AI data center financing in 2026: who is lending and who carries the risk

The five hyperscalers have turned the investment-grade bond market into a core funding pipe for AI capex. From 2020–2024 they were a rounding error; in 2025–2026 they became a primary-market driver, terming out 20- to 100-year money while order-book coverage fell and new-issue concessions rose. Microsoft is the exception that proves the rule: it has barely issued senior unsecured debt and is pushing capacity into lease and project vehicles instead.

The regime change

Until 2024 these companies funded servers, data centers, and chips from operating cash flow. J.P. Morgan Asset Management puts combined gross bond issuance at about $17 billion in 2024, then $109 billion in 2025 and $194 billion in the first half of 2026 alone. Vanguard, using Bloomberg data, shows a 2020–2024 average of roughly $35 billion a year, $93 billion in 2025, and about $132 billion of U.S. bonds through July 31, 2026. The gap between those series is definitional—currency, whether related project debt is included, and cutoff date—not a dispute about the direction. [1] [2]

The mechanism is a cash-flow crossover, not a distress event. UBP estimates 2026 capex of about $820 billion against operating cash flow of about $750 billion. Goldman Sachs, cited by Reuters in late July, had capex near $750 billion versus operating cash flow near $778 billion, and expected the five to issue roughly $250 billion of bonds in 2026 and $400 billion in 2027. Either way, multi-year GPU and power contracts no longer fit inside a single year’s free cash flow, so treasurers are locking long-dated unsecured debt while the rating is still high. [3] [4]

They are also leaving the dollar market on purpose. CPR Asset Management counted about $223 billion of hyperscaler bonds since the start of 2026, of which about $132 billion was in dollars, plus €23.5 billion, 5.9 billion Swiss francs, and £9.75 billion. Dollar IG remains the workhorse; euros, sterling, Swiss francs, Canadian dollars, yen, and Australian dollars are how they avoid saturating one buyer base. [5]

Publicly reported deals

Figures below are gross proceeds of publicly reported senior bond sales. Off-balance-sheet project bonds (Meta’s Hyperion vehicle, Oracle’s Project Jupiter loans, Microsoft-tenanted QTS deals) are noted separately and are not the issuer’s own unsecured debt. Sources disagree on some 2025–2026 aggregates; deal-level amounts are from contemporaneous reports.

Issuer	Date	Size	Tenors / structure	Spread at issuance	Order book
Group	2024	~$17B combined	Light, mostly routine	Not a market event	—
Oracle	Sept. 24, 2025	$18B, 6 parts	5s to 40s (2030–2065)	+75 / +90 / +105 / +115 / +125 / +137 bp	Peak ~$88–90B (~5x)
Meta	Oct. 30, 2025	$30B, 6 parts	5s to 40s	+50 / +70 / +78 / +88 / +98 / +110 bp	~$125B (then a record)
Alphabet	Nov. 3, 2025	$17.5B USD + €6.5B (~$7.5B)	USD 3s to 50s	50-year at +107 bp (IPT +135)	~$90B on the USD leg
Amazon	Nov. 17, 2025	$15B, 6 parts	Out to 40 years	40-year tightened to +85 bp from IPT +115	Peak ~$80B
Oracle	Feb. 2–4, 2026	$25B, 8 parts	3s to 40s, plus a $500M floater	+95 / SOFR+111 / +115 / +130 / +145 / +170 / +180 / +195 bp	Peak ~$129–155B; final books ~$109B
Alphabet	Feb. 9–10, 2026	~$32B equivalent	$20B USD (7 parts, to 40 years); £5.5B (~$7.5B) incl. £1B century; ~$4.0B Swiss francs	USD 3-year +27 bp, 40-year +95 bp; century +120 bp over gilts, 6.125% coupon	USD book ~$95–100B+; century ~10x (£9.5B orders)
Meta	Apr. 30, 2026	$25B, 6 parts	To 2066	2066 at +147 bp (IPT up to +180); October 40-year had been +110	Peak $96B (~3.8x)
Amazon	Mar. 10, 2026	$37B USD + €14.5B (~$16.6B)	USD 2s to 50s (2076); euro to 38 years	2076 at +130 bp (IPT +155)	USD peak ~$126B (~3.4x)
Amazon	July 7, 2026	$25B, 8 parts	3s to 40s (2066)	Longest tightened ~20 bp to about +125 bp; 18–21 bp concession on the long end	Peak $62B, settled near $41B (~1.6–2.5x)
Alphabet	Aug. 6, 2026	$25B, 10 parts	2s to 40s (2066), incl. floaters	Coupons 4.50% (2028) to 6.50% (2066); final spreads not fully public	Peak ~$115B
Amazon	Sept. 9, 2026	£4.25B ($5.76B)	3s, 6s, 12s, 19s	+53 bp (3-year) to +93 bp (19-year) over gilts; yields ~5.2% to 6.7%	>£10.65B (~2.5x)
Microsoft	2025–2026	No jumbo senior unsecured sale	—	—	Still the AAA outlier

Oracle’s September 2025 coupons were 4.45% (2030, $3B), 4.80% (2032, $3B), 5.20% (2035, $4B), 5.875% (2045, $2.5B), 5.95% (2055, $3.5B), and 6.10% (2065, $2B). The February 2026 print landed 40–58 bp wider than that curve. Meta’s October 2025 coupons were 4.20% ($4B fives), 4.60% ($4B sevens), 4.875% ($6.5B tens), 5.50% ($4.5B twenties), 5.625% ($6.5B thirties), and 5.75% ($4.5B forties), with final books of $15.8B, $20.5B, $24.3B, $15.4B, $23.8B, and $16.5B and concessions of 10–15 bp. [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25]

One reporting conflict: some secondary write-ups call Oracle’s February deal $30 billion. Primary syndicate accounts and the filing detail a $25 billion note offering, paired with an equity and mandatory-convertible plan that together targeted $45–50 billion for 2026. [26] [27]

Not in the table, but part of the same funding stack: Meta’s October 2025 Blue Owl joint-venture bond of about $27 billion for the Louisiana Hyperion campus (debt sits in a vehicle, not on Meta’s unsecured curve); roughly $18 billion of Oracle Project Jupiter loans quoted in September 2026 at 89–91 cents; and Microsoft-linked QTS paper—$4.6 billion in April 2026 at Treasuries plus 137.5 bp (Baa2, peak demand $12.5 billion) and a Project Odyssey deal marketed around $3.9 billion in August at a high-6% coupon and a low-to-mid-7% yield, investment-grade on rating but junk-like on yield. [28] [29] [30] [31]

Amazon is the volume leader. By early September 2026 it had sold the equivalent of more than $92 billion across dollars, euros (a record €14.5 billion corporate deal), Swiss francs, Canadian dollars (a record C$14 billion, about $10.3 billion), and sterling. Alphabet set local records in sterling, Swiss francs, yen (¥576.5 billion, about $3.6 billion, the largest yen bond by a non-Japanese issuer), and Australian dollars. [32] [33] [34]

How the paper is priced, and why books are shrinking

The product design is deliberate duration. Across 2025 and 2026 supply, hyperscalers placed 41% of issuance beyond 15 years, versus 16% for the ICE BofA U.S. Corporate Index, and their weight in the 26-year-plus bucket is about 31% versus 13% for the broad market. Amazon has gone to 2076; Alphabet sold a 2126 sterling century, the first technology century bond since Motorola in 1997. That matches a 15- to 20-year data-center lease and power contract better than a 5-year bond, and it parks supply with insurers and pensions rather than total-return credit funds. [25] [35]

The cost of that access has risen inside a still-tight investment-grade market. Reuters’ read of LSEG data showed median spreads for Amazon, Alphabet, Meta, and Oracle moving from 30 bp to 40 bp on 2- to 4-year bonds, 50 bp to 60 bp on 5- to 7-year bonds, and 108.5 bp to 118 bp on bonds longer than 20 years, comparing 2025 with 2026 through late July. The median deal-level concession rose from 2.25 bp in 2025 to 12 bp in 2026. Of 91 hyperscaler bonds issued in 2026 with comparable pricing, 78 were trading at higher yields on July 28 than at issue; the median gap was about 22 bp. UBP’s count as of September 3 was that just over half of 2026 bonds from Oracle, Meta, Amazon, and Alphabet were at least 5 bp wider than the new-issue spread, while the broader IG market sat near its tightest levels since the financial crisis. [4] [36]

Coverage tells the same story in sequence. Meta’s October 2025 book was about 4.2 times. Oracle’s February book peaked near 5 times. Amazon’s March dollar book was 3.4 times. Alphabet’s August book was still about 4.6 times ($115 billion on $25 billion). Amazon’s July deal fell to roughly 1.6–2.5 times, and its September sterling book was about 2.5 times, versus roughly 5 times on Alphabet’s February sterling sale. Apollo data cited in September showed hyperscaler cover ratios falling from about 5 times in February to about 2 times in July. The market is still clearing jumbo deals; it is no longer clearing them for free. [1] [23]

Oracle is where secondary trading has detached from the rating. On October 2, 2026, the 6.55% bonds due February 2046 were at 84.91 cents, a yield to maturity of 8.11%, about 247 bp over the 20-year Treasury (5.64% on October 1). Every Oracle bond due after 2044 was yielding between 7.85% and 8.44%. The 6.70% 2056s crossed 8%. MarketAxess noted 30-year spreads had widened about 150 bp since the September 2025 deal and were trading wide of several BB names. Oracle’s own August 31, 2026 10-Q marked $125.0 billion of senior notes at a fair value of $105.7 billion, about 84.6 cents. Five-year CDS hit roughly 203 bp in July, an 18-year high, and set fresh records in late September. [37] [38] [29] [39]

Ratings: four fortresses, one cliff

Microsoft is AAA/Aaa with stable outlooks, one of two U.S. public companies still triple-A at both S&P and Moody’s, and rated above the U.S. sovereign. It is also the only hyperscaler that, as of early September 2026, had not termed out 2026–27 capex in its own bond curve. The implication is mechanical: the next large long-end print the market does not yet have in the price is Microsoft’s, if and when cash flow no longer covers the build. [40] [25] [41]

Moody’s affirmed Alphabet at Aa2, outlook stable, on March 30, 2026. Meta’s October 2025 deal was Aa3/AA−. On the April 2026 sale, S&P kept a stable outlook and said leverage should stay “well below” the downgrade threshold for at least two years, while noting that AI investment was “starting to affect credit metrics.” UBP groups Alphabet, Amazon, and Meta as AA− or better, with more than $500 billion of cash among the four highest-rated names, and argues they could add another $400 billion of debt over 18 months without a rating event, and about $600 billion before a slide to single-A. That is capacity, not a forecast of issuance. [42] [43] [16] [3] [36]

Oracle is the break in the complex. S&P cut the issuer from BBB to BBB− on July 9, 2026, outlook stable, and cut short-term and commercial paper from A-2 to A-3. The agency said it had underestimated the scale of AI investment, that the infrastructure business is diluting a previously strong business-risk profile, and that adjusted leverage stays above 4x. It forecast a fiscal 2027 free-operating-cash-flow deficit of nearly $42 billion and flagged OpenAI as a key credit risk—about half of a roughly $638 billion remaining-performance-obligation backlog at the time of the review. Moody’s is Baa2 with a negative outlook; Fitch is BBB. A further notch at S&P would push a stock of bonds that secondary commentary puts near $120 billion out of investment-grade indexes. That forced-selling channel, not the coupon, is the systemic piece. [44] [45] [46] [29]

Moody’s July 2026 note is the cross-issuer view: capex of $785 billion in 2026 and about $1 trillion in 2027 is eroding free cash flow, and stock sales plus off-balance-sheet structures “threaten credit quality” at Microsoft, Amazon, Alphabet, Meta, Oracle, and CoreWeave. It also said the four strongest balance sheets are unlikely to lose investment-grade status imminently. Pressure is concentrated at Oracle and at high-yield neoclouds. [47]

Index math is already moving. By August, Amazon was about 1.4% of the Bloomberg U.S. Corporate Index (fifth-largest issuer) and Oracle about 1.3% (eighth). UBP puts the five hyperscalers plus Nvidia at 4.6% of the U.S. IG index. That is large enough to reprice the long end of credit, not large enough to “break” a multi-trillion market—Pimco found no statistically significant rise in 10-year Treasury yields around the past six mega AI offerings. [48] [3] [49]

What share of AI capex this channel funds

Gross bonds are not the same as net capex funding. Companies still pay for most of the build with operating cash flow; bonds also refinance, support buybacks, and pre-fund years that have not been spent. Leases, project finance, and chip-backed structures sit outside the unsecured tally. UBP puts off-balance-sheet leases, purchase commitments, and guarantees near $3 trillion; Oracle alone disclosed $288 billion of additional data-center lease commitments not yet on the balance sheet. [3] [29]

The cleanest sourced ratio is Goldman’s, via Reuters on July 29, 2026: hyperscaler debt issuance equivalent to about one-third of 2026 capital spending and about 35% in 2027, on an expected $250 billion of bond issuance this year against roughly $750 billion of capex, and $400 billion of issuance in 2027. [4]

A range, not a point estimate:

• 2026 capex. Goldman via Yahoo Finance in late September: about $800 billion this year and $1.2 trillion next year. UBP: about $820 billion. Moody’s: $785 billion, then about $1 trillion in 2027. Use $750–820 billion for 2026. [49] [3] [47]
• 2026 bond issuance by the five. Realized totals depend on the cutoff. Capital Group counted $240.7 billion year-to-date through August 31 across currencies, of which $66 billion was non-dollar, and expected more than $25 billion more from the five in the last four months of 2026. CPRAM had about $223 billion since the start of the year. Goldman’s full-year marker is about $250 billion. A full-year gross figure around $250–280 billion is the reasonable band if late-year supply arrives as those desks expect. [50] [5] [4]
• Implied share. $250–280 billion of gross bonds against $750–820 billion of capex is about 30–35%. That matches Goldman’s one-third. It is an upper-bound description of this channel’s role in the funding mix, not a claim that a third of every dollar of servers was borrowed.

The cash gap is smaller than the bond print. On UBP’s $820 billion versus $750 billion, the operating-cash shortfall is about $70 billion—so a large slice of 2026 issuance is terming out future spend, replacing commercial paper, or sitting as liquidity, not plugging a same-year hole. Goldman’s earlier July snapshot even had operating cash flow slightly above capex. The strategic point is forward: 2027 capex estimates of $1.0–1.3 trillion (UBP) or $1.2 trillion (Goldman) against issuance forecasts of about $400 billion (Goldman) to $420 billion (Goldman, cited by Reuters as a 60% rise from 2026) keep the bond share in the mid-30s even if cash generation grows. [51] [1]

Wider AI-linked borrowing is a different numerator. The IIF said global bond issuance by AI-linked companies had already passed $400 billion in 2026 and was running above a $500 billion annualized pace, with U.S. companies about 90% of that. Goldman has put investor financing to AI-linked groups near $500 billion this year, of which the five hyperscalers are about $200 billion. Those figures include chipmakers, data-center developers, utilities, and private credit. They should not be divided into hyperscaler capex. [49] [51]

What this means if you are trying to fund, or lend against, the same build

The unsecured IG bid is still open for AA balance sheets, but the clearing price is now set by supply technicals, not by the last deal’s spread. Each jumbo print has cheapened the prior one: Amazon’s July $25 billion sale pushed an existing 30-year about 20 bp wider. Issuers that can still diversify currency—and Alphabet and Amazon have shown they can set records in five markets in one year—will pay less than issuers that have to keep returning to the same 30-year dollar buyer. [52]

Microsoft’s absence is the residual supply risk. A first large, long-dated Microsoft deal would land into a long end that is already 2.5 times overweight hyperscaler paper versus the corporate index. The market has demonstrated it can take $25–37 billion from AA names; it has not demonstrated it can take that from the AAA name on top of another Oracle or Amazon print in the same month.

Oracle is the template for what happens when the same capex story meets a BBB− rating, customer concentration, and negative free cash flow. The bonds still clear—February’s book was the largest of the cycle—but they now trade with high-yield yields and high-yield liquidity behavior while remaining inside IG portfolios. The binding constraint is no longer coupon capacity. It is the index bid that disappears if S&P moves one more notch, and the lease and project-finance markets that are being asked to hold what the unsecured curve will not.

---

Recent Findings Supplement (October 2026)

Alphabet and Amazon have led a surge in multi-currency, multi-tranche investment-grade bond issuance since mid-2026 to fund AI-driven capex, with deals often oversubscribed but showing signs of cooling demand and wider spreads amid rising Treasury yields and heavy supply. Microsoft has signaled a major upcoming issuance, while Meta tapped the market again and plans euro entry; Oracle has faced secondary-market pressure and ratings pressure without prominent new IG bond prints in the period. Bonds represent a growing but partial share of the financing mix for the five companies’ combined ~$820 billion 2026 capex (versus ~$750 billion operating cash flow), supplemented by equity, project finance, and other channels.[1][2]

Alphabet’s Record Multi-Tranche Deals

Alphabet executed two large USD offerings in 2026 (post-April), extending its curve to ultra-long tenors while maintaining strong (though not record) demand. These support its elevated capex guidance (raised to $195–205 billion for 2026).[3]

• October 2026 ($20 billion, 7 tranches, 2029–2066 maturities): Priced with yields from ~3.829% (3-year, +27 bp spread) to 5.794% (40-year, +95 bp spread). Net proceeds ~$19.85 billion. Ratings Aa2/AA+ (stable). Joint coordinators BofA, Goldman, JPM.[4]
• August 2026 ($25 billion, 10 tranches including floaters, 2028–2066): Coupons 4.5%–6.5%; spreads 33–130 bp over Treasuries (tightened 25–27 bp from IPTs). Strong demand supported the size; part of ~$88 billion+ issuance since late 2025.[5][3]

Implication for competitors: Alphabet’s ability to issue across currencies and tenors (including prior euro, sterling, etc.) demonstrates broad investor appetite for AA-rated hyperscaler paper, but repeated large prints are testing absorption capacity and contributing to modestly wider AI-issuer spreads (~115 bp vs. ~78 bp broader IG).[6]

Amazon’s Diversified Currency Push and Upcoming Mega-Deal

Amazon has been the most prolific issuer, adding sterling and planning simultaneous USD/euro raises while its 2026 capex guidance sits around $200–220 billion.[7]

• October 2026 launch (up to $42 billion total): Targeting $25–30 billion USD (up to 11 tranches, 2–50 years) plus up to €10 billion euro (up to 8 tranches, 2–38 years). One of the largest-ever corporate offerings; initial talks for longest tranche ~+155 bp over Treasuries.[8]
• September 2026 (£4.25 billion / ~$5.76 billion sterling debut, 4 tranches, 3–19 years): Demand >£10.65–12 billion (2.5x+ coverage). Yields ~5.2% (3y) to 6.7% (19y). Brings 2026 borrowing past $92–98 billion equivalent.[9][7]
• July 2026 ($25 billion USD, 8 tranches, 2029–2066): Peak demand $62 billion (~2.5x); fixed coupons ~4.6%–6.25%. Used for general corporate purposes including capex.[10][11]

Implication: Amazon’s multi-currency strategy (euro, Swiss franc, CAD, now sterling) diversifies funding and taps local demand, but lower coverage ratios versus earlier 2026 deals signal potential investor fatigue with sustained hyperscaler supply.[6]

Meta’s Continued USD Access and Euro Plans

Meta followed its large 2025 issuance with another sizable USD deal and is preparing its first euro outing.[12]

• May 2026 ($25 billion, 6 tranches): Demand peaked at $96 billion. Spreads wider than its prior $30 billion deal, reflecting some investor caution on capex ramp (2026 guidance $130–145 billion). Ratings stable (S&P noted leverage expected to stay below downgrade thresholds).[13][14]
• Autumn 2026 euro debut planned: First foray into euro market amid broader hyperscaler shift.[12]

Implication: Meta demonstrates that even with equity-market volatility around AI spending, bond investors remain receptive to its AA-/Aa3 paper, though concessions are increasing.

