Markets: Who is actually funding the AI buildout now?

2026-07-02 · long-form

Executive summary

For two years the reassuring answer to "who pays for the AI buildout" was: the strongest balance sheets in market history, out of their own cash flow. That answer is going out of date. In 2026 the five largest US hyperscalers plan roughly $600–690B of capital spending, close to 90% of their combined operating cash flow, against a ten-year average near 40% T3. The group now holds more debt than cash for the first time, and Morgan Stanley expects hyperscaler borrowing above $400B in 2026, more than double the roughly $165B of 2025 T3. Below that layer sits a second one the cash-flow story never included: off-balance-sheet vehicles, private-credit data-center loans, GPU-collateralized debt, and vendor deals in which the chip seller funds the chip buyer.

The question this report takes is not whether AI demand is real. It is who is now on the hook if the demand curve bends, and what that does to how this cohort draws down. The answer: the marginal dollar of the buildout has crossed from equity and cash flow into debt and circular financing, and that changes the shape of the risk. A cash-funded boom slows gently when demand disappoints — you simply stop spending. A debt-and-vendor-funded boom has refinancing dates, forced sellers, and a reflexive loop between the funder's stock and the customer's order book.

Our read: this is the mechanism that turns the kit's standing late-cycle valuation call into a drawdown call. It does not tell you the month. It tells you that when the cohort de-rates, it will de-rate faster and with more contagion than the fortress-balance-sheet framing implies. Monday and this morning gave the first tape-level rhyme — the memory complex sold off hard with SK Hynix down 14.57% into the record equity raise it is about to price T3.

House view reconciliation

The standing Markets view holds a structural late-cycle posture: patience and cash, the AI cohort's multiple over-extrapolated toward "permanent structural," the supply curve identified as the mispriced variable rather than demand, and credit named as the last unpriced leg of the stagflation read _house-view, 2026-06-25-record-raise-supply-curve, 2026-06-11-credit-nine-basis-points. A dedicated theme dossier on AI buildout financing was opened last week 2026-06-26-ai-buildout-financing-turn-v1.

This report extends that view along one axis it had not yet made explicit: the funding mix, and what it does to drawdown mechanics. The prior notes established that the cohort is expensively priced (the equity-risk-premium note), that supply is being summoned at a margin peak (the supply-curve note), and that headline credit spreads are quiet while quality decompresses underneath (the nine-basis-points note). This adds the plumbing — that the incremental buildout is increasingly debt-financed, off-balance-sheet, and circular — which is the reason a de-rate here would not be an orderly repricing. It confirms and sharpens; it does not reverse. The confidence band on the cycle-position call is unchanged, consistent with the kit's own record that late-cycle calls have been early by twelve months and more.

One honest cross-check runs the other way this morning. The June jobs print landed soft — 57,000 jobs against a roughly 110,000 consensus, with the unemployment rate falling to 4.2% only because participation dropped to 61.5%, its lowest since March 2021 T1. That eased the discount-rate pressure the bear case has leaned on — the two-year yield fell about 3.5 basis points to 4.13% and equity futures rose T3. But a debt-funded, cash-burning buildout is not obviously safer in a softening labor market. Lower rates help the refinancing math; weaker final demand hurts the thing the debt was borrowed against. The funding-structure argument is largely independent of which way the rate path breaks.

The setup

Start with the number that changed. Through 2023 and most of 2024, hyperscaler capital spending ran comfortably inside operating cash flow, the way a mature, cash-generative business funds growth. In 2026 that gap has closed. The combined capex of roughly $600–690B is about 90% of operating cash flow, up from a ten-year norm near 40%, and about three-quarters of it — some $450B — is tied directly to AI infrastructure T3. Meta guided roughly $125B, Alphabet around $180B, Amazon near $200B, Microsoft about $105B T3. Epoch AI's tracking puts the group's capex above cash flow outright by the third quarter of 2026 T3.

When capex exceeds cash flow, the difference is funded — by drawing down cash, by issuing equity, or by borrowing. All three are happening, and borrowing is the fastest-growing. Hyperscalers issued over $121B of bonds in 2025, and the group crossed into a net-debt position for the first time T3. That is the on-balance-sheet layer. The more interesting layer is the one designed not to appear there.

The analysis

The buildout moved off balance sheet on purpose

The signature financial-engineering move of this cycle is the special-purpose vehicle. Rather than carry the data center on its own balance sheet, the operator forms a joint venture, fills it with private-credit and institutional money, and signs a long lease. The scale is now large. Reporting tallies roughly $120B of AI infrastructure debt moved off the books of Oracle, xAI, Meta, and CoreWeave through such vehicles, funded by PIMCO, BlackRock, Apollo, Blue Owl, and the large banks T3. JPMorgan expects annual data-center securitization to run $30–40B in 2026 and 2027 T3.

