The Long End Has Two Sellers Now

The case that the United States is approaching a fiscal limit is no longer a fringe position, and it no longer rests on projections alone. The Congressional Budget Office’s February baseline put debt held by the public at 101% of GDP this year, on its way to 120% by 2036, and net interest at more than $1.0 trillion — larger than the defense budget. By August, CBO had revised the FY2026 deficit to roughly $2.1 trillion after the Supreme Court’s February ruling on IEEPA tariffs cut projected customs receipts by roughly $250 billion — about 60% below earlier projections, though the net deficit effect is smaller once stronger income and payroll receipts are counted; July alone ran a $432 billion deficit, a record for the month — though roughly $99 billion of that is a calendar artifact, since August 1 fell on a weekend and payments shifted into July; CBO puts the underlying figure at $333 billion. Net interest through ten months of the fiscal year is $963 billion, up 14% year on year, and CBO attributes the increase to a larger debt stock and higher long-term rates, with declines in short-term rates partially mitigating it — an attribution that contains the whole transmission problem this piece is about.

The market has started to price it. The 30-year Treasury touched 5.31% on August 17 and ~5.33% on August 18, a 19-year high; the August 13 thirty-year auction cleared at 5.216%, the highest since 2001, and the August 19 twenty-year drew weak demand at 5.204%. The rise is almost entirely real-rate and term-premium — the July FOMC minutes attribute the intermeeting move to “corresponding increases in real interest rates,” with breakevens little changed — and the NY Fed’s ACM 10-year term premium, near zero in late 2024, printed ~0.84% at the end of July. Treasury’s response on August 19 was to announce a doubling of long-end buybacks, to “at least $4 billion per operation” from September 9, with Secretary Bessent saying the next day “we are going to make a market in these” and that yields “don’t reflect the underlying fundamentals of this Iran conflict” — the Secretary attributing the move to geopolitics, not the deficit. Yields fell on the announcement and gave most of it back within a day.

We are not going to adjudicate whether this becomes a crisis. Ray Dalio says the US is “past the point of no return” and expects a 1930s-style yield-suppression outcome; Jamie Dimon says “the way it’s going now, there will be some kind of bond crisis, and then we’ll have to deal with it” — while insisting, separately, “I’m not that worried… We’ll be able to deal with it”; the Penn Wharton Budget Model puts an outer bound on sustainable debt at ~210% of GDP, likely reached within two decades on its higher healthcare-cost scenario, with a 25% chance inside fourteen years — while stressing this is a bound rather than a forecast, and that a self-fulfilling run could arrive well below it. Reasonable people hold each position. What we will do is insist on precision about the mechanism, because “debt crisis” is doing a lot of work in those sentences and it means different things for a data center depending on which of three paths it takes.

The three are not exclusive; a real episode usually starts as Path I and is resolved into II or III by the policy response. The Japan 2025–26 super-long selloff (40-year JGBs from under 3% to a record 4.10% on MOF’s reference series in August 2026, with life insurers briefly turning net sellers in May before returning as the largest buyers in three years in June) and the UK’s July 2026 gilt move (30-year at 5.75%, a two-month high and close to the September 2025 peak; the 10-year at 5.04% is the G7 high) are both Path I episodes that have so far been contained without a fork. Treasury’s August 19 buyback announcement is the first explicit US step toward Path II; Chair Warsh’s stated preference for a narrower Fed footprint and the July FOMC’s tightening bias are the institutional resistance to it.

Here is the conflict stated plainly. The Treasury needs to place roughly $2 trillion of net new debt a year into a market whose traditional anchors are retreating: total foreign holdings of Treasuries fell $72 billion in June to $9.30 trillion, with China at $633 billion, its lowest since 2008, and Japan, the largest holder, also a net seller. The IMF’s April Fiscal Monitor says the quiet part directly: the growth in Treasury supply “is compressing the safety premium that US Treasuries have traditionally commanded,” and “private (and often leveraged) investors have become the marginal buyers of government bonds.” The BIS’s June annual report adds that hedge funds’ sovereign exposure has more than doubled since 2022, that ~70% of their bilateral dollar repo runs at zero haircut, and that high public debt alone raises the three-month probability of a GFC-type stress event roughly tenfold (3.8% versus 0.3%).