Microsoft’s Planned Large Issuance and Linked Financing

Microsoft has not completed a major direct IG bond print in the immediate post-April window highlighted in results but announced a substantial upcoming deal and participated in linked project financing.[15]

• September 2026 announcement (at least $14 billion, 7 tranches, 3–40 years): Expected to be one of the year’s largest non-M&A deals; underwriters Barclays and HSBC. Proceeds for general corporate purposes amid expected rate rises. Cash position ~$123 billion.[15][15]
• August 2026 Microsoft-linked QTS “Project Odyssey” data-center bond (~$3.9 billion upsized from initial target): Demand >$8 billion; marketed at junk-like yields (~6% coupon, low-to-mid 7% yield, single-B equivalent) despite expected IG rating (Baa2 from Moody’s on related paper). Finances Georgia facility with Microsoft as tenant.[16][17]

Implication: Microsoft’s pristine credit (often AAA/Aaa) supports low-cost direct issuance when needed, while linked structures allow specialized data-center financing at higher yields.

Oracle’s Secondary-Market Pressure and Mixed Financing Approach

Oracle has relied more on prior issuance, equity, and project structures amid capex intensity and ratings scrutiny, with no prominent new large IG bond deals detailed post-April.[18]

• Secondary trading (as of late Sept/early Oct 2026): Long-dated bonds (e.g., 2046–2056) yielding 7.85%–8.44% (e.g., 6.55% 2046 at ~8.11% YTM); 2056 bonds breached 8% for the first time. 5-year CDS at records. Fair value of ~$125 billion senior notes/borrowings estimated at ~84.6 cents on the dollar.[19][20]
• Ratings: S&P downgraded to BBB– (stable) in July 2026, citing high capex ($90–95 billion FY2027 guidance) and negative FCF trajectory; one notch above junk. Fitch affirmed BBB/stable on earlier notes.[21]
• Financing plans: Expects ~$40 billion in debt + equity raises (including $20 billion ATM equity); earlier $25 billion notes in Feb 2026 referenced. Project Jupiter loans trading at discounts.[19][18]

Implication: Oracle illustrates the limits for lower-rated (BBB) issuers in the group—strong backlog supports the buildout, but market pricing reflects leverage and execution risks more acutely than for AA/AAA peers.

Estimated Bond Share of AI Capex Funding

The five hyperscalers’ combined 2026 capex (~$820 billion) exceeds operating cash flow, creating a funding gap filled partly by bonds.[1]

• Hyperscaler bond issuance: ~$194 billion in H1 2026 (or through July/Aug per varying reports); part of ~$400–500 billion total AI-related financing/debt raised in 2026 so far, with hyperscalers accounting for roughly $200 billion.[6][22][23]
• Bond channel share: Described as covering “most” of the dollar IG financing need, with the balance from other currencies, project finance, leveraged loans, securitization, equity (~$115 billion announced in one period), and off-balance-sheet leases/commitments (~$3 trillion noted in one analysis). One estimate positions hyperscaler USD IG issuance at a substantial portion (~40% of AI debt supply in some breakdowns). Goldman Sachs projects hyperscaler gross debt issuance could reach ~$420 billion in 2027.[6][1][2]

Overall implication: Bonds have become a core, scalable channel (especially IG USD and multi-currency) as capex outpaces cash flow, but absorption depends on continued investor appetite amid wider spreads and competition from Treasuries/other issuers. Companies with stronger ratings (Microsoft, Alphabet, Amazon, Meta) retain more flexibility than Oracle. Future supply could pressure spreads further if capex guidance rises. Data is drawn from post-April 2026 reporting; exact secondary spreads and full order-book details for every tranche are not uniformly disclosed in public sources.

The market has split AI data-center finance into two layers that do not match the accounting. Project debt and equity sit in bankruptcy-remote vehicles that own the shell, the power, or the GPUs. The hyperscaler (or a vendor standing behind a weaker tenant) keeps the economic risks that actually make the paper investment-grade: construction completion, rent, and residual value. Rating agencies largely price the tenant. GAAP often does not consolidate the debt, and often does not even book the full lease or the guarantee.

The template: Meta–Blue Owl Hyperion

Hyperion is the cleanest public example of how form is separated from substance. Meta and Blue Owl funds formed a joint venture to own the Richland Parish, Louisiana campus. Blue Owl holds 80 percent and the formal governance; Meta holds 20 percent, contributed land and construction-in-progress, and took a roughly $3 billion cash distribution at close. A bankruptcy-remote issuer, Beignet Investor LLC, sold $27.294 billion of 6.581 percent senior secured notes due May 30, 2049, priced at about Treasuries plus 225 basis points and rated A+ by S&P (single agency). PIMCO anchored roughly $18 billion; BlackRock took more than $3 billion. The notes are fully amortizing 144A project bonds that fund construction from day one, with proceeds parked and drawn as the campus is built. Rent is structured to start around June 1, 2029.
The financing channel works by splitting ownership from the risks that actually make the paper investable. A bankruptcy-remote vehicle owns the shell, the power infrastructure, or the GPUs and borrows against them. The hyperscaler—or a vendor standing behind a weaker tenant—keeps construction completion, rent, and residual value. Rating agencies largely price the tenant. US GAAP often consolidates neither the debt nor the full economic commitment.

Meta–Blue Owl is the template, and the accounting is the product

Hyperion in Richland Parish, Louisiana is the deal the rest of the market is copying because it shows how to keep a multi-decade, single-tenant campus off the operator’s balance sheet without giving up the credit that lenders require.

Funds managed by Blue Owl own 80 percent of the joint venture and the formal governance. Meta owns 20 percent, contributed land and construction-in-progress, and took a roughly $3 billion cash distribution at close. A bankruptcy-remote issuer, Beignet Investor LLC, sold $27.294 billion of 6.581 percent senior secured notes due May 30, 2049, priced at about Treasuries plus 225 basis points and rated A+ by S&P. PIMCO anchored roughly $18 billion; BlackRock took more than $3 billion. The notes are fully amortizing 144A project bonds that fund construction from day one. Rent is structured to begin around June 2029. [1] [2] [3] [4]

Three contractual features do the work:

• Lease slicing. Occupancy is written as four-year non-cancellable blocks with renewal options that can run total occupancy to about 20 years. The economic life of the campus is closer to 25–30 years. The short initial term is what Meta uses to argue the lease is an operating lease, not a finance lease, under ASC 842: renewal is asserted to be “not reasonably certain,” and the present value of four years of rent plus the residual-value guarantee is asserted to be less than “substantially all” of fair value. [1] [5]
• Residual value guarantee. For the first 16 years, if Meta does not renew, terminates early, or defaults, it owes a guaranteed minimum value sized so the related bonds can be repaid from sale proceeds or a Meta cheque. The threshold starts near $28 billion and declines. The last four years of the 2049 bond are covered by a termination fee equal to the outstanding balance, not the RVG. Meta’s parent also guarantees rent. [1] [6] [4]
• VIE control test. Under ASC 810, Meta argues it is not the primary beneficiary because Blue Owl holds the equity majority and the power to direct the activities that most significantly affect economic performance. Meta therefore does not consolidate the campus or the $27.3 billion of notes. It books a minority equity investment and, once leases commence, a short operating-lease liability. As of December 31, 2025, Meta disclosed maximum exposure to loss on the venture of about $45.95 billion—equity carrying value, lease commitments, future funding, and the RVG threshold—while recording no RVG liability because a payout was judged “not probable.” Ernst & Young flagged the VIE determination as a critical audit matter. [7] [8] [9]

S&P treats the paper as Meta credit in project-finance clothing: A+, one notch below Meta’s AA−, because of structural separation (no direct recourse to Meta assets, legal distance between the JV and the bond issuer). During construction, S&P’s view is that contractual protections shift cost-overrun and delay risk to Meta. In operation, a debt-service coverage ratio of about 1.12x plus the RVG supports the same A+. S&P has not added the RVG to Meta’s adjusted debt, citing headroom between the guarantee and a third-party appraisal. It has said it could change that view. [4] [10]

Moody’s is less willing to stop at the footnotes. In February 2026 it said short initial terms plus RVGs can leave reported lease liabilities well below plausible cash outflows, and that it stands ready to make a quantitative debt adjustment for the likely renewal period, the likely RVG payout, or both. [10] [9]

A second Meta vehicle followed the same pattern at a wider price. Reporting in mid-2026 describes Sopaipilla Investor notes of about $12.5 billion for an El Paso campus, issued around a 7.5 percent yield (roughly Treasuries plus 290 basis points) versus Hyperion’s 6.58 percent. Same 16-year backstop logic; the market charged more for the second use. [11] [12]

What this means for a competitor. The one-notch concession is cheap only while appraisal headroom and a AA− tenant exist. The structure fails the moment the tenant’s rating, the residual appraisal, or the “not reasonably certain” renewal judgment is challenged. Equity sponsors are paid to hold governance and first-loss equity, not to underwrite AI demand.

Stargate is an Oracle lease wrapped around an OpenAI contract

Stargate sites do not put OpenAI on the project debt. Developers and their lenders own the buildings. Oracle signs the long lease. OpenAI is the economic user of the compute Oracle sells. Lenders are underwriting Oracle; Oracle is underwriting OpenAI. That double layer is why construction loans can look like project finance and still trade as a hyperscaler credit—until Oracle’s own spread widens.

Site	Developer / owner vehicle	Reported project debt	Tenant chain	Notable risk allocation
Abilene, Texas (flagship, ~1.2 GW)	Crusoe with Blue Owl Real Assets and Primary Digital Infrastructure	Phase construction loans of about $2.3 billion then $7.1 billion, JPMorgan-led; FT counted roughly $13 billion invested by Blue Owl and JPMorgan, including about $10 billion of debt, into the SPV	Oracle lease → OpenAI	Build-to-suit on Lancium’s campus; equity in the JV, banks in construction debt
Shackelford County, Texas and Port Washington, Wisconsin	Vantage Data Centers	About $38 billion in two senior secured facilities (~$23.3 billion Texas, ~$14.8 billion Wisconsin); four-year maturity plus a two-year extension option; about SOFR plus 250 basis points; interest-only in construction, then amortizing	Oracle for OpenAI	Bank syndicate (JPMorgan, MUFG, Wells Fargo, BNP, Goldman, SMBC, Société Générale); syndication stretched by single-tenant concentration limits
Doña Ana County, New Mexico (Project Jupiter)	STACK Infrastructure (Blue Owl)	About $18 billion of bank construction loans; roughly $3 billion of Blue Owl equity underneath	Oracle; cannot terminate the lease	Power delivery is Oracle’s job. A force-majeure notice can defer rent, but Oracle still owes carry costs covering debt interest and a Blue Owl equity return for up to about three years. Loans quoted around 89–91 cents
Saline Township, Michigan (“The Barn,” >1 GW)	Related Digital; Blackstone-affiliated equity	About $16 billion package; roughly $14 billion of bonds at a 7.5 percent coupon, maturity around 2045, PIMCO anchoring about $10 billion; Blackstone equity about $2 billion	Oracle for OpenAI	Blue Owl declined an earlier version. Banks retreated; a bond fund set long-dated terms matched to the lease. Power customer is an Oracle subsidiary

[13] [14] [15] [16] [17] [18] [19] [20] [21] [22]

Oracle’s own filings show how large the lease layer has become relative to on-balance-sheet debt. As of May 31, 2026, Oracle reported $260 billion of additional lease commitments, substantially all data-center related, expected to commence between fiscal 2027 and 2029 for terms of 15 to 19 years, not yet on the balance sheet, including one lease with a lessor-borrowing guarantee of up to $3.3 billion. By August 31, 2026, the uncommenced lease figure cited from the 10-Q was $288 billion, against about $125 billion of notes and borrowings already on the books. [23] [24]

Default recourse on these SPV loans is to the project—land, building, and in some structures the chips—not to the developer’s corporate parent. It is not, in form, recourse to Oracle either. In substance, date-certain rent, non-termination clauses, and carry-cost obligations put delay and tenant risk back on Oracle. Jupiter is the stress test: permitting slippage on a gas pipeline did not release Oracle from funding the capital structure. [18] [15]

What this means for a competitor. A neocloud or developer cannot replicate Stargate by copying the SPV. The rating and the syndication depend on an investment-grade lessee who will pay even if power is late. Weaker tenants clear only with a hyperscaler or chip-vendor backstop, and they clear at much wider spreads. A September 2026 survey of single-campus bonds showed hyperscaler or investment-grade tenant coupons roughly in a 5.70–7.875 percent band, Google-guaranteed Fluidstack leases around 6.19–7.75 percent, and CoreWeave-leased paper around 7.00–9.875 percent. [25]

xAI’s SPV finances the chips, not the building

xAI’s public structure is a hardware lease, which matters because GPU residual value depreciates on a three-to-five-year product cycle, not a 25-year shell life.

The disclosed transaction is Valor Compute Infrastructure, managed by Valor Equity Partners. Apollo-managed funds led a $3.5 billion capital solution for a $5.4 billion purchase of compute, including Nvidia GB200 GPUs, leased to an xAI subsidiary on a triple-net basis to support training for Grok. Nvidia invested in the vehicle as an anchor limited partner. Apollo described the deal as downside-protected and asset-based. A later fund document describes a related VCI vehicle leasing GPUs to SpaceX on a triple-net basis, with a SpaceX guarantee after SpaceX’s acquisition of xAI, and a private investment-grade rating on $3.5 billion of debt at a 7.375 percent coupon maturing in 2031. [26] [27]

Earlier reporting, before that close, described a larger raise of about $20 billion for Colossus compute—roughly $7.5 billion of equity and as much as $12.5 billion of debt—in an SPV that would buy Nvidia processors and rent them to xAI for about five years, with xAI’s commitment limited to rent. Those figures were pre-close and should not be stacked on top of the $5.4 billion Apollo transaction. [28] [29]

Risk sits differently than at Hyperion. Lenders are secured by the GPUs and a five-year triple-net lease, not by a 16-year residual-value guarantee sized to repay 24-year bonds. Equity (Valor, Nvidia, and other LPs) eats obsolescence and re-lease risk after the term. Nvidia’s equity check also creates a round-trip: the chip vendor funds the buyer of its own GPUs, which supports shipments but concentrates residual risk in a vehicle whose only customer is one Musk company. [30]

CoreWeave has used a related pattern at smaller scale. In March 2025 it said an SPV created to fulfill an up-to-$11.9 billion OpenAI contract would “incur indebtedness to finance its obligations,” and later borrowed further against OpenAI contracts. CoreWeave’s own credit is below investment grade, so project lenders are underwriting the offtake contract and the hardware, not a AA tenant. [29] [31]

Vendor guarantees are the other off-balance-sheet channel

Where the tenant is not itself investment-grade, the chip vendor or cloud incumbent writes the credit support and accounts for it as a derivative or a contingent guarantee, not as debt.

• Alphabet. As of June 30, 2026, Alphabet reported about $43.8 billion of credit-derivative notional tied to data-center backstops, up from $6.5 billion at the end of September 2025, plus $7.6 billion of maximum potential payments under financial guarantees and an estimated $24.1 billion of additional future backstops still subject to final terms. The fair value recorded on the balance sheet was about $815 million. The mechanism is a promise to cover lease payments if a tenant such as Fluidstack defaults, which lets the landlord’s lenders price closer to Google than to the tenant. Alphabet also had about $85.2 billion of yet-to-commence data-center leases at the end of June 2026. [32] [33] [32]
• Broadcom. Maximum potential liability under one AI-rack lease backstop for Anthropic is about $29 billion. The exposure is not a full cheque: if the customer defaults, Broadcom owes the gap between 85 percent of outstanding backstopped lease amounts and whatever the racks fetch. No payments had been made at the last filing cited, and fair value was described as not material. [34] [32]
• Nvidia. The largest single disclosed cap is guarantees of about $105 billion (about $108.5 billion maximum gross exposure in one primary-source compilation) supporting an OpenAI campus that SB Energy is developing in Ohio, roughly 4.25 GW. The guarantee phases in as leases commence, runs against 20-year OpenAI leases, and declines as rent is paid. Nvidia has also described limited residual-value support—up to 25 percent of an opportunity, project by project—as a complement to independent underwriting, not a full wrap. [34] [34] [32]

These are not homogeneous liabilities. A filing compilation of primary-source maximums put Nvidia, Broadcom, Meta’s RVG threshold, and Alphabet’s guarantees and derivative notionals at roughly $217 billion currently disclosed, or about $241 billion including Alphabet’s not-yet-final backstops. That is a ceiling stack, not debt outstanding, and not the same number as SPV principal. [32]

How big the channel is depends on which number you mean

Three published aggregates measure different things. They should not be added.