The template is Meta's Hyperion campus in Louisiana. Meta and Blue Owl closed roughly $30B of financing — about $27B of A+-rated debt plus $2.5B of equity, anchored by an $18B PIMCO commitment — with Meta owning just 20% of the joint venture and Blue Owl-managed funds owning 80% T3. The structure is legal and rated investment grade. It is also a way to build a $30B asset while consolidating a fraction of it, which means the headline balance sheets understate how much of the buildout is debt-supported.

The structural mismatch inside these deals is worth stating plainly. The GPUs that fill the buildings depreciate on something like a seven-year life; the buildings and the debt against them run twenty to thirty years T3. A loan secured partly by fast-depreciating chips, in a business where the chips are the whole point, carries a collateral profile that has never been tested through a demand downturn.

The vendor is funding the customer

The second break from the cash-flow story is circular financing — the chip and cloud vendors taking equity stakes in, or extending purchase commitments to, the customers who then buy their product. Bloomberg's mapping of the named commitments across Nvidia, Broadcom, Oracle, Microsoft, AMD, Amazon, and CoreWeave totals north of $800B T3. Nvidia has committed up to $100B to OpenAI and holds roughly a 7% stake in CoreWeave, which builds its business on Nvidia chips T3. CoreWeave has opened facilities collateralized by Nvidia chips to buy more Nvidia chips, including a $2.3B facility, and in early 2026 closed the first investment-grade-rated GPU-backed financing at $8.5B T1.

The reason this matters is not that it is fraudulent — it is not. It is that vendor financing manufactures the appearance of demand. When the seller funds the buyer, the order gets booked before the end market has proven it needs the capacity. The incentive shifts from testing demand to closing deals. And the customer at the center of the loop is not yet a going concern on its own cash flow: OpenAI is on track to lose about $14B in 2026, roughly triple its 2025 loss, while it anchors hundreds of billions in purchase commitments and targets $100B of revenue by 2029 T3.

Hyman Minsky gave the taxonomy for this. He divided borrowers into three units: hedge finance, where income covers interest and principal; speculative finance, where income covers interest but the principal must be rolled; and Ponzi finance, where income covers neither and the position depends on the asset appreciating or on new financing arriving T2. A firm losing $14B a year and funding hundreds of billions of commitments out of fresh capital is, in Minsky's precise and non-pejorative sense, a Ponzi unit — solvent only so long as the next financing round shows up. The buildout's demand anchor is a Ponzi-financed borrower, and the anchor's suppliers are lending it the money to place the orders.

Why the funding mix changes the drawdown, not just the valuation

The kit's late-cycle call has always rested on valuation — a thin equity risk premium, a record CAPE, an over-extrapolated cohort multiple 2026-06-18-equity-risk-premium-three-measures. Valuation tells you the expected return is poor and the ten-year distribution is compressed. It does not tell you the path. Funding structure is what governs the path.

A cash-funded buildout has a graceful exit. If demand softens, the hyperscaler simply spends less next quarter; nobody forces the sale of an asset, because nobody lent against it. A debt-and-vendor-funded buildout does not exit gracefully. The SPV debt has maturities that must be refinanced regardless of where AI revenue sits that quarter. The GPU-backed loans have collateral that falls in value exactly when demand falls. The circular deals reverse: if the vendor's stock falls, its ability to fund customers shrinks, which cuts the customers' orders, which cuts the vendor's revenue — a reflexive loop in the direction the cash-flow story never had. This is the difference between a boom that fades and a boom that breaks.

The telecom build of 1999–2001 is the cleanest base rate, because it broke on financing rather than on the absence of demand — internet traffic did grow. Equipment makers financed their own customers to book sales: Lucent extended on the order of $1.5B of customer financing with commitments running much higher, Nortel roughly $3.1B committed, Cisco about $2.4B in customer loans T3. The customers were competitive carriers that had raised about $82B by the end of 1999 and were spending it on gear; industry capex ran from $56B in 1997 to $120B in 2000 T3. When the carriers could not fund themselves, the vendor loans went bad and the equipment makers fell 80–90%. The demand for bandwidth kept growing the whole way down. The financing is what broke.

Credit is still not pricing any of it

The tell that this is late and not yet acute is in the spread. The ICE BofA US High Yield index option-adjusted spread sat at 2.74% on July 1, 2026 — 274 basis points — barely off the June 17 cycle low of 2.63% and among the tightest readings of the expansion T1. Investment-grade sits near 80 basis points T3. Spreads did not widen through the memory rout of the past two sessions. The AI-linked private credit and GPU-backed paper is rated investment grade and priced as such.

