AI share of long-dated IG supply
Nearly50%
US non-financial IG gross supply with maturities beyond 15 years, year to date; PIMCO analysis published Nov 10, 2026
AI share of 15-year-plus IG issuance
40%
2026 year to date, issuers funding the AI theme; Goldman Sachs, Aug 5, 2026
Hyperscaler bond supply, 2026 estimate
$279B
Up from $109B in 2025 and $17B in 2024; J.P. Morgan Asset Management, Aug 5, 2026
AI credit issuance forecast, 2027
$685B
Up 31% from 2026, including public and private debt; Bank of America estimate reported Nov 6, 2026
CoreWeave 2032 bond yield
Above 16%
9.625% coupon; five-year CDS near 1,155bp after a spike above 1,289bp, Nov 6, 2026
The primary market
Half of Everything Long-Dated Is Now an AI Credit Decision
The striking number in the AI credit cycle is no longer the total amount borrowed. It is where the borrowing lands. PIMCO's November 2026 analysis of Bloomberg data shows AI-related borrowers taking nearly half of year-to-date gross US non-financial investment-grade supply in maturities beyond 15 years, and Goldman Sachs independently puts the AI share of 15-year-plus IG issuance at 40% in 2026 year to date. Long duration was once the diversifier inside an IG portfolio; it is now the concentrated expression of one trade.
The supply behind that share is not marginal. Hyperscaler corporate bond issuance went from $17B in 2024 to $109B in 2025 and is tracking toward $279B in 2026, with J.P. Morgan Asset Management estimating another $220B–$300B in 2027. Goldman's broader tracker puts total AI-related debt issuance near $500B so far in 2026, of which hyperscalers are only about 40% — the rest sits with data center developers, utilities and the supply chain.
Hyperscaler corporate bond issuance: $17B to a projected $279B in three years
J.P. Morgan Asset Management estimates hyperscaler bond issuance of $279B for 2026 and $220B–$300B for 2027. The 2027 bar uses the midpoint of that range.
Unit: $ billions
2024
Actual
17
2025
Actual
109
2026E
Estimate
279
2027E
Midpoint of $220B–$300B
260
Who is absorbing it
Duration Holders Did Not Volunteer for the Risk They Now Own
- Hyperscalers placed 41% of their 2025–2026 issuance beyond 15 years, against 16% for the ICE BofA US Corporate Index — the long end of IG is being re-weighted by a handful of issuers.
- Technology's weight in the Bloomberg US Corporate Index rose from 2.67% at end-2024 to 4.85% by early August 2026, with Amazon the largest single name at 1.39%.
- Oversubscription on hyperscaler deals averaged 3.5x–4.5x in the first half of 2026, then collapsed to roughly 1.4x on the first surprise deal of the second half — the marginal buyer has thinned.
- New-issue concessions widened to as much as 20bp on a very large hyperscaler deal, versus two to three basis points a few years earlier.
- Even the long tail is being bought: Alphabet's $1B 100-year sterling bond was roughly 10x oversubscribed at about 120bp over 10-year gilts, anchored by pension and insurance money.
- Technology spreads now trade wide of the broad IG benchmark — a reversal from their historic status as index leaders.
Read together, these are the fingerprints of a market where the marginal buyer of duration has changed. Index funds and liability-driven investors are the ones who own the risk whether or not they have an AI view, because new AI paper mechanically replaces maturing paper in the benchmark they track. Goldman Sachs frames the structural version of the same point: hyperscalers could approach 10% of the US IG market by 2030, and AI-related bonds typically carry maturities three to four years longer than the average issuer's.
