Japan is repricing. Foreign demand for Treasuries is thinning. Washington is borrowing at a record pace, and the AI buildout now leans on the same debt markets. This brief maps how those pieces connect, who is exposed, and lets you stress test the funding rate yourself.
Run the stress testFor two decades, near-zero rates at home sent Japanese savings abroad, and a large share landed in U.S. Treasuries. That flow is reversing at the margin, and the margin is what sets price.
The 10-year JGB yield sits at 3.11%. Domestic bonds now compete with foreign assets for Japanese savings.
Borrowing yen to buy higher-yielding assets pays less as Japanese rates rise. Leveraged positions carry unwind risk.
Japan is the largest foreign holder at about $1.1T. China has fallen from $1.3T in 2013 to about $618B. Slower reinvestment alone matters.
With fewer natural buyers, the market demands more yield. A higher 10-year raises the discount rate on every long-dated project.
U.S. versus its 52-week low. Japan versus about a year ago.
Trillions of U.S. dollars. Figures vary by source and date.
The federal government and the AI industry are now reaching for the same long-duration buyers: insurers, pensions and foreign accounts. When both borrow heavily at once, the clearing price is a higher yield for everyone.
With the policy rate near 3.6% and the 10-year at 5.3%, the gap is term premium: the extra yield investors demand to hold long debt through fiscal and inflation uncertainty. AI projects are priced off the long end, so this is the rate that matters for the buildout.
Record issuance meets a thinner foreign bid. Someone has to buy every bond, and the marginal buyer sets the yield.
More supply and more fiscal uncertainty lift the premium for duration, independent of Fed policy.
Stimulus into a tight economy and tariff-driven input costs push expected inflation, and long yields with it.
A Fed intent on shrinking its bond holdings is a seller too, stacked on top of Treasury issuance.
For two decades the U.S. could count on buyers who did not haggle over price: the Fed, and foreign central banks led by Japan and China. Both have pulled back. The debt still gets sold every week. It just goes to buyers who want to be paid more to hold it, and that is where the higher rate comes from.
Pick a view. Billions of U.S. dollars.
Dollar stablecoins are backed one for one with cash-like assets, and the big issuers have put most of that money into T-bills. Tether held $115.0B of T-bills at June 30, plus $25.6B of Treasury-backed reverse repo. Circle's USDC reserves held $13.2B of Treasuries and $49.6B of Treasury repo at August 31. Together that is about $128B held directly, or about $203B counting repo.
The San Francisco Fed puts that in context. Over the past five years, stablecoin issuers added about $200B of Treasuries, more than 40% of China's decline over the same period. If the trend holds, their demand could reach about $400B by the end of 2030.
The GENIUS Act, signed July 18, 2025, limits payment stablecoin reserves to cash, insured deposits, Treasuries with 93 days or less to maturity, overnight Treasury repo, and government money funds that hold the same things. It takes effect 18 months after signing, on January 18, 2027, or 120 days after regulators issue final rules, whichever comes first.
That turns stablecoin growth into a standing bid for bills. Each new dollar of regulated stablecoin needs a dollar of short Treasury paper or its equivalent behind it, whether rates are high or low.
The New York Fed's estimate of the 10-year term premium reached 0.98% on October 7, the highest since 2014 and up from 0.62% a year earlier. The 3-month bill yields 4.22% and the 10-year yields 5.28%. The front end found new buyers. The long end, where AI projects borrow, has to pay up to find them.
Bills are 22.8% of marketable debt. Treasury plans to hold coupon auction sizes steady "for at least the next several quarters" and says it is watching Fed bill purchases and private demand for bills. That plan works because money funds, the Fed and stablecoins want bills.
Hedge funds, offshore funds and money funds buy when the yield is right and leave when it is not. A buyer base built on repo funding can pull back fast in a stress event, which is exactly when Treasury needs it most.
Stablecoins filled part of the hole China and Japan left, but at the short end of the curve. The long end has to clear on price, and that price is the 10-year the stress test below runs on.