Into that market, the AI build-out has become the second seller. Moody’s now puts 2026 capex for six US names (the five hyperscalers plus CoreWeave) at $785 billion, approaching $1 trillion in 2027. Morgan Stanley’s framework has ~$1.4 trillion of the $2.9 trillion 2025–28 global data-center bill funded from operating cash flow and a $1.5 trillion gap filled externally — $200 billion of investment-grade bonds, $150 billion of ABS and CMBS, $800 billion of private credit and asset-backed finance. The bond leg is running ahead of that schedule: hyperscaler-linked issuance was $194 billion through July 7, up 79% on the $108 billion issued across all of 2025 by the same four issuers; Goldman projects roughly $250 billion for 2026 and $400 billion in 2027 for a five-issuer universe that adds Microsoft, so the two are not quite like for like. Nine deals above $20 billion have priced this year; seven were hyperscaler-linked.

What matters for the conflict is not the volume but the tenor. Meta’s $30 billion October 2025 deal included a 40-year tranche at 5.75%. Oracle’s $25 billion February deal included a 40-year at Treasuries +195bp. Alphabet sold a 100-year sterling bond in February at 6.125%, ten times covered — and by late July it traded below 90 pence, a loss of more than 10% on an AA-rated credit in five months. Amazon’s $25 billion July deal paid new-issue concessions of 18–21bp on its longest tranches by Bank of America’s tally (other counts put it nearer 12bp), on the weakest hyperscaler book since Meta’s. These are not coincidences. A 40-year corporate bond and a 30-year Treasury are sold to the same pension, the same insurer, the same Japanese lifer now retreating from its own super-long curve. When the sovereign’s supply of duration rises, the corporate’s price of duration rises with it; when the corporates flood the long end, the sovereign’s auctions tail.

Treasury’s August refunding held coupon sizes flat “for at least the next several quarters” and pushed incremental borrowing into bills; the bill share of marketable debt is already ~22%, against a long-standing TBAC guideline of 15–20%, and dealers project it rising further if coupons stay frozen. The IMF notes that “systemic economies such as the United States” have shortened their maturity profiles “to manage near-term interest bills.” The compute complex made the same choice for the same reason: CoreWeave’s GPU facilities are five- to six-year amortizers at SOFR-plus, not 30-year fixed, because no one would lend 30 years against a depreciating chip. Two borrowers with very different balance sheets have converged on the same maturity structure, which means they now roll into the same windows — and a funding shock that closes the long end forces both onto the short end at once.

The intuitive model — “rates go up, projects financed with debt get more expensive, some default” — is mostly wrong for the capital already in the ground, and it is wrong in an instructive way. Work through where the money actually sits.

Channel 1 · The long end into fixed-rate hyperscaler debt — fast for holders, slow for issuers

A +150bp move in long yields takes roughly 18–20% off the price of a 30-, 40- or 100-year 5.6–6.1% coupon bond. That loss is real, but it belongs to BlackRock, PIMCO, the Japanese lifer and the pension that bought Meta’s 40-year — not to Meta, which locked 5.75% for four decades and is, if anything, a winner from the move. The issuer’s exposure is to the next deal: Oracle’s February 40-year priced 40–58bp wider than its September tranche, its five-year CDS reached a record 215bp in late July, S&P cut it to BBB− on July 9, and the company said in February it does not expect to issue further bonds in calendar 2026. That is the channel working as designed: the marginal leveraged builder is priced out of the long end first. Alphabet and Amazon, by contrast, borrow at 2.5–3.5% yields with $100 billion-plus of cash; their issuance is a capital-structure choice, not a necessity, and a shock changes the choice rather than the capex.

Channel 2 · The short rate into floating GPU debt — the channel that doesn’t fire on Paths I and II

This is the most important and least appreciated point. Neocloud and project debt is overwhelmingly floating against SOFR: CoreWeave’s DDTL 4.0 at SOFR+225, its older facilities at SOFR+400, construction loans at SOFR+150, CoreWeave-tenant deals underwritten at SOFR+325–425. SOFR follows the federal funds rate. A term-premium shock that takes the 30-year from 5.2% to 6.7% while the Fed holds at 3.50–3.75% — which is exactly the Path I shape, and exactly what CBO described happening this year — does not raise the coupon on a single existing GPU loan. Under Path II (fiscal dominance) the Fed is, if anything, cutting. The stock of GPU debt is insulated from the long end by construction. What it is not insulated from is the spread, and the refinancing date.

