§ 03 REPORTS / TRADFI / 2026-08-30

Four Ways the AI Credit Stack Ends

FILED UNDER TRADFI § 2026-08-30

This report maps the twelve layers financing the AI buildout, sets out the four ways it can end, and puts odds on each.

tl;dr

Note: This report was co-written with Claude — Opus 5 drafted the first version, Fable 5 edited, fact-checked and wrote this expansion. Research briefs were generated by Moonshot's Kimi K3; two independent model reviews (xAI's Grok, and a separate Claude instance) were applied before publication. Every number carries its source in the Facts list. Nothing in this report is investment advice.

Contents

  1. The short version
  2. What the words mean
  3. The system as it stands
  4. What history says
  5. Private credit and the BDC scare of 2026
  6. Where the VCs and LPs sit, and the chain
  7. Why the money was there in the first place
  8. Scenarios with probabilities
  9. What Washington does
  10. The feedback loop
  11. Bans and slowdowns
  12. Positioning
  13. Facts used

The short version

Big technology companies are building computer warehouses the size of small towns, and the five biggest plan to spend around seven hundred billion dollars on them this year.

Almost none of them pay for it entirely out of pocket. Four of them, Microsoft, Google, Amazon and Meta, earn enough from other businesses that they mostly could, although only one of them currently has cash left over after paying for the building. Oracle borrows a lot. Smaller companies that rent out computing power borrow far more relative to their size, at high interest rates, using the chips themselves as security. The AI labs that use the computing raise fresh money from investors about once a year, a short leash for promises that run fifteen or twenty years.

The money comes from ordinary people, mostly without their knowing it. Some of it is a pension, some the reserve behind a life insurance policy, some a university's savings, and some is Japanese households about to buy a bond from SoftBank, a large Japanese investment company that put money into OpenAI. It ended up here because interest rates were held near zero for about a decade, so anyone who needed income had to go somewhere riskier to find it, and because rules written after 2008 pushed risky lending out of banks and into investment funds that face fewer rules.

Four things can happen from here.

The most likely, a little under half, is that people do want the computing but the price of it falls faster than the borrowed money can be repaid, so the heavily borrowed small builders fail while the giants are fine. A bit under a quarter is that towns, states and electric utilities slow the building down, because power bills are rising and voters have noticed. A bit over two in ten is that the lenders stop lending before demand ever falls, which hits the edges fast even though the funds themselves pay out slowly. And about one in ten is that everything works at once, which needs interest rates to fall, and they have not.

If it goes wrong, the order of pain is predictable. Shareholders in the small borrowers lose everything first, then their lenders. Then the investment funds that made those loans tell their own investors they cannot have their money back this quarter. Then pensions and insurers lower the value of what they hold, quietly, over several years, which is the slowest loss and the one nobody votes on. People who own ordinary index funds, the kind that simply buy every big company, lose money on a bet they never chose to make.

Washington probably does less than people expect. The law written after 2008 says the central bank may rescue a whole market but not one named company, and Congress has shown no appetite to vote for one. Almost every comparable building spree in history, from British railways to American electricity to nuclear plants to satellite phones, ended with the builders wiped out and no rescue. The one exception was the banks in 2008, rescued because their money ran through the payments system. This money does not.

Three things are worth watching. The rental price of computing power. Whether an investment fund refuses to give its investors their money back. And whether the towns keep saying no.

What the words mean

Every finance word used later is here.

Also used, and worth knowing: ABCP is short-term commercial paper backed by a pool of assets, the overnight funding of 2007; an SIV was a vehicle that funded long assets with it. Amortization is repaying a loan's principal on a schedule rather than at the end. Basel III is the post-2008 international bank capital standard. Float is the premiums an insurer holds before it pays claims, invested meanwhile. Flexed, of a loan, means repriced wider to get it sold. GSEs are Fannie Mae and Freddie Mac. High-beta means moves more than the market, in both directions. Holdco / opco are a parent holding company and the operating business under it. Loan-to-value is the loan divided by the collateral's mark. Multiple is the price paid per dollar of earnings; multiple compression is that number falling while earnings do not. Securitization is turning a pool of loans or leases into bonds; a tranche is one slice of those bonds, senior slices paid first and the equity tranche last. SOFR is the overnight dollar rate that floating loans are priced off. TALF was the 2008 and 2020 Fed program lending against asset-backed bonds. TARP was the $700B rescue fund Congress passed in October 2008. Vintage is the year a fund started investing.

The system as it stands

Twelve layers, each funded differently and each breaking differently. The right-hand column is what the market charges to insure that layer's debt. Where it says no price exists, that is the interesting part.

Layer Players How it is funded Betting on Fragility Price of that risk
1. IG hyperscalers Microsoft, Alphabet, Amazon, Meta Operating cash, IG bonds ($182–200B issued YTD 2026 against ~$100–121B in all of 2025, facts #6), leases and part-owned vehicles that miss the debt line AI revenue arrives before the cash pile runs down; Bloomberg Intelligence models the turn in 2028 (facts #9) 2026 capex ~$690–770B against ~$125–130B in 2021 and $162B in 2022; only Microsoft generates free cash flow today (facts #9) Meta ~90–93 bps, Alphabet 67, Amazon ~65–70, all records 28–29 Jul. Group ex-Oracle ~49 bps on 24 Jul, before the 28 Jul records, widest since at least 2018 (facts #4)
2. Oracle Oracle A $45–50B debt-and-equity plan plus a $25B bond (facts #7) A backlog concentrated in a few AI customers turns into cash ~$100B of debt plus ~$250B of 15- to 20-year lease commitments. Barclays saw cash running out by Nov 2026 without financing (facts #7) CDS ~215 bps, an 18-year high, ~4x the IG index at ~53. 2054 bonds yield 7.8% (facts #1, #7)
3. Vendor-financiers Nvidia, Broadcom Own cash plus the "over $500B" platform announced 10 Aug 2026 with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR (facts #10) The non-IG buyers they stand behind stay solvent long enough to pay Over $540B of deals announced in 2026, some promising what used GPUs will be worth (facts #10). Broadcom backstops the $6B A1 and $24B A2 notes in the $35B Anthropic vehicle; the $4.5B B notes carry no guarantee (facts #16). No Nvidia partner has disclosed a dollar commitment and no first project is named Nvidia CDS 82 bps, a record on 27 Jul, against ~41 in mid-June. Trades like BBB on an AA rating (facts #4)
4. Neoclouds CoreWeave, Nebius, Core Scientific, TeraWulf, Applied Digital Loans secured on GPUs, convertibles, high-yield bonds Signed backlog converts to cash before the debt comes due CoreWeave Q1 2026 interest expense $536M, free cash flow −$4.71B. Its $2.6B loan cleared only after +100–125 bps, a 1.35x DSCR covenant and full amortization (facts #2). Nebius liabilities went $5.3B to $17.6B in three quarters (facts #11) CoreWeave ~855 bps, about a 50% risk-neutral five-year default probability (facts #2). HY data-center composite ~220 bps wide of high yield (facts #11)
5. Frontier labs OpenAI, Anthropic, xAI Equity roughly once a year. OpenAI's Mar 2026 round at $852B took $30B from Nvidia, $50B from Amazon, $30B from SoftBank (facts #10). xAI raised $20B at ~$230B on 6 Jan 2026 (v2 facts #17) Compute spending becomes revenue faster than the next round is needed Nvidia's CFO drew the line on the 27 Aug 2026 call. Hyperscalers are IG, frontier labs are not (facts #10) No traded credit. The only read is SoftBank at ~360 bps and its $10B margin loan on OpenAI shares at SOFR+425 (facts #3)
6. Private credit, insurers, BDCs Blackstone, Blue Owl, Apollo, PIMCO, BlackRock Insurance float, BDC borrowing, locked-up fund capital, retail money through brokers, bank fund lines Data-center collateral holds value across a full cycle $200–250B put into data centers and over $450B of AI-linked debt arranged since 2024 (facts #8). Meta's Hyperion vehicle alone is $27B of debt maturing 2049, 80% Blue Owl, PIMCO anchoring $18B. The same managers lent over $500B to software firms, ~19% of all direct loans, into AI disruption (v2 facts #3) No public spread. The price is the quarterly mark, and in Q2 2026 the clearing price to get out was about 62 cents on the dollar requested (v2 facts #2)
6b. Banks lending to the stack JPMorgan, Goldman, Morgan Stanley, regional fund-finance desks Deposits and their own IG paper The borrowing fund stays solvent and its investors keep meeting capital calls US banks carry ~$300B of private-credit exposure inside more than $1.2T of loans to non-bank financial firms. Fund finance passed $1T in 2026. The Fed began collecting bank-by-bank data only in April 2026 (facts #17) Not priced as a category. It shows up in individual bank spreads after the fact
7. IG bond market and index funds Index funds, pensions, foreign buyers Passive inflows into broad bond products This paper stays investment grade AI-linked IG bonds were $127B traded in July, 14.5% of US IG volume against a 4.5% 2025 average (facts #6) Hyperscaler cash-bond composite ~35 bps wide of the IG bond index, peaked near 50 in H1; this is a bond spread, not CDS (facts #11). 10-year Treasury 4.65% (facts #12)
8. Equity concentration The same names plus Nvidia Retail and passive flows Multiples hold while capex is heavy Buybacks cut to zero at most of the group (facts #9), so the bid under the shares is flows rather than the companies Alphabet raised $85B of equity (facts #6). CoreWeave stock roughly doubled off its lows while its CDS stayed near 855 bps (facts #2)
9. Sovereign and policy Fed, Treasury, Congress, Commerce n/a Policy stays permissive and the buildout stays popular The AI Kill Switch Act, the FRONTIER Act push and the Casar letter landed inside five weeks. Data centers were named a top target of the 2026 midterms (facts #14). OpenAI's CFO floated a federal backstop on 5 Nov 2025 and retracted it the next day (facts #15) Not priced anywhere
10. Physical supply TSMC, utilities, turbine and transformer makers Customer prepayments, utility rate base Chips and power arrive on schedule TSMC holds over 90% of advanced capacity. A 25% Section 232 tariff took effect 14 Jan 2026 with relief tied to building US capacity (facts #13). PJM's 2027/28 capacity auction cleared at $333.44/MW-day, ~1,053% above pre-surge, with data centers nearly all of a 5,250 MW projected peak-load increase (v2 facts #22) Relief is conditional on construction milestones, so the risk sits in the schedule rather than in a spread
11. State and local permitting State legislatures, county councils, governors, courts n/a Local politics stays quiet while household power bills rise Over 300 data-center bills in 30-plus states in 2026, at least 12 states considering moratoria, more than 100 local moratoria adopted, roughly $64B blocked or delayed (v2 facts #9). New York's legislature passed a one-year moratorium at 20 MW on 4 Jun 2026, unsigned as of this writing; the operative freeze is Hochul's EO 62 at 50 MW from 14 Jul. Oklahoma banned over 100 MW until 1 Nov 2029. PJM household bills projected up ~$70 a month by 2028 (v2 facts #22) Not priced anywhere. The only layer whose calendar is a legislative session