That calm is the point, and it is consistent with the nine-basis-points note's finding that headline spreads are a composition artifact while quality decompresses underneath 2026-06-11-credit-nine-basis-points. Credit reprices late and fast. In a cycle where much of the marginal leverage sits in private vehicles and vendor commitments rather than in the public high-yield index, the public index is even less likely to lead. The equity cohort will telegraph the stress first — as it started to this week — and the credit confirmation will come after, in a hurry.

Variant perception

Consensus holds that AI capex is the safest large capital cycle in history because it is funded by the most profitable companies ever assembled, out of cash flow, and the debt that exists is investment-grade and well-covered. The evidence is not weak: the hyperscalers do generate enormous cash flow, the Meta and CoreWeave deals did earn investment-grade ratings, and the demand — Micron sold out through 2026, HBM allocated quarters ahead — is real 2026-06-25-record-raise-supply-curve.

Our variant is that the safety is a stock claim being applied to a flow that has already changed. On the marginal 2026 dollar, the buildout is no longer cash-funded — capex is roughly 90% of cash flow and crossing above it, the group is in net debt, borrowing is set to double, $120B has moved into off-balance-sheet vehicles, and the demand anchor is a Ponzi-financed borrower whose orders are underwritten by its own suppliers. The load-bearing consensus assumption is that this financing is a bridge — that AI revenue scales fast enough to term it out and the circular deals resolve into real end-demand. The base rates argue the other way: record equity issuance predicts below-average forward returns T2, the capital cycle resets when high returns summon supply T2, and manias run on credit expansion that looks investment-grade until the displacement fades T2. "This time the funding is different" is the same wager every prior cycle lost on the one variable it got wrong T2.

What falsifies the variant. If hyperscaler AI revenue reaccelerates enough to pull capex intensity back toward historical norms, and OpenAI's revenue trajectory closes on its commitments while the SPV and GPU-backed debt terms out at stable investment-grade spreads, then the financing was a bridge and the circular deals were pre-funding of genuine demand. Concretely: high-yield OAS holding near 275 basis points through a full year of this issuance, GPU-backed securitizations pricing at investment grade without widening, and OpenAI's losses narrowing rather than tripling again. If instead the cohort begins to sort — the funders' equity falling while the funded names' order books shrink in step — that is the reflexive loop turning, and it confirms the variant. This week's memory sell-off, with the name raising record equity falling hardest, is the first data point on that axis.

Implications for AlphaSteve

The top-down implication is that the standing valuation call now has a mechanism attached that governs the shape of the eventual de-rate. The cohort is not only expensive; the marginal buildout behind it is increasingly funded by debt, off-balance-sheet structures, and vendor commitments, which means a demand wobble converts into a financing event rather than a spending pause. That is a reason to hold the patience-and-cash posture with more conviction, not to short — late-cycle calls have been early in this kit's own record, and the soft-jobs print this morning is a live reminder that the rate path can cut against the bear case even as the funding fragility builds. The durable expression stays the same: refuse to chase the cohort, watch the funding markers rather than the demand headlines, and keep the credit and issuance tells on the daily scan.

• Portfolio: No position change. The book stays fully in cash by construction; no AI-cohort exposure is added on the memory rout. Nothing on the watchlist sits in the complex under stress.
• Watchlist: SK Hynix's roughly $29.65B Nasdaq ADR, now drifting from July 10 toward August, held as the live issuance-appetite test — a name raising record equity while its stock falls 14.57% is the issuance-into-weakness marker in its cleanest form T3. CoreWeave GPU-backed paper added as a live private-credit price test.
• Theses on the workbench: AI buildout financing dossier 2026-06-26-ai-buildout-financing-turn-v1 extended with the funding-mix crossover (capex ≈ 90% of cash flow, net-debt turn, ~$400B borrowing, ~$120B off balance sheet) and the Minsky Ponzi-unit framing of the demand anchor.
• Sectors: AI infrastructure / memory held at "bottleneck rent on a defined clock — do not chase." AI-linked private credit and GPU-backed securitization flagged as an "avoid" for any future credit sleeve. Neocloud equity (CoreWeave, Nebius) flagged as the sharpest expression of the circular-financing fragility.
• House view updates: Equity-market cycle and AI buildout financing sections updated; cycle-position confidence band unchanged.
• Daily-scan adjustments: Add hyperscaler capex-to-operating-cash-flow ratio as a standing funding-mix read (threshold: above 100% = flow fully debt/equity-funded). Add the funder-vs-funded dispersion check (do the vendors and their customers move together on down days — the reflexive loop). Carry HY OAS 275 bps and the CCC−BB gap from the June 11 note as the credit-confirmation tells.