• SPV principal already raised. A Financial Times analysis in December 2025 found more than $120 billion of debt and equity supplied by PIMCO, BlackRock, Apollo, Blue Owl, JPMorgan and others into SPVs for Meta, xAI, Oracle, and CoreWeave. The named pieces—Hyperion’s ~$30 billion, Abilene’s ~$13 billion, Vantage’s $38 billion, Jupiter’s $18 billion, xAI’s chip SPV, CoreWeave’s OpenAI vehicle—are the core of that figure. Later 2026 closes (Michigan’s $16 billion, Meta’s reported El Paso deal) sit outside that December snapshot. [29] [15]
• Contingent guarantees. A September 2026 Financial Times report, as relayed by secondary write-ups, put Big Tech AI-infrastructure exposure kept off balance sheets through guarantees and SPVs at roughly $300 billion. That scope includes residual-value guarantees and vendor backstops, not just SPV debt. [35] [34]
• Signed leases not yet on the balance sheet. Moody’s, in February 2026, put undiscounted future data-center lease commitments of Amazon, Meta, Alphabet, Microsoft, and Oracle at $969 billion at the end of 2025, of which more than two-thirds—about $662 billion—had not commenced and therefore was not yet a balance-sheet liability. Moody’s said that uncommenced amount was about 113 percent of the five companies’ combined adjusted debt. Many of those leases are the offtake that supports the SPVs above; some are ordinary colocation. They are obligations of the tenant, not a measure of project-finance principal. [10] [36] [37]

Forward estimates are larger still and should be labeled as forecasts. Morgan Stanley said in October 2025 that roughly $1.5 trillion of external financing would be needed across the AI ecosystem, of which as much as $800 billion could come from private credit in asset-tied deals, including SPVs, by 2028, against a $2.9 trillion global datacenter spend estimate through 2028. [38] [39]

Who actually bears each risk

Construction risk is contractually pushed to the tenant or a guaranteed-maximum-price contractor wherever the rating depends on it. On Beignet, S&P’s explicit premise is that Meta absorbs overruns and delays so the project is not exposed to ordinary construction risk. On Jupiter, Oracle remains liable for carry costs if the site is not ready, including if power slips. Equity sponsors (Blue Owl, Vantage, Crusoe, Related, Blackstone) bear completion risk only to the extent the lease does not start paying and the tenant’s delay protections do not cover them. Banks that cannot syndicate—Jupiter loans stuck with underwriters, quoted in the high 80s to low 90s—are already wearing that execution risk at a mark. [4] [18] [24]

Tenant credit risk is the rating. Clifford Chance’s 2026 review of the data-center bond wave describes three buckets: the bond tracks the tenant when an investment-grade hyperscaler takes a date-certain lease through maturity and stands behind rent with an RVG or absolute guaranty; the bond is notched down for structural gaps; the bond is high-yield when hyperscaler support is deferred or absent. Every agency treats tenant credit as the ceiling. In Stargate, that tenant is Oracle, not OpenAI. If OpenAI cannot pay Oracle, the project lenders still look to Oracle’s lease—until Oracle itself is stressed, which is what the Jupiter marks and the Michigan bank retreat already price. [40] [41]

Residual-value risk is the risk the structures were built to move, and it has not fully moved. On Hyperion, bondholders are protected for 16 years by Meta’s declining guarantee and thereafter by a termination fee; Blue Owl’s equity is the cushion above that floor and the owner of whatever value remains if Meta stays. S&P currently assigns the RVG zero incremental corporate debt because it believes stressed asset value still covers the notes. Meta records nothing because a payout is “not probable.” Those two judgments can diverge from cash reality at the same time: the guarantee is most likely to be called in the scenario where the appraisal headroom disappears. On GPU SPVs, there is often no equivalent 16-year floor. Lenders have a five-year lease and hardware collateral; equity and, where offered, a capped Nvidia residual-support slice take obsolescence. On Alphabet and Broadcom backstops, the vendor takes the gap between contractual coverage (or 85 percent of remaining rent) and recovery value, and books only a modeled fair value. [10] [9] [34]

The practical map is therefore not “debt has left the hyperscaler.” Construction risk sits with whoever signed the completion and delay protections, usually the tenant. Tenant risk sits with the lessee and, one level down, with that lessee’s unrated customer. Residual-value risk sits with the guarantor for as long as the guarantee is in the money, with project equity for the first loss above that floor, and with lenders only if both the tenant and the guarantee fail or the collateral is worth less than the notes after enforcement. Accounting keeps the first two largely off the balance sheet until leases commence and a payout is probable. Rating agencies are already debating whether to put them back into adjusted leverage. The spread between Hyperion’s A+ at Treasuries plus 225 basis points in October 2025 and wider prints on the second Meta deal, CoreWeave-tenant bonds, and sub-par Oracle project loans is the market’s running vote on how much of that risk was actually transferred.

---

Recent Findings Supplement (October 2026)

Meta-Blue Owl Hyperion (Louisiana) remains the clearest, largest, and most documented template for off-balance-sheet AI data center project finance, with fresh 2026 confirmations on scaling, lease terms, guarantees, and accounting treatment.[1][2]

The structure uses Beignet Investor LLC (a bankruptcy-remote SPV/JV): Blue Owl funds hold an 80% interest and Meta 20%. The vehicle owns the assets and issued ~$27.3 billion in senior secured notes (due 2049, ~6.58% coupon, priced at ~T+225 bp, A+ rated by S&P—one notch below Meta’s AA-). PIMCO anchored (~$18B), with BlackRock and others participating; Morgan Stanley arranged. Meta contributed land/CIP (~$5.7–5.8B equity/in-kind) and receives a one-time distribution (~$3B); Blue Owl contributed ~$7B cash equity. Total committed capital base ~$28.8B.[3][4]

Meta leases via operating leases (4-year initial term per building, with renewal options that can extend to ~20 years total). Rent obligations begin June 1, 2029 (even during construction delays, with narrow force majeure abatements). Meta provides a payment guaranty plus a Residual Value Guarantee (RVG, initial threshold ~$28B, declining over time). If Meta does not renew/terminates and asset value falls below the threshold, it covers the shortfall. Creditors have no direct recourse to assets.[1][5]

Accounting/rating treatment: Structured as operating leases (short initial term + options help avoid finance lease classification under GAAP). Meta does not consolidate the VIE/SPV. The A+ rating reflects Meta’s credit with a modest notch for structural/execution risk. This keeps tens of billions off Meta’s balance sheet while converting capex into lease obligations.[3]

Project updates (mid-2026 onward): Scaled to 5 GW across ~3,200 acres (first 2 GW targeted online by 2030, full by ~2032–2036); total investment expectations rose significantly from the initial ~$10B announcement (now often cited in the $27B+ financing range or higher aggregate). Similar structures are viewed as replicable templates.[1][6]

Stargate-linked sites (notably Oracle/Blue Owl/Stack Project Jupiter in New Mexico) highlight execution and tenant-credit risks in construction-phase project finance.[7]

~2.45 GW campus (part of the broader OpenAI/Oracle/SoftBank Stargate initiative). Developer: Stack Infrastructure (Blue Owl portfolio). Financing includes ~$18B in bank construction loans (syndicated by ~20 banks, e.g., Santander/Jefferies involvement; quoted at 89–91 cents on the dollar pre-notice) plus Blue Owl equity (~$3B). Oracle is the tenant under a long-term lease (power responsibility on Oracle; limited termination rights).[8]

September 2026 development: Oracle sent a force majeure notice to Blue Owl/Stack citing permitting delays on a needed natural gas pipeline, seeking to defer rent start (originally targeted ~2028). This does not eliminate ultimate obligations but extends timelines and carry costs. Loans traded distressed; equities reacted negatively. Underscores construction/permitting risk borne primarily by the developer/tenant side, with heavy reliance on Oracle’s credit (recently downgraded).[7][9]

Broader Stargate financing involves multi-party JVs, partner-level debt, and vendor commitments across sites (e.g., Michigan “The Barn” campus with Related Digital/Blackstone equity and bond financing).

Crusoe/Abilene (Texas) and related sites use JV/project-level debt with hyperscaler tenants (OpenAI/Oracle historically; Microsoft expansion).[10]

Abilene campus: JV among Crusoe, Blue Owl, and Primary Digital Infrastructure (~$15B total; construction financing in the $7B+ range reported earlier). Leased initially to OpenAI/Oracle (1.2 GW phase); Microsoft confirmed for 900 MW expansion (adjacent buildings + on-site generation/BESS). Power mix: grid primary + gas backup/turbines.[11]

Recent updates (2026): Crusoe filed for additional Jayton, TX “spur” buildings (~$4.8B investment, part of broader Project Hyper/Childress-linked plans; construction start ~early 2027). Crusoe closed initial $3.9B Series F equity (~$30.9B post-money valuation, Sept 2026). Abandoned a $1.25B Boom Supersonic turbine deal (Sept 2026) while maintaining other gas generation plans.[12][13]

Vantage developments center on corporate/development-stage facilities rather than pure tenant-specific project finance SPVs.[14]

September 14, 2026: Closed a $2B 5-year revolving development credit facility (extension options) for early-stage North American projects. Initial collateral: three unnamed development assets (expandable pool). Funded by a dozen insurance/institutional investors (arrangers: Evercore, Wells Fargo). Functions as a “warehouse” for land/sitework/early construction before permanent financing. Part of >$40B total capital raised by Vantage in 2026. Separate securitizations of operating leased assets (tenant leases as collateral) carry A-(sf) ratings.[15]

This supports developer balance sheet flexibility but does not shift tenant credit or residual risks in the same SPV/lease-guarantee manner as Hyperion.

xAI structures appear more equity- and direct-build focused, with limited public detail on large SPV/project-finance leases or RVGs.[16]

xAI has pursued rapid own-build Colossus clusters (Memphis-area and others) with mobile/gas turbine power and GPU leasing to third parties (e.g., Anthropic multi-year full capacity; Google ~110k GPUs starting Oct 2026; short termination clauses in some deals). Equity raises (e.g., $20B+ Series E). Broader reporting groups xAI among users of SPVs to move data center/compute debt off-balance, but specific legal structures, leases, or guarantees are not detailed in recent sources. One related vehicle (Valor/Apollo-backed) uses triple-net leases for GPUs supplied to xAI.[17]

Broader channel size and risk allocation (new 2026 estimates and disclosures):

• Single deals like Hyperion (~$27B debt) and Jupiter (~$18B loans) are prominent. Aggregate: reports cite >$120B in data center financing debt shifted via SPVs; Big Tech guarantees/RVGs supporting ~$200–300B total AI infrastructure exposure (e.g., Meta ~$28B RVG; Nvidia $105B cap on Ohio PORTS-Pike/OpenAI campus via SB Energy; Broadcom ~$29B; Alphabet guarantees rising to $43.8B).[18][16]
• Who bears what: Construction/permitting/execution risk often on developer or tenant (e.g., Oracle power obligations; delays triggering notices). Tenant credit (hyperscaler IG or near-IG) is central to pricing and ratings (one-notch concessions common). Residual-value risk shifted to Big Tech guarantors via RVGs (covers shortfall on re-lease/sale if tenant exits and values fall due to obsolescence). Payment guarantees ensure cash flow continuity. Accounting keeps most debt/leases off sponsor balance sheets (operating classification + VIE non-consolidation); guarantees disclosed with low probability assessments in some cases.[7]

Implications for entrants/competitors: These structures lower sponsor capex intensity and leverage metrics but embed long-term tenant dependence and residual risk on Big Tech balance sheets. Execution surprises (permitting, power) can pressure secondary debt pricing and highlight the limits of credit support. New development facilities (e.g., Vantage) and equity raises provide alternative capital but at earlier, riskier stages. No major regulatory or policy shifts on lease accounting or SPV treatment noted in post-April 2026 sources.[7]

Sources are drawn from 2026-dated reporting (primarily July–October) on the specific transactions and market commentary; earlier foundational announcements are referenced only for context where needed to explain current structures.

Banks are the origination machine for the AI data-center buildout, but single-name tickets of $15–38 billion have already pushed them against internal concentration limits. The response is not to stop lending; it is to underwrite, then push risk into syndication, project bonds, CMBS/ABS, and experimental single-asset significant risk transfers (SRTs). Public data do not support a precise global bank-exposure number. Tracked retained books at the named lenders are on the order of $10–17 billion each, while system-wide estimates of bank credit tied to AI physical infrastructure run from the low hundreds of billions (narrow regulatory definitions) to roughly $800 billion (broader academic tallies).

Mega construction loans, and who actually leads them

The binding constraint is no longer finding a bank willing to look at a hyperscale campus. It is finding a syndicate that can hold a four-year construction loan whose size exceeds what any one balance sheet wants to keep. Japanese project-finance banks and U.S. bulge brackets now share the lead on the largest tickets, with pricing that started tight and has widened as Oracle-linked supply piled up.

The landmark package is the Vantage Data Centers financing for Oracle-leased campuses in Texas and Wisconsin. JPMorgan and MUFG led a roughly $38 billion senior secured package—about $23.25 billion for Texas and $14.75 billion for Wisconsin—with Wells Fargo, BNP Paribas, Goldman Sachs, SMBC, and Société Générale in the underwriting group. Both facilities were structured as four-year loans with two one-year extension options, priced around 250 basis points over the benchmark (SOFR), interest-only during construction and amortizing once operations start. [1] [2] [3] An earlier report framed a related Vantage Texas campus loan near $22–23 billion under the same lead banks. [4] [5]

A second Oracle-tied campus, in New Mexico, drew about $18 billion from roughly 20 banks, with SMBC, BNP Paribas, Goldman Sachs, and MUFG as administrative agents. Pricing was discussed at the same SOFR+250 level, again with a four-year tenor plus extensions, and the leads planned a retail syndication. [6] [7]

Other publicly reported construction and term deals show the same cast:

• JPMorgan led a $7.1 billion financing for Crusoe and Blue Owl’s Abilene, Texas campus (Stargate), where construction began in June 2024. [8]
• Bank of America was structuring agent and a major lender on about $14 billion of debt for a Related Digital / Oracle campus tied to Stargate; the bank later announced a broader $250 billion, 18-month (January 2026–July 2027) digital-infrastructure financing initiative covering on-balance-sheet loans, project bonds, and project capital. [8] [9]
• Natixis, MUFG, and Société Générale, with RBC and SMBC as arrangers, launched a $3 billion Meta-backed Ohio package ($2.1 billion plus $900 million) at SOFR+250, four years plus a one-year extension. [10]
• Morgan Stanley and MUFG arranged a $3.1 billion broadly syndicated term loan for CoreWeave to buy and install Nvidia GPUs—the first GPU financing done in that format. [11]
• SMBC says it structured and led about $25 billion of North American hyperscaler data-center construction financing from 2022 through 2024, excluding deals where it was only a participant. [12]

What this means for entrants: Lead roles now require both project-finance underwriting (power, construction, lease assignment) and a distribution franchise large enough to move multi-billion hold positions. Pure balance-sheet lenders without a takeout story are being asked to step back.

Pricing and syndication: from FOMO to a distribution problem

Initial clearing levels for investment-grade-tenant construction loans clustered near SOFR+250. That level has not held for incremental Oracle-linked paper, and the time required to place the $38 billion package is the clearest evidence that syndication capacity, not origination appetite, is the scarce resource.

More than a dozen banks lent against Oracle’s long-term leases on the $38 billion Texas/Wisconsin facilities and the $18 billion New Mexico campus at SOFR+250. By late January 2026, TD Cowen reported that borrowing costs on newer, not-yet-sold Oracle-linked data-center debt had widened to 300–450 basis points over SOFR—closer to high-yield than to classic infrastructure. [13] The $38 billion package, underwritten from around August 2025, took until around April 2026 to approach full distribution across more than two dozen banks and other investors, with less than $1 billion still seeking a home. Retail lenders needed to commit at least $300 million to earn the top fee tier. [14] [15] Business Insider reported diminished incremental interest as JPMorgan worked through the tail of that syndication, even as bankers described the projects as fully financed. [16]

Secondary marks have already slipped on at least one large ticket. Syndicate banks, including Santander and Jefferies, were quoted around 89–91 cents on the dollar on roughly $18 billion of loans tied to the Oracle-leased New Mexico campus, according to reporting cited by Disruption Banking. [17]

Weaker or construction-stage credits clear in the bond market at much higher coupons. Morgan Stanley arranged a $3.2 billion TeraWulf construction bond, Google-backed, at a 7.75% yield, explicitly to avoid staged bank drawdown reviews. [11] Applied Digital sold 6.75% and 7.00% senior secured notes due 2031, with Goldman as representative. [18]

Demand is no longer uniformly strong. JPMorgan CFO Jeremy Barnum said in July 2026 that the bank had passed on some data-center deals over power-supply and tenant questions. [19] The Information reported on September 30, 2026 that Société Générale and MUFG had pulled back from certain data-center loan deals. [20] Apollo’s Torsten Slok has noted order books on some AI-linked offerings thinning from about five times oversubscribed in February to less than two times by July. [21] By contrast, SoftBank’s separate $15 billion one-year AI bridge—led by Mizuho, SMBC, and JPMorgan, with Goldman and MUFG among participants—was more than twice oversubscribed, showing that relationship bridge capacity still exists when the credit is the sponsor rather than a single campus. [22]

What this means for entrants: Underwriting at SOFR+250 and distributing six to nine months later into a SOFR+350 market is a negative-carry problem. New lenders should assume flex language, delay-draw mechanics, and a pre-committed bond or private-credit takeout, not a hope that the bank market will absorb the hold.

How concentrated the bank books actually are

No major U.S. bank has disclosed a standalone data-center loan number. Third-party reconstructions and regulatory snapshots therefore diverge by definition—retained mortgages versus commitments versus “AI-adjacent” commercial and industrial (C&I) loans that include software borrowers.

AtriumData.ai, as reported by Bisnow in August 2026 (rankings as of around March), estimated that U.S. data centers and related infrastructure carry at least $1.3 trillion of debt. That total is broader than bank loans: utility and infrastructure loans about $428 billion, syndicated facilities about $225 billion, and hyperscaler corporate credit about $225 billion. The top 15 lenders accounted for about $196 billion, or 15% of the total. Named exposures in that cut:

• Goldman Sachs about $16.5 billion (likely understated given its arranger role; at least 48 syndicated deals)
• JPMorgan about $13.3 billion ($11 billion syndicated, $2.3 billion mortgages)
• MUFG about $11.8 billion
• SMBC about $11.5 billion across roughly 60 syndicated deals
• Morgan Stanley about $9.3 billion
• Mizuho about $8 billion
• Bank of America about $12.8 billion; Citi about $10.5 billion; Wells Fargo about $16.4 billion

The three Japanese banks together were about $31.6 billion. [8] [8] These figures are retained or tracked exposures, not peak underwriting commitments, and they likely lag the late-2025 mega-deals still being distributed in 2026.

Infralogic’s 2025 project-finance rankings show the same shift in flow. The five U.S. bulge-bracket banks lent about $13.28 billion on data-center project finance in 2025 versus about $600 million in 2024. JPMorgan ranked first in that slice at $6.72 billion. MUFG participated in 146 project-finance deals overall in 2025, far more numerous if smaller than the U.S. houses’ tickets. [23]

Federal Reserve Bank of Chicago staff, using supervisory data through late 2025, put large-bank C&I commitments to “AI-adjacent” borrowers (software, infrastructure, data-center construction, and loans secured by data centers) at about $450 billion, of which about $150 billion was outstanding. That commitment share rose from about 9% of C&I commitments in 2015 (about $250 billion) to 13% in late 2025. Direct outstanding exposure to AI-adjacent industries was about 0.8% of total assets. Across large banks, C&I outstanding to that group averaged about 9% of Tier 1 capital, with committed exposure closer to 25%. MSCI Real Capital Analytics counted only $14.9 billion of bank lending to data centers in the year through the third quarter of 2025—a property-transaction measure, not project-finance commitments. [24] [17]

A Columbia Business School estimate cited by American Banker puts bank exposure to AI physical-infrastructure debt—mortgages, syndicated loans, and project debt—at about $800 billion, with many banks described as near internal concentration limits and leverage on some projects as high as 90% debt. [25] [26] That figure is an academic aggregation, not a regulatory total, and it is wider than the Chicago Fed’s C&I snapshot.