Three things connect these layers.

The maturity of the money against the maturity of the promise. At the core they match or run long. Meta's Hyperion vehicle is $27B of debt maturing 2049 against a building Meta leases, so the money is longer than the lease (facts #8). The four IG hyperscalers issue five- to thirty-year bonds against servers written off over five or six years, which is borrowing longer than the asset. Nothing at the core routes through a frontier lab's next round. The mismatch lives at the edges: a lab raises equity once a year and signs multi-year compute contracts, a neocloud takes three- to five-year GPU loans against contracts with those labs, and Oracle owes fifteen- to twenty-year lease commitments against the revenue of one customer that funds itself annually and that Nvidia's finance chief says is not investment grade (facts #7, #10). That is the only spot where the longest promise sits on the shortest money, and it is also the only investment-grade name whose CDS trades like junk. The market found the mismatch. It is narrower than the headline debt numbers suggest.

Money moves in a circle. Nvidia put $30B into OpenAI's March round (facts #10). OpenAI buys compute with money like that. The compute is bought from hyperscalers and neoclouds who use it to buy Nvidia chips. Nvidia also assembles outside capital for buyers who cannot borrow on their own name and makes promises about what the hardware will be worth later. Each step is a normal commercial arrangement. Stacked, a chunk of Nvidia's revenue is funded by Nvidia's own balance sheet and by capital Nvidia arranged, and a chunk of the collateral behind the loans is chips whose resale value Nvidia is guaranteeing.

Accounting. Hyperscalers extended assumed server useful life from three or four years to six between 2020 and 2024, lowering reported depreciation by roughly $18B a year across the group, and Amazon then shortened a subset back to five in early 2025 because AI hardware moves too fast (facts, historical list). If the real life is closer to three years than six, the deferred cost shows up later and the residual value backing GPU loans in layer 4 is worth less than the loan documents assume. About $1.65 trillion of obligations sits off the balance sheets of the five big spenders, roughly eight times the 2022 level, against about $1.35 trillion of on-balance-sheet debt (facts #8). Nikkei builds it from long-term lease commitments, take-or-pay GPU contracts and vehicle financing, with undiscounted future lease commitments alone at $969B (v2 facts #13). A twenty-year lease on a Meta building is underwritten on the tenant's credit and the life of a building, not on a server life or a GPU resale price. The two accounting assumptions sit in the smaller slice: neocloud GPU loans, vendor residual-value promises, and unguaranteed B notes like the Anthropic vehicle's (facts #16). That slice is where the 2007 analogy is real. The rest of the $1.65 trillion is where it gets manufactured by people who have not read the footnote.

Layer 11 and layer 6b change the shape of the problem. Every other layer is a financial constraint that money can solve. A county council vote is not. It is also the only layer that gets more hostile as the buildout gets more successful, because the mechanism running it is household electricity bills.

What history says

Ten build-outs over 180 years. One warning first. Seven of them come from a research brief that went looking for capacity financed ahead of demand that broke, so the sample is selected on the outcome. It shows what breakage looks like. It does not tell you how often breakage happens, and the frequency question is handled in the probabilities instead.

Telecom and fiber, 1999 to 2002

Carriers borrowed to lay fiber on the belief that internet traffic would keep doubling every few months, and equipment makers lent them the money to buy the equipment. Lucent committed about $8.1B of vendor financing, gave one customer, Winstar, a $2B facility, and took roughly $2.2B of bad-debt provisions in fiscal 2001 and $1.3B more in fiscal 2002 after Winstar went bust owing it over $800M (facts, historical list). WorldCom filed with $107B of assets and $41B of debt in July 2002, the largest US bankruptcy at the time. By 2004 only about a tenth of the installed fiber was lit, which is the weakest-sourced number here.

2000 2026 Why they match
Fiber capacity GPU compute capacity Capital-intensive capacity built ahead of demand on an extrapolated traffic curve
Lucent and Nortel vendor financing Nvidia credit support and residual-value promises (facts #10), Broadcom's $30B guarantee on the Anthropic vehicle's senior notes (facts #16) The vendor funds the customer, so vendor revenue and customer solvency become one risk. Broadcom's is a signed tranched guarantee; Nvidia's $500B platform is still a press release
WorldCom, Global Crossing CoreWeave, Nebius, Applied Digital Levered pure-plays with no other business to fall back on
Winstar OpenAI behind Oracle's backlog One customer's funding round supports another company's twenty-year lease
AT&T, Verizon Microsoft, Alphabet, Amazon, Meta Survive with damaged multiples because they have cash flow from something else
"Traffic doubles every 100 days" Compute demand curves in capex decks An unaudited demand claim everybody underwrote
Nasdaq telecom concentration S&P concentration in the AI complex The index becomes a single-sector bet without telling holders
Dark fiber Idle or resold GPU clusters Capacity that exists physically and earns nothing

Mortgages and the funding stop, 2006 to 2008

Long assets funded with much shorter borrowing. Asset-backed commercial paper peaked around $1.2 trillion in July 2007, structured investment vehicles held about $400B, and AIG's Financial Products unit had written $513B notional of credit derivatives by 30 Sep 2007, $78B on multi-sector CDOs. Two Bear Stearns funds that held this paper on leverage filed on 31 Jul 2007 with about $1.8B of investor losses, after Bear had put $1.6B into one of them in June to meet margin calls. The bond insurers lost their AAA ratings through 2008 and the GSEs went into conservatorship on 7 Sep 2008 with a Treasury pledge of up to $200B (facts, historical list).

2007 2026 Why they match
Mortgages Data-center leases and GPU contracts Long-dated contractual cash flows treated as safe collateral
ABCP and SIVs Private-credit funds and insurance general accounts Assets in vehicles that report periodically and are funded by capital that can leave
Bond insurers Nvidia's residual-value promises and Broadcom's guarantee on $30B of Anthropic vehicle notes (facts #10, #16) A guarantor standing behind paper that is only investment grade because of the guarantee
Bank off-balance-sheet conduits Meta's Hyperion vehicle and its peers Risk that legally sits outside the sponsor and practically does not
AAA on subprime tranches A+ on Hyperion debt A rating that rests on one assumption holding: for subprime, that defaults were uncorrelated; for Hyperion, that one tenant honors one lease for twenty years
Housing appraisals GPU residual values and six-year server lives The collateral value is an assumption, not a market price
Bear Stearns funds The first data-center private-credit fund to gate The first visible crack is a fund that cannot meet withdrawals
AIG Nvidia, only if the residual-value promises turn out to be concentrated AIG had $513B of written protection with collateral calls; Nvidia's $500B platform has no disclosed commitment, so this seat is hypothetical until disclosure says otherwise

The eight others, briefly

UK railways, 1844 to 1850. Parliament authorized 263 railway acts in 1846 covering roughly 9,500 miles. Investment peaked at about £44m in 1847, somewhere between 5 and 7% of British GDP, a larger share of the economy than anything else in this report. It was funded almost entirely by middle-class equity subscribed on a 10% deposit with the rest called later, and that structure is the whole story: when the Bank of England raised rates in October 1845 and the calls kept coming, subscribers sold everything else to meet them. About a third of authorized mileage was never built (R1).

US railroads, 1868 to 1893. Debt-funded, sold to American retail and to European buyers, alongside land grants and local aid. The forecast that failed was that transcontinental lines could be financed on traffic from territory not yet settled. Jay Cooke failed on 18 Sep 1873 unable to place Northern Pacific bonds and the NYSE closed for ten days. The 1893 repeat put about a quarter of US mileage into receivership. The only federal-adjacent intervention runs backwards: in February 1895 a Morgan and Belmont syndicate lent the Treasury $65m in gold (R1).