Spreads are where Path I bites. CoreWeave’s net interest was $640 million in Q2 against $1.51 billion of adjusted EBITDA, and lenders already price its tenancy at 175–275bp over a hyperscaler’s. A sovereign-led widening of IG and HY spreads — from today’s unusually tight 81bp and 273bp — would reprice every DDTL at its next draw and every 2027–29 refinancing at once. CoreWeave has described its near-term maturities as self-amortizing contract-backed debt and vendor financing — but the facilities that do mature earlier are precisely the ones collateralized by chips whose rental value is falling 20–30% a year. The first debt-crisis casualty in this complex is unlikely to be a hyperscaler. It is far more likely to be a floating-rate, non-hyperscaler-tenant, partially merchant neocloud whose 2028 refinancing arrives in a spread regime that didn’t exist when the loan was written.

Channel 3 · The dollar into the cost stack — hits the flow, not the stock

About 78% of data-center spend is IT equipment, and the US imports 70–90% of its computing equipment — Taiwan supplied $86 billion of servers in 2025 and “100% of US-brand AI servers” by CCIA’s count, increasingly routed through Mexico under USMCA. Put those together and roughly 60% of a new project’s capex is import content (our derivation; the GPU share alone is 45%+ by CSIS’s estimate). Non-petroleum import prices are already up 4.5% year on year, the most since 2022, with capital goods leading. A further 10–20% dollar decline — the 2025 move was 9.5%, the worst year since 2017 — passes through at 30–35% in the first year and more over time, which puts 6–12% on a project’s capex before supplier margins absorb any of it. Amazon already raised its 2026 guidance by ~$20 billion citing memory costs; transformer prices are up ~80% in five years with four-year lead times; GE Vernova’s gas-turbine slots run into 2031. None of this touches a project that has already taken delivery. All of it touches the next one.

Channel 4 · Inflation into opex and revenue — the channel that decides the winner

Under Path II, the economy runs hotter in nominal terms. Power, the dominant operating cost, inflates — PJM’s capacity auction has now cleared at the cap three years running, at $325–333/MW-day, with the 2028/29 auction 6.8 GW short of its reliability requirement before any macro shock. But nominal rental prices also hold better in an inflationary world than a deflationary one, and a fixed-rate 40-year liability gets cheaper in real terms every year. Fiscal dominance is, perversely, the kindest path for the most leveraged long-dated borrowers and the cruelest for their bondholders. Path I is the reverse. Path III is bad for everyone, but it is the only one in which the floating-rate stock breaks.

To make the channels concrete we built an illustrative 100 MW AI facility and ran it through the four regimes. The inputs are drawn from published ranges, not from any one deal: $13 million per MW for shell, power and cooling and $25 million per MW for the IT stack ($3.8 billion all-in, ~80,000 accelerators at ~1.2 kW each), 70% debt on each layer, a 20-year-amortizing shell loan with a seven-year mini-perm priced off the 10-year Treasury plus 150bp, a five-year fully amortizing GPU facility priced off SOFR plus 300bp, opex of $1.4 million per MW-year including power, 80% utilization, and — the parameter that turns out to matter most — a rental curve that starts at $2.50 per GPU-hour and declines 22% a year, which is the gentle end of the technology-deflation range our listing-window work and the Merchant’s Dilemma case study both arrive at (ln 2 over a 2.5–3-year performance-per-dollar doubling). Two variants: a merchant project that sells at the index, and a contracted project with a five-year take-or-pay at a flat $1.80 and 95% utilization.

Three things fall out of the table, and they reorganize the debt-crisis question.

First, the merchant project dies of deflation, not rates. Under today’s curve, with no macro shock at all, a facility selling at the index and carrying a five-year GPU amortizer drops below 1.0x coverage in year five. The rate shock moves that breach forward by a year and takes the equity from a respectable 16% to negative — material, but a second-order effect next to the 22% annual decline in what the asset earns. This is the case study’s point that “long-dated fixed-price receivables are implicitly short technological progress,” seen from the lender’s side: a long-dated fixed-rate liability against a depreciating asset is implicitly long technological progress, and the debt crisis only changes the price of that bet, not its direction.