Working estimate (inference, triangulating the sources above, not a single published total): retained bank credit specifically against data-center real estate and construction is most plausibly in the low-to-mid hundreds of billions globally. Adding AI-adjacent C&I commitments and loans to private-credit vehicles that themselves finance campuses pushes the banking system’s economic exposure into the mid-hundreds of billions, and possibly toward the $800 billion academic estimate if utility lending and broadly defined infrastructure are included. The $1.3 trillion Atrium figure is system debt, not bank debt. Banks remain the origination bottleneck even when they are not the final holder.

What this means for entrants: The visible $8–17 billion books understate risk because the same hyperscaler tenants (Oracle, Meta, Microsoft, Google, Amazon, and neoclouds such as CoreWeave) recur across campuses. Single-name and single-tenant limits bind before sector limits do.

Risk transfer: SRTs, CLOs, and ABS/CMBS takeouts

Banks are not waiting for construction loans to season. They are trying to sell the riskiest slice of a single campus, refinance delay-draw loans into bonds as soon as they are drawn, and use CMBS and ABS as the permanent takeout once a lease is in place. The transfer is real, but it is incomplete, and some of the buyers are already long the same assets in equity or private credit.

The Financial Times reported on May 3, 2026 that JPMorgan, Morgan Stanley, and SMBC were seeking ways to distribute data-center exposure, including a variant of significant risk transfer aimed at large, concentrated loans rather than a diversified reference portfolio. JPMorgan and MUFG had already spent more than six months distributing the $38 billion Oracle-linked construction debt; some banks tried to sell pieces at a discount to non-bank lenders. Linklaters had seen deals in the $500 million range backed by a single borrower, with the bank typically retaining a slice. [27] [27] Octus reported in July 2026 that data-center-specific SRTs were pricing in the high single digits (about 7–9%), tighter than a broader infrastructure SRT range of 8–12%, even though sources said development and construction risk often made up 50–70% of the referenced portfolios. Critics described buyers as investors who already hold equity in the same assets—an “asset-gathering” loop rather than clean risk transfer. One cited example was a TD Bank repeat SRT on a $2 billion pool (0–8% first-loss bought by Blackstone) that included data centers plus other assets and ramp features. [28]

Bond takeouts are the preferred exit when a hyperscaler backstop exists but does not cover construction. A Morgan Stanley-led consortium planned to refinance about $15 billion of debt on a Google-backed, Anthropic-leased Texas campus into the bond market as soon as delay-draw loans were funded. The bonds were expected to be speculative-grade because Google’s support activates only after completion, leaving investors with overrun and delay risk. Banks had also spent months seeking buyers for more than $50 billion of construction debt on several Oracle-leased projects. [29] [21]

Securitized takeouts are scaling but remain a fraction of the construction pipeline:

• About $17 billion of data-center CMBS has been issued since the start of 2025, more than triple the prior two years, and data centers are roughly 8% of new commercial-property bond deals. Most are single-asset, single-borrower. Citigroup expects about $18–20 billion of data-center CMBS next year. AAA data-center bonds have averaged about 165 basis points over their floating-rate benchmark, wider than office, retail, or industrial. Blackstone closed a $2.05 billion CMBS loan on a QTS portfolio in February 2026. [30] [31]
• Barclays counted $11.9 billion of data-center ABS issued so far in 2026. Spreads on hyperscaler ABS were about 150 basis points over the five-year Treasury, and the paper has started trading more like hyperscaler corporates than like generic ABS. [32]
• Apollo tallied $81.4 billion of U.S. data-center securitization (ABS and CMBS) since 2018, of which $18 billion was in the first half of 2026. Europe and the UK remain negligible by comparison ($1.7 billion and $2.3 billion since 2018). [33]
• Morningstar DBRS recorded $9.51 billion of U.S. data-center securitization issuance in the first quarter of 2026 alone, about three-quarters ABS historically, with CMBS regaining share via large SASB deals. [34]
• On the CLO side, Bank of America strategists estimated in September 2026 that CLO exposure to chip and data-center term loans could reach $100 billion, versus roughly $4.5 billion today. PGIM anchored a $500 million Allstate CLO, arranged by BNP Paribas, that caps AI collateral at 15%—the first explicit cap of its kind. [35]

Morgan Stanley has been the most visible architect of the non-bank structures: about $65 billion of corporate bond deals for data centers or other AI investment led or co-led since October 2025, including the Meta–Blue Owl Hyperion project bond tied to a roughly $27 billion Louisiana joint venture. [36] [11]

What this means for entrants: Construction-phase capital is a bridge, not a home. The economic spread is earned by whoever can commit the construction loan and simultaneously pre-place the SRT tranche or the post-completion bond. Holding the residual first-loss slice of a single Oracle or neocloud campus is a concentration decision, not a yield decision.

What regulators and analysts are actually saying

Official commentary has moved from “the stock of AI debt is still modest” to “the flow is now large enough to matter,” with the sharpest warnings aimed at opacity, circular financing, and the re-entry of transferred risk through non-bank channels—not at a classic bank capital shortfall today.

Bank for International Settlements. Bulletin No. 120 (January 7, 2026) argued that AI investment is large enough to require a shift from operating cash flow to debt, with private credit playing a rapidly increasing role. Outstanding private-credit loans to AI-related companies had risen from near zero to over $200 billion; risks looked moderate but hinged on earnings expectations that equity markets had priced more aggressively than debt markets. [37] [38] The March 2026 Quarterly Review (Eren, Krohn, and Todorov) documented hyperscaler gross bond issuance topping $100 billion in 2025, mostly longer than five years, and described joint-venture and special-purpose-entity financings with private credit as “shadow borrowing”—economically like debt, largely off corporate balance sheets. [39] The BIS annual report, as summarized in late June 2026, put AI-related capital expenditure by the five largest hyperscalers at more than $1 trillion in 2025–2026 combined, flagged poorly disclosed lease structures that could allow the same asset to be pledged more than once, and noted that private-credit funds had quadrupled lending to AI and IT companies over five years, with AI around 15% of those loan books, while banks are increasingly exposed by lending to those funds. [40]

Bank of England. The July 2026 Financial Stability Report said AI-related use of public markets, private credit, leveraged finance, and structured finance had accelerated and was set to rise further, from a stock that was still modest at the start of 2026. It highlighted off-balance-sheet vehicles, securitized data-center structures, and the risk that long-dated debt is financing buildings that may obsolesce if they cannot support newer hardware. On SRTs, the Bank noted that transferred credit risk can re-enter the banking system through banks’ exposures to non-banks that hold the SRT positions, making the distribution of risk more complex and less transparent. [41] [17] On September 30, 2026 the Financial Policy Committee said the likelihood that interconnected vulnerabilities crystallize had risen, citing a rapid increase in AI-related debt. It referenced Morgan Stanley’s estimate of about $450 billion of global AI-related debt issuance by early September 2026, roughly double 2025, and JPMorgan’s projection of about $4.1 trillion of AI-related debt issuance from 2026 to 2030. It also cited Morgan Stanley’s view that about $700 billion of data-center capital expenditure in 2026–2028 could be financed by private credit, inside a roughly $2.9 trillion global data-center spending forecast through 2028. The countercyclical capital buffer was left at 2%. [42] [43] [44]

IMF. The April 2026 Global Financial Stability Report treated AI-related equity concentration and interconnected financing as an amplification channel, but judged the near-term financial-stability impact modest. It noted hyperscalers had raised more than $100 billion in bonds since January 2025, supplemented by leveraged loans and sometimes circular intercorporate arrangements, against a projected AI-related capital-expenditure path that some estimates put near $3.4 trillion by the end of the decade. [45] [46] A secondary citation of Box 1.3 in that report, in a 2026 academic draft, says the IMF recorded $46 billion of data-center securitizations since 2018 (about 70% ABS, 30% CMBS) and an expectation that data-center securitization could reach $150 billion by 2028. That box figure was not independently re-extracted from the IMF PDF in this research pass, so it should be treated as reported rather than re-verified. [47]

Federal Reserve system. There is no single Board of Governors concentration limit or public aggregate for data-center loans. The Chicago Fed’s April 2026 note is the clearest supervisory snapshot: AI-adjacent C&I is a rising but still moderate share of large-bank balance sheets, and a small fraction of all AI-related debt (JPMorgan analysts’ roughly $1.2 trillion of AI-company-issued debt was cited as context). The note frames the issue as tail risk and as hard-to-measure indirect exposure through loans to private-credit lenders. [24] The Kansas City Fed’s October 2, 2026 Economic Bulletin showed AI firms’ share of investment-grade bond issuance rising from 2% in 2023 to 10% year-to-date 2026, with about $330 billion of AI-related investment-grade issuance already in 2026—ten times full-year 2023—and noted that data-center construction is a growing but still modest share of bank commercial real estate. [48]

Analyst and rating-agency overlay. Morningstar DBRS (August 27, 2026) said concentration is one of the most significant credit issues for banks in the sector, and that syndication, securitization, and risk-transfer structures shift exposure to institutional investors while reducing transparency and complicating system-wide assessment. [49] Goldman Sachs analysts tracked nearly $500 billion of AI-related debt sold through early August 2026. [21] The Reserve Bank of Australia’s October 2026 stability review, relevant as an outside central-bank read, warned that off-balance-sheet data-center funding and circular financing could create hidden exposures, citing external estimates of $1–1.5 trillion of related financial obligations. [50]

What this means for entrants: Regulators are not, on the public record, telling banks to stop originating data-center loans. They are watching whether distribution creates a less transparent version of the same concentration—in insurers, private-credit funds, CLO equity, and SRT buyers who are financed by the same banks. A lender that cannot show both a hold limit and a credible, non-circular takeout will be the one supervisors ask about first.

---

Recent Findings Supplement (October 2026)

Recent data-center financing shows continued large-scale lending by JPMorgan, Goldman Sachs, MUFG, and others, alongside emerging pullbacks, tougher terms, and growing use of risk transfers amid rising scrutiny.[1][2]

Major recent deals (announced or advanced in summer-fall 2026): JPMorgan and MUFG are in advanced talks to underwrite/lead a ~$22 billion loan for Vantage Data Centers’ planned $25 billion Texas campus (1,200 acres, phased completion mid-2026 to 2028), with Goldman Sachs in the broader syndicate and $3 billion equity from Silver Lake/DigitalBridge.[1][1] Goldman Sachs is leading/premarketing a $3.5 billion high-yield bond for Applied Digital’s Louisiana campus (15-year lease, potential Meta/Amazon tenant) and a ~$1.15 billion secured note for Digital Drive’s Virginia campus (CoreWeave tenant).[3] Other pipeline items include ~$15 billion Nexus bridge financing (Texas, Anthropic lease) and Morgan Stanley-led bond takeouts for Crusoe’s $5.5 billion bridge plus a $509 million Goldman-led GPU SPV loan.[3] An $18 billion syndicated facility (BNP Paribas/Goldman-led) supports a major New Mexico Oracle-related project.[2]

Bank of America’s $250 billion 18-month initiative (Jan 2026–July 2027) includes a closed April 2026 $16 billion Michigan Oracle/OpenAI data center financing. Morgan Stanley announced a parallel $1.5 trillion 10-year infrastructure push, while Goldman participates in a $500 billion AI infrastructure collaboration.[4]

Lender pullbacks and syndication dynamics have shifted since mid-2026. Société Générale and MUFG stepped away from certain data center loan deals (reported late September 2026).[5][2] Order books for some offerings thinned sharply (5x oversubscribed in February 2026 to <2x in July). Lenders are demanding tougher terms amid power delays, local opposition, and project risks (e.g., Oracle force majeure notice on a New Mexico site, flagged by Morgan Stanley analysts in late September as weighing on debt markets).[2][6] Pricing examples include JPMorgan-led deals discussed around SOFR+500 bps with 98–98.5 OID and rising bond yields (e.g., certain Meta-backed from ~6.58% earlier to 7.53% in July; CoreWeave-linked slices near 10%).[3][7]

Risk transfer tools (SRTs, securitizations) are seeing increased activity to manage concentration. Deutsche Bank is structuring its first project-finance SRT on a ~€2 billion ($2.3 billion) book including data center loans (September 2026).[8] JPMorgan, Morgan Stanley, and SMBC have explored synthetic risk transfers (SRTs) and private sales of data center exposures to free capacity.[9][10] Securitization/ABS markets are absorbing volumes, with ~$11.9 billion data center ABS YTD (Barclays data) and expectations of $30–40 billion annual absorption; CMBS and other structures are also used for takeouts.[11][2]

Aggregate exposure estimates and concentration remain elevated. Goldman Sachs analysts tracked nearly $500 billion in AI-related debt issuance through early August 2026.[2][12] Morgan Stanley forecasts $2.9 trillion global data center capex through 2028 (with a $1.5 trillion financing gap after hyperscaler cash flows).[2] AtriumData.ai estimates at least $1.3 trillion in U.S. data center debt (utility/infrastructure loans the largest slice).[13] Japanese megabanks (MUFG, SMBC, Mizuho) hold significant project finance exposure, including gigawatt-scale AI data center deals; Japan’s FSA highlighted concentration risk management in its July 2026–June 2027 priorities (published mid-September).[14] Chicago Fed estimates place direct AI-adjacent exposure at ~0.8% of large-bank assets (C&I ~9% of Tier 1 capital; committed closer to 25%).[10]

Regulatory and analyst commentary has intensified on concentration and systemic risks. BIS head Pablo Hernández de Cos (September 10, 2026) flagged AI infrastructure debt/private credit financing as opaque and interconnected, with top-five tech firms investing >$1 trillion in 2025–2026 and potential vulnerabilities from valuations and concentration.[15] Bank of England’s Financial Policy Committee (late September 2026 record) noted rising likelihood of interconnected vulnerabilities crystallizing, citing rapid AI debt growth (~$450 billion by early September per Morgan Stanley) and broader cyber/operational risks.[16] Reserve Bank of Australia’s October 2026 Financial Stability Review warned of hidden risks from off-balance-sheet structures, opaque interlinkages, and circular financing in data centers (roughly one-third debt-funded globally), though leverage remains contained for now.[17][18]

These developments indicate banks are scaling lending while actively managing balance-sheet pressure through syndication, pullbacks from marginal deals, and risk transfers—yet regulatory bodies are increasingly highlighting concentration and opacity as potential fault lines. Newer data (post-August 2026) centers on pullbacks, specific large Texas/Vantage talks, SRT execution examples, and coordinated warnings from BIS/BoE/RBA/FSA.

The AI data-center buildout is no longer financed mainly on hyperscaler balance sheets. Alternative managers now own the buildings, the power, and increasingly the chips, then lease them back under long contracts that keep the capex off Big Tech’s books. The risk does not stay with Blue Owl, Apollo, Blackstone, Brookfield, or KKR. It is passed through to insurers (including their own captive life companies), pension LPs, sovereign wealth funds, bond buyers, banks on construction loans, and retail vehicles that promise periodic liquidity against assets that cannot be sold on that schedule.

How the capital stack actually works

The template was set in October 2025, when Meta and Blue Owl closed the Hyperion campus in Richland Parish, Louisiana. Funds managed by Blue Owl took an 80% stake; Meta kept 20%. Blue Owl put in about $7 billion of cash; Meta contributed land and construction-in-progress and received a one-time distribution of roughly $3 billion. Total development cost is about $27 billion for a campus of more than 2 gigawatts. Morgan Stanley arranged the financing through Beignet Investor, which priced $27.3 billion of 6.581% senior secured bonds due 2049, rated A+, sold privately to institutions including PIMCO. Meta signed operating leases with a four-year initial term and options out to 20 years, plus a residual-value guarantee covering the first 16 years of operations that effectively starts near $28 billion and declines over time. Meta’s own 10-Q treats the venture as an unconsolidated variable-interest entity. The structure is the product: construction and residual risk are contractually pushed back onto Meta, the debt looks close to Meta’s credit, and Meta does not consolidate the full project. [1][2][3][4]

The same pattern now covers Oracle’s Stargate sites. On the Abilene, Texas campus, Blue Owl invested about $3 billion of equity and borrowed roughly $10 billion from JPMorgan; debt is repaid from Oracle’s 15-year lease, and people close to the deal said Blue Owl’s targeted equity return was as high as 25%. On the New Mexico campus (Project Jupiter), a bank consortium led by BNP Paribas, Goldman Sachs, MUFG, and SMBC provided about $18 billion of construction debt. Oracle has issued a force-majeure notice over power permitting, but the contract still requires it to pay a “carry cost” that covers interest to debtholders and a minimum equity return to Blue Owl funds for up to three years if power is late. Blue Owl has said the notice does not change its financial commitments. Loans tied to the project have been quoted under 90 cents on the dollar. [5][6][7]

A second layer is emerging around the hardware itself. In June 2026, Apollo and Blackstone anchored a roughly $35 billion financing (AI XPV) secured against Broadcom-designed chips leased to Anthropic, with Google backstopping site leases and Broadcom supporting the senior notes. Senior tranches were described as investment grade. Athene, Apollo’s insurance arm, was expected to take a sizable share, with the rest syndicated to insurers and banks. Anthropic gets about a gigawatt of compute without putting the debt on its own balance sheet. [8][9]

Firm-by-firm commitments

Blue Owl is the purest “landlord to the hyperscalers” franchise. Firm-wide AUM was $319 billion at June 30, 2026, of which digital infrastructure was $18.4 billion of AUM ($12.4 billion fee-paying). The broader digital-infrastructure strategy is described at about $39 billion, including Digital Infrastructure Fund III, which had raised $7.17 billion. Fund III’s since-inception gross IRR was 25.0% and net IRR 11.3% at June 30, 2026—early, mostly unrealized marks, not cash distributions. Blue Owl’s private-credit sleeve also lends against the same campuses; on the Meta deal, reporting has described about $7 billion coming from private-credit funds for the 80% stake. It also owns STACK Infrastructure and has been tied to a Crusoe/Primary Digital joint venture around the Abilene build (figures in the $15 billion range appear in industry trackers and should not be added to the JPMorgan equity/debt split above). In December 2025, talks for Blue Owl to arrange up to $10 billion for an Oracle site in Michigan stalled; Blackstone was reported in discussions to step in, and Oracle later said a different equity partner was selected. [10][5][11][12][13]

