dynamics among computing operators in a dedicated note^[2]. Joyce Chang, the report's author, writes that investment booms rarely end in an orderly fashion.

From cash flow to credit markets

The critical point concerns the source of funds.

The bank estimates total cycle requirements at $5.5 trillion. Of this, $1 trillion comes from operating cash flow, $400 billion from new equity, $300 billion from structured products markets, $2.1 trillion from investment-grade markets, and $400 billion from leveraged finance. An additional $1.4 trillion remains to be covered by alternative capital.

The high-quality bond market thus becomes the primary external source, with $2.1 trillion expected over the next five years. The cost of capital for these infrastructures now depends on credit conditions more than on computing demand.

The report also signals a crowding-out effect. Long-dated issuances by hyperscalers displace other investments and push real rates higher, while leverage and momentum amplify corrections.

Risk changes hands

This is a transfer of risk, more than an investment boom.

Technology companies convert industrial risk, the uncertain return of a data center over ten years, into credit risk distributed among insurance companies, pension funds, bond funds, and bank balance sheets. The substance of the risk remains identical. What changes is who holds it during stress.

The ultimate counterparty in this chain is the central bank. In 2008 the Federal Reserve bought commercial paper; in 2020 it expanded intervention to corporate bonds and bond ETFs through two dedicated facilities.

Each time the sequence has been identical: first spreads widen, then primary market illiquidity, finally public backing. Three episodes (1847, 2008, 2020) are enough to call it a pattern. The category is structural, with a multi-decade horizon.

This desk's position

The thesis is explicit: the AI cycle's breaking point lies in the debt financing chips, not in technology stock valuations.

Consensus looks at technology sector multiples. I instead focus on liability structures: maturities, covenants, concentration among buyers of that paper. Equity absorbs losses through price decline; credit transmits them through liquidity channels, and transmission runs faster.

The BIS and International Monetary Fund have converged for months on the same warning regarding leverage supporting computing infrastructure. Concentration of private credit on this topic remains the less-mapped channel of the two.

I would change my view facing two measurable facts. First: a shift in financing back toward internal cash flow, with bond issuance falling for four consecutive quarters. Second: spreads in the data center sector stable below the five-year average.

Three implications for capital

Three implications for capital, each with its own horizon.

First, 12-month horizon. For family offices and sovereign wealth funds, the relevant risk lies in indirect bond exposure: a generic investment-grade fund today contains a growing share of paper tied to computing. It's worth measuring this by issuer, beyond sector analysis.

Second, 24-month horizon. For CEOs and boards, cost of capital rises even for those staying out of AI, because crowding-out acts across the entire curve and hits those refinancing in 2027 and 2028. Industrial plans built on 2024 real rates need redoing now.

Third, 36-month horizon. For the chief risk officer, the scenario missing from VAR models is primary investment-grade market shutdown with sector concentration exceeding 20%. Historical variance describes a market of diversified issuers, while today's market rests on a handful of names.

The forecast

By December 31, 2027, annual investment-grade US dollar issuance tied to AI capex exceeds $400 billion. Over the same period at least one major sector issuer (a hyperscaler or data center operator) suffers a downgrade of at least one notch from Moody's or S&P.

Confidence: 65%. Horizon: 473 days, namely December 31, 2027. Verification: quarterly high-grade issuance data and rating actions published by both agencies.

The signal that disproves the thesis is equally precise. If issuance related to this spending closes 2027 below $250 billion, with hyperscaler ratings flat or improving, the thesis falls entirely.

What to watch

What to watch: three leading indicators drive direction over the next four quarters.

• The spread between hyperscaler bonds and ten-year Treasuries.
• The share of private credit allocated to data centers and computing infrastructure.
• The ratio of announced capex to operating cash flow in quarterly reports from the five largest issuers.

This is a regime change, more than a cycle. The market has already priced in growth from computing-related revenues, while pricing little the map of who holds the debt financing it. The divergence between AI stock prices and credit quality supporting them always resolves: the question is how.

Debt-financed capex follows a constant historical rule. Creditors discover actual asset returns last, and pay first.