AI Bond Sales Now Share the Treasury Market’s Fuse

U.S. net interest hit $1.05 trillion in fiscal 2026, while AI builders sold about $200 billion of high-grade bonds into the same long end. Hedge funds still fund a record book of cash Treasuries in overnight repo, so debt and AI spending now press on one set of pipes.

The warning is not three shocks waiting to line up. Dallas Fed economists said on Oct. 16 that the cash-futures basis trade is already losing its other side to swaps used to finance data centers. A jump in yields now reprices the government’s coupon, the buildout’s debt, and the leveraged funds in between.

A $1.05 Trillion Interest Bill

Treasury figures for the year ended Sept. 30 put net interest at $1.05 trillion, up $80 billion, or 8.2%, from $970 billion in fiscal 2025. That print landed on Oct. 17. The average rate on marketable Treasuries was 3.52% at the end of September, more than double the 1.47% of five years earlier.

The Congressional Budget Office’s February outlook already had debt held by the public at $32.1 trillion for 2026, or 100.6% of GDP. On that path, net interest rising toward $2.1 trillion by 2036 would equal 4.6% of GDP. The Peter G. Peterson Foundation had marked the 2025 bill at 3.2% of GDP, matching the prior high from 1991.

THE 2026 INTEREST BILL

• Net interest: $1.05 trillion in fiscal 2026, up from $970 billion the year before.
• Average coupon: 3.52% on marketable Treasuries at the end of September, versus 1.47% five years earlier.
• Public debt stock: $32.1 trillion, or 100.6% of GDP, in the CBO’s 2026 outlook.
• Long-run path: $2.1 trillion of net interest, 4.6% of GDP, in 2036.

Deputy Assistant Secretary for Federal Finance Brian Smith said on Aug. 5 that Treasury would keep nominal coupon and floating-rate note auction sizes unchanged for at least the next several quarters. The refunding on offer that week was a $58 billion 3-year, a $42 billion 10-year, and a $25 billion 30-year. The next refunding is set for Nov. 4. The government is rolling a $1.05 trillion interest bill without adding long coupons, which leaves the long end open to whoever else wants to sell duration.

AI Borrowers Moved Onto the Long End

Srini Ramaswamy, a senior financial economist at the Federal Reserve Bank of Dallas, put the remaining-decade buildout on the order of $5 trillion and the cost of a campus at $30 billion to $40 billion per gigawatt. Power is part of that capital fight, which is why Amazon’s off-grid gas plant in Pecos is being framed as a way around the grid.

The funding mix has shifted. For the first years after ChatGPT, the large cloud companies paid for chips and halls mainly from operating cash. In 2026 they have been selling long bonds. Ramaswamy said in an Oct. 7 interview that the high-grade calendar for this financing was already large, and that most of it had been sold in the United States.

This year, for instance, there’s been about $200 billion in high-grade bond issuance to finance this. And about 70 percent of that is actually in the U.S.

Srini Ramaswamy, senior financial economist, Federal Reserve Bank of Dallas

He also said data-center operators prefer longer-term debt, that Treasury has not been raising long-end supply and has bought some of it back, and that the extra paper at those maturities is coming from the new AI borrowers. A February Dallas Fed note had already flagged that a year of this issuance could add as much as an eighth of the duration supply that comes from Treasury itself.

That is the second-order turn. The buildout is sold as the growth that makes a $32.1 trillion debt stock easier to carry. The bonds that pay for the halls are now the extra supply at the maturities the government uses most carefully. Higher yields then lift the $1.05 trillion interest bill and the coupon on the next data-center note in the same session.

The large issuers still carry high grades and still fund a big share of spending from operations, so a bust would more likely show up as capex cuts and write-downs than as a bank-loan crisis. The fuse is the duration, not a hidden pile of junk inside the core dealers.

$2 Trillion of Cash Treasuries on Overnight Repo

Ted Berg and Daniel Stemp at the Office of Financial Research wrote on Aug. 19 that hedge funds’ cash Treasuries at $2 trillion as of the end of 2025 were nearly three times the level of five years earlier. Marketable Treasury debt outstanding rose 29% over that span, to $28.9 trillion, so the funds’ share of the cash market hit a record 7%. Short futures against those holdings totaled $1.4 trillion.