Blackstone is the largest owner-operator. After Q1 2026, Stephen Schwarzman said the data-center portfolio had topped $150 billion owned or under construction, with a prospective pipeline of $160 billion. By the July 23, 2026 earnings call, the firm said the platform had expanded to $185 billion including facilities under construction, up from $130 billion at the start of 2026, and that the market for long-term ownership of stabilized data centers could eventually exceed $1 trillion. Those figures overlap by definition (owned vs. pipeline vs. development potential) and should not be summed. The platform is built on QTS, acquired in 2021 for $10 billion. In May 2026, Blackstone Digital Infrastructure Trust (BXDC) completed an IPO of up to about $2 billion to buy newly built, income-producing centers leased to investment-grade hyperscalers—an attempted public exit for stabilized assets the private market was not absorbing at the pace of development. Retail vehicle BREIT had data centers at 27% of portfolio fair value at June 30, 2026, and invested $3.3 billion into data-center development via QTS in the second quarter. Blackstone is also a partner on the $35 billion Anthropic chip facility and on Nvidia’s compute-financing platforms. [14][15][16][17][18]

Apollo reports about $1.05 trillion of AUM at June 30, 2026, and says it has deployed more than $40 billion into next-generation infrastructure since 2022. Named data-center equity includes a majority stake in Stream Data Centers (more than 4 GW of pipeline; Anthropic has been reported in early talks to lease up to 1 GW, which developers told The Information could require at least $40 billion of capital) and the acquisition of STACK Infrastructure’s European business, cited around $4.3 billion. The economic engine is Athene: insurance liabilities fund origination, and Athene was lined up as a large buyer of the investment-grade slice of the Anthropic chip deal. Apollo was also in the running to lead Meta’s Louisiana financing before Blue Owl and PIMCO won it. [19][20][21][8][22]

Brookfield launched the Brookfield Artificial Intelligence Infrastructure Fund in November 2025 with a $10 billion equity target and $5 billion already committed by Brookfield, Nvidia, and the Kuwait Investment Authority. With co-invest and debt, the program is designed to acquire up to $100 billion of assets across energy, land, data centers, and compute. Nvidia has been described as a $2 billion anchor LP. Operating platforms include Compass Datacenters (acquired with Ontario Teachers’ Pension Plan in 2023 at a reported $5.5 billion enterprise value) and stakes in Data4, Ascenty, and others. Geographic commitments include up to about $10 billion in Sweden (Reuters converted 95 billion Swedish crowns to $10.01 billion at the time) and a framework of 20 billion euros in France—about $22.6 billion at the October 3, 2026 euro reference rate of roughly 1.13—plus a nonbinding agreement to provide up to $9 billion for Naver’s Sejong expansion in Korea alongside a separate $1 billion Nvidia investment in Naver. [23][24][25][26][27][28]

KKR has invested approximately $34 billion of equity in digital infrastructure across 24 investments and cites a 12 GW data-center pipeline (its own materials have also cited $31.3 billion of equity over six years; treat these as overlapping cumulative figures, not additive). In June 2026 it launched Helix Digital Infrastructure with more than $10 billion of committed capital from KKR, the Kuwait Investment Authority, Nvidia, and Vistra, led by former AWS CEO Adam Selipsky, with Waldemar Szlezak as CIO. Samsung and five affiliates added $1 billion in September 2026. Helix’s pitch is integrated power plus data centers, with Vistra as preferred power partner—explicitly a response to interconnection queues. Other positions include a stake in assets of Brookfield-backed Compass (expected to raise several billion dollars), a minority stake in Korea’s SK Horizon, and, with Singtel, the acquisition of STT GDC at about $10.9 billion. [29][24][20][21][30][31]

Others in the same trade. Ares is raising a global data-center equity fund and already closed a $2.4 billion Japan development fund in 2025, of which Canada Pension Plan Investment Board committed about $1.3 billion. SoftBank agreed to acquire DigitalBridge for about $4 billion. S&P Global put private-equity investment in U.S. data centers at $45.70 billion in 2025, 72% of the $63.35 billion total—driven overwhelmingly by one transaction, the Aligned deal below. [32][33][34][35]

The Nvidia and AIP joint ventures

Two platforms sit above individual campuses.

The AI Infrastructure Partnership (originally Global AI Infrastructure Investment Partnership) was launched in September 2024 by BlackRock, Global Infrastructure Partners, Microsoft, and Abu Dhabi’s MGX, with Nvidia as technical adviser. It seeks $30 billion of equity, scaling to as much as $100 billion including debt. Nvidia and xAI joined in March 2025; Kuwait Investment Authority joined in June 2025; Temasek anchored an initial closing in October 2025. The first major deployment closed July 21, 2026: AIP, MGX, and BlackRock’s GIP bought 100% of Aligned Data Centers from Macquarie at about a $40 billion enterprise value, funded with equity (trackers put the equity cheque near $21 billion), plus $5 billion of additional growth capital. Aligned has 51 campuses and more than 6.4 GW of operational and planned capacity. That is the largest private digital-infrastructure deal on record and the clearest example of sovereign, strategic, and asset-manager capital buying an operating platform rather than a single lease. [36][37][38][39][40]

On August 10, 2026, Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create independent “compute financing platforms” intended to mobilize more than $500 billion of third-party capital over time. Nvidia is not writing the checks. The firms would finance hardware and facilities and lease compute to labs, enterprises, and clouds. Jensen Huang said Nvidia may provide residual-value support of up to 25% of an opportunity, project by project, as a complement to independent underwriting—not a blanket guarantee. Twenty-five percent of a $500 billion aspiration is a large contingent number; Morgan Stanley credit analysts estimated that, under a scenario of repeated similar platforms, Nvidia’s all-in credit exposure could reach about $200 billion by the end of 2028, of which roughly $170 billion would be adjustments and contingent obligations rather than drawn debt. Those are analyst estimates, not company guidance. [41][42][43][44][45]

What the capital is being paid

Returns split cleanly by where an investor sits in the stack. There is no single “AI data-center IRR.”

• Investment-grade, hyperscaler-backed project bonds have priced in the mid-5s to high-7s. Beignet printed at 6.581% in October 2025; by early October 2026 those bonds were indicated around a 7.5% yield (price about 92), wider than the 6.6% issue yield. A later Meta-linked vehicle, Sopaipilla Investor ($12.5 billion), was indicated near 7.7%. A selected set of 2025–26 data-center bonds shows hyperscaler or investment-grade tenants at coupons of 5.70% to 7.875%, versus CoreWeave-backed paper at 7.00% to 9.875%. [1][46][47]
• Weaker tenants price like high-yield. A Blue Owl-affiliated, CoreWeave-leased 76 MW building outside Richmond priced a $1.1 billion five-year deal at a 9.25% coupon in late September 2026; CoreWeave is rated below investment grade, and the collateral is one building plus a 15-year lease. CoreWeave’s own 9.75% bonds due 2031 have traded near a 13.4% yield. [48][46]
• Development equity is underwritten much higher and is mostly unrealized. The Abilene Blue Owl equity target of as high as 25% is the aggressive end, reported by people close to the deal, not a fund-level result. On Jupiter, Bloomberg-sourced reporting put Blue Owl’s equity return at about 9% during construction, stepping up to about 11% once the campus is finished—closer to a contracted net-lease yield than a venture return. Industry commentary puts levered development IRRs commonly in a 12–18% underwriting range and stabilized unlevered yields historically around 5–6%, but those are market ranges, not audited fund results. Canada’s IMCO has described single-B to BB data-center credit at high-single to low-double digits, and mid-teens when spreads were wider a year earlier—sometimes above what the equity beneath it was earning, because debt capital was scarcer than equity. [48][5][49][50]

The spread between a 6.6% Meta-backed 2049 bond and a 25% targeted development equity return is the entire business model: managers use insurance and institutional debt to lever a contracted lease, and keep the residual and development upside (and the construction/power miss) in the equity sleeve.

Valuation, redemption, and concentration concerns

The market is already marking the risk, even where the buildings are not finished.

Credit has cheapened. Beignet’s move from a 6.6% new-issue yield to about 7.5% is a repricing of the same A+ Meta-linked paper, not a default. Jupiter construction loans under 90 cents, and Oracle’s Michigan financing falling through on stricter lender terms, show banks and credit funds will not automatically roll the next campus at the last campus’s spread. Oracle’s disclosed lease commitments jumped from $100 billion to $248 billion in the three months to the end of November 2025; Morgan Stanley has forecast net debt including leases rising toward about $290 billion by 2028. S&P has cut Oracle to BBB-. [5][9][46][51]

Power is the binding constraint, and contracts do not make electrons appear. Oracle’s New Mexico force majeure is the live case: the tenant can delay rent, but still owes carry that protects lenders and a floor return to the equity. Local opposition and interconnection queues are now a credit event, not just a development delay. Brookfield’s own infrastructure head has said AI is large enough to over-concentrate a traditional infrastructure fund, which is why they carved out a dedicated vehicle. [6][52]

Residual value of GPUs is unpriceable at this scale. Nvidia is in early talks, via Howden Re, to have insurers cover lenders if a neocloud defaults and pledged chips cannot be resold for enough to repay the loan. It has shared depreciation data with at least one insurer. An insurance executive told CNBC that insuring a single $20 billion campus was nearly impossible to price as recently as 2023. If those policies are written, obsolescence risk moves from the lender to the insurer’s balance sheet; it is not eliminated. Critics compare the web of Nvidia equity, purchase commitments, and residual support to late-1990s vendor financing. [53][54][55]

Concentration is the underwriting problem. An Ares alternative-credit note argues that despite different issuers, structures, and rating agencies, the risk converges on roughly eight names—Meta, Oracle, Microsoft, Amazon, Google, Nvidia, and on a look-through basis OpenAI and Anthropic. It cites more than 25% of the investment-grade credit market now constituting AI/tech exposure and 42% of year-to-date investment-grade private placements as digital infrastructure. A diversified book of project bonds can still be one credit story. [56]

Retail liquidity is the mismatch. Reuters reported in April 2026 that retail accounts for about 24% of Blackstone’s assets and around 40% of Blue Owl’s. Private-credit funds faced record redemption requests in early 2026, driven by software-credit worries and AI disruption fears as much as by data centers specifically. Blue Owl’s flagship credit funds saw withdrawal requests of $4.2 billion in the third quarter of 2026, down from $4.7 billion the prior quarter—easing, not cleared. Its technology-focused BDC, OTIC, saw tenders equal to 39% of shares, mostly resubmissions of previously unfilled requests. BREIT, which gated redemptions for 15 months from late 2022, has been meeting requests again and raised data centers to 27% of the portfolio; that concentration is a feature if AI demand holds and a liquidity problem if marks on development assets are challenged while retail investors want out. Blackstone has also explored a secondary sale for investors in an $11 billion property fund whose largest exposure is data centers and digital infrastructure. [11][57][58][59][17]

Who ultimately holds the risk

The managers are agents. The economic exposure sits in five places, in roughly this order of loss absorption.

1. Equity and first-loss in the project SPVs and development funds. That is Blue Owl Digital Infrastructure funds, Blackstone real-estate and infrastructure funds (including BREIT’s QTS development), Brookfield’s AI fund, KKR’s Helix commitments, and AIP’s equity in Aligned. LPs here are pensions, endowments, sovereigns (KIA, MGX, Temasek), and the managers’ own balance sheets and employee capital. A severe re-rating hits these holders first. A Chicago Booth analysis of AI-linked debt argues that permanent losses would land first on private-credit funds and dedicated AI-infrastructure credit vehicles, then on their limited partners—principally pensions and endowments—with PE-owned insurance platforms next, and BDC retail investors to a lesser extent. [60]

2. Captive and affiliated insurers. Apollo’s Athene is the clearest case: long-duration annuity liabilities are the bid for investment-grade project and chip paper. Blackstone Credit & Insurance and KKR’s insurance affiliates are in the same trade on energy and digital assets. Ares has argued publicly that insurance is the only pool deep enough to fill the financing gap, which is precisely why venture-like technology risk is arriving on insurance balance sheets dressed as investment-grade bonds. If those insurers fail, U.S. life-insurance insolvencies are absorbed by state guaranty funds funded by assessments on other insurers—not by a federal bailout of the asset manager, but not by the policyholder either, up to statutory caps. That is the political transmission mechanism critics have flagged. [61][8][56]

3. Third-party bondholders and ABS buyers. PIMCO and other institutions own the Beignet 2049s; BlackRock ETFs were reported as large buyers of associated Meta-linked debt. QTS is refinancing stabilized centers into ABS (a roughly $1 billion deal was in the market in September 2026, inside a trust Fitch sized around $3.73 billion). These holders have contractual cash flows and, on Hyperion, a Meta residual-value guarantee. They do not have Meta’s unsecured guarantee on every project. [2][62][63][64]

4. Banks, still, on construction. Abilene’s JPMorgan facility and Jupiter’s $18 billion bank club show that “private credit replaced the banks” is incomplete. Banks originate, hold through construction, and try to distribute. Loans quoted below par mean some of that risk is stuck, or will be distributed at a loss. [5][6]

5. Retail vehicles with gated liquidity. BREIT (data centers 27%), Blue Owl’s net-lease and digital REITs and BDCs (ORENT, ODIT, OTIC, OCIC), and peers such as Blackstone’s BCRED. These investors were sold income and periodic redemption windows. The underlying assets are 15- to 20-year leases, multi-year construction, and chip collateral whose resale value is what Nvidia is trying to insure. When redemption requests exceed caps, the manager either gates, injects its own capital (Blackstone put $400 million into BCRED to clear a quarter), or seeks a secondary buyer. The retail holder discovers the duration only when they try to leave. [57][65][17]

Hyperscalers have not fully exited the risk. Meta’s residual-value guarantee, Oracle’s non-terminable leases and carry costs, Google’s discussed credit support on Anthropic leases, and Nvidia’s optional 25% residual support mean the strategic tenants remain the economic backstop on the largest deals. The accounting achievement is deconsolidation and delayed cash, not a true transfer of demand risk. If AI revenue does not cover the leases, the loss still runs through Meta, Oracle, Microsoft, Amazon, Google, and Nvidia earnings—and only then through the insurers, pensions, and retail funds that financed the buildings those companies no longer show as capex.

For anyone allocating into this, the competitive question is not “who has the biggest headline fund.” It is which contracts still pay if power is three years late, which equity is levered against a single sub-investment-grade tenant, and whether the LP base can be asked for capital calls—or redemptions—on the same day the marks move.

---

Recent Findings Supplement (October 2026)

Blue Owl has aggressively scaled its data center lending and ownership while facing redemption pressure in its credit vehicles. In 2026, the firm closed a $7 billion digital infrastructure fund (May), contributed $7 billion in private credit for an 80% stake in Meta’s $27 billion Hyperion JV in Louisiana, and pursued a planned public data center REIT seeded with ~$6.5 billion of existing assets (reported September 4–5). Its Stack subsidiary also sought a $5.9 billion loan (July) and maintains a ~$39 billion digital infrastructure strategy.[1][2][3]

• Blue Owl’s private credit arm provides large-scale lending to developers and structures long-term triple-net leases with hyperscalers (Google, Meta, Amazon AI arms), creating a dual lending/ownership model.[2]
• As of early October 2026, redemption requests hit 39% on its ~$5 billion Blue Owl Technology Income Corp. (tech-focused) and 16.8–18.8% on the $35.1 billion flagship OCIC, prompting caps at 5%; pressure eased slightly quarter-over-quarter but remains elevated amid AI lending concerns.[4][5]
• Oracle invoked force majeure in September 2026 on a New Mexico Stack project (Project Jupiter, part of broader leases), seeking to delay rent start (potentially by a year+ due to power delays); related debt traded below 90 cents on the dollar.[6]

Implications for competitors/entrants: Blue Owl’s integrated credit + lease approach creates a financing moat for large projects but exposes retail/investor vehicles (BDCs) to redemption risk if AI deployment timelines slip. New entrants need strong LP relationships (pensions, insurers, SWFs) or public market access to match scale.

Nvidia’s August 10, 2026, MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR established dedicated AI compute financing platforms targeting mobilization of over $500 billion in third-party capital over time. These platforms aim to turn Nvidia compute and full-stack AI infrastructure into an investable asset class, providing dedicated capital pools at attractive rates for frontier labs, enterprises, and AI clouds, with potential residual-value support from Nvidia (up to 25% on select projects, project-by-project).[7][8][9]

• The partnerships build on existing firm-specific efforts and position compute as productive, long-duration infrastructure with usage-linked revenue potential.[7]
• No specific per-firm allocation sizes or return targets disclosed in the announcements; capital deployment remains independent per opportunity.

Implications: This shifts financing from ad-hoc bank or hyperscaler balance-sheet funding toward repeatable institutional platforms. Competitors without similar Nvidia ties or scale may struggle to access “AI factory” deals at competitive rates.

KKR launched Helix Digital Infrastructure (June 2026) with >$10 billion committed capital (anchors: KKR, Nvidia, Vistra, Kuwait Investment Authority), led by former AWS CEO Adam Selipsky; Samsung and affiliates added $1 billion in September. Brookfield is backing its own Artificial Intelligence Infrastructure Fund (targeting $10 billion equity with Nvidia and KIA) as part of a broader $100 billion AI infrastructure push.[10][11][12]

• Helix focuses on integrated financing, building, and coordination of data centers, power, fiber, and networks to address bottlenecks; KKR closed a separate $19.2 billion infrastructure fund with data center focus.[13]
• Blackstone maintains a >$150 billion data center portfolio (+$160 billion pipeline) and launched its Digital Infrastructure Trust REIT (May 2026 IPO raising $2 billion as a blind pool).[14]

Implications: Integrated power + compute platforms (e.g., KKR-Vistra, Brookfield’s energy/land holdings) differentiate winners by mitigating grid/power risks that pure real estate plays face. Public REIT vehicles provide liquidity but require stabilized assets for investor appeal.

BlackRock’s GIP, MGX, and the AI Infrastructure Partnership (AIP, including Microsoft) closed the $40 billion Aligned Data Centers acquisition (July 21, 2026; ~$21 billion equity check) and committed an additional $5 billion growth capital—the largest data center transaction to date. AIP targets $30 billion equity (scaling to $100 billion with debt). The consortium is also in exclusive talks (as of late September) for Stack Infrastructure’s Asia-Pacific assets at a potential $20–25 billion valuation.[15][16][17]

• BlackRock/GIP executed ~$57 billion in recent data center deals, including the Aligned close and $12 billion debt for a Meta Texas campus.[13]

Implications: Sovereign-backed capital (MGX/Abu Dhabi) and tech partnerships (Microsoft) accelerate platform-scale deals. Entrants targeting hyperscale or international assets will compete with these well-capitalized consortia.

Capital ultimately flows from a mix of institutional LPs (pensions, insurers, SWFs like KIA), private credit/BDC vehicles (exposing retail/investors to redemptions), and emerging public markets (REITs). Blue Owl’s BDC redemptions and Oracle-related valuation pressure highlight near-term risks around deployment timelines and concentration.[5][6]

• Insurers and pensions appear as key LPs in Blue Owl’s digital fund and similar vehicles; public REITs shift some risk to broader investors.[1]
• No major new regulatory updates or independent research publications surfaced in the period; developments center on deal announcements and execution.