The borrowers' own books
The Leverage Is Real, but It Sits Against Cash Flows Most Buyers Ignore
The AI borrowers are not leveraged corporates in the classic sense. Their balance sheets show why their bonds stayed investment grade while the supply grew: the debt is being added on top of extraordinary operating cash flow, and the absolute leverage remains modest relative to revenue for most of them. That is precisely why the market's problem is duration and supply, not near-term default risk.
| Issuer | Total debt | Cash + short-term investments | Latest reported revenue | Reporting document |
|---|---|---|---|---|
| NVIDIA | $38.9B | $99.4B | $96.2B in Q2 FY2027 | 10-Q for the quarter ended Jul 26, 2026 |
| Oracle | $155.9B | $37.1B | $67.4B in FY2026 | 10-Q for the quarter ended Aug 31, 2026 |
| Meta Platforms | $112.1B | $82.4B | Reported Sep 30, 2026 | 10-Q for the quarter ended Sep 30, 2026 |
| Amazon | $271.9B incl. finance leases | $114.2B | Reported Sep 30, 2026 | 10-Q for the quarter ended Sep 30, 2026 |
NVIDIA is the cleanest illustration of how fast this reprices the cost of capital. Long-term debt rose from $7.5B at the end of Q1 FY2027 to $32.4B at the end of Q2 FY2027, a $25B increase that maps directly onto the seven-tranche bond sale priced in June 2026 with maturities as late as 2056. Reuters reported that NVIDIA capped the deal at $25B specifically to keep new-issue spreads tight, with demand of about $85B — a deliberate choice to establish a benchmark cost of credit rather than to raise money it did not need.
- Oracle carries $119.6B of net debt against $67.4B of fiscal 2026 revenue, and on Nov 2, 2026 said it would raise a further $20B through bond and share sales in six tranches maturing between 2029 and 2046.
- Meta Platforms grew total debt from $83.9B at the end of 2025 to $112.1B at Sep 30, 2026, and S&P cut its outlook to AA- negative on Oct 30, 2026 as it marketed a $20B–$25B deal split two-thirds dollars, one-third euros.
- Amazon increased total debt including finance leases from $153.0B at the end of 2025 to $271.9B at Sep 30, 2026, and now carries the highest duration weight in the IG index.
- Alphabet keeps funding long-dated paper offshore, including the roughly $1B 100-year sterling issue that pension and insurance buyers absorbed at about 120bp over gilts.
Pricing
Dispersion Is the Trade: Widening AI Spreads, Not a Credit Event
Morgan Stanley's framing for2026 was that AI infrastructure requirements would become an important driver of credit dispersion, with spread performance increasingly sensitive to capex guidance and issuance expectations rather than to default probability. The market has now delivered that dispersion. In the week to Nov 6, 2026, five-year CDS widened76% in a month for NVIDIA to roughly 139bp, 49% for Broadcom, 44% for Amazon and about 298bp for Oracle, while CoreWeave's five-year CDS traded above 1,289bp before easing near 1,155bp.
| Marker | Level | Move or context |
|---|---|---|
| CoreWeave 2032 bond | Yield above 16% | 9.625% coupon, Nov 6, 2026 |
| CoreWeave 5-year CDS | About 1,155bp | Peaked above 1,289bp |
| Oracle 5-year CDS | About 298bp | Level cited Nov 6, 2026 |
| NVIDIA 5-year CDS | About 139bp | Up 76% in one month |
| Amazon 5-year CDS | Not disclosed | Up 44% in one month |
| Broadcom 5-year CDS | Not disclosed | Up 49% in one month |
The forward calendar keeps the supply story alive rather than resolving it. Bank of America expects AI-related credit issuance to rise 31% to $685B in 2027, including public and private debt, before falling 23% to $531B in 2028. Goldman's counterpoint is that a US moratorium on AI infrastructure could strand 45% of planned capacity, which is why the market is unwilling to pay up for paper whose collateral is a data center under construction.
Upstream
The Suppliers of the Buildout Get Paid in Cash, Not in Concessions
The asymmetry in the AI credit chain is that the long-dated paper is issued by the companies buying the infrastructure, while the cash lands with the companies selling it. Those suppliers are largely funding themselves with operating cash flow rather than bonds, which is why they do not need the new-issue concession the issuers just paid.
- Vertiv reported Q3 2026 revenue of $4.31B against $2.68B a year earlier, roughly 61% growth, and ended with $15.0B of backlog at the end of 2025, up 109% on the year, with fourth-quarter 2025 organic orders up 252%.