You can raise capital for a data center in months. Getting it energized takes years. That gap is where financial risk and physical risk meet.
Average wait for a grid connection in primary U.S. markets. Northern Virginia runs about seven years.
Substation transformer lead times in 2026, up from about 140 weeks in 2023.
U.S. data center power demand, 2024 to 2030, roughly tripling.
Average interconnection queue time today, up from under two years in 2008.
Debt raised before the power arrives accrues interest with no revenue against it. At a 7.8% all-in cost, a project that waits two extra years for power spends about 16 cents of interest for every dollar borrowed before it earns anything. Higher rates and longer queues compound each other.
Illustrative arithmetic: 2 years x 7.8%, before compounding.
Public infrastructure programs and private AI campuses compete for the same turbines, transformers, transmission and high-voltage engineers. Developers are turning to on-site generation to skip the queue. Maine banned new data center construction outright in April 2026. Closer to home, Siemens is expanding a transformer plant in Charlotte to $421 million.
"Is AI a bubble" is the wrong question. The useful one: if rates stay high and revenue lags the buildout, who can turn off the spigot without dying?
GPU-backed debt has gone from roughly 11% to 14% all-in private credit in 2023 to investment-grade structures at SOFR plus 225 basis points in 2026. Lambda closed a $926M loan at SOFR plus 300. Sub-investment-grade pools still price around 450 to 550 over. Lenders are underwriting the customer contract, not the chip. Nebius shows the alternative: customers prepay at least half its buildout.
The BIS counts more than $200B of private credit loans outstanding to AI companies and projects, with exposure running through insurers, pensions and bank lending to nonbanks. In a Bank of America survey, 48% of fund managers named AI hyperscaler capex as the most likely source of the next systemic credit event. Of 91 hyperscaler bonds issued this year, 78 were trading at higher yields than at issue by late July.
Each tier's new borrowing reprices at the 10-year plus its credit spread. Debt already locked stays at today's rate. Coverage is cash flow divided by interest. Below 1.25x, lenders typically balk; below 1.0x, the business cannot pay its interest.
| 10-year | Self-funders | Stretched middle | Fragile tier |
|---|
How the math works. Per $100 of AI capital: debt equals leverage x $100. Cash yield is annual operating cash flow (EBITDA) per $100. The share that reprices is debt that rolls or is newly raised over roughly three years; it costs the 10-year plus the tier's spread plus any shock. The rest stays at today's rate. Coverage equals cash flow divided by interest.
Why these defaults. Self-funders: 30% leverage reflects incremental debt at 32% of hyperscaler capex; spread of 0.80% fits AA to A credit; long-dated bonds mean only 35% reprices. Stretched middle: 60% leverage and a 1.75% spread fit BBB issuers and project finance. Fragile tier: 80% leverage reflects CoreWeave-style balance sheets (debt to equity near 9x) tempered by peers such as Nebius that use customer prepayments; a 2.50% spread over the 10-year brackets SOFR plus 225 to 550; 85% reprices because most contracted capacity is not yet financed.
Simplifications. Floating loans actually key off SOFR, which follows the Fed; spreads are expressed over the 10-year so one dial moves every tier. EBITDA ignores the cost of replacing GPUs, which flatters the fragile tier. CoreWeave's own trailing figures imply coverage nearer 1.1x to 1.2x today because much of its debt funds capacity not yet earning revenue. This is a framing tool, not a forecast.
The fiber buildout of the late 1990s is the right parallel. The fiber got used. Most of the companies that borrowed to lay it did not survive to see it. The infrastructure was right; the capital structures were wrong.
So the question is not bubble or no bubble. It is which players are funding the buildout with earnings versus debt, how much of that debt reprices before the revenue arrives, and how entangled it is with the banks, insurers and private credit funds that made the loans. Contained, it is a shakeout where the strong buy distressed assets cheap. Entangled, a digestion period becomes a credit event.