Overall for market participants: The sector has moved from opportunistic lending to structured platforms and public vehicles, but high redemption activity and project delays signal that ultimate risk holders (especially in credit/BDC structures) face liquidity and valuation sensitivity if AI capex or utilization slows. New entrants should prioritize power integration, diversified LP bases, and structures that limit redemption exposure.

GPU-collateralized debt is mostly customer-contract credit with chips as a recovery floor, not a loan against resale value. Pricing, ratings, and who is on the hook if demand or residual values fall all track the offtaker—and, increasingly, Nvidia’s partial guarantees—not the GPU itself.

The loan is underwritten on the tenant, not the chip

The structure that scaled after CoreWeave’s August 2023 facility treats GPUs as pledged equipment inside a special-purpose borrower, but senior sizing is driven by take-or-pay (or equivalent) customer contracts. Lenders get paid if the named customer pays, whether or not that customer fully uses the cluster. The chips matter mainly if the customer defaults or the contract ends before the loan does.

That is why the same borrower can print investment-grade and junk paper weeks apart. CoreWeave’s March 2026 DDTL 4.0—$8.5 billion, borrower is bankruptcy-remote CoreWeave Compute Acquisition Co. VIII, non-recourse to the parent except customary carve-outs—was rated A3 by Moody’s and A (low) by DBRS because it sits on an investment-grade customer contract (widely reported as Meta, with press figures of at least $14–19 billion of contract value behind it). It priced at SOFR + 2.25% floating, or about 5.9% on a fixed tranche, and matures in March 2032. Ten weeks later, DDTL 5.0—$3.1 billion, backed by two large non-investment-grade customers (press identified OpenAI and Cohere)—was rated Ba2 / BB+, priced at SOFR + 4.50% at 99, and is guaranteed by CoreWeave and key subsidiaries. The July/August 2026 DDTL 5.5, $2.6 billion for Anthropic, Jane Street and others, cleared at SOFR + 5.50% with a 97 OID, implying about a 10.44% yield, after talk had been tighter. Same hardware class; the spread is the tenant. [1] [2] [3] [4] [5]

Contract tenor and loan tenor do not match. On DDTL 5.5 the customer contracts average about three years; the loan runs about five. The gap is refinancing and residual-value risk, not current cash-flow coverage. [6]

What this means for a new entrant: a GPU fleet without an investment-grade offtaker does not get IG pricing. The bankable asset is the contract assignment plus a depreciating recovery floor. Spot capacity is a different, much more expensive credit.

Pricing, advance rates, and the depreciation disagreement

Public 2026 prints sort into two bands. Loans on investment-grade offtakers have cleared around SOFR + 2.25% to + 3.00%, or fixed coupons in the mid-6s when sold to insurers. Loans on unrated or non-IG customers have cleared at SOFR + 4.50% to + 5.50%, with all-in yields near 10% when OID is included. Early 2023 private GPU facilities were described around 15%. Industry commentary still quotes a 9–18% range once weaker sponsors and shorter contracts are included. [1] [2] [7] [8]

Facility (2026 unless noted)	Size	Offtaker quality	Pricing	Rating
CoreWeave DDTL 4.0 (Mar)	$8.5B	IG (Meta, per press)	SOFR+225 / ~5.9% fixed	A3 / A (low); non-recourse SPV
Nebius senior secured (Jul)	$775M	IG	SOFR+250	—
Lambda TLB (Aug)	$926M	IG (Nvidia, per press)	SOFR+300 (tracker)	Baa2
Lambda fixed DDTL (Oct 1)	$1.008B	Two IG offtakers	6.78% fixed, fully amortizing to May 2033	A (low) / Baa1
Nscale Ward County DDTL	up to $1.85B	Microsoft take-or-pay	~SOFR+238	Baa1
CoreWeave DDTL 5.0 (May)	$3.1B	Two large non-IG	SOFR+450 at 99	Ba2 / BB+; parent guarantee
CoreWeave DDTL 5.5 (Aug)	$2.6B	Mixed, non-IG	SOFR+550 at 97 (~10.4% YTM)	Ba2 / BB+
Sharon AI SPV (Oct)	$365M	Not IG-rated in release	9.95% fixed	—
CoreWeave 2023 DDTL 1.0	$2.3B	Early mix	SOFR+9.62%	Unrated private credit

Sources for the table rows: CoreWeave filings and deal reporting, Lambda’s October 1 release, Nebius release, Nscale and Sharon coverage, and the CCIR instrument tracker. [1] [9] [10] [11] [8] [12]

Advance rates are less transparent than coupons. Sector commentary clusters at 50–70% of appraised or depreciable cost, with borrower equity of roughly 20–40%, and loan-to-value described by analysts in the 60–70% region. Newer CoreWeave credit agreements size draws off depreciable GPU-server cost, projected debt-service coverage, and project conditions rather than a published loan-to-value or third-party appraisal. One close reading of the DDTL 5.5 agreement found defined “GPU Depreciated Amount” language and no appraisal, residual-value, orderly-liquidation, or borrowing-base mechanics. Treat 50–70% as market practice, not a term you can read off every indenture. [13] [14] [15] [7] [16]

Depreciation assumptions are the unresolved collateral variable. CoreWeave extended computing-equipment life from five to six years effective January 1, 2023, and still uses six. Lambda uses five. Nebius uses four. Microsoft, Alphabet, and Amazon have settled near six; Meta extended to 5.5 years. Banks, per a Reuters source in October 2026, typically underwrite GPUs on a 3–4 year schedule. Nvidia argues top-tier systems can earn for much longer and has pointed to a Barkr valuation study putting GB300 NVL72 useful life at 9–10 years. S&P has said it takes a conservative view of chip value even while acknowledging GPUs have worked beyond five years so far. [1] [17] [18] [19]

Rental rates show why lenders will not take Nvidia’s economic-life argument at face value. H100 on-demand rates that ran about $7–10 per GPU-hour in early 2024 were described in 2026 coverage in a roughly $2.35–$3.50 range for budget capacity. Older A100s still earn above cash operating cost and CoreWeave signed an A100 contract running into 2029, so physical life and earning life are not the same thing—but the rent decay is what determines foreclosure value. There is still no aircraft-style public record of realized GPU disposals across generations. [1] [20] [15] [21]

Implication: a five-year loan on a three-year contract is a bet that either the customer renews, a new offtaker appears, or the hardware refinances above the remaining principal. Operators who book six-year lives and lenders who underwrite three-to-four-year lives are not describing the same collateral. The basis risk sits with whoever holds the loan after the contract expires.

How large the books actually are

A specialist filing tracker, CCIR, counted about $63 billion of GPU-collateralized commitments signed across 33 instruments as of early October 2026. That is committed amount at signing, not drawn balance, and it excludes campus lease bonds and non-GPU chip leases (it separately notes about $35 billion of non-GPU accelerator lease paper, chiefly TPUs). Earlier 2026 analyst estimates of “more than $20 billion outstanding” are now stale against CoreWeave alone. A secondary report that U.S. GPU-backed debt was about $46 billion as of August, attributed to Federal Reserve researchers via Nikkei, could not be tied here to a primary Fed document and should not be treated as confirmed. [22] [23] [10]

CoreWeave is the only fully public book. Total debt was about $35 billion at June 30, 2026 (press citing the 10-Q; one write-up uses $35.6 billion), up from $22.7 billion of long-term debt at the end of Q1. Principal debt at March 31 was $25.1 billion, of which $7.5 billion was current, against $2.2 billion of cash. Q2 net interest expense was $640 million. Revenue backlog was about $104 billion. 2026 capex guidance was $35–39 billion. Named GPU facilities include the 2023 $2.3 billion Magnetar/Blackstone deal (first H100 collateral facility), later DDTLs that took committed capacity well into the tens of billions, the $8.5 billion IG DDTL 4.0 (only $2.8 billion drawn at June 30), the $3.1 billion and $2.6 billion 2026 leveraged DDTLs, and $2.75 billion of 9.75% senior notes due 2031. Not all of the $35 billion is GPU-secured; converts and unsecured notes sit in the same total. Principal due was described as $4.4 billion for the rest of 2026 and $6.2 billion in 2027. [13] [24] [25] [16] [26]

Lambda has layered a 2024 $500 million Macquarie GPU facility, an August 2026 $926 million term loan B, and an October 1, 2026 $1.008 billion investment-grade delayed-draw term loan at 6.78% fixed, fully amortizing to May 30, 2033, secured by the funded GPU servers and contracts with two investment-grade offtakers across three deployments. [9] [27] [8]

Nebius closed its first secured facility in July 2026: $775 million at SOFR + 2.50%, maturing October 31, 2030, against deployed GPUs and an IG customer contract. It has said Microsoft and Meta contracts exceed $40 billion of additional contracted revenue and that it expects to repeat the structure. That $40 billion is contracted revenue, not debt. [11]

Crusoe is not a CoreWeave-style disclosed GPU-loan issuer at public scale. Early GPU debt was on the order of $200–225 million (Upper90). In February 2026 AMD agreed to rent GPUs if other customers were not found, supporting a roughly $300 million Goldman-arranged loan near 6%. A Jane Street contract of roughly $13 billion over five years was reported as having been pledged as chip-loan collateral before it was announced. The larger numbers—$3.9 billion Series F at a $30.9 billion valuation in September 2026, more than $140 billion of total contracted value, more than 6 GW contracted and 1 GW operating—are equity and long-dated campus commitments, not a single GPU borrowing base. Abilene campus construction debt (a reported $7.1 billion loan inside a much larger Blue Owl joint venture) is real-estate project finance with Oracle/OpenAI as the economic tenants, a different risk stack. [28] [29] [30] [31]

Fluidstack is the least transparent. Macquarie has provided GPU-secured senior debt; multiple secondary write-ups say borrowing capacity of up to $10 billion. The company has not published a drawn balance. What is documented is Google standing behind pieces of Fluidstack’s miner-colocation leases—about $1.8 billion of TeraWulf obligations, plus additional backstops elsewhere—in exchange for warrants. Forbes reported in September 2026 that Fluidstack and partners had loaded more than $15 billion of project debt and that Fluidstack raised $1.5 billion of equity at a valuation above $18 billion. Those debt figures are partner-project totals, not a Fluidstack corporate GPU facility. [32] [33] [34] [35]

Adjacent prints that belong in the same market: Nscale’s Microsoft-backed facilities (up to $1.85 billion Ward County, Baa1; a separate up to $1.2 billion), IREN’s Microsoft-linked package rated A / A (low) around $3.6–3.65 billion, and a September 2026 $22 billion bank chip loan for Blackstone and Alphabet’s Crux AI venture secured by Google TPUs and contracts—not Nvidia GPUs, but the same contract-plus-silicon template. [10] [12] [36] [37]

Implication: CoreWeave’s recourse and unsecured stack is large enough that a sector stress would show up in public credit first. Private neocloud exposure is real but not consolidatable from filings. Anyone quoting a single “GPU debt” number should say whether they mean committed facilities, drawn balances, or all neocloud liabilities.

Nvidia’s $500 billion initiative is a target, not a funded book

On August 10, 2026 Nvidia signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build compute-financing platforms aimed at mobilizing more than $500 billion of third-party capital over time. The MoUs are subject to final agreements. As of the week of the announcement, no money had been raised under them. Jensen Huang said Nvidia has the option to backstop up to $125 billion, or 25% of potential deals, via residual-value support assessed project by project and capped at 25% of a given opportunity. Nvidia’s stated design is that it contributes no upfront cash and takes on no debt; independent vehicles buy hardware and sell compute, and Nvidia’s support is a residual-value floor so lenders will treat chips as collateral. Huang framed the structure as a response to circular-financing criticism, because the equity in the vehicles is supposed to come from pensions, insurers, and sovereign funds rather than from Nvidia’s balance sheet. [38] [39] [40] [41] [42]

Wall Street has not accepted the 25% cap as sufficient. An October 1, 2026 Reuters report said bankers and asset managers doubt chips work as long-term collateral on Nvidia’s terms and want stronger guarantees. Sources said pipeline deals are likely to carry stronger guarantees and customer contracts, and that banks underwrite GPUs over 3–4 years, not the decade-scale earning life Nvidia argues for. The $500 billion figure remains a mobilization target. The only number Nvidia has actually optioned, and not yet contracted deal by deal, is up to $125 billion of residual-value support. [17] [17]

The instrument that did get signed is different, and larger in headline terms than the entire early history of GPU loans. On August 17, 2026 Nvidia filed residual-value guarantees with SB Energy covering leases for about 4.25 GW of IT load at the PORTS campus in Pike County, Ohio, where OpenAI is the 20-year tenant on a campus sized up to about 8 GW. Aggregate payment obligation on the initial guarantees is capped at $105 billion. Nvidia may, at its option, extend support to roughly another 3.8 GW. The guarantee pays the shortfall between a contractual minimum lease value and what SB Energy recovers by reletting or selling, and only after a trigger: OpenAI insolvency that defaults the lease, or OpenAI failing to pay. It is not a promise to pay OpenAI’s rent as it comes due, and it is not a guarantee of the chips inside the building. Nvidia is the exclusive AI-compute provider on the supported capacity and is investing $1.5 billion in SB Energy. The guarantees generally take effect when leases commence and ready-for-service conditions are met, expected beginning in 2028, and can terminate if OpenAI reaches a satisfactory credit rating. July press had described talks around a roughly $250 billion backstop; the filed cap is $105 billion. [43] [44] [45] [46]

Nvidia’s own quarterly disclosure put gross commitments and credit guarantees above $530 billion as of July 2026, more than triple the level three months earlier. Bucket labels differ slightly across write-ups of the same filing, but the pieces that recur are: supply and capacity commitments on the order of $279 billion (IFR grouped a larger $366 billion with supplier agreements, factory support, leases, and R&D); about $56 billion to secure land, power, and data-center capacity for customers committed to Nvidia chips; $108.5 billion of credit guarantees, of which $105 billion is the Ohio residual-value cap; about $36 billion of AI-cloud take-or-pay floors; and about $20 billion of data-center leases Nvidia signed as tenant and expects to reassign. Only a small slice of the guarantee book was in force at quarter-end ($3.5 billion of land/power/shell guarantees); the $105 billion phases in as Ohio leases start. These are maximum gross exposures, not expected losses, and not on-balance-sheet debt. [47] [48] [49] [50]

Implication: competing with this platform means competing with a supplier that can cheapen the residual-value tranche by up to a quarter of project cost. Until final agreements exist, that cheapening is a term sheet, not a market. Lenders who will not take 25% are already pushing deals back toward IG-contract structures that do not need Nvidia at all.

Equity loops: OpenAI, xAI, CoreWeave

The circular-financing critique is that Nvidia invests in, or guarantees obligations of, customers whose spending comes back as Nvidia revenue, so reported demand and Nvidia’s equity book can fall together.

OpenAI. The September 2025 “up to $100 billion” investment, deployed as OpenAI built Nvidia systems, did not close on those terms. In a February 2026 round Nvidia invested $30 billion (alongside SoftBank at $30 billion and Amazon at $50 billion, in a round reported around $110 billion). The live credit exposure is the $105 billion Ohio residual-value cap plus the $1.5 billion SB Energy equity check, with exclusivity so the site runs Nvidia hardware. OpenAI has agreed to reimburse Nvidia for amounts actually paid under the guarantees, which only helps if OpenAI can pay. [51] [52] [53] [54]

xAI. In October 2025 Bloomberg reported xAI raising about $20 billion, split as roughly $7.5 billion of equity and up to $12.5 billion of debt in an SPV that would buy Nvidia processors and lease them to xAI for five years for the Memphis Colossus 2 site, with the debt secured by the GPUs rather than xAI corporate assets. Nvidia was reported as investing as much as $2 billion on the equity side while also being the chip supplier. xAI said in January 2026 that it had completed a $20 billion round including Nvidia, without breaking out debt versus equity or confirming the $2 billion. Apollo and Diameter were reported on the debt. This is the cleanest “chips as collateral, supplier as equity investor” loop in the set: if the lease pays, debt investors are repaid from xAI’s rent; if it does not, they own depreciated GPUs, and Nvidia has already recognized chip revenue and holds equity that is first-loss. [55] [56] [57]

CoreWeave. Nvidia invested $2 billion in January 2026 at $87.20 per share (about 22.9 million shares). By the June 30, 2026 holder data it held about 47.2 million shares, roughly 8.7–11% depending on the share-count denominator used. Separately, a September 2025 agreement obligates Nvidia to purchase up to $6.3 billion of unsold CoreWeave cloud capacity through April 2032. A CoreWeave spokesperson told Reuters the January cash would not be used to buy Nvidia processors. The loop is still commercial: Nvidia is supplier, shareholder, and buyer of last resort for unused capacity. Nvidia also put a reported $2 billion into Nebius in March 2026, so the equity support is not exclusive to one neocloud. [58] [59] [60] [61] [62]

Critics (including comparisons to Lucent-style vendor financing and, on the xAI SPV, to off-balance-sheet structures) argue the ecosystem can show demand that is partly Nvidia-financed. Nvidia’s rebuttal is that it does not require investees to buy its chips with the proceeds, that its own cash generation is unlike 1990s telecom vendors, and that the $500 billion platform is meant to replace vendor financing with third-party capital plus a capped residual guarantee. Both descriptions can be true of different deals at once. The Ohio guarantee and the CoreWeave capacity backstop are contingent liabilities that only pay in a downturn—the scenario in which Nvidia’s equity stakes in the same customers are also impaired. [51] [63] [17] [41]

Who absorbs the loss

Losses do not land in one place. They follow the contract, then the recourse package, then any third-party backstop.

