The Bond Market Gets the Bill: Goldman Counts $500 Billion of AI Debt in 2026 Alone

The AI boom has found its invoice. On September 30, the Financial Times reported that Goldman Sachs strategists now count roughly $500 billion of AI-linked debt issuance so far in 2026 — and expect the total to top $750 billion by year-end and approach $1.2 trillion in 2027. The same week, Bloomberg reported that Goldman’s own asset management arm has turned underweight hyperscaler credit, citing the coming surge of debt supply.

For two years the AI capital-expenditure story was told in equity terms — mega-rounds, soaring market caps, a $1.4 trillion OpenAI valuation rumor. The Goldman tally makes official what credit desks have been watching all year: the buildout is now being financed substantially with borrowed money, and the bond market is starting to price what equity investors have mostly waved through.

The numbers

The core figures, from Goldman Sachs Research:

• ~$500 billion: AI-related debt issued globally in 2026 year-to-date. Hyperscalers account for only about 40% of it — the rest comes from the broader AI ecosystem: chipmakers, data-center operators, energy projects, and the financing vehicles built around them.
• $194 billion: hyperscaler debt issued so far in 2026, nearly double the $108 billion the group produced in all of 2025.
• ~$250 billion: expected direct hyperscaler bond supply for full-year 2026, with roughly one-third of hyperscaler capex debt-financed this year, rising to 35% at the 2027 peak.
• ~$1.2 trillion: projected AI-linked issuance in 2027, with the flood now spreading from US dollars into euro and Canadian dollar markets.

Behind those aggregates sits a simple mechanical fact that Amanda Lynam, Goldman’s head of credit strategy research, laid out in the firm’s Exchanges podcast: capex and cash flow from operations at these companies are converging. The internal cash engine that used to fund expansion is no longer sufficient, so companies that spent years hoarding investment-grade balance sheets are now deliberately spending that ratings headroom.

Why cash-rich giants are borrowing anyway

The obvious question — why would companies sitting on tens of billions in cash issue debt? — has a colder answer than “they need the money.” As Lynam frames it, hyperscalers are running a “waterfall of capital”: exhaust internal cash flow, then debt, then equity, positioning for a multi-year investment cycle whose endpoint nobody can price. Because the sector spent decades under-levered and highly rated, there is enormous runway to add debt while staying comfortably investment grade — Goldman’s exercise putting the group at two times net leverage implies around $2 trillion of incremental debt capacity.

The binding constraint, in Goldman’s view, is not what hyperscalers can issue. It is what the US investment-grade market can absorb. Compare the three largest US bank issuers — none carries more than $175 billion of index-eligible debt or 2.3% of the index. Lift the hyperscalers to that same ceiling and you get roughly $510 billion of additional capacity in the US IG market alone. After that, the supply has to go somewhere else.

That “somewhere else” is already visible. The ECB documented in an August 31 blog post that US big tech now accounts for just shy of 10% of gross new euro-denominated non-financial corporate issuance, with Amazon and Alphabet the largest issuers in the euro market this year — the Amazon deal setting an all-time size record. Meta is now preparing its first-ever euro bond, after a $25 billion six-tranche dollar deal in April. Reverse-Yankee issuance by hyperscalers almost doubled between 2025 and 2026, and the euro already makes up close to 10% of their outstanding bonds.

The market is already pushing back

The most telling evidence that this is not a costless exercise came from Zach Ablon, who runs the credit sales desk at Goldman. The firm’s AI leadership bond basket has widened from 74 basis points to nearly double that over twelve months — the market’s word for indigestion. In Q1, roughly 15 insurance clients were coming into 30-year AI-linked paper with $50-million-plus orders; by late Q2, about half that. AI-related issuance has gone from 1% of IG supply in 2024, to 7% in 2025, to about 18% this year — and 40% of all 15-year-plus IG issuance now comes from the AI theme. Duration weight is shifting accordingly: Amazon is now the highest duration weight in the IG index (it ranked 20th a year ago); Google jumped from 86th to 18th.

Reuters reported the same dynamic on September 22: corporate bond buyers are getting picky, demanding richer concessions as gross hyperscaler issuance is expected to hit a record $420 billion next year, up 60% from 2026 estimates.

There is a legitimate bull case, and the ECB makes it: hyperscalers are lengthening maturities, lifting average ratings (often AA- or higher in a market dominated by A/BBB), and diversifying euro credit benchmarks that carry roughly a third of the tech weight of US indices. Issuance of this scale is itself evidence the euro market can absorb giant deals — a capability it was historically doubted to have.

But the ECB also flags the risk that matters most: AI is a new sector, and rating methodologies rest on assumptions about future revenue growth that “may not stand the test of time” — a polite way of saying nobody actually knows if $1-trillion-a-year capex earns its cost of capital. Credit investors, who don’t share in the equity upside, are the ones being asked to hold that risk.

Why this matters beyond the bond desk

Three implications worth tracking:

The AI debate just moved to fixed income. Equity markets can hold multiple contradictory AI narratives for years. Credit markets clear at a price every day. Goldman going underweight hyperscalers — while its research arm simultaneously says there’s no access-to-capital problem — is the kind of disagreement that surfaces first in bond spreads.

The financing tail is wagging. When a third of hyperscaler capex is debt-funded, the marginal buyer of AI risk is no longer a venture capitalist or a retail index fund — it’s an insurance portfolio in Hartford or a pension fund in Frankfurt with concentration limits and a mandate to get paid back. Their appetite, not technology roadmap, now sets the pace of the buildout.

Policy exposure is rising. Meta’s federal tax bill reportedly fell 71% in 2025 after it classified AI data centers as experimental research eligible for immediate write-offs — and three US senators have already demanded accounting from Meta, Amazon, Alphabet, and Microsoft. When the debt tally, the tax treatment, and the midterm politics of “who pays for the AI buildout” converge in the same news cycle, the financing layer becomes the political layer.

The Goldman numbers do not say the AI capex boom is a bubble about to burst. They say something more precise: the boom has become a credit event in progress — $500 billion committed this year, $1.2 trillion likely next, absorbed by markets that are already charging more for it, in currencies the issuers never used before. The infrastructure being financed may or may not earn its return. The interest payments start regardless.