A FEDS Note published June 22, using Form PF through September 2025, put large funds’ gross Treasury exposure at $4 trillion, with $2.4 trillion long and $1.6 trillion short. Repo cash borrowing had reached $3 trillion. Holdings were about 8.5% of privately held Treasuries, up from 4.5% at the start of 2023, and the 50 largest funds by gross exposure accounted for about 90% of the book, up from 84%.

The same note put the cash-futures basis trade near $830 billion, about double its early-2020 peak and 35% of the long book, or 3.5% of privately held Treasuries against a 2.5% peak in 2020. Swap-spread trades were about $305 billion. Basis and swap-spread together were nearly half of the $2.4 trillion long book. OFR, which does not split cash from derivatives on Form PF, said outside estimates of the basis trade run from $350 billion to $1.5 trillion.

The trade is small per dollar and large in borrowed money. Funds buy a cash note, sell the future, and roll the note in repo, often overnight because overnight is cheaper than term. Haircuts on the repo can be low or zero, and futures margin is thin, which is why a few basis points become a business. It is also why a jump in repo rates or a futures margin call can force sales of the same notes the government is trying to place.

OFR put the vulnerability in one line: a disorderly unwind of highly leveraged basis trades could amplify volatility in cash and futures. Dealers, boxed in by post-2009 capital rules, cannot always absorb the other side. Asset managers have been taking duration in futures rather than cash, which is the structural bid the funds fade. When that bid moves, the fade has to move with it.

Why the Basis Trade Is Losing Its Other Side

Matthew McCormick, Ramaswamy, and Hugo De Vere wrote on Oct. 16 that estimates of the basis trade have edged lower, and that the reason is not a sudden flood of naive new shorts. Higher yields make the notes eligible for delivery look more alike, which cuts the option value that shorts used to collect. Hedge funds also have swap-spread trades that pay a similar risk-adjusted return. And asset managers, the funds’ usual counterparties in futures, are taking duration in other derivatives, including swaps tied to data-center debt.

The mechanics are ordinary. A data-center borrower who wants fixed-rate money, or a lender who originated a floating loan, pays fixed on a swap. Asset managers can receive fixed on that swap instead of rolling a Treasury future. The Dallas Fed team called that receive-fixed flow a candidate for the missing futures bid, and noted that the new issuance is large and unusually long.

According to new analysis asset managers have increasingly found the cash-futures basis trade less attractive relative to alternatives, including swaps linked to new debt issuance tied to data center construction in support of artificial intelligence (AI). https://t.co/VEnPteZias pic.twitter.com/MzPzNWovd8

— Dallas Fed (@DallasFed) October 16, 2026

Once those swaps are the place duration lives, the Treasury market and the AI complex are not adjacent stories. They clear through the same dealers, the same futures commission merchants, and the same repo screens. A wider swap spread, a failed long-bond auction, or a margin increase on the note future all hit the same desks.

The cost of a chip-hour already tracks that pipe. A hyperscaler borrowing a thin spread over Treasuries can carry a GPU at a price a smaller cloud firm, paying a junk-like coupon, cannot match. When the 10-year backs up, the gap does not show up first as a software-revenue miss. It shows up as a higher break-even on the same rack.

April 2025 Showed How Fast Swap Trades Unwind

The Fed note is blunt about the last fire drill. After the April 2025 tariff announcements, swap spreads fell, yields jumped, and liquidity in Treasuries got worse. Market participants told researchers that funds dumping swap-spread trades made the conditions worse. The decomposition showed about $60 billion, or 20%, of those positions came off in April and another $40 billion in May. By September 2025 the book had been rebuilt to about $305 billion.

March 2020 is the older rhyme. The FEDS Note, following other Fed and academic work, partly attributes that Treasury stress to a rapid unwind of basis trades. Authorities then bought paper in size and backstopped funding. The 2026 version of that choice is harder, because a purchase program that pins the long end also cheapens the AI debt that is competing for the same buyers.