Electrification and Insull, 1919 to 1935. Samuel Insull financed the American grid by selling utility bonds and preferred shares to the public, including to his own customers and employees, over a pyramid of holding companies. He controlled a roughly $500m empire on about $27m of equity. Generation fell around 15 to 20% between 1929 and 1932, a modest decline that the leverage turned into an extinction event. Middle West Utilities could not roll its short-term debt, the receiverships came in April 1932, and a young Arthur Andersen audit found the structure retroactively insolvent (v2 facts #15). The transferable detail is that holding-company paper was a near-total loss while the operating companies' own bonds mostly recovered.

Nuclear and WPPSS, 1965 to 1983. The AEC projected 1,000 to 1,200 GW of US nuclear capacity by 2000 against 7% annual demand growth. Growth fell to 2 or 3% after 1973, over 100 ordered units were cancelled, and completed plants cost two to four times their estimates. WPPSS planned five plants, finished one, and defaulted on $2.25bn on 25 Jul 1983, the largest municipal default until Detroit's in 2013 (v2 facts #14). The financing detail matters because the same instrument runs through the AI stack: the bonds were backed by take-or-pay contracts with dozens of Northwest utilities, and in June 1983 the Washington Supreme Court voided those contracts. The guarantee that made the bonds investment grade turned out not to be enforceable, and they went from safe to defaulted without the demand forecast changing that year.

Satellite phones, 1998 to 2002. Iridium spent roughly $5bn, launched service on 1 Nov 1998, and filed Chapter 11 nine months later. The forecast that failed was millions of travellers paying about $3 a minute on $3,000 handsets. Terrestrial cellular arrived instead, which is the closest match to the second version of scenario A: the underlying demand was real and was met by a cheaper substitute while the capital-intensive version was still being built. The satellites worked and sold out of bankruptcy for $25m, under 1% of invested capital. Motorola took a $994M pretax Iridium charge in October 1999 and carried $1.8B of Iridium reserves at year-end 1999, per its 10-K (R1; Motorola filings).

US shale, 2010 to 2020. Roughly $300bn of cumulative negative free cash flow and $450bn of impaired capital from 2010 to 2019, funded by high-yield bonds that took energy to something like 15 to 20% of the US high-yield index by 2014. WTI went from $107.26 in June 2014 to $26.21 in February 2016 and to minus $37.63 on 20 Apr 2020. Roughly 110 producer bankruptcies in 2015 and 2016 carried about $74bn of debt. Capacity came back fastest of any case here, two to three years per bust, because a well is cheap to restart and the bust was a price event. Reserve-based bank loans mostly recovered in full while bondholders and equity took the loss (R1).

US merchant power, 1998 to 2005. After electricity deregulation, independent generators built gas turbines on project finance against load forecasts and offtake contracts, with General Electric as the turbine vendor and a lender to its customers. Enron's collapse in December 2001 shut the market for merchant paper, capacity gluts appeared in PJM and Texas, and the sponsors went: NRG filed in May 2003, Mirant in July 2003, Calpine in December 2005 with about $22bn of debt. The plants kept running under new owners within a few years. Nobody was rescued. This is the closest case to layers 4, 6 and 11 at once: a levered builder, a lender holding project paper, and a power market that turned political.

China property, 2015 to 2024. Property was about 29% of GDP including indirect activity and Evergrande's liabilities were ¥2.43tn at the end of 2022. Funded largely by homebuyer presales, which is vendor financing pointed the other way, plus bonds, bank loans and trust credit. The trigger was a policy choice, the three red lines of August 2020. Equity fell about 99% and offshore bonds trade at one to five cents. This is the only case where the capacity has not been absorbed as of 2026. Beijing eased rates, whitelisted projects and lent ¥300bn to buy unsold homes, all of it aimed at finishing buildings for people who prepaid rather than at making creditors whole (R1).

The ten in one table

Build-out Years Funded mostly by Forecast that failed Years to absorb Equity loss Debt loss Policy rescue
UK railways 1844–50 Retail equity on staged calls Every line earns pioneer-line dividends 10–20 −66% to −85% Minimal, little debt issued No. Bank Charter Act suspended Oct 1847 for the banking system
US railroads 1868–93 Bonds to US and European retail, land grants, local aid Traffic from territory not yet settled 5–7 per cycle R1 marks unsure ~a quarter of mileage in receivership in 1893, ~$1.8bn No. Morgan lent gold to the Treasury in 1895, not to the railroads
Electric utilities 1919–35 Public bonds and preferred over a thin equity pyramid Demand doubles every decade forever 10–15, finished by the war ~−100% at holding companies, 600,000 holders Holdco paper near-total, opco bonds mostly recovered No. Securities Act 1933, SEC 1934, PUHCA 1935 broke it up
Nuclear and WPPSS 1965–83 Utility bonds on the rate base, munis on take-or-pay contracts 7% annual demand growth, 1,000+ GW by 2000 10–15 for completed units, never for cancelled ones Heavy, via regulatory disallowances $2.25bn default, partial recovery after a 1988 settlement No
Satellite phones 1998–2002 Motorola equity and guarantees, vendor financing, HY bonds Millions of travellers at $3 a minute Mass market never came; a niche from 2001 ~−100% Assets sold at under 1% of invested capital No. Chapter 11, then a DoD procurement contract
Telecom and fiber 1999–2002 HY bonds, bank debt, vendor financing Traffic doubles every 100 days ~a decade, and the gains went to users Over $2T of market value, Nasdaq −78% 23 bankruptcies, WorldCom $41B of debt No. Sarbanes-Oxley was an accounting response
US mortgages 2006–08 Overnight and short-term wholesale funding of long assets House prices do not fall nationally Housing stock absorbed by ~2013 Bank equity destroyed or diluted, bond insurers wiped out Trillions, socialized in part Yes, and this is the exception. The funding ran through the payments system
US merchant power 1998–2005 Project finance, HY bonds, vendor (GE) financing Load growth and merchant power prices hold 2–5, plants ran under new owners ~−100% at NRG, Mirant, Calpine Calpine ~$22bn of debt restructured No
US shale 2010–20 HY bonds, reserve-based bank loans, repeated equity $80–100 oil defended by OPEC 2–3 per bust ~−70%, 2014–16 Energy HY default rate ~30% in 2016 Not in 2015–16. In 2020 broad facilities caught it incidentally
China property 2015–24 Presales, bonds, bank and trust credit 1.7bn sqm of annual sales indefinitely Not absorbed as of 2026 ~−99% at Evergrande Offshore bonds at 1–5 cents No for creditors. Support went to finishing prepaid units

What the set actually says

The base rate is two things that get run together. Capacity built ahead of demand is almost always used eventually. Eight of the ten absorbed the physical capacity, in windows of two to twenty years, and only satellite phones and Chinese property have not. The second half is harsher. In nearly all of them, the entities that built the capacity did not survive to earn from it; the exceptions are the ones with a second business (AT&T and Verizon in 2002), the secured lenders (shale's reserve-based banks recovered in full, Insull's operating-company bonds mostly recovered), and the 2008 banks, which were rescued. Railway shareholders lost two thirds to five sixths. Insull's holding companies went to zero and the wartime electricity boom was harvested by the reorganized operating companies. Iridium's satellites worked and sold for half a cent on the dollar. Shale's wells came back and the equity that drilled them did not. The asset survives and the owner does not, and that is the most reliable thing in this report. AI demand being real is entirely consistent with every current owner of AI capacity being wiped out. Those are separate questions and only the first one is being argued about in public.

Which of today's layers is which. Layer 1 is the incumbent telcos and the surviving operating utilities, with a second business that prints cash, so they eat a multiple compression and keep the assets. Layer 2, Oracle, is the transcontinental railroad, with long-dated commitments underwritten on traffic from territory not yet settled. Layer 3, Nvidia and Broadcom, is Motorola and Lucent, the vendor whose revenue and whose customer's solvency turn out to be the same number. Layer 4, the neoclouds, is the shale producer and the satellite operator, a levered pure-play funded in high yield with nothing to fall back on, and the historical outcome for that seat is close to uniform. Layer 5 has no clean analog, because the demand story and the borrower are the same entity. Layer 6, private credit and insurers and retail BDC buyers, is the most under-discussed match in the set: it is the 1920s public buying Insull's preferred and the Northwest municipalities signing WPPSS take-or-pay contracts, long-dated paper sold to people told it was safe because a guarantee sat behind it. And layer 11 is the nuclear case, where what turned an engineering programme into the largest municipal default in American history was not the technology and not the demand. It was a court, a cost curve and a regional power surplus arriving together.

The policy base rate is the cleanest number here. Of the ten, one produced a rescue that reached the builders, and that was 2008, where the funding ran through the payments system and the central bank already had tools pointed at it. The other nine got no rescue for the people who put up the money, and several got new regulation instead, which arrives after the losses. WPPSS bondholders litigated for years and settled in 1988. Iridium's creditors got under a cent. Evergrande's offshore bondholders are still waiting. Any argument that today's stack gets rescued has to explain why it looks more like 2008's plumbing than like the other nine, and the honest answer is that only layer 6b, the bank fund lines, has that character.