Charts / data

Table 1 — The funding mix has turned (2026)

Table 2 — The circular loop, named deals

Table 3 — Telecom 2000 vs AI 2026 (financing, not demand)

Source rows as tagged inline. Telecom figures T3.

Sources

• T1 — payrolls +57,000; unemployment 4.2%; participation 61.5%. https://www.bls.gov/news.release/empsit.nr0.htm
• T1 — 2.74% (2026-07-01); 2.63% cycle low (2026-06-17). https://fred.stlouisfed.org/series/BAMLH0A0HYM2
• T1 https://investors.coreweave.com/news
• T2 — hedge / speculative / Ponzi finance taxonomy.
• T2 — top-quartile issuance → below-average forward returns.
• T2 — supply-side capital-cycle discipline.
• T2 — displacement and credit expansion in manias.
• T2 — the "this time is different" failure mode.
• T3 — capex ≈90% of operating cash flow. https://www.mufgamericas.com
• T3 https://techblog.comsoc.org
• T3 https://introl.com/blog
• T3 https://epoch.ai/data-insights/hyperscaler-capex-vs-cash-flow
• T3 https://www.bloomberg.com/graphics/2026-ai-circular-deals/
• T3 https://www.bloomberg.com/news/articles/2025-10-16
• T3 https://global.morningstar.com
• T3 https://www.theinformation.com
• T3 — JPMorgan securitization estimate; SPV structures. https://www.quinnemanuel.com
• T3 https://qz.com
• T3 https://www.cryptopolitan.com
• T3 https://americanaffairsjournal.org/2020/08/
• T3
• T3 https://www.cnbc.com/2026/07/02/
• T3 https://www.cnbc.com/2026/07/02/jobs-report-june-2026-.html
• T3 https://finance.yahoo.com/economy

See sources-policy for the citation discipline applied. Wikipedia not used. Where a figure originates with sell-side or FT reporting reached through a secondary outlet, the underlying source is named and the claim is tagged T3 rather than T2.

House view changes this run

• AI buildout financing (dossier v1) — extended. Added the funding-mix crossover as a named marker: 2026 hyperscaler capex ≈90% of operating cash flow (10-yr avg ~40%) and crossing above cash flow by Q3, the group's first net-debt position, expected borrowing >$400B (≈2× 2025), ~$120B moved off balance sheet via SPVs, and JPMorgan's $30–40B/yr securitization run-rate. Added the Minsky taxonomy — the demand anchor (OpenAI, ~$14B 2026 loss) is a Ponzi financing unit whose orders are underwritten by its own suppliers. Logged the drawdown-mechanics distinction (cash-funded booms fade; debt-and-vendor-funded booms break). No confidence-band change.
• Equity-market cycle position — extended. The funding structure is logged as the mechanism attaching to the standing valuation call: it governs the shape of the de-rate, not its timing. The July 1–2 memory rout (SK Hynix −14.57% into a record raise) logged as the first tape-level rhyme of the funder-vs-funded reflexive loop. Cycle-position confidence band unchanged.
• Credit — carried and reinforced. HY OAS 2.74% (2026-07-01), near the June 17 cycle low of 2.63%, did not widen through the memory rout — consistent with the June 11 note that the public index will lag a cohort that levered privately. HY-linked private credit / GPU-backed securitization moved to "avoid" for any future credit sleeve.
• US rate path — one honest counter logged. The soft June jobs print eased the discount-rate leg of the bear case (2Y −3.5bp to 4.13%, futures up). Noted that the funding-fragility argument is largely independent of the rate path, and that weak final demand is not benign for a debt-funded, cash-burning buildout. No band change.
• Daily-scan additions: hyperscaler capex-to-operating-cash-flow ratio (funding-mix read, >100% threshold); funder-vs-funded dispersion check (reflexive loop); HY OAS 275 bps and CCC−BB gap carried as credit-confirmation tells.
• No weight changes. No confidence-band changes. No portfolio changes.

Linked

• _house-view
• 2026-06-26-ai-buildout-financing-turn-v1
• 2026-06-25-record-raise-supply-curve
• 2026-06-18-equity-risk-premium-three-measures
• 2026-06-11-credit-nine-basis-points
• 2026-06-04-stagflation-signature-cross-asset
• 2026-06-05-ai-infrastructure-capacity-dossier-v1
• 2026-07-02-AM
• sources-policy
• voice-and-style