- GE Vernova told investors in April 2026 that it expected at least 110GW of combined gas turbine backlog and slot reservations on data center demand, reaching 116GW by the second quarter, and it still runs net cash.
- NVIDIA shows the mirror image of a supplier: trailing twelve-month revenue of $303.0B against total debt of $38.9B and only $463M of trailing interest expense.
- SK hynix is funding the memory buildout from operations and equity rather than investment-grade bonds — FY2025 revenue of ₩97.1T with capital expenditure at 29.4% of revenue — and filed with the SEC in 2026 for a US ADS offering.
That is the practical answer to who gets paid for dispersion. Vertiv and GE Vernova sell the turbines, switchgear, cooling and grid equipment that the bond proceeds fund, and their revenue growth arrives as cash. Bondholders, by contrast, absorb the concession. Over the next one to three years, the more AI paper is issued, the more of the buildout's cash flow accrues upstream and the more of the spread bill accrues to whoever bought the paper.
Downstream
Downstream, the Duration Becomes a Refinancing Problem
The other end of the chain is where duration turns from an allocation question into a solvency question. CoreWeave leases compute rather than selling software, so its capital expenditure runs at roughly 279% of revenue on a trailing basis, its trailing free cash flow is negative about $17.8B, and its net debt runs about 17.6x trailing EBITDA against a current ratio of 0.44. It raised more than $30B in debt markets this year and carries roughly $50B of debt outstanding.
- CoreWeave now prints a bond yield above 16% and five-year CDS near 1,155bp, the market's clearest verdict on AI infrastructure debt that depends on refinancing rather than on operating cash flow.
- Oracle is the middle case: $119.6B of net debt against $67.4B of fiscal 2026 revenue, with a $20B combined bond and equity raise announced Nov 2, 2026 and existing 6.85% bonds due 2066 falling more than 1.2 cents on the dollar.
- Meta Platforms and Amazon fund AI capital expenditure from operating cash flow, but both now refinance at the long end, which is where duration ownership becomes a cost line as rates and supply move.
This is why PIMCO's diversification advice is about portfolios rather than about defaults. Agency mortgage-backed securities offer about a 33bp pickup over US IG corporate spreads with less rate sensitivity than in 2023, and dollar EM investment-grade semi-liquid infrastructure deals have picked up a median of roughly 40bp this year. Moving modest allocations into EM external and local debt raised ex-post Sharpe ratios by as much as 0.13 over the past three years. None of that is a forecast of AI failure; it is a hedge against owning one theme through the only asset class that now expresses it.
Horizons and risks
What Moves First, and What Breaks the Thesis
- Days to quarters: new-issue concessions on the next hyperscaler deal, and whether second-half oversubscription holds near 1.4x, will set the marginal cost of AI duration faster than any rating action.
- Days to quarters: Meta Platforms's pricing on the Nov 3, 2026 deal and Oracle's six-tranche sale will test whether euro and sterling buyers still absorb 20-year-and-longer paper from these credits.
- One to three years: Goldman expects hyperscaler supply of roughly $250B of direct issuance and up to 35% of capex debt-financed at peak, which points to a heavier IG index presence, not a lighter one.
- One to three years: Bank of America's forecast of $685B of AI credit in 2027 before a23% drop in 2028 makes 2027 the year the market learns whether supply is structural or cyclical.
- Thesis breaker: a US moratorium could hit 45% of planned AI infrastructure capacity, which would hit suppliers' backlogs before it hit issuers' coupons.
- Thesis breaker: faster-than-expected monetization of AI capacity would compress the duration premium and remove the dispersion trade that now pays the cost of the long end.
Synthesis
Own the Theme Through Equity and Cash Flow, Rent the Duration
Facts: AI-related borrowers supply nearly half of US non-financial IG paper beyond 15 years and 40% of 15-year-plus issuance in 2026; hyperscaler supply runs toward $279B this year; new-issue concessions widened to as much as 20bp; oversubscription fell from 3.5x–4.5x to about 1.4x. Inference: because the marginal buyer of long-dated IG paper is now an index or a liability-driven fund rather than a discriminating credit analyst, the AI theme is expressed through duration, and duration risk is no longer optional for buyers who never chose it.