Private credit and the BDC scare of 2026

Roughly 140 BDCs exist, about 50 trade on an exchange, and industry assets went from about $30B in 2010 to $127B in 2020 to $451B in 2025 (v2 facts #4). The non-traded ones are the interesting half. Apollo, Blackstone, Blue Owl and others sell them to retail through wirehouses and broker-dealers, they raised about $35B in 2025, and the only exit is a quarterly tender. Two rule changes made the wrapper what it is. The 2018 Small Business Credit Availability Act let a BDC cut required asset coverage from 200% to 150%, doubling allowed leverage from one turn to two, and 14 had shareholder approval by the end of that year (v2 facts #16). The 2013 leveraged-lending guidance capped what banks would lend and left funds unconstrained (v2 facts #5). One rule let the wrapper borrow more and the other handed it the borrowers.

Date Event
Oct 2025 – Feb 2026 Software stocks fall ~30%. BDC shares fall ~10% and discounts to NAV widen (v2 facts #3)
Feb 2026 Blue Owl permanently halts redemptions at Blue Owl Capital Corp II, switches to returning capital, and sells $1.4B of assets from three credit funds to North American pensions and insurers. OWL shares fall (v2 facts #1)
Q1 2026 Blue Owl Technology Finance takes a $490M markdown, its largest ever, over 80% of it attributed to spread moves rather than credit events (v2 facts #1)
2 Apr 2026 Blue Owl caps redemptions at 5% across its private credit funds (v2 facts #1)
Q2 2026 $15.6B requested from non-traded BDCs. Ten of the sixteen Fitch tracks cannot meet requests in full. Average request 10.3% of shares against 9.7% in Q1, range 1.3% to 38.1%. New inflows down ~56%. Investors receive about 62 cents per dollar requested (v2 facts #2)
Q2 2026 Apollo, Blackstone, Ares and Blue Owl all cap. Goldman Sachs Private Credit Corp narrowly avoids a gate (v2 facts #2)
H1 2026 $12.7B pulled from non-traded BDCs. Blue Owl's first-half loss is $77.3M on software markdowns (v2 facts #16, #1)
Jul–Aug 2026 Blue Owl request rates ease from 40% to 38.1% and from 22% to 18.8% at two vehicles. Lower, not gone (v2 facts #2)

What it proved. The gates work. Every fund did what its offering documents said, nobody was defrauded, and there was no fire sale. It also proved that a wrapper sold to retail on the promise of bond-like income can leave most of the people who ask for their money short in one quarter without any borrower defaulting. Over 80% of the largest single markdown was spreads moving.

What it did not prove. This was not a data-center event. The losses came from software lending. Loans to software firms grew from about $8B in 2015 to over $500B by end-2025, about 19% of all direct loans, with a third of private credit funds in the sector (v2 facts #3). What is marking them down is AI eating the borrowers' revenue. So the first large AI-driven private credit loss of this cycle did not come from the buildout. It came from the same funds' loans to the companies AI is displacing, a channel almost nobody was modelling and one that runs whether or not the buildout works.

Data-center exposure sits in the same books. Private credit has put $200 to $250B into data centers and arranged over $450B of AI-linked debt since 2024 (facts #8). Blue Owl owns 80% of Meta's $27B Hyperion vehicle with PIMCO anchoring $18B of the bonds, and Blue Owl is also the manager that halted redemptions in February and capped at 5% in April. Nothing in the public record connects those two facts causally. It is a concentration observation, and it is why the 2026 scare matters even though it was about software: the manager with the largest single data-center position in the market spent this year demonstrating that its retail funding base can ask for money back faster than the funds can pay. Meta's next vehicle, a $12B deal in July 2026, priced above 7% (R5). The concession is widening while the wrapper is under stress.

The insurance leg is larger and quieter. Moody's put US life insurers' private and illiquid credit holdings at about $807B in June 2026, roughly a fifth of a $4T bond book; the larger $1.5 to $2T figures that circulate count all private placements, about half of insurers' bonds, which is a different and mostly investment-grade category. Private-equity-owned insurers hold about 20% of US life industry assets (R5; Moody's Jun 2026). Athene is roughly half of Apollo's assets under management and roughly half its pretax earnings, and the model is to originate the loan and hold it on the insurer (v2 facts #16). A general account has no quarterly tender. It has a fifty-year liability and an annual statutory filing, which is the slowest exit in this report and therefore where the largest loss can sit longest without anyone acting on it.

Where the VCs and LPs sit, and the chain

AI took about two thirds of US venture dollars in 2025, $222.1B against $73.0B in 2022, and roughly 80% of first-quarter 2026 dollars (v2 facts #17, #6). Fortune reported 87.5% for a later period whose measurement window is not stated clearly enough to use (period not confirmed). A venture portfolio raised in 2024 or 2025 is therefore not diversified in any meaningful sense. It is one bet with different tickers on it.

The checks are not coming from venture funds any more either. OpenAI's March 2026 round at $852B took $30B from Nvidia, $50B from Amazon and $30B from SoftBank (facts #10). xAI's $20B Series E at about $230B on 6 Jan 2026 took money from Nvidia, Cisco, Fidelity, QIA, MGX, Valor and StepStone (v2 facts #17). Corporates, sovereigns and crossover funds. The traditional limited partners hold the exposure through older funds at lower prices, and they are already sellers. Harvard explored a roughly $1B stake sale in April 2025. Yale mandated Evercore to sell up to $6B, about 15% of its endowment, and completed $1.8B through Lazard in February 2026. The top ten endowments sold $12.5B in the first quarter of 2026 (v2 facts #19). Venture stakes priced around 78 cents in the fourth quarter of 2025, with first-half 2026 discounts of 10 to 25% for 2020 and 2021 vintages and 20 to 35% for tail-end pools (v2 facts #7). Endowments over $1B average 9.8% in venture, pensions 2 to 5%, sovereign funds 5 to 10% across alternatives. Selling at 78 cents while marks are at record highs is what people do when they need cash, not when they are worried.

The daisy chain, in order, with what transmits at each step.

  1. A frontier lab takes a down round or misses a plan. Nothing happens to the lab that quarter. It has the money in the bank. The effect is dilution and a harder conversation next year.
  2. The secondary market prices it within weeks. Nasdaq Private Market marks put Anthropic above $1.1T in August 2026 (v2 facts #8). Those prints are public and they are what auditors point at.
  3. Crossover and late-stage funds mark down, because a recent third-party transaction is the strongest evidence a valuation committee can be shown.
  4. Venture fund NAV falls. Distributions were already slow, so investors have a lower mark and no cash.
  5. If public equities fell at the same time, the denominator effect bites and allocation policy forces a sale.
  6. The investor sells in the secondary at a discount, and that discount is a public data point that pushes marks down again for everyone holding the same names. This loop is what makes the process self-reinforcing rather than one-directional.
  7. New commitments slow. Funds counting on them to meet capital calls turn to fund finance. NAV finance deployed roughly $70B in 2025 on 17Capital's estimate (v2 facts #7).
  8. A NAV loan is sized off the marks. When the marks fall, loan-to-value breaches and the fund sells what has a bid, which is the good assets. That turns a mark problem into a portfolio-quality problem.
  9. Direct leverage on private shares reprices immediately. SoftBank has a $10B margin loan secured on OpenAI shares at SOFR plus 425 (facts #3). A mark cut is a collateral call, not a paper loss.
  10. SoftBank is funding its OpenAI commitments with a ¥1 trillion, roughly $6.3B, seven-year retail bond, the largest ever by a Japanese issuer, pricing 4 Sep 2026 at 4.3 to 4.9% (v2 facts #18). Japanese households are the last link and they hold a fixed-rate bond whose repayment depends on a private American company's equity value.
  11. Employees are the link people forget. Tenders are how they get paid. OpenAI completed a roughly $7B employee tender at $852B in August 2026, and Anthropic's April 2026 tender at $350B fell short of $6B of investor demand because employees would not sell (v2 facts #8). If the tender price stops rising, retention has to be paid in cash, and cash is what the labs are short of.
  12. Banks sit behind all of it, with about $300B of private-credit exposure inside more than $1.2T of loans to non-bank financial firms, and fund finance above $1T (facts #17).

The chain is slow at the top and fast at the bottom. Steps 1 through 6 run on quarterly reporting and take a year or more. Steps 9 and 10 run on daily collateral values. That inversion is the thing to watch, because the earliest and loudest signal in the whole chain is a margin call at a Japanese holding company, several links away from where the problem actually is.

Why the money was there in the first place

Rates near zero for most of a decade pushed anyone who needed income into riskier assets, and rules written to make banks safer moved the risky lending to people the rules did not cover. Both are measurable.