Second, the contracted project barely notices. With a five-year take-or-pay from a creditworthy tenant, the worst regime takes minimum coverage from 2.1x to 1.6x and the IRR from 48% to 29%. That is why the 15-year hyperscaler lease, the Hyperion four-year block with a residual-value guarantee, and the utility’s 15-year minimum-demand tariff exist: they convert a technology-deflation exposure into a credit exposure on the tenant. Which means the question “does a debt crisis cause data-center defaults” reduces, for the contracted majority of the build, to “does a debt crisis impair the tenant” — and for Meta and Alphabet the answer under Paths I and II is no; for Oracle, at BBB− with mid-4x leverage and negative free cash flow, it is the live question; for a neocloud tenant it is the whole game.

Third, the real casualty is the flow. Debt service per MW rises 13–32% and capex 6–13% across the shock regimes. A project that cleared its hurdle at A does not clear it at B or C. That is where the macro shock reaches physical reality: not in defaults on existing capacity but in the cancellation of capacity that was going to be built in 2027–29, on top of the $130 billion already blocked or delayed in the first quarter by local opposition. And cancelled supply is the stabilizer for everyone already in the ground — it slows the deflation curve that is actually killing the merchant case. A debt crisis is bad for the builder and, with a lag, good for the incumbent.

Scenario thinking beats point forecasts here, so we lay the episode out as a sequence with the markers that would confirm each phase. The shape borrows from the gilt crisis of September 2022, when the 30-year moved ~130bp in three sessions — by the Bank’s own account more than twice the largest move since 2000 — and the Bank of England was buying within five days; from the 1994 bond massacre, in which the 30-year went from under 6% in late 1993 to over 8% by late 1994 and the casualties (Orange County) were leveraged duration holders nobody had been watching; and from Japan’s 2025–26 super-long episode, which shows that a reserve-currency sovereign can lose its long-end bid slowly rather than suddenly.

Two features of the sequence deserve emphasis because they run against the intuitive story. The first is that Phase 2 precedes any default by quarters, and it is deflationary for compute prices before it is inflationary for anything else: shelved projects and a closed issuance window reduce future supply, which is precisely what an asset class in 20–30% annual rental deflation needs. The second is that Phase 3 is a policy choice, not a market outcome, and the choice flips the sign of the effect on the two main borrower types. A Fed that holds (Path I) protects the bondholder’s real return and sacrifices the refinancing neocloud; a Fed that caps (Path II) does the opposite. Anyone positioning for “the debt crisis” without a view on which fork the Warsh Fed takes is positioning for two opposite outcomes at once. The July FOMC minutes — three dissents for a hike, a stated view that tightening “would likely be necessary if inflation did not decline” — and Warsh’s own call for a modernized Fed–Treasury accord that narrows the Fed’s balance-sheet footprint both argue the institution is leaning toward Path I. Treasury’s buybacks argue the administration is leaning toward II. That disagreement is the single most important input to the compute credit question over the next two years, and it is not a compute question at all.

Our companion case study on positioning a compute business, The Merchant’s Dilemma, specifies a risk architecture built around one common factor — AI demand — and a set of pre-commitments designed to keep a merchant out of the Enron quadrant: aggregated contingent-liquidity stress, a wrong-way capital charge on any financing where collateral value, counterparty revenue and firm equity load on that factor, and cleared-first credit. The debt-crisis scenario does not invalidate any of it. It adds a second common factor that the architecture was not written to see.

The sovereign factor loads on the same positions. The GPU-collateralized loan whose collateral, borrower revenue and lender equity all correlate to AI demand also correlate — through the refinancing spread, the import bill and the residual-value put — to the term premium and the dollar. The case study’s wrong-way charge, exposure at default times (1 + λρ), should be computed against both factors, and the cross-term is the dangerous one: a Path I shock that widens spreads at the moment rental deflation reaches the covenant is two bad draws from one underlying cause, because the sovereign shock causes the supply cancellation that would otherwise have slowed the deflation, and it does so only with a lag.