THE UNWINDS THAT ALREADY HAPPENED

1. March 2020: Treasury cash and futures seize up in the pandemic dash for cash; later research ties part of the move to a basis-trade unwind, and the Fed buys bonds in size.
2. April 2025: Tariff headlines spike yields; about $60 billion of swap-spread trades unwind that month and $40 billion more in May, then the book is rebuilt by September.
3. Year-end 2025: OFR prints hedge-fund cash Treasuries at $2 trillion, 7% of the cash market.
4. June 22, 2026: Fed researchers put the basis trade at about $830 billion and repo borrowing at $3 trillion as of September 2025.
5. August 5, 2026: Treasury holds coupon auction sizes steady and schedules the next refunding for Nov. 4.
6. October 7, 2026: Ramaswamy cites about $200 billion of high-grade issuance for the AI buildout, mostly sold in the United States.
7. October 16, 2026: Dallas Fed economists argue the basis trade is losing its asset-manager bid to AI-linked swaps.
8. October 17, 2026: Treasury data show a $1.05 trillion net interest bill for fiscal 2026.

The lesson from those dates is not that every spike becomes 2020. It is that the unwind channel is tested, and that the book got larger after each test. Concentration at the top 50 funds makes a correlated exit more plausible, not less.

Mortgages, Pensions, and the Chip-Rental Spread

A disorderly sale of cash notes does not stay inside hedge-fund NAV. Treasuries price other credit, sit in collateral accounts, and set the floor under fixed-rate products. The people on the other side of that floor did not sign up for 50-to-100 times leverage on a few basis points of gross basis.

WHERE A YIELD SPIKE LANDS NEXT

• Household credit: Mortgage and corporate borrowing rates move with the long end, so a forced sale of notes shows up in monthly payments, not just in dealer chats.
• Pensions and money funds: They feel it through collateral values and the cost of rolling short-term paper when repo haircuts gap.
• Smaller cloud borrowers: Their funding spread over Treasuries is already wide; a higher risk-free rate plus a wider spread prices them out of the same chips the large issuers still carry.
• The government itself: Each backup in yields adds to a $1.05 trillion interest line that already compounds on a $32.1 trillion public debt stock.

Crypto executive Arthur Hayes has already placed a public wager on an AI credit break as the event that would reprice risk assets. You do not have to take his side of that bet to see the transmission. If the buildout’s paper cheapens in a hurry, the same real-money accounts that left Treasury futures for AI swaps have to mark that book, and some of them will want cash duration again at the same moment leveraged funds are selling it.

Regulators can demand more reporting, more central clearing, and more term funding. OFR and the Fed have been writing that script for years. Tighter rules also raise the cost of making markets in Treasuries, which is the trade-off Smith’s refunding language is already dancing around: hold coupon sizes, watch bill demand, buy back off-the-runs, and hope private leverage keeps absorbing what dealers cannot.

Cutting Rates Would Juice the Same Buildout

The awkward policy choice in the original warning still holds. A cut eases the $1.05 trillion interest bill and the repo bill on $200 billion of high-grade issuance. It also re-steepens the incentive to lever the basis and to launch the next campus. A hold, or a hike, contains that incentive and presses public and private borrowers at once.

That is why the three-risk framing undersells the 2026 market. The interest bill, the AI bonds, and the overnight-repo book are not waiting on a coincidence. Asset managers are already taking the AI side of the duration trade, Treasury is holding long coupons flat, and the funds that have been the marginal cash buyers still fund in a market that can demand cash before the sun comes up.

Watch the next refunding on Nov. 4, the next round of cloud-company earnings later this month, and whether the basis trade’s other side keeps migrating into those data-center swaps. Each print is a reading on the same pipe.

Disclaimer: This article is news reporting and analysis of public market, fiscal, and research data. It is for information only and is not investment advice, a solicitation to buy or sell any security or derivative, or a recommendation of any trading strategy, including Treasury basis trades or technology-sector bonds. Readers should consult a licensed financial adviser, fiduciary, or qualified investment professional who can consider their own objectives, time horizon, and risk limits before acting. Interest costs, issuance calendars, position sizes, and credit conditions are taken from the cited official and research sources on their stated dates and will change with new auctions, filings, and policy decisions.