Pool Then Now Source
Global private credit AUM ~$158B (2010) ~$2T (mid-2024) v2 facts #5
BDC industry assets ~$30B (2010) $451B (2025) v2 facts #4
US venture AUM ~$0.2T (2008) ~$1.2T (2024) R5
Nonbank share of leveraged-loan commitments About half (2007) 80–85% (2024) R5
Insurer private and illiquid credit ~$0.4–0.5T (2015) ~$807B (Jun 2026), about a fifth of a ~$4T life-insurer bond book R5; Moody's Jun 2026
Pension allocation to alternatives ~7% (2003) ~26% (2023) R5
Bank exposure via lines to private-credit funds n/a ~$300B facts #17

The academic finding underneath that is specific and it is not a metaphor. Becker and Ivashina showed US insurers systematically buy the highest-yielding and riskiest bonds within each rating bucket, that the tilt is stronger when rates are low, and that it maps to regulatory capital buckets rather than fundamentals. Choi and Kronlund found the same in corporate bond funds, concentrated in low-rate periods, generating apparent outperformance that reverses under stress. Di Maggio and Kacperczyk found money-market funds doing it at the zero lower bound (R5). That literature explains exactly who bought $18B of A+ rated bonds issued by a vehicle that owns a building leased to Meta. It is the highest yield available inside the rating bucket the buyer is allowed to hold, which is the trade those papers describe.

On the rules. The March 2013 interagency leveraged-lending guidance effectively capped bank appetite at around six times debt to earnings and required senior debt to be repayable in five to seven years. Banks were bound and funds were not, and private credit growth accelerated from there (v2 facts #5). On 5 Dec 2025 the OCC and the FDIC rescinded it outright, effective immediately, replacing it with eight supervisory principles and no bright-line test, on the stated grounds that it had pushed the lending outside the regulatory perimeter (R5). The Federal Reserve had not joined the rescission as of mid-December 2025. The timing is the part to sit with. The rule that created private credit was repealed in the same quarter private credit's retail funding base started asking for its money back.

Off-balance-sheet needs one correction that most commentary misses. ASC 842 already fixed the old version of this. Since 2019 a lease that has commenced sits on the balance sheet. The $1.65 trillion is three things that survive the rule: leases signed but not yet started, where recognition waits until the tenant controls the building; take-or-pay purchase obligations for GPUs, power and capacity, disclosed in the commitments footnote and becoming liabilities only as performance occurs; and debt inside vehicles the sponsor does not control enough to consolidate (R5, v2 facts #13). Undiscounted future lease commitments alone are $969B, and Meta's share of the total is about $420B, roughly three times its reported debt (R5). None of it is hidden. All of it is in the notes. It is simply not in the ratio anyone quotes.

How much of each pool exists because of rates and rules. These are estimates and the source labels them as such. Under 2005-style rates and pre-2013 rules, private credit would plausibly be 40 to 55% of its current size, roughly $0.7 to $0.9T against $1.7T, with about half the gap attributable to reach-for-yield demand from insurers and pensions, about a third to banks retreating under the 2013 guidance and Basel III, and the rest to borrowers preferring the speed. Venture AUM would plausibly be 50 to 65% of actual, and regulation barely matters there because the driver is what a zero discount rate does to long-dated equity. BDC assets would be about 40% of today, because the wrapper predates ZIRP and its scale does not. The hyperscaler off-balance-sheet stack is different in kind: it is an AI capex phenomenon from 2023 onward and perhaps 60 to 75% of it would exist under any rate regime, but the buyer base for A+ project debt is a reach-for-yield inheritance, so without it the financing would be smaller and more expensive (R5, all estimates).

The short answer. Roughly half of private credit, roughly half of the BDC wrapper and roughly half of venture are artifacts of rates and rules rather than of demand for the underlying assets. The buildout itself is not. What rates and rules supplied was not the idea. It was the buyer.

Scenarios with probabilities

Four paths, and the numbers sum to one hundred. They were set two ways and then checked against each other. The first way is the base rate from the history section: capacity built ahead of demand is almost always used eventually, and the companies that built it usually do not survive to use it. The second way is what the market is already charging: about 50 bps a year of CDS premium to insure the four investment-grade names, four times the index for Oracle, and 855 bps for CoreWeave (facts #1, #2, #4). Both point the same direction. The likeliest outcome is that the top clears and the bottom does not.

Scenario Probability What it is Runs on
A. Demand and price disappointment 45% The computing gets used but the price of it falls faster than the borrowed money can be repaid Earnings calls, quarterly
B. Funding stop 22% Lenders and fund investors stop before demand falls Redemption windows, quarterly
C. Policy and power stall 23% Permits, power and politics slow the build without a credit event Legislative sessions and rate cases, yearly
D. It clears 10% Revenue arrives before the money runs out, at every layer Through 2028

Scenario A, 45%

Trigger. A frontier lab misses a revenue plan or takes a down round, and a hyperscaler answers by cutting capex guidance on an earnings call. Or the quieter version: smaller models keep crossing the good-enough line, the price of a token falls faster than the hardware depreciates, unit demand keeps rising, and revenue per GPU falls anyway. That second version is exactly what happened to fiber. The internet delivered everything it promised and the owners of the bandwidth went bankrupt (facts, historical).

Sequence. Quarter 1: a guidance cut, or a signed lease that becomes an option. Neocloud equity falls, neocloud CDS spreads widen, the four IG names barely move. Quarters 2 and 3: GPU rental prices fall as capacity built for one buyer is resold into the spot market; CoreWeave-style covenants like the 1.35x DSCR test (facts #2) start to bind; new neocloud issuance stops. Quarters 3 to 5: the first neocloud restructuring, and lenders discover that GPU collateral in a falling rental market is worth much less than the appraisal, and that Nvidia's residual-value support covers some contracts and not others. Nvidia takes provisions the way Lucent did. Quarters 4 to 8: Oracle refinances at a worse level or sells assets; multiples compress across the complex; the hyperscalers protect free cash flow and restart buybacks a year late. Quarters 6 to 12: the write-down phase, slow, because a data center can be repurposed and a chip can be resold at some price.

Layer Outcome
IG hyperscalers Survive. Multiples compress hard, CDS spreads widen 30 to 60 bps, no default risk
Oracle Damaged. A downgrade from BBB- into high yield is the live question
Vendor-financiers Survive as companies, break as growth stories. Provisions hit earnings
Neoclouds Break. This is where the bankruptcies are
Frontier labs Consolidate into the two or three best-funded
Private credit, insurers, BDCs Wounded. Losses land in equity tranches and 2025 to 2026 vintage funds
Banks Hit one step removed, through fund lines; whether the lines get pulled decides if A turns into B
IG bond market Widens, does not seize
Equity concentration Breaks. The largest dollar loss in the scenario
Policy Mostly does nothing, because equity losses do not force a response
Physical supply Excess capacity; TSMC cuts advanced-node utilization; interconnection queues shorten
State and local permitting Moratoria become moot as projects cancel themselves

Who eats it. Neocloud equity holders lose everything. Neocloud bondholders recover something on hard assets. Nvidia shareholders eat provisions and a multiple reset. Private-credit LPs, mostly pensions and insurers, take marks in specific vintages. Index-fund holders take concentration losses on a bet they never chose.

Signals. A hyperscaler cuts capex guidance or converts a signed lease into an option. Spot GPU rental prices fall two quarters running while announced capacity rises. Nvidia discloses a provision against a customer receivable or a residual-value guarantee. An announced partner in the $500B platform declines to fund, or the platform still has no named first project a year after announcement (facts #10). Neocloud high-yield issuance stops for a full quarter. The HY data-center composite widens through 400 bps over high yield, from about 220 (facts #11).

Why 45. It is the modal historical outcome, the majority of the ten cases in the history section once you sort them by mechanism, and it is the residual: the base rate for the capacity being used is high, the base rate for the builders surviving is low, and the difference between those two is this scenario. It is also what the market is pricing when it leaves the top four near 50 bps and puts the fringe at junk levels. It is also the outcome that needs nothing undisclosed to be true. What moves it up ten points: two quarters of falling rental rates, or one capex cut. What moves it down ten: a neocloud refinancing at a tighter spread than its last deal, which would be the opposite of the August CoreWeave experience.

Scenario B, 22%

Trigger. A funding event, not a demand event. A private-credit fund with data-center exposure gates, or a data-center securitization fails to place at any price. The 2007 question follows: if this one cannot be sold, what is everything else worth?

Sequence. Quarter 1: a fund gates; peers are asked to disclose data-center exposure and cannot answer quickly, because the assets are not traded and the marks are quarterly. Quarter 2: the marks move. An insurer writes down a data-center loan and every holder of a similar loan now has a comparable it did not want; state insurance regulators start asking about concentration inside general accounts. Quarters 2 and 3: new financing stops for anyone below investment grade. Quarters 3 and 4: three- to five-year construction loans come up against fifteen- to twenty-year leases, and if the lender will not roll, the building stops mid-construction. Quarters 4 to 6: Oracle is the pressure point in public markets, where a downgrade forces index-driven selling into a market that has stopped buying (facts #7). Quarters 5 to 8: contagion by label. Anything called AI-linked gets sold, including the paper that is fine, because 14.5% of July IG volume was AI-linked and that is a lot of sellers reaching for the same exit (facts #6).