During a performing take-or-pay, demand risk sits with the offtaker. Meta, Microsoft, or another IG tenant pays whether utilization is high or low. That is why those loans are rated on the tenant. If the offtaker keeps paying and simply does not renew, the neocloud and its lenders hold renewal and residual-value risk for whatever principal is still outstanding. On DDTL 5.5 that gap is built in: three-year contracts, five-year debt. [6] [5]

If the offtaker defaults, recovery is the GPUs plus any replacement contract, minus the cost of pulling racks out of someone else’s data center. Orderly-liquidation estimates in market commentary run well below going-concern value—plausibly 30–50% in a distressed sale versus 50–70% of new for hardware that is still running in place—and those ranges are estimates, not transaction prints. On non-recourse paper such as DDTL 4.0, a shortfall stops at the SPV; CoreWeave’s parent is not the lender’s general credit. On DDTL 5.0, CoreWeave and key subsidiaries guarantee the debt, so parent equity and unsecured creditors are in the stack. Unsecured CoreWeave notes (9.75% coupon) have no GPU claim at all. [3] [5] [15]

Nvidia absorbs loss only inside a written cap, and only after other recovery is exhausted:

• Ohio: shortfall versus guaranteed minimum lease value, after relet or sale, capped at $105 billion in aggregate, and only on an OpenAI payment default or insolvency default. Chip residual value inside the halls is not what that instrument guarantees. Cash Nvidia has committed there is the $1.5 billion SB Energy equity, which is first-loss equity, not the $105 billion headline. [64] [45]
• $500 billion platforms: at most 25% residual-value support per deal, and only if a final agreement includes it. The other 75% or more is lender and equity risk. Reuters reporting indicates lenders are trying to push that share higher before they fund. [17] [65]
• CoreWeave unsold capacity: Nvidia is on the hook up to $6.3 billion through April 2032. Above that cap, CoreWeave equity and any unsecured or recourse lenders hold merchant exposure. Nvidia’s equity stake (about 47 million shares) is wiped out before any of those debt claims. [66] [59]
• xAI SPV: debt investors hold a five-year lease and the GPUs. Nvidia’s reported up-to-$2 billion is equity in the round—first loss on that slice—while chip sale proceeds are already Nvidia revenue. If both the lease and residual values fail, debt takes the hardware loss; Nvidia takes the equity loss and a demand hit, not the SPV’s senior principal. [55]

Other named backstops work the same way at smaller scale. Google has backstopped specified Fluidstack lease obligations (including $1.8 billion at TeraWulf) and receives warrants; if Fluidstack misses those payments, Google pays and can end up with miner equity, not with the GPUs. AMD’s Crusoe arrangement shifts unsold-GPU risk to AMD on a roughly $300 million loan by agreeing to rent the chips, which is why that loan priced near 6%. [31] [35]

Spot-exposed capacity—clusters without a take-or-pay, or capacity after the contract ends and outside any Nvidia or hyperscaler backstop—is equity risk first and lender risk only to the extent the facility is recourse or the borrowing base has not amortized. That is the slice where an H100 rent collapse actually impairs principal. Contract-backed senior debt is impaired only if the tenant fails or the residual after amortization is below the remaining loan, which is exactly the scenario the 5-year loan / 3-year contract mismatch creates.

The practical map: IG offtakers absorb utilization risk while their contracts run; neocloud equity absorbs first loss at the operating company; recourse lenders and unsecured bondholders absorb the next layer at CoreWeave; non-recourse SPV lenders absorb collateral shortfall only inside that vehicle; Nvidia absorbs residual or capacity shortfalls only up to written caps ($105 billion Ohio maximum, $6.3 billion CoreWeave capacity, up to 25% and up to $125 billion if the Wall Street platforms are ever signed on the announced terms); everyone else holding the other 75% of a chip-backed loan is holding depreciation and renewal risk that the market has not yet observed through a full default cycle.

---

Recent Findings Supplement (October 2026)

CoreWeave has shifted more residual-value and re-leasing risk onto lenders in its latest facilities while scaling total debt to ~$21.8B, with pricing now clearly differentiated by customer credit quality rather than GPU hardware value alone.[1][2]

• On August 10, 2026, CoreWeave closed a $2.6B delayed-draw term loan (DDTL 5.5) via JPMorgan and MUFG at SOFR + 5.50% (Ba2/BB+ ratings), with ~5-year maturity exceeding the average ~3-year customer contracts; the structure explicitly allows re-leasing of GPUs post-contract (subject to credit agreement criteria), marking a deliberate transfer of renewal risk to lenders compared to prior deals where contracts matched or exceeded debt tenor. The facility was meaningfully oversubscribed and dedicated to committed customer deployments.[2][3][4]
• This follows the March 31, 2026, $8.5B DDTL 4.0 (A3/A(low) ratings, SOFR + 2.25% floating or ~5.9% fixed tranche, non-recourse SPV secured by GPUs and a large Meta contract initially reported at ~$14.2B, later expanded), and a ~$3.1B DDTL 5.0 at SOFR + 4.50%. CoreWeave reported securing >$30B in debt and equity capital YTD 2026 (including these facilities) and holds a $98.8B revenue backlog as of March 31.[2][5][3]
• Broader data shows GPU-backed loans to IG customers priced at SOFR + 2.25–3.00% in 2026 versus +4.50–5.50% for unrated customers; CoreWeave’s own corporate credit remains speculative-grade, so the SPV/contract structure drives the rating uplift. Total CoreWeave debt reached $21.84B as of early October 2026 (up 172% YoY).[6][1]

This evolution means lenders now underwrite a combination of contracted cash flows and expected re-leasing/residual GPU value, increasing sensitivity to sustained AI demand and secondary-market GPU pricing. New entrants or competitors must demonstrate either hyperscaler/IG contracts or accept wider spreads and higher equity cushions to access similar capital.

Lambda has repeatedly accessed investment-grade GPU-backed debt at scale by anchoring facilities to named or implied IG offtakers and contracted cash flows, with its October 2026 deal highlighting full amortization and multi-year tenors.[7]

• On or around October 1, 2026, Lambda closed a $1.008B investment-grade delayed-draw term loan (Baa1/A(low) ratings from Moody’s/DBRS) at a fixed 6.78% semi-annual coupon, fully amortizing to May 30, 2033; it is secured by GPU servers/infrastructure and contracted cash flows supporting three customer deployments with two IG offtakers. The deal was oversubscribed and marketed to insurance companies and fixed-income investors.[8][7][9]
• This follows an August 2026 ~$926M senior secured facility (Baa2, SOFR + 3.00%, amortizing to end-2030) for GPU deployment to an IG customer (reportedly linked to Microsoft leasing in some coverage) and an earlier $500M 2024 facility. Lambda has now executed multiple ~$1B-scale institutional GPU facilities in 2026.[10]

The emphasis on contracted cash flows (rather than spot GPU values) and IG counterparties has enabled tighter pricing and longer, amortizing structures. Competitors without similar IG anchors face higher costs or reliance on shorter-term/private placements; the model rewards proven contract execution over raw hardware collateral.

Nvidia’s August 2026 $500B compute financing initiative introduces third-party institutional capital with capped residual-value guarantees, explicitly positioned to mitigate circular-financing concerns while exposing Nvidia to up to 25% downside on select deals.[11]

• Announced around August 10, 2026, via MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create “compute financing platforms” targeting >$500B in third-party capital for AI infrastructure (data centers, chips); Nvidia itself commits no cash or balance-sheet debt but offers optional residual-value support of up to 25% per deal (potentially up to $125B aggregate). The structure is lease-like: platforms buy/own hardware, lease to customers, with Nvidia backstopping a portion of collateral value decline.[12][13][11]
• As of early October 2026, reports indicate no fully signed large-scale agreements yet, with ongoing Wall Street skepticism over long-term GPU residual values and demands for potentially higher guarantees; the initiative is framed as bringing independent capital to reduce perceptions of circularity (Nvidia financing its own chip sales).[14][15][16]

If GPU values or demand fall materially, Nvidia absorbs the guaranteed portion (capped), while platforms/lenders bear the balance; the 25% cap limits but does not eliminate Nvidia’s exposure. Pure-play GPU lenders or neoclouds without such backstops face fuller residual risk, potentially requiring higher advance rates or equity.

Nvidia has made substantial recent equity investments in key customers (including CoreWeave and OpenAI), amplifying both alignment and circular-financing scrutiny, with portfolio holdings in buyers of its chips reaching tens of billions.[17]

• Nvidia invested $2B in CoreWeave (announced January 2026, with ongoing references), becoming a major shareholder to support data-center expansion (targeting >5 GW by 2030) and integration of Nvidia tech (including Rubin/Vera platforms); the cash was directed toward land, power, R&D, and workforce rather than direct chip purchases.[17][18]
• Nvidia participated in OpenAI’s large funding round (with tranches including $10B each from Nvidia and SoftBank paid around October 1, 2026, as part of a ~$122B raise where Nvidia/SoftBank/Amazon provided the bulk); its overall equity bets in AI-related companies (including frontier labs and neoclouds) have grown significantly, with public holdings alone in the tens of billions (e.g., Intel, SpaceX/xAI-derived stakes, CoreWeave).[19][20]

These stakes create direct economic alignment but fuel criticisms of circularity, as Nvidia’s investments help customers buy more of its chips. Losses from customer or GPU-value declines would hit Nvidia via both equity write-downs and any guarantee payouts.

Other neoclouds and data-center players show equity raises and contract-backed debt or leases, with limited new public GPU-collateralized loan details; risk pricing increasingly tracks tenant/anchor credit, and newer structures expose lenders to more residual value.[21]

• Crusoe raised $3.9B Series F equity at $30.9B valuation (September 2026, Nvidia among investors), citing >$140B contracted value and >6 GW contracted capacity (~1 GW operational); separate reports reference potential debt/bridge facilities (e.g., $5.5B Morgan Stanley-led) and a large Jane Street contract, but no granular new GPU SPV loan terms disclosed post-April 2026.[22][23]
• Fluidstack closed a $1.5B equity round led by Jane Street at >$18B valuation (September 2026, up sharply from prior $7.5B mark), alongside mentions of >$15B debt capacity for projects (including Google-guaranteed leases); focus remains on data-center development for anchors like Anthropic, with limited specific GPU-collateralized debt updates.[24][25]
• Sharon AI secured $365M senior secured GPU-backed SPV debt at 9.95% fixed (October 2026 announcement).[21]

Market analyses note that bond and loan pricing now heavily weights the ultimate tenant/anchor (hyperscaler/IG tighter than CoreWeave-style or unrated), with GPU hardware serving more as a secondary or re-leasing backstop.[6]

Implications for entrants or competitors: Success hinges on securing IG or hyperscaler contracts to access tighter pricing and IG ratings; pure spot-exposed or weaker-tenant structures command wider spreads and require stronger equity or guarantees. The Nvidia platform and residual guarantees provide one path to de-risk lender exposure, but cap Nvidia’s participation and leave the majority of downside with platforms/lenders. If AI demand or secondary GPU markets weaken, losses would primarily fall on lenders (via non-recourse SPVs and re-lease shortfalls), with partial absorption by Nvidia (guarantees/equity) and neocloud equity holders. Ongoing repricing (e.g., wider spreads on some CoreWeave paper) signals lenders are already adjusting for these risks.[26]

Data is drawn exclusively from sources published after April 3, 2026. Specific advance rates, depreciation schedules, and exact loss waterfalls remain largely opaque in public disclosures, with emphasis instead on contract coverage and structural protections.

The financing stress is real, but it is concentrated. Credit markets are charging a rising premium for Oracle, CoreWeave, and single-tenant project vehicles, while Microsoft, Alphabet, Amazon, and Meta still fund most of the build from operating cash and investment-grade paper that the bond market has so far absorbed. The fragile part is the mismatch between multi-year, often non-cancelable leases and the cash generation of the tenant who sits behind a large share of those leases—OpenAI—plus physical delays that keep interest accruing before rent starts.

1. Credit is repricing the structure, not declaring a default

Oracle has become the credit bellwether because it is financing a hyperscale build with a thinner cash cushion than Microsoft or Alphabet. Its three-year CDS rose from about 20 basis points in early 2025 to more than 140 basis points by early October 2026, even as the AI demand narrative held. Five-year Oracle CDS prints in late September clustered around 215–237 basis points—several times the investment-grade index near 53–55 basis points—and its 6.7% bonds due 2056 yielded above 8% for the first time, wider than the average for lower-rated B2/B paper. Apollo’s Torsten Slok tracks a basket of Amazon, Google, Microsoft, and Oracle whose five-year CDS reached about 100 basis points in September 2026, versus below 40 a year earlier. That is a price-of-capital signal, not a near-term default signal. [1] [2] [3] [4]

The mechanism is layered debt, not one corporate bond. Oracle’s on-balance-sheet borrowings were about $125 billion at August 31, 2026, with a disclosed fair value near 85 cents on the dollar. Separately, additional data-center lease commitments not yet on the balance sheet had swollen to $288 billion, starting mostly in fiscal 2027–2029 for 15–19 years—up from $260 billion at May 31. S&P cut Oracle to BBB− on July 9, 2026, one notch above junk, citing a fiscal 2027 free-operating-cash-flow deficit it forecast near $42 billion and leverage heading into the mid-4x range. A further cut would push a large slug of Oracle bonds—reports cite on the order of $120 billion—out of investment-grade indexes. [5] [6] [7] [8]

Project finance is already marked down. On September 24, Oracle sent a force majeure notice on Project Jupiter in New Mexico (about 2.45 GW, Blue Owl/Stack), citing a gas-pipeline slip to February 2027 and air-quality permitting, seeking to defer rent if the 2028 target is missed. The related ~$18 billion bank loans were quoted at 89–91 cents. Japan’s three megabanks are in that syndicate; Nikkei reported the loans already imply unrealized losses for lenders. [4] [9] [10]

CoreWeave shows the same logic one notch down the credit spectrum. A Blue Owl affiliate’s $1.1 billion five-year notes for a Virginia campus fully pre-leased to CoreWeave (76 MW, 15-year contract valued at $2.94 billion, operations 2027–2028) priced at a 9.25% yield—about 270 basis points over similarly rated bonds—with S&P at BB− explicitly citing single speculative-grade tenant risk. In July, a $2.6 billion delayed-draw term loan cleared at SOFR+550 and a 97 original-issue discount, roughly a 10.4% yield to maturity, after investors forced a 100–125 basis point widening and a cash lockbox. CoreWeave’s five-year CDS was reported near 855 basis points in late July. By contrast, a Meta-backed CoreWeave facility earlier in 2026 was rated A3 and priced near SOFR+225. The tenant’s rating, not the megawatts, sets the coupon. [11] [12] [13] [14]

What this means for entrants: investment-grade hyperscaler paper still clears, but new single-tenant or neocloud-backed deals are being underwritten as construction-plus-counterparty risk, with milestone protections and yields in the high single digits to low teens. Supply itself is warping spreads. Hyperscalers issued roughly $220 billion of debt over the prior year; Capital Group counted $241 billion from five majors year-to-date through August, and overall U.S. investment-grade spreads were unchanged at 78 basis points even as AI-linked names widened. Goldman data cited in September put new AI-related issuance near 115 basis points versus 78 for the broader investment-grade market—about 37 basis points of extra cost. Barclays has warned the investment-grade market cannot absorb the full pipeline without concessions. [15] [16] [17] [18]

2. Tenant concentration: commitments are real; OpenAI’s cash is not

The load-bearing risk is not “AI demand is fake.” It is that a private company with a multi-hundred-billion-dollar burn path is the anchor tenant on assets financed for 15 years.

S&P said OpenAI is roughly half of Oracle’s then-$638 billion remaining performance obligations. Oracle’s RPO later reached $664 billion at August 31, 2026, with only about 13% expected to become revenue in the next 12 months. Microsoft booked $24.1 billion of revenue from OpenAI arrangements in fiscal 2026 and was still owed $6.0 billion at June 30. Its commercial remaining performance obligations hit $678 billion, up 84%; excluding OpenAI, that backlog grew 25%. Amazon disclosed that OpenAI expanded an existing $38 billion AWS commitment by $100 billion over eight years. CoreWeave has about $22.4 billion of OpenAI contracts—an estimated third of its forward book—with no disclosed OpenAI credit support, while its top three customers were 72% of second-quarter revenue. In 2025, Microsoft alone was 67% of CoreWeave revenue. [6] [19] [20] [21] [22] [23]

OpenAI’s own numbers, as reported from an internal presentation, do not cover that stack from operations. The Financial Times, confirmed by Reuters, said OpenAI projects $278 billion of negative free cash flow from 2026 through 2030, revenue rising from $36 billion this year to $350 billion in 2030, and about $856 billion of compute and infrastructure spending. A $122 billion March raise at an $852 billion valuation is projected to be exhausted by 2028. Cumulative revenue through 2030 ($840 billion) is roughly in line with compute spend, not ahead of it. Altman has spoken of on the order of $1.4 trillion of computing commitments over eight years; later reporting put the through-2030 compute plan nearer $600 billion. Those figures are company projections, not audited SEC numbers—OpenAI does not file. [24] [25] [26]

The circularity matters. Microsoft both sells Azure to OpenAI and owns roughly 27% of it (stake valued around $135 billion after the October 2025 recapitalization). Amazon holds large contracts and equity. If OpenAI slows, the same firms lose a customer and mark down an equity stake. Columbia’s Stijn van Nieuwerburgh estimates the broader build needs more than $10 trillion through 2032—about 3.6% of U.S. GDP a year—and that the industry would need roughly $3.7 trillion of annual revenue by 2032 to earn the expected return, implying something like 80% annual revenue growth from a combined OpenAI-plus-Anthropic base he put near $100 billion. That is the overstatement risk: contracted backlog is being treated as cash-like when the payer’s plan is still a venture-scale burn. [20] [27]

What this means: lenders to Oracle landlords and to CoreWeave are underwriting OpenAI’s ability to raise capital through 2028–2030, not OpenAI’s current P&L. Investment-grade clouds can re-lease to internal workloads. Specialist clouds and single-purpose campuses cannot.