The contingent-liquidity table gains a row. The case study’s stress template enumerates triggers — firm credit events, index dislocation of ±40%, counterparty downgrade clusters, vendor or architecture shocks — and aggregates the contingent outflow against unencumbered liquidity. A sovereign-rates event belongs on that list with a timing of days: it hits cleared futures margin (variation and initial-margin step-ups scale with rate volatility even when the compute index itself hasn’t moved), it hits the value of any Treasury collateral posted, and it hits the repo rate on that collateral at exactly the moment it is needed. The 2022 LDI episode is the template: the damage was not that gilts fell but that the fall generated margin calls whose settlement required selling gilts. A compute desk that posts Treasuries as margin against compute futures is running a small version of the same loop.

The ratings keystone reappears in a new costume. Enron’s fatal dependency was an investment-grade rating that, once lost, converted contingent liabilities into cash calls in weeks. The AI complex has rebuilt that dependency one notch removed: the Hyperion bonds are rated “Meta-minus-one” on the strength of four-year lease blocks and a residual-value guarantee; CoreWeave’s DDTL 4.0 carries an A3 because of who the offtaker is, not what the collateral is; NVIDIA’s financing platforms and residual-value support are only as good as NVIDIA’s own balance sheet, which is now the hidden seller of duration in the whole structure. A sovereign shock that migrates ratings — and Oracle has already been cut to BBB−, one notch above high yield, on its own capex — pulls on every one of those keystones at once. The case study’s remedy, cleared where clearable and every ratings trigger aggregated into a single reported number, is the right one; the debt-crisis case is the reason to run it now.

And the dislocation stage arrives on schedule. The case study’s Scenario C — a shock “before the market matures” that either kills the derivatives or proves them necessary, and in which franchises are bought from distressed sellers — is the Phase 3–4 sequence above with a macro trigger rather than a compute one. Its positioning implication transfers intact: build the balance sheet and the dry powder now to be the buyer of shells, power and interconnection in that moment rather than a seller of silicon into it. The Phibro pivot — own the durable asset, flow the depreciating one — is also, it turns out, the correct rates hedge, because the shell is the only part of the stack whose value rises under fiscal dominance.

1. Long-end auction tails and the buyback size. Bessent said the $4 billion per operation “could be more.” Each increase is a step toward Path II; each tail that isn’t met is a step toward Phase 1.
2. The Fed’s reserve-management purchase maturity. The directive currently allows bills “and, if needed, other Treasury securities with remaining maturities of 3 years or less.” Those purchases were themselves cut to zero on August 14; the markers are whether they resume and at what maturity. The day that ceiling moves is the day the fork is chosen.
3. Hyperscaler new-issue concessions and book coverage. Amazon’s July deal — 2.5x covered and 18–21bp NIC by BofA’s count — is the baseline; the next sub-2x book is Phase 1 for the corporate leg.
4. Oracle’s CDS and the Hyperion spread. The two best-priced proxies for “leveraged builder” and “hyperscaler-minus-one” credit; both already at series highs.
5. Neocloud DDTL repricing and 2027–28 refinancing terms. The first facility that rolls at SOFR+500 or can’t roll is the canary; watch the self-amortizing, contract-backed paper before the 2029 unsecured.
6. Non-petroleum import prices and server import values from Taiwan and Mexico. Already +4.5% year on year; the dollar channel shows up here quarters before it shows up in a capex guide.
7. Project cancellations versus the rental index. If the blocked-and-delayed tally keeps rising while the H100 and B200 indices stop falling, Phase 2 is stabilizing incumbents on schedule. If the indices keep falling anyway, the deflation curve is the whole story and the macro is noise.
8. Compute-futures open interest after October 5. Lender-mandated hedging is the mechanism that converts a macro shock from a balance-sheet event into a priced one. Its arrival would be the first evidence that the market is doing what gas and power did after their crises.

The question we began with — does a US debt crisis threaten the AI build-out — has a precise answer once the mechanism is specified. It does not threaten the capacity that exists, most of which is either fixed-rate and long or floating against a rate the crisis doesn’t move. It threatens the capacity that was going to exist, through the hurdle rate and the import bill; it threatens the refinancing of the most leveraged tier, through spreads; and it threatens the tenant credit on which the contracted majority depends, in proportion to how leveraged that tenant already is. The shape of the damage depends on a policy fork the compute market has no influence over. And the asset that already carries a 20–30% annual deflation curve will, in every regime, find the macro shock to be the smaller of its two problems.