Layer Outcome
IG hyperscalers Survive on cash; their SPVs and leases get pulled back onto the balance sheet in every analyst's model
Oracle Breaks first in public markets
Vendor-financiers The critical unknown. If the backstops are concentrated, Nvidia is the AIG seat
Neoclouds Break, faster than in A, because funding stops before revenue falls
Frontier labs Break hardest. Annual equity against multi-year commitments is the shortest liability in the chain
Private credit, insurers, BDCs The epicenter. Gates, queues, forced sales into no bid
Banks The junction. Pulled fund lines are what turns a queue into a fire sale
IG bond market Seizes for weeks. Index funds are forced sellers on downgrade
Equity concentration Breaks, and moves with credit rather than against it
Policy Forced to act, inside the limits described below
Physical supply Half-built data centers, stranded interconnection agreements, litigation with utilities
State and local permitting Irrelevant; the money stopped before the permits mattered

Who eats it. Policyholders and pensioners, slowly and invisibly, through general accounts. Private-credit LPs get gated. Non-traded BDC shareholders take mark-to-market losses on something they were sold as a yield product, and the 2026 episode already showed what that looks like at 62 cents per dollar requested (v2 facts #2). Oracle bondholders take a downgrade. Banks are exposed through roughly $300B of private-credit lending inside more than $1.2T of loans to non-bank financials (facts #17).

Signals. Any data-center or private-credit fund gates or extends its queue. An insurer discloses a data-center write-down or a state regulator asks about concentration. A bank cuts, reprices or refuses to renew a subscription or NAV line to a fund with heavy data-center exposure. A data-center securitization is pulled. Oracle CDS through 300 bps or the 2054 bonds above 8.5% (facts #7). Nvidia is required to disclose the size and concentration of its residual-value guarantees. The hyperscaler IG composite past 75 bps over the index, from about 35 (facts #11).

Why 22. The 2026 episode cuts both ways and the number sits between them. It showed that the retail exit from these funds is real, that marks move on public comparables before any borrower defaults, and that the manager holding the largest data-center position in the market was the one gating, even though the trigger was software loans and not data centers (v2 facts #1, #2). It also showed a queue rather than a fire sale, the same thing BREIT showed in 2022, because the capital is locked and pays out quarterly (R6). So the mechanism exists and it is slow. What would make it fast is the bank fund-line channel, and nobody outside can see it. Up ten points: a fund with data-center exposure gates, or a bank reprices a fund line. Down ten: Q3 tender requests fall back under 5% across the big non-traded BDCs, or a data-center securitization prices tight.

Scenario C, 23%

Trigger. Not a market event at all. A second state follows New York with a statewide moratorium, or a third PJM capacity auction clears at its cap, or a court upholds a county's right to say no against the federal preemption order. The build slows for reasons no spread can see coming.

Sequence. The first year is already happening: more than three hundred data-center bills in thirty-plus states, over a hundred local moratoria, about $64B of projects blocked or delayed, New York's one-year freeze on permits for anything at 20 MW and above, Oklahoma's freeze on anything over 100 MW until November 2029, Indianapolis barring new construction through 2027 (v2 facts #9). Year 2: capacity prices in PJM, already up about 1,053% for the 2027/28 delivery year with data centers accounting for nearly all of the projected load growth, show up in household bills at about $70 a month, and the midterm politics of that are not subtle (v2 facts #22). Year 2 to 3: the federal preemption order of December 2025 is litigated state by state; some sites win, some do not; interconnection queues stretch, and a signed lease on a site that cannot get power is a commitment with no building under it.

Layer Outcome
IG hyperscalers Slow the build and keep the cash. The scenario is good for their free cash flow
Oracle Depends entirely on which of its sites are affected. Lease commitments on stalled sites become an argument with the landlord
Vendor-financiers Order timing slips; the growth story stretches rather than breaks
Neoclouds The ones with unbuilt, financed capacity break; the ones already running benefit from scarcer supply
Frontier labs Compute stays scarce and expensive, which is bad for them and good for whoever owns running capacity
Private credit, insurers, BDCs Construction loans on stalled sites are the exposure; operating assets hold value
Banks Exposure through construction lending, small relative to B
IG bond market Barely notices
Equity concentration Modest; the market reads a slower build as capital discipline
Policy Is the scenario
Physical supply Utilities and generators are the winners; TSMC loses order timing
State and local permitting The binding constraint

Who eats it. Developers holding land, permits in process, and three- to five-year construction loans on sites that cannot be finished. Their lenders. Tenants who signed leases that never commence, who argue about whether they owe anything. Ratepayers, who pay for the generation that got built for load that did not arrive.

Signals. A second statewide moratorium. PJM clearing at the cap again. A court ruling against federal preemption of a local data-center ban. A hyperscaler disclosing a site cancellation on permitting or power grounds rather than demand grounds. FERC or DOE fast-tracking interconnection, which cuts the other way.

Why 23. This path runs on a different clock and has already started. The evidence is not a forecast; it is a count of ordinances, and PJM's 2027/28 and 2028/29 auctions both cleared at the cap. It is held below A because the federal executive is pushing hard the other way, with the permitting order, the four federal sites, and the preemption order (v2 facts #20, #21), and because a stall at the top is partly self-correcting: the hyperscalers keep the cash and the market rewards them for it. Up ten: the second statewide moratorium, or preemption losing in court. Down ten: preemption holds, or a federal interconnection fast-track that actually clears queues.

A and C blur at the edges. A power-driven stall and a demand-driven price fall look identical to a neocloud lender, because in both cases the borrower cannot service the loan and the collateral has no bid.

Scenario D, 10%

For it to clear, four things have to be true at once. Frontier-lab revenue grows into the compute already contracted, so the annual equity round becomes optional. The hyperscalers reach the 2028 free-cash-flow inflection that Bloomberg Intelligence models against roughly $6T of cumulative spending (facts #9). Oracle completes its financing plan and holds BBB- (facts #7). And rates come down, because at a 10-year of 4.65%, the 98th percentile of its twelve-month range, every refinancing in the stack reprices higher (facts #12).

The numbers that prove it: free cash flow positive at three of the five names by end-2027; Oracle debt peaking under about $130B with the CDS back under 150 bps; GPU rental prices flat or rising for two quarters; one neocloud refinancing tighter than its last deal.

Why 10. The top of the stack clearing is close to certain, and that is not this scenario; that is A. This scenario needs all four legs at once, and the rate leg is already failing as this is written, the free-cash-flow leg runs a year ahead of the model it cites, and the neocloud leg requires a company priced at a 50% risk-neutral default probability to refinance clean. Conjunctions with a visibly failing conjunct are not twenty-percent events. Ten is what is left after the residual goes to A. Up ten: the 10-year under 4% and Oracle under 150 bps. Down ten: any one leg visibly breaking.

What Washington does

The frame for every branch is the same. After 2008 the law was rewritten so that the government can rescue a market but not a company. Dodd-Frank Section 1101 says a Fed emergency facility must be broad-based and have the Treasury Secretary's approval, and the Fed's 2015 rule implementing it defines broad-based as an identifiable market with at least five eligible participants and bars lending to insolvent borrowers (v2 facts #23). That single rule shapes what follows.

The Fed. In scenario A it does what it did in 2001: cuts rates and otherwise stays out, because a stock-market loss on identifiable balance sheets is not a liquidity crisis. In scenario B it has three templates and one hard limit. The template for investment-grade bonds is the 2020 corporate credit facility, announced at $250B of capacity with Treasury equity from the CARES Act behind it, which bought bonds rated at least BBB- and broad IG ETFs and in the end held only about $14B because the announcement did most of the work (v2 facts #10, #23); pointed at 2026 it would buy hyperscaler paper and, until the downgrade, Oracle. The template for data-center securitizations is TALF, which in 2020 lent only against AAA asset-backed paper; Hyperion's debt is A+, so reaching it would need the facility widened, and the unrated B notes and the private loans are out of reach on any version. The template for the bank junction is the 2023 Bank Term Funding Program, lending to banks at par against collateral, which would stop banks pulling fund lines to protect their own liquidity. The hard limit is the one that matters most: the Fed cannot buy private-credit fund units, cannot lend to an insolvent neocloud, and cannot rescue Oracle by name. The epicenter of scenario B sits exactly where the Fed is not allowed to go. In scenario C it does nothing, because there is nothing for a central bank to lend against.

Treasury. It has less room than people assume. The Exchange Stabilization Fund guaranteed money-market funds for up to $50B in September 2008 (v2 facts #23), and Congress then barred that use, so there is no pool the Secretary can point at a non-traded BDC redemption queue without a vote. The loan-guarantee authorities that exist sit in other departments: the Energy Department's loan office, whose statute reaches energy projects rather than compute; the Commerce Department's CHIPS money; and Title III of the Defense Production Act at the Pentagon. The July 2025 permitting order already directs loan guarantees and tax incentives toward AI data centers over 100 MW (v2 facts #21). So the strange truth is that the executive is already guaranteeing this build, on the way up, through the energy and industrial-policy door, and none of that machinery is built to catch it on the way down. Anything bigger, a TARP for AI, needs Congress, and the political ceiling on that was set in November 2025 when OpenAI's finance chief said the word "backstop" and retracted it inside a day (facts #15).

Two supervisors outside Washington matter more than the Fed in scenario B. The SEC sets the BDC leverage and tender rules, and in April 2020 it issued temporary relief letting BDCs exceed their asset-coverage limits and co-invest with affiliates; that is the fastest and cheapest lever in B, and it needs neither Congress nor Section 13(3). State insurance commissioners and the NAIC set the capital charges and the permitted-practice marks on general accounts, which is where the largest silent loss sits; a regulator that lets an insurer carry a data-center loan at cost for two extra years is doing more for scenario B than any facility the Fed can legally build.