3. Power and pulled deals: the asset is late, the debt is not

Shells can be built in two to three years. Energization often cannot. PJM data show AI infrastructure projects that entered service in 2025 took more than seven years on average—over three years to an interconnection agreement, then about four more years after approval. Permitting, “other,” and supply chain dominated milestone slips; substation transformer lead times are reported above 160 weeks. Goldman has said only about half of data-center capacity scheduled for the next one to two years is expected online on time. [28] [29]

Texas is the acute policy shock. Governor Abbott ordered an audit of data centers in an ERCOT large-load queue he put at about 474 GW—more than five times peak demand, with roughly 90% data centers—and froze progress on new connections. BloombergNEF said the pause affects almost 49.8 GW and puts about 20% of the U.S. data-center pipeline at risk of delay, with revenue losses that could reach $8 billion by the first quarter of 2027 if 60% of delayed capacity is AI-related (BNEF’s implied AI earning rate was about $1.76 billion per gigawatt per month). FERC, on September 30, accepted but suspended PJM’s plan to procure new capacity until February 28, 2027, with Chair Laura Swett refusing an “11th-hour” mechanism with billion-dollar consumer implications. In June, FERC had already issued show-cause orders to all six U.S. regional grid operators on large-load interconnection. [30] [31] [32]

Several headline “Stargate” sites were not built as announced. OpenAI and Oracle canceled an roughly 800–900 MW expansion beside the Abilene, Texas campus over financing and a revised demand forecast; Microsoft leased that capacity from Crusoe instead. OpenAI paused Stargate UK in April 2026, citing power costs and regulation; a Guardian investigation found no construction and little evidence OpenAI had even visited a key site, and described much of the government’s £30 billion figure as hypothetical. Norway capacity initially branded Stargate was contracted by Microsoft. The Information reported in February 2026 that the Stargate joint venture itself had not staffed up and was not developing OpenAI’s data centers. Capacity often moved to a better-capitalized tenant rather than disappearing—which is evidence of demand, and also evidence that OpenAI could not carry the long lease at the offered price. One person involved called the Abilene expansion price “insane.” [33] [34] [35] [36] [37]

Balance sheets already show the lag. Construction-in-progress not yet in service totaled about $250 billion at the latest year-ends for Alphabet ($78.6 billion), Amazon ($71.7 billion), Meta ($50.5 billion), Oracle ($40 billion), and CoreWeave ($9.4 billion). Interest and lease clocks run while that capital earns nothing. [38]

4. Depreciation is a collateral and earnings-quality debate, not a settled fraud

Michael Burry’s claim is that a two-to-three-year Nvidia product cycle cannot support five-to-six-year book lives, and that hyperscalers will understate depreciation by about $176 billion from 2026 to 2028, overstating Oracle earnings by roughly 27% and Meta’s by about 21% by 2028 if those lives hold. Jim Chanos has argued GPU-cloud is a commodity business and that if economic life is closer to five years than ten, “there’s going to be debt defaults” on chip-secured loans to loss-making neoclouds. Nvidia, by contrast, has pointed to A100s still in commercial service six years after the 2020 launch, and CoreWeave’s CFO said the company signed an A100 contract extending into 2029 and is largely sold out of older generations. Bernstein’s Stacy Rasgon has argued five-to-six-year accounting is reasonable because even older A100s still earn comfortable margins. [39] [40] [41] [42] [43] [44] [45]

The accounting record is mixed, which weakens the “everyone is extending lives to flatter earnings” version. Effective January 2025, Meta lengthened most servers and network assets to 5.5 years, cutting depreciation by about $2.9 billion. Amazon shortened a subset from six years to five, adding about $1.4 billion of depreciation, and said the reason was the faster pace of AI technology. Microsoft has said about two-thirds of capex is short-lived assets, mostly GPUs, depreciated over six years, and that capacity is already sold for the entirety of that life. Nobody discloses GPU-only useful lives; filings are at the “servers and network equipment” level. Resale data (Silicon Data and others) show third-year H100 systems still worth a large fraction of new, and six-year-old A100s still worth roughly a quarter of cost—above a five-year straight-line residual of zero, but well below par. Useful life is a stack of workloads (training, then inference, then cheaper batch), not one number. [38] [46] [47]

The financing implication is narrower than the earnings debate. GPU-backed loans to neoclouds assume residual values and rental rates that have to survive the next two Nvidia generations. If rental rates on H100-class hardware fall faster than debt amortizes, the first losses sit in private credit and high-yield project bonds, not in Microsoft’s income statement.

5. Who is warning, and how hard

Rating agencies are explicit and differentiated. Moody’s projects hyperscaler capex (including Oracle and CoreWeave) of $785 billion in 2026 and about $1 trillion in 2027, and says lease commitments across the group have reached $1.2 trillion, more than $820 billion of which has not started. It treats those as debt-equivalent, says the shift from asset-light to asset-heavy “threatens credit quality,” and locates immediate pressure at Oracle and CoreWeave—not at Microsoft, Alphabet, Amazon, or Meta, whose ratings it does not see as imminently at risk. S&P’s Oracle downgrade named OpenAI concentration as a key credit risk and contrasted Oracle with AWS, Google, and Microsoft, which can absorb spare capacity internally. [48] [49] [50]

Regulators are watching the lenders, not banning the loans. On September 25, Japan’s Financial Services Agency stepped up scrutiny of megabank and life-insurer exposure to AI data-center project finance, mainly in the United States, including concentration-risk management. The agency said it is not trying to shut off funding. Singapore’s central bank has flagged uncertainty over sustaining AI investment as a macro-financial risk. [51] [52]

Short sellers are focused on off-balance-sheet stock and depreciation. Burry estimates nearly $1.2 trillion of uncommenced lease commitments and more than $1.5 trillion of purchase commitments at the hyperscalers, above $3 trillion once guarantees and SPV backstops are included, and has disclosed a short in Oracle. Chanos has called this the first “four-in-one” bubble—policy, technology enthusiasm, credit, and data-center real estate—and has said Microsoft and Meta can finance from cash flow while others need external capital. These are arguments, not findings of fact; the lease and purchase figures overlap with what Moody’s and company filings already disclose. [40] [53] [54] [41]

6. The counterargument: cash engines, absorbed supply, and bad historical rhymes

The strongest balance sheets are still funding a large share of the build from operations, and the bond market has not seized up.

Microsoft is the cleanest case. Forward free cash flow was still positive at about $33 billion in Apollo/FactSet figures cited by CNBC in September, with debt-to-equity near 7%. Fiscal 2026 operating cash flow was about $183 billion against capex of about $116 billion. Alphabet, Amazon, and Meta are the pressure cases inside the investment-grade complex: Alphabet posted a negative quarterly free cash flow of $5.9 billion and raised 2026 capex guidance to $195–205 billion; Amazon’s trailing-12-month free cash flow fell to negative $7.6 billion as property and equipment spending rose $66 billion; Meta’s second-quarter free cash flow fell 91% to $784 million as capex rose 83% to $31.1 billion. JPMorgan Asset Management’s Charles Wu said hyperscalers are still funding much of the need from operating cash flow, plus about $250 billion from the bond market and a similar amount from bank loans—foreign pensions, insurers, and Gulf and Asian capital included. Epoch AI’s June analysis had aggregate cash capex on trend to overtake operating cash flow around the third quarter of 2026, with Oracle already across that line and Microsoft not until around 2028 on then-current trends. S&P Global Market Intelligence’s point is the transition risk: the “Hyper 5” spent $1.1 trillion of capex over five years and are estimated to spend another $5.3 trillion through 2030, and the danger is debt-funded investment before returns show up. Capex-to-depreciation intensity already exceeds dot-com and Great Recession peaks. That is a funding-mix warning, not a claim that Microsoft is about to lose investment grade. [55] [56] [57] [58] [59] [60]

Historical comparisons cut both ways, and the differences are the point.

• Telecom/fiber, 1996–2002. Carriers laid on the order of 80 million miles of fiber; by 2002 only about 2.7% was lit in commonly cited figures, and Global Crossing, WorldCom, and others failed under debt. The infrastructure later became cheap capacity for the cloud. Differences that weaken a straight replay: utilization of AI clusters is high, not 95% dark; wavelength multiplexing multiplied fiber capacity on the same glass, whereas GPU performance gains are slowing and power per rack is rising; fiber lasts decades with almost no operating cost, while GPUs must earn their keep in a few years and then be replaced. Similarities that still bite: revenue was assumed, debt was contracted, and the equity of the builders was wiped out even though the technology was real. [61] [62] [63]
• Shale, 2010–2015. The large producers spent themselves into negative free cash flow; aggregate free cash flow stayed negative into the 2015 price collapse and recovered only years later. The business was not fraudulent. The pace of investment exceeded internal funding. Aggregate hyperscaler free cash flow has been described as down more than 90% from a late-2024 peak, with nearly $500 billion of debt and equity raised or announced since then. The shale lesson is about the funding gap and the creditor, not about whether the resource existed. [64]
• Railways. Nineteenth-century railroad manias overbuilt, bankrupted promoters, and left track that later users exploited. Social returns and investor returns diverged. That is the right frame for “the buildout can be historically important and still produce a credit cycle.” It is a poor frame for assuming GPU campuses have railroad-like residual lives. [65] [63]

A fair synthesis: the thesis that AI infrastructure will be used is better supported than the 1999 bandwidth thesis was at the peak. The thesis that today’s financing structure—15-year leases, GPU-secured loans, and OpenAI-backed backlog—will earn equity-like returns for every layer of the capital stack is what credit markets are marking down.

Ranked stress indicators and the channels they hit

Ranked by how directly they threaten the financing structure, not by how loud the equity debate is.

1. Oracle rating path and CDS (BBB−, 5-year CDS ~215–237 bp, long bonds yielding 8%+). Channel: investment-grade index eligibility, cost of Oracle’s own bonds, and every project loan or lease that prices off Oracle as tenant. A junk downgrade is the single cleanest observable that would force mechanical selling. Watch S&P/Moody’s outlooks and whether 5-year CDS holds above 200 bp. [4] [6]

2. OpenAI cash runway versus contracted take-or-pay. Channel: Oracle RPO quality (~half of the backlog, per S&P), CoreWeave forward book (~$22 billion, unsecured by OpenAI credit), Microsoft growth optics (most of the RPO increase), and private rounds that have to refill a burn projected at $278 billion through 2030. The observable is not a CDS—OpenAI has none public—but missed funding steps, IPO timing (Altman has ruled out 2026), and any renegotiation of cloud minimums. [24] [6] [50]

3. Stressed project-loan marks and force majeure. Jupiter loans at 89–91 cents, with Japanese and U.S. banks in the syndicate, are an early loss channel for construction finance before corporate bonds gap. Further notices, or secondary marks below 85, would tighten terms on the next Oracle-tenant campus (the separate large Vantage Texas/Wisconsin package has been discussed in the same ecosystem). Channel: bank project finance and Blue Owl-style equity sponsors. [5] [4] [51]

4. Neocloud all-in yields and CDS. CoreWeave-linked bonds at 7–9.9% and loans near 10%, versus Meta-backed paper near 6%, are the spread between “IG tenant” and “speculative tenant.” A sustained CoreWeave CDS above several hundred basis points raises refinancing risk on floating-rate GPU debt (the company has said each 100 bp of rates adds about $30 million of interest expense on the June 30 floating-rate balance). Channel: high-yield bonds, private credit, and Nvidia-linked collateral values. [11] [14] [66]

5. Energization slippage versus lease start dates. Texas audit (20% of the U.S. pipeline at risk, per BNEF), PJM’s seven-year clock, transformer lead times over 160 weeks, and ~$250 billion of construction-in-progress are the physical constraint. Channel: developers’ interest during construction, customer rent-start dates, and any financing that assumed 2027–2028 commercial operation. Goldman’s “half of near-term capacity on time” is the base case to track, not a tail. [30] [28] [29] [38]

6. Lease-commitment growth versus free cash flow at the weaker IG names. Moody’s $1.2 trillion lease stock, Oracle’s $288 billion uncommenced leases, and negative free cash flow at Amazon, Meta, Alphabet, and Oracle are the slow-burn version of the same risk. Channel: future fixed charges and rating headroom, not day-to-day liquidity. Microsoft remaining free-cash-flow positive is the control variable; if that flips while capex guidance rises, the “cash-funded build” counterargument narrows to Alphabet’s and Amazon’s non-cloud franchises. [48] [5] [59]

7. GPU rental-rate and residual-value breaks. Still mostly a debate. The observable that would make it a stress indicator is a break in multi-year pricing for prior-generation chips (A100/H100 contract rates and secondary prices), not another argument about six-year versus three-year accounting. Channel: earnings quality at clouds that extended lives, and recovery rates on GPU-secured loans. CoreWeave’s 2029 A100 booking is the bullish print; a failed refinancing of a chip-backed facility would be the bearish one. [43] [47]

8. Broad investment-grade spread contagion. Not happening yet. Aggregate IG spreads were flat through August while AI issuers paid a premium, and dealer commentary describes indigestion and concessions, not a buyers’ strike. This indicator matters only if AI supply starts widening the whole corporate index, which would raise the hurdle rate for every data-center bond at once—especially with the 10-year Treasury near multi-decade highs around 5.2% in early October. Channel: the entire debt-funded half of the build. Morgan Stanley has estimated about $3 trillion of AI spend through 2028, roughly half debt or debt-like. [16] [67] [68]

Bottom line. The overstated thesis is that backlog, investment-grade ratings, and “this time the assets will be used” make the capital structure safe. The assets can be used and the marginal lender can still lose money if rent starts late, the anchor tenant must keep raising equity to pay, and the debt was sized to a six-year chip life and a 15-year lease. The well-funded counterargument holds for Microsoft and, with more strain, for Alphabet, Amazon, and Meta. It does not hold, on current market prices, for Oracle-tenant project loans, CoreWeave-backed junk, or any structure whose cash flow depends on OpenAI meeting a plan that still burns cash through the end of the decade.

---

Recent Findings Supplement (October 2026)

Oracle’s credit markets flashed red in September 2026 as CDS spreads hit records and long bonds breached 8% yields, triggered by a force-majeure notice on its 2.4 GW Project Jupiter data center in New Mexico.[1][2]

This reflects investor repricing of AI-related debt amid rising capex, power delays, and leverage concerns. Oracle’s 5-year CDS reached 227–240.7 bps (record highs as of late September), up sharply from earlier 2026 levels; its 6.7% 2056 bonds yielded over 8% for the first time. S&P downgraded Oracle to BBB- in July, putting ~$120 billion of bonds at risk of IG index exclusion on further cuts. Broader hyperscaler CDS (Amazon, Alphabet, Microsoft, Meta, Nvidia) also widened, with Apollo’s basket exceeding 100 bps—the highest in its eight-year series—versus ~40 bps for major banks. Goldman Sachs estimates hyperscaler debt issuance could hit $420 billion in 2027 amid ~$700 billion 2026 capex (projected >$1 trillion in 2027).[3][4]

• Oracle sent the force-majeure notice citing regulatory/infrastructure (power) hurdles that could delay the 2028 opening; the company maintains alignment with partners but the signal rattled markets.[1]
• CoreWeave’s data-center-backed bonds (e.g., $1.1 billion BB- notes leased solely to it) priced at 9.25% (270 bps above comparables) due to single speculative-grade tenant risk; its total debt reached ~$35 billion by June (up from $21 billion end-2025), with quarterly interest expense at $640 million (2.4x YoY).[5][6]
• Earlier 2026 saw some CoreWeave-linked yields ease (e.g., to 7%), but September pricing and convertibles ($4.2 billion upsized 2.875% notes due 2033) highlight persistent high costs for non-IG tenants.[7]

For competitors or new entrants: Rising borrowing costs and index exclusion risks raise the bar for non-hyperscaler or lower-rated players; project finance now explicitly prices tenant credit quality, favoring those with diversified or IG-backed offtake.

Power and grid constraints intensified in September 2026, with Texas halting new data-center permits and PJM facing capacity shortfalls that threaten timelines.[8][9]

Texas Gov. Abbott directed an audit of data-center interconnections (ERCOT queue: 474 GW, ~90% data centers), pausing permits until mid-October and exposing ~50 GW (~20% of U.S. pipeline). Goldman Sachs estimates only 50–60% of planned capacity will come online in the next two years due to delays/cancellations. PJM’s large-load procurement plan was accepted but suspended five months by FERC; interconnection queues in hot markets run 4–7 years, compounded by 2–3+ year transformer lead times and rising local opposition (45 U.S. projects worth $68 billion blocked/delayed in Q2 2026 alone, following $130 billion in Q1).[10][11]

Oracle’s Project Jupiter force-majeure explicitly tied to power delays underscores the mechanism: without grid connections, GPUs sit idle and revenue commitments slip, pressuring debt service.[12]

For competitors: Site selection now requires proven power headroom or on-site generation; delays amplify capex timing risk and favor operators with existing grid access or flexible offtake that can absorb slippage.

Tenant concentration around OpenAI (and Anthropic) emerged as a focal risk, with Microsoft disclosing heavy reliance and analysts flagging single-point-of-failure dynamics.[13]

Microsoft booked $24.1 billion in FY2026 revenue from OpenAI arrangements (~70% of its AI sales and a material share of Azure growth to >$100 billion); OpenAI’s $250 billion Azure commitment (part of Microsoft’s $678 billion commercial backlog) and revised revenue-share terms (capped through 2030) highlight the linkage. OpenAI’s ARR approached $70 billion by late September (up >70% since July, driven by enterprise). Steve Eisman warned that ~70% of hyperscaler AI revenue traces to OpenAI/Anthropic (potentially 25–35% of their cloud revenue). Oracle’s compute backlog is reportedly ~50% tied to OpenAI.[14][15]

For competitors: Diversifying beyond a handful of frontier labs reduces offtake risk; long-term contracts provide visibility but concentrate renewal and credit exposure.

Accounting debates intensified around GPU useful lives and off-balance-sheet liabilities, with short seller Michael Burry estimating $176 billion in understated depreciation through 2028 and >$3 trillion in hidden commitments.[16]

Burry argues hyperscalers’ 5–6 year schedules (some extended, e.g., Meta to 5.5 years) understate obsolescence given 2–3 year generational leaps; Amazon shortened some server lives to 5 years. Counter-evidence includes CoreWeave’s A100 rentals extended into 2029 and active resale markets where older GPUs hold or gain value due to memory/workload fit. Off-balance-sheet items (non-cancellable leases, purchase commitments, SPVs, guarantees) add substantial leverage not yet on balance sheets.[17][18]

For competitors: Conservative depreciation and transparent residual-value assumptions improve credibility with lenders; over-reliance on extended lives risks future earnings hits.

Rating agencies and analysts issued direct cautions on credit quality and financing complexity, while short sellers targeted specific names.[19]

S&P warned hyperscaler credit quality is “gradually weakening” amid rising capex (> $7 trillion to 2030 for top six), complicated/off-balance-sheet structures, and delayed ROI; negative free operating cash flow projected for 2026–2027 at several firms. Apollo’s Torsten Slok highlighted CDS widening as repricing of fundamentals (debt-financed capex, neg FCF, depreciating assets). Goldman Sachs turned cautious on hyperscaler bond supply. Burry shorted Oracle and flagged hidden liabilities.[20][21]

Counterarguments center on investment-grade or near-IG balance sheets and revenue growth providing runway, though capex is consuming nearly all operating cash flow.[22]

Top hyperscalers retain strong ratings (Microsoft AAA; others AA-/A range) and large cash piles, allowing additional debt issuance without immediate downgrade. Revenue momentum (e.g., OpenAI ARR surge, Azure growth) supports commitments, and long-term contracts provide contracted cash flows. Some project yields eased earlier in 2026 amid strong demand for certain structures. However, aggregate capex is projected to consume 90–100% of operating cash flow in 2026 (gap widening to hundreds of billions cumulatively), forcing external financing.[23]

Ranked observable stress indicators (most recent/acute first) and primary channels affected:

1. Oracle CDS/force-majeure (Sep 2026) — Credit markets + project execution (power).
2. Texas permit pause + PJM/FERC actions (Sep 2026) — Grid interconnection + timelines.
3. CoreWeave tenant-specific bond pricing (Sep 2026) — Non-IG / single-tenant project finance.
4. Microsoft/OpenAI concentration disclosures + Eisman warning (Aug–Sep 2026) — Revenue concentration + offtake risk.
5. S&P/Apollo/Goldman cautions + Burry short (Jul–Sep 2026) — Rating agency + analyst/short sentiment.
6. GPU depreciation/off-balance-sheet debate (ongoing, Sep spikes) — Accounting/earnings quality.
7. Opposition-driven project blocks ($68B Q2) — Permitting + local/regulatory.

These primarily pressure credit costs, project delivery timelines, and earnings visibility for leveraged or concentrated players, while IG hyperscalers retain more flexibility via balance-sheet strength and revenue scale. Historical parallels (e.g., telecom overbuild) suggest demand can eventually absorb supply, but near-term friction in power, credit, and concentration is elevating execution risk.