• Fiscal: CBO, Budget and Economic Outlook 2026–2036 (Feb 11, 2026) · CBO Monthly Budget Review, July 2026 (Aug 2026) · GAO on federal debt management (Mar 2026) · Penn Wharton Budget Model fiscal-limit update (Jun 2026) · IMF Fiscal Monitor (Apr 2026) · BIS Annual Economic Report, ch. II (Jun 2026).
• Rates and Treasury: FRED DGS30 / DGS10 · CRFB on the Aug 13 30-year auction · Treasury August 2026 refunding statement · Treasury doubles long-end buybacks (Aug 19) · Bessent: “make a market” (Aug 20) · FOMC minutes, Jul 28–29, 2026 · Fed Monetary Policy Report (Jul 2026) on reserve-management purchases · Warsh’s Fed–Treasury accord proposal · BofA bill-share projection and dealer pushback · TIC June 2026 data.
• Voices: Dalio, “past the point of no return” (Jun 4, 2026) · Dalio, How Countries Go Broke summary · Dalio House Budget Committee statement (Mar 26, 2026) · Dimon, Oslo (Apr 28, 2026) · Rogoff (Jan 2026).
• Analogs: Bank of England gilt-market case study (2023) · Fed note on the 2013 taper tantrum · Japan 40-year JGB · Japanese lifers net sellers of super-longs (May 2026).
• Build-out scale and issuance: Moody’s $785B / ~$1T (May 2026) · Moody’s on balance-sheet structure and $821B off-balance-sheet commitments (Jul 2026) · Morgan Stanley, Bridging the Data Center Gap · Amazon $25B July deal and 2026 issuance tally · Oracle $25B February deal (IFR) · Meta $30B October 2025 deal · Alphabet century bond below 90p · Oracle leverage, CDS and ratings (Reuters, Aug 4, 2026) · SFA on data-center ABS/CMBS (Jul 2026).
• Structures and terms: Meta–Blue Owl Hyperion release · IFR on Hyperion structure · CoreWeave Q2 2026 results · CoreWeave DDTL stack · Norton Rose Fulbright, Cost of Capital 2026 · NVIDIA financing-platform MOUs (Aug 10, 2026) · Meta Q2 2026 call transcript.
• Imports, power and costs: CCIA on import content and tariffs · CSIS on GPU imports · Import prices (Jul 2026) · PJM 2028/29 BRA results · GE Vernova backlog · Transformer lead times · $130B blocked or delayed (2026) · Silicon Data H100 index.
• Revenue: OpenAI run-rate >$40B (Aug 2026) · Anthropic $65B run-rate (Aug 2026) · Bain $2T / $800B gap (Sep 2025).
• Kinetic Alpha: Compute Became Collateral. NVIDIA Sold the Floor. · The Contract Got a Date. · The Contest Can’t Be Hedged. The Financing Can. · Bandi & Su on the compute risk premium · The Merchant’s Dilemma: Phibro, Enron, and the Making of a Compute Business (case study, Aug 2026).
• Model notes: the 100 MW project uses $13M/MW shell and $25M/MW IT (within the $30–45M/MW full-stack range in JLL/Turner & Townsend syntheses), 80,000 accelerators at ~1.2 kW node-level draw, 70% debt on each layer, a 20-year-amortizing shell loan with 7-year mini-perm at UST10 + 150bp (a market estimate for a non-hyperscaler-leased shell; Norton Rose reports ~SOFR+150 construction pricing for contracted projects), a 5-year fully amortizing GPU facility at SOFR + 300bp, $1.4M/MW-yr opex, 80% utilization (95% contracted), a $2.50/GPU-hr starting rental (Silicon Data H100 neo-cloud index $2.53), and 22% annual rental deflation (17% in the inflation regime) — the gentle end of the ln 2/N range; a 28% curve would bring the merchant breach forward a year in every regime. Import content is 80% of IT and 35% of shell/power for the dollar pass-through. Dollar pass-through is applied fully to import content, which overstates first-year effects (observed pass-through is ~0.3–0.35) and is closer to a multi-year figure. Every figure is illustrative and should be re-cut against actual term sheets before use.