The Executive. This is the branch with the most reach in scenario C and the least in B. It has the permitting order, four federal sites, the December 2025 order directing agencies to challenge state AI laws, tariff relief tied to building US capacity, and export controls it can loosen or tighten (v2 facts #20, #21; facts #13). What it cannot do is override a county council or a state legislature by decree; the preemption order is a litigation strategy, not a switch. In B it can jawbone, convene banks the way Treasury did for First Republic in March 2023 (R6), and lean on regulators to go slow on marks, which is what happened with commercial real estate in 2023. It cannot make a lender roll a construction loan.

Congress. In A it holds hearings and passes nothing, which is the 2002 precedent: Sarbanes-Oxley, no telecom rescue, and WorldCom's creditors on their own (facts, historical). In B a bailout vote is possible only if the payments system is threatened, and the current mood is the opposite: the AI Kill Switch Act arrived within weeks of the Hugging Face incident, the Senate struck the state-law moratorium 99 to 1, and data centers are a named midterm target (v2 facts #20; facts #14). Before November 2026 a vote to rescue AI lenders is toxic; after it, possible, and only with strings. In C, Congress is where the fight actually is: a federal preemption statute for data-center siting, which the 99 to 1 vote says will not pass as written, or permitting and interconnection reform, which has a bipartisan constituency because utilities want it too.

The most likely package. A: rate cuts and hearings, no money. B: an IG bond facility and a bank funding facility from the Fed with Treasury equity behind them, SEC leverage relief for BDCs, state regulators going slow on insurer marks, and a Congress that does nothing until after the midterms. C: an executive branch litigating preemption site by site, a permitting bill that moves slowly, and ratepayer politics that move fast.

The feedback loop

A fair question is whether social media makes this worse than the CNBC era of 2008. The evidence says yes, in a specific way, and no in another.

Yes for anything that can be withdrawn with a tap. Silicon Valley Bank went from a capital-raise announcement on the afternoon of March 8 2023 to seizure on the morning of March 10: $42B withdrawn on March 9, about a quarter of deposits, with roughly $100B more queued for the next day, and the regulator took the bank mid-session because waiting for Friday close was no longer safe (v2 facts #11; R6). Cookson and co-authors measured it: banks with high pre-existing Twitter exposure lost about 4.3 percentage points more than similar banks during the run, and hour-by-hour spikes in Twitter attention preceded the next hour's losses. Bear Stearns, by comparison, took four days to drain, from about $18B of liquidity on Monday March 10 2008 to about $2B by Thursday, with Jim Cramer telling a viewer on the Tuesday that "Bear Stearns is not in trouble" (R6). The medium changed the speed by an order of magnitude.

Now the part that is new in 2026 and did not exist in 2023. Odds are quoted. Kalshi and Polymarket carry tradeable markets on Fed emergency meetings and on credit events, so a rumor arrives with a price attached and a screenshot of the price is the rumor. CDS levels, which in 2008 were a Bloomberg screen a few thousand people could see, circulate as images: "CoreWeave CDS hits 50% default odds" was a headline on July 30 2026 and Nebius and CoreWeave stock fell 9 to 10% on CDS-cost stories the day before (R6). And the market never closes. Crypto trades all night, tokenized stocks trade on weekends, and the exchanges are building 23-hour trading, so there is now an overnight price for the panic that Sunday-night policy meetings in 2008 never had to face. Anyone whose portfolio leans crypto should notice that it is the instrument that prints that overnight price.

No, for the part of the stack that cannot be withdrawn. Private-credit capital is locked for five to ten years. Non-traded BDCs and interval funds pay out on a schedule, typically capped at 5% a quarter, and have no daily price. When BREIT was hit with withdrawal requests in late 2022 it gated for months and there was no fire sale (R6). Deutsche Bank's CDS spiked and its shares fell 14% on no news in March 2023, and nothing happened, because nothing could be withdrawn (R6). Noise needs a runnable liability to become a run.

So the loop does not change what breaks. It changes where the slow loss becomes a fast one. The runnable liabilities in this stack are neocloud equity, Oracle bonds through forced index selling on a downgrade, and, above all, the bank fund lines to private-credit vehicles, which a bank can pull on a phone call after reading a screenshot. That junction was the swing factor in scenario B before social media entered the picture. Social media makes it faster. The locked-up pools still take quarters. What is worse than 2008 is the speed at which the middle of the stack can be pushed into the epicenter; what is not worse is the epicenter itself, which was built to be slow.

Bans and slowdowns

A moratorium is a temporary ban on approving new data centers. The count as of this writing: more than three hundred data-center bills in over thirty states in 2026, at least twelve states that considered statewide moratoria, more than a hundred local ones adopted, and about $64B of projects blocked or delayed. New York's legislature passed a one-year freeze on state permits for anything at 20 MW and above on June 4 2026, exempting projects already under construction; the governor's July 14 executive order set the line at 50 MW with carve-outs for manufacturing, research, education and medical uses. Oklahoma froze anything over 100 MW until November 2029. Indianapolis voted 23 to 1 to bar new construction through the end of 2027. Maine's legislature passed one and the governor vetoed it. Hill County, Texas, passed a one-year ban in May and rescinded it on June 4 after a developer sued (v2 facts #9). The pressure underneath is the power bill: PJM's capacity price for 2027/28 is about eleven times the pre-surge level and the market monitor put data centers at the center of it (v2 facts #22).

What a stall does to a project depends on where the project is. A site with a permit in process and a construction loan is the worst place to be. The loan is three to five years and it amortizes; the building does not earn until it is finished; the developer is paying interest on a hole in the ground. A tenant who signed a fifteen-year lease that has not commenced is in a strange spot: the lease is an off-balance-sheet commitment, disclosed in a footnote rather than counted as debt (R5), and whether the tenant owes anything on a building that cannot be built is a question for the lease's force-majeure clause and a courtroom. The lender to the developer is the one who actually loses, and that lender is, increasingly, a private-credit fund or an insurer rather than a bank, which is how a zoning vote in Marion County ends up as a mark in a general account.

That is the link to the other scenarios. A stall that leaves construction loans unable to roll is a scenario B trigger by another name. A stall that raises the cost and delay of every new site raises the price of running capacity, which helps whoever already owns it and hurts whoever financed capacity they have not built. The neoclouds with the largest backlogs and the least built capacity are on the wrong side of that.

Positioning

A map, not advice, for anyone whose portfolio already leans crypto and tech and therefore already owns the high-beta version of this trade.

Hold the four IG hyperscalers through all four scenarios. They survive A and B as credits, keep their cash in C, and win D outright. Their equity is the biggest dollar loss in A, because index concentration compresses the multiple even though the business does not, and holding them is a bet that the cash flow outlasts the multiple, which it did for the incumbent telcos in 2002. Do not trim the top to express a view about the bottom.

Crypto is not a hedge for any of the four. It trades as the highest-beta leg of the same liquidity that funds the build, so it falls with equity in A, with everything in B, and it is the instrument that prints the overnight panic price. Anyone who holds it already holds the levered version of D.

Avoid neocloud equity and neocloud credit. CoreWeave 5-year CDS at about 855 bps is priced as a binary on refinancing. The "50% default probability" is the risk-neutral number, always higher than the real-world odds, but the credit market has the covenants and the equity market, which doubled the stock off its lows in the same weeks, does not (facts #2).

The traded hedge is Oracle 5-year CDS protection at about 215 bps, or the 2054 bonds at 7.8% (facts #1, #7). It carries both exposures that A and B hit, at a rating where a downgrade forces selling. Say the cost plainly: about 430 bps over a two-year hold, more than the 30 to 60 bps of widening it protects the top four against, in the most crowded name in the complex (facts #5). It is a bet on a downgrade, priced like one.

The hedge for B is refusal. Do not buy non-traded BDCs with data-center concentration, do not commit to 2025 or 2026 vintage private-credit vehicles in this sector, and treat any insurance product with an AI-lending sleeve as an unpriced position. The 2026 episode is the proof that the exit is 5% a quarter at best. The traded proxy is Nvidia CDS at about 82 bps, which has to widen if the residual-value backstops turn out to be concentrated, and which already trades like a BBB credit on an AA rating (facts #4).

C has a different shape. The winners are the owners of running capacity and the owners of generation and transmission in constrained markets, which is where the PJM price goes. The losers are developers and lenders on unbuilt, financed sites. There is no clean instrument for this; the practical version is to prefer names with built capacity over names with backlog.

Watch four spreads and three non-spreads. Oracle 5-year CDS against the IG index, about 4x. The hyperscaler IG composite against the index, about 35 bps wide. The HY data-center composite against high yield, about 220 bps wide. CoreWeave outright. None of the four tells you which scenario you are in. The discriminators are outside the CDS market: GPU rental rates and capex guidance for A, a fund gate or a bank line for B, and a county council for C. Nothing here is investment advice; it is one person's map of a system everyone is inside of.