That points to a specific way to be paid for the theme. The suppliers of power, cooling and silicon convert the buildout into cash without issuing long-dated paper, and the issuers that own the duration are paying a widening concession to do it. In equities, the durable exposure is the cash-flow side of the chain; in credit, the durable exposure is the spread between the best and worst AI credits, not the average of them. Speculation worth flagging: if the capex cycle slows rather than breaks, both the dispersion and the concession disappear together, and the hedges PIMCO recommends stop paying.
Where the AI duration actually lands
- funded $25B of seven tranches out to 2056 and still holds $99.4B of cash and short-term investments
- Total debt of $38.9B against $303.0B of trailing revenue keeps its own credit cost minimal while it sets the benchmark curve for AI duration.
- Its five-year CDS widening 76% in a month to about 139bp shows even the strongest AI credit now prices supply risk.
- Next one to three years: every new benchmark deal it prices resets the cost of capital for the whole AI complex.
- carries $119.6B of net debt against $67.4B of fiscal 2026 revenue and plans a further $20B raise
- Five-year CDS near 298bp and a 1.2-cent drop in the 2066 bonds show long-dated paper is repricing fastest for the most levered issuer.
- Days to quarters: the Nov 2, 2026 six-tranche sale prices whether 20-year paper from a 1.8x-levered borrower still clears.
- One to three years: repeated refinancing at wider spreads turns AI capacity growth into an equity-return problem.
- now carries the highest duration weight in the US IG index with total debt of $271.9B including finance leases
- Its five-year CDS widened 44% in a month, so the same duration that widens the index also widens its own funding cost.
- Days to quarters: demand for its long-dated paper is the cleanest read on whether insurers and pensions still absorb AI duration.
- One to three years: at $114.2B of cash and short-term investments, it can self-fund, but the index weight concentrates the exposure.
- grew total debt from $83.9B to $112.1B in nine months and lost its stable AA- outlook on Oct 30, 2026
- A $20B–$25B sale split two-thirds dollars and one-third euros adds currency and duration exposure at the same time.
- Days to quarters: pricing on the Nov 3, 2026 deal sets the concession that other AI issuers will then pay.
- One to three years: a negative outlook converts AI capex funding from a non-event into a recurring credit cost.
- grew Q3 2026 revenue 61% to $4.31B and ended 2025 with a $15.0B backlog, up 109% on the year
- It is funded by operating cash flow, not by the long-dated paper that is repricing, so it captures the buildout without paying the concession.
- Days to quarters: backlog conversion is the direct read on whether AI bond proceeds are being spent.
- One to three years: supplier cash flows compound while issuance-side spreads widen — the cleanest way to own the theme.
- expects at least 110GW of gas turbine backlog and slot reservations on data center demand, reaching 116GW by Q2 2026
- It runs net cash with trailing free cash flow near $19.7B, so it funds turbine slots without issuing long-dated debt.
- Days to quarters: turbine slot reservations are the leading indicator of AI power buildout, ahead of data center energisation.
- One to three years: the constraint on AI capacity — and therefore on AI borrowing — becomes power equipment, not chips.
- trades at a bond yield above 16% with five-year CDS near 1,155bp after a spike above 1,289bp
- Trailing capex runs about 279% of revenue and trailing free cash flow is negative roughly $17.8B, so growth is funded by issuance.
- Days to quarters: any new issue or refinancing terms set the market's price for AI infrastructure risk beyond the hyperscalers.
- One to three years: at 17.6x trailing net debt to EBITDA, dispersion becomes permanent for this credit rather than cyclical.
- funds a 29.4% of revenue capital program from operations and a US ADS offering instead of investment-grade bonds
- FY2025 revenue of ₩97.1T with operating income of ₩47.2T means memory cash flow, not credit markets, carries the buildout.
- Days to quarters: memory pricing drives the cash that funds the AI supply chain above the hyperscalers.
- One to three years: the equity route avoids spread risk but adds dilution if AI credit conditions tighten.