Facts used

  1. Oracle 5-year CDS ~215 bps (Jul 27–28 2026), an 18-year high; ~4x the IG CDX at ~53; rated BBB-. Path: 40 bps Sep 2025 to 215 Jul 2026. [IFR/S&P; Bloomberg; ASIC via ABC]
  2. CoreWeave 5-year CDS ~855 bps (Jul 28 2026), ~50% risk-neutral 5-year default probability; Aug 3 $2.6B loan flexed +100–125 bps with a 1.35x DSCR covenant and full amortization; Q1 interest expense $536M, FCF −$4.71B; Q2 revenue +200% y/y, backlog ~$104–129B. [press-reported]
  3. SoftBank CDS ~360 bps; $10B margin loan against OpenAI shares at SOFR+425; ¥1tn (~$6.3B) 7-year retail bond pricing Sep 4 2026 at 4.3–4.9%, the largest Japanese retail bond, to fund OpenAI commitments. [Bloomberg Aug 24 2026; Nikkei]
  4. Meta ~90–93 bps, Alphabet 67, Amazon ~65–70, Nvidia 82 (Jul 27–29 2026 records); hyperscalers ex-Oracle average ~49 bps, highest since at least 2018; Nvidia trades "like BBB" on an AA rating. [CNBC; Bloomberg Intelligence Aug 27]
  5. Tech single-name CDS ~$650M average daily notional Q2 2026, +~600% y/y; net notional on major tech $12.5B, Oracle $6.5B. [DTCC via Reuters Jul 29; Bloomberg/LSEG]
  6. Big Tech IG issuance $182–200B YTD 2026 vs ~$100–121B in all of 2025; AI-linked IG bonds 14.5% of July US IG volume vs 4.5% in 2025. [IFR]
  7. Oracle 2031s ~6.23%, 2054s 7.8% (Jul 29); $45–50B financing plan and $25B bond (Feb 2026); ~$100B debt plus ~$250B of 15–20-year lease commitments; Barclays: cash out by Nov 2026 absent financing. [Barclays Nov 2025; press]
  8. Off-balance-sheet obligations across the five ≈ $1.65T, ~8x 2022, vs ~$1.35T on-balance-sheet debt; undiscounted lease commitments alone $969B; built from uncommenced leases, take-or-pay chip contracts and SPV/JV debt. Meta Hyperion: $27B debt to 2049, Blue Owl 80%, PIMCO anchor $18B, rated A+. Private credit in data centers $200–250B. [Nikkei Jul 2026; CNBC Oct 21 2025]
  9. Five-hyperscaler 2026 capex ≈ $690–770B vs ~$125–130B in 2021; Bloomberg Intelligence ~$6T cumulative 2025–2030, FCF inflection 2028; only Microsoft FCF-positive now. [BI; company filings]
  10. Nvidia CFO Kress, Aug 27 2026: "hyperscalers are IG, frontier labs are not IG"; "over $500B" third-party financing platform announced Aug 10 2026, no partner commitment disclosed, no first project named; $30B into OpenAI's Mar 2026 round at $852B post. [Nvidia; CNBC; Bloomberg]
  11. Hyperscaler IG composite ~35 bps wide of the IG index; HY data-center composite ~220 bps wide of HY; Nebius liabilities $5.3B to $17.6B in three quarters. [Penn Mutual Aug 6 2026]
  12. 10-year Treasury 4.65% (Jul 29 2026), 98th percentile of its 12-month range.
  13. Section 232: 25% tariff on advanced logic chips from Jan 14 2026 with relief tied to US construction; TSMC >90% of advanced capacity. [USTR; Commerce]
  14. OpenAI/Hugging Face incident (Jul 2026); AI Kill Switch Act (Lieu/Moran, Jul 23 2026) lets DHS order a slowdown or shutdown; data centers a named 2026 midterm target. [CNBC Jul 23; Axios Aug 23]
  15. OpenAI CFO Friar, Nov 5 2025, floated a federal "backstop" for chip debt and retracted it Nov 6. [CNBC]
  16. Broadcom–Anthropic SPV (Jun 10 2026): $35B, Apollo and Blackstone; $30B of notes backstopped by Broadcom, $4.5B B notes unguaranteed. [Broadcom IR; Axios]
  17. US banks' private-credit exposure ~$300B inside >$1.2T of loans to non-bank financials; fund-finance market >$1T; Fed began bank-by-bank collection Apr 2026. [Moody's; OFR; FSB]
  18. Blue Owl: Feb 2026 halted quarterly redemptions at OBDC II, moved to return-of-capital, sold $1.4B of assets; Apr 2 capped redemptions at 5%; OTF took a $490M Q1 markdown, largest since inception. [Reuters; CNBC; S&P Global]
  19. Fitch, Jun 1 2026 and Q2 data: $15.6B requested from non-traded BDCs; 10 of 16 tracked could not meet requests in full; average request 10.3%, range 1.3% to 38.1% (OTIC); inflows down ~56%; ~62 cents per dollar requested paid. [Fitch; trade press]
  20. SaaS loans ~$8B (2015) to >$500B, ~19% of direct loans (end-2025); a third of private-credit funds lend to SaaS. [BIS Quarterly Review Mar 2026]
  21. BDC industry assets ~$30B (2010) to $451B (2025), ~140 BDCs; SBCAA (2018) cut required asset coverage from 200% to 150%. [Mayer Brown; Houlihan Lokey; Proskauer]
  22. Private credit ~$158B (2010) to ~$2T (mid-2024); 2013 leveraged-lending guidance with its ~6x line rescinded by the OCC and FDIC on Dec 5 2025. [PitchBook; BIS; Mayer Brown]
  23. AI took about two-thirds of US VC dollars in 2025, $222.1B vs $73B in 2022; xAI $20B Series E at ~$230B (Jan 6 2026). [PitchBook; press]
  24. Venture secondaries ~78c on the dollar Q4 2025; venture and tail-end pools 20–35% off in H1 2026; VC secondary volume ~$152B in 2025; NAV finance ~$70B deployed in 2025. [Chronograph; IMD; 17Capital]
  25. OpenAI ~$7B employee tender at $852B (Aug 10 2026); Anthropic's Apr 2026 tender at $350B fell short of $6B demand; Nasdaq Private Market marks Anthropic above $1.1T. [Bloomberg; TechCrunch]
  26. Endowments: Harvard ~$1B PE-stake sale (2025); Yale up to $6B via Evercore (2025) and $1.8B via Lazard (Feb 2026); top ten endowments sold $12.5B in Q1 2026; endowments over $1B average 9.8% in venture. [CIO; Crimson; NACUBO 2024]
  27. Moratoria: NY S10642/A11560 passed the legislature Jun 4 2026 (≥20 MW, one year, construction exempt), not signed as of this writing; the operative freeze is NY EO 62 of Jul 14 2026 (≥50 MW); Oklahoma SB 1488 (>100 MW to Nov 1 2029); Indianapolis 23–1 on Aug 10 2026 through 2027; Maine vetoed; Hill County TX rescinded Jun 4 after suit; >300 bills, >100 local moratoria, ~$64B blocked or delayed. [DLA Piper; Davis Polk; Good Jobs First]
  28. PJM 2027/28 capacity auction $333.44/MW-day, ~1,053% above pre-surge; data centers nearly all of a 5,250 MW load increase; ~$70/month household impact by 2028. [IEEFA; Utility Dive; PJM IMM]
  29. Dodd-Frank §1101: 13(3) facilities broad-based (identifiable market, ≥5 participants), Treasury approval, no insolvent borrowers; ESF $50B money-fund guarantee Sep 19 2008; CARES §4003(b)(4) up to $454B, $195B committed by Nov 15 2020. [Fed 2015 rule; CRS R46329; GAO-21-180]
  30. SMCCF, announced Mar 23 2020 with $250B of capacity (PMCCF $500B), bought IG bonds rated ≥BBB-/Baa3 and broad IG ETFs and held about $14.2B when purchases stopped. [NY Fed]
  31. Executive: EO "Accelerating Federal Permitting of Data Center Infrastructure" Jul 23 2025 (>100 MW; federal land; loan guarantees); DOE named INL, Oak Ridge, Paducah, Savannah River Jul 24 2025; EO "Ensuring a National Policy Framework for AI" Dec 11 2025 (preempt state AI laws); Senate struck the 10-year state moratorium 99–1 Jul 1 2025. [White House; White & Case]
  32. SVB: $42B withdrawn Mar 9 2023, ~$100B queued for Mar 10, seized mid-session; Cookson et al. (SSRN 4422754): 4.3 pp excess loss for high-Twitter-exposure banks; Bear Stearns liquidity ~$18B to ~$2B Mar 10–13 2008; Cramer Mar 11 2008. [Fed Barr review; FCIC; R6]
  33. History: WorldCom Ch.11 Jul 21 2002 ($107B assets, $41B debt); Lucent ~$8.1B vendor financing, Winstar $2B; ~10% of fiber lit by 2004; WPPSS (five planned plants, one finished) $2.25B default Jul 25 1983, the largest municipal default until Detroit 2013, no federal rescue; Insull collapse 1932, 600,000 shareholders, PUHCA 1935; Iridium Ch.11 Aug 13 1999, ~$5B system sold for $25M; shale ~110 bankruptcies with ~$74B of debt 2015–16; Evergrande ¥2.43tn liabilities, default Dec 2021; merchant power: NRG Ch.11 May 2003, Mirant Jul 2003, Calpine Dec 2005 with ~$22B of debt. [CRS; Newsweek; UPI; McDonald 1962; Reuters; Haynes & Boone; R1]
  34. Moody's (Jun 2026): US life insurers' private and illiquid credit ~$807B, about a fifth of a ~$4T bond book; private placements overall ~48% of bonds. SEC temporary relief for BDCs (Apr 2020): asset-coverage flexibility and affiliate co-investment. [Moody's; SEC]