kangwijen
Back to posts

More Debt. More AI. Who Pays for It?

kangwijen

26 min read

On Wednesday, 2 September 2026, the U.S. 10-year Treasury yield traded as high as 4.818% (opens in a new tab) in New York, its highest level since November 2023. The 10-year yield is effectively the rate the U.S. government pays to borrow for a decade. Mortgage rates, car loans, and the math investors use to value fast-growing stocks all take their cue from it. FRED (opens in a new tab), the St. Louis Fed's data service, puts Wednesday's close at 4.79% and Thursday's at 4.77%.

The Treasury market is raising the price of long-term capital at exactly the moment Washington and the biggest technology companies both need more of it. The Congressional Budget Office (opens in a new tab) puts this year's federal deficit at 5.8% of GDP, meaning Washington is spending about 6 cents more than it collects in taxes for every $1 the economy produces. Microsoft, Amazon, Alphabet, and Meta (opens in a new tab) told investors in April that they plan to spend roughly $725 billion this year, mostly on AI data centers. Everyone refinancing a loan sits in the same queue. Inflation is what stops the Federal Reserve from making that money cheap again. Core prices, the Fed's preferred measure after stripping out food and energy, were 3.3% higher in July than a year earlier (opens in a new tab). The Fed's target is 2%.

There's a finite supply of investable capital. When the public sector and private sector both demand more of it, the price can rise even when no Treasury dollar walks straight into a data center. On Wednesday, when the 10-year hit 4.818%, the S&P 500 still closed up 0.46% (opens in a new tab).

Same yield, different setup

Yields fell below 1% during the pandemic, peaked near 4.80% in late 2023, and are back near those highs. But the setup around that level has changed.

In 2023, the Fed was still raising rates to fight inflation, so long-term yields were rising alongside an intentionally tightening monetary policy. The overnight rate banks charge each other is at a different point now. The Fed's rate-setting committee held its target at 3.50% to 3.75% on 29 July (opens in a new tab). Three regional Fed presidents, Beth Hammack of Cleveland, Neel Kashkari of Minneapolis, and Lorie Logan of Dallas, dissented in favor of a quarter-point hike. The Congressional Budget Office's February forecast (opens in a new tab) still assumed the Fed would be cutting rates in 2026 as the labor market weakened.

Friday's jobs report made an emergency cut look even less likely. On 4 September, the Bureau of Labor Statistics (opens in a new tab) reported that employers added 162,000 jobs in August, while unemployment held at 4.1%. That was well above the 31,000 average monthly gain over the prior twelve months. Hiring had been running cold. July was initially reported as a 23,000 decline and later revised to a 21,000 gain. August was a bounce.

John Williams (opens in a new tab), president of the New York Fed, told CNBC on the same day as the 4.818% high that he saw the move as a sign of economic strength. U.S. Bank's 1 September outlook (opens in a new tab) expects a quarter-point hike this month and sees core inflation averaging 3.3% in the second half of 2026. Christopher Waller, one of the seven Fed governors, offered a different explanation the next day at Reuters NEXT. He pointed to an ongoing deficit around 6% of GDP (opens in a new tab), a debt load around $40 trillion, and what he called the disappearance of the premium investors once accepted for holding safe, liquid U.S. government debt. In other words, investors may no longer be willing to accept as low a return for owning Treasuries. Reuters also reported Waller saying that competition for capital from AI infrastructure investment was pushing yields higher.

Why inflation isn't going away

The 3.3% core number matters because of what is producing it. Different sources of inflation give the Fed different room to cut if the bond market sells off.

Start with services and wages, the sticky part. July's Consumer Price Index (opens in a new tab) had services excluding energy up 3.0% from a year earlier and shelter up 3.2%. Average hourly earnings in the August jobs report were up 3.1%. If labor costs and rents keep running there, one cool month of oil will not let the Fed declare victory.

Energy is already pushing up the headline inflation number even though it is excluded from the core measure the Fed prefers. In July, the CPI showed energy prices up 14.7% from a year earlier and gasoline up 24.6%. Core inflation strips those categories out, but a buyer of a 10-year Treasury still has to live with the broader inflation picture. Brent crude (opens in a new tab) was $96.02 a barrel on 1 September. Slovenia's central bank (opens in a new tab), using data through 31 August, said oil had risen 24.1% since the start of July after the U.S.-Iran ceasefire broke down. July's inflation data do not yet reflect that full move in oil prices.

Then goods that rate cuts can't manufacture. A Minneapolis Fed note on 28 August (opens in a new tab) estimated that tariffs were adding about 0.2-0.4% points to core inflation as of July. Even without the tariff effect, the staff analysis put core inflation about 1% point above the Fed's 2% target. Clothing prices in the same data had risen from 0.3% year over year in December 2025 to 3.5% in July. The note also identified a separate shock from AI hardware. Video and information-processing equipment prices were up 12.2% year over year, contributing about 0.4 percentage points to core inflation in the staff's estimates.

Higher interest rates can restrain inflation caused by excess demand. They are much less effective against oil shocks, tariffs, or shortages of semiconductors. The Fed can weaken demand, but it cannot produce more oil, repeal a tariff, or create more chips. That makes the composition of the 3.3% core inflation rate important. The more of it comes from supply-side shocks rather than demand, the less room the Fed has to cut rates if the bond market comes under pressure.

Washington needs the money

A 4.8% 10-year yield can rise for several reasons. Investors may expect more inflation, demand a higher return after inflation, or want more compensation for locking their money into a long-term bond. Those are different forces, and they matter for different reasons.

On 2 September, FRED's 10-year yield closed at 4.79%. The inflation-protected 10-year, TIPS (opens in a new tab), yielded 2.45%. The difference between the two, the 10-year breakeven rate (opens in a new tab), was 2.34%. In other words, investors were pricing roughly 2.3% average inflation over the next decade, well below the 3.3% increase in core prices over the past year. The 4.79% nominal yield therefore reflected both a 2.45% real yield and additional compensation for holding a 10-year bond with interest-rate risk.

Frances Cheung (opens in a new tab), who runs currency and bond strategy at OCBC, told Reuters that the rise in U.S. long-term yields was still being driven more by higher real yields than by a sharp rise in inflation expectations, even as those expectations had edged higher. Bond investors were asking for more compensation to hold long-dated debt through changes in interest rates.

Those two inflation numbers do different jobs. The 3.3% core reading matters to the Fed when it decides whether it can cut the overnight rate. The 2.34% breakeven tells you what bond investors are pricing for inflation over the next decade. Meanwhile, the real yield and the compensation investors demand for taking on long-term interest-rate risk help explain why long-term borrowing can remain expensive even if inflation expectations stay relatively contained.

Heavy government borrowing is one possible reason those long-term premiums are rising. Treasury Fiscal Data (opens in a new tab) puts total federal debt at just over $40.10 trillion. Economists usually focus on a smaller number, debt held by the public, which excludes money the federal government owes to itself. The Congressional Budget Office (opens in a new tab) puts that figure at 101% of GDP this year and 120% by 2036.

CBO's full-year projection for debt held by the public is 101% of GDP this year. The FRED quarterly series above sits a bit lower at the start of 2026, around 99%. Both point the same way: debt held by the public is roughly the size of annual GDP, and CBO expects that ratio to keep rising.

New borrowing is what matters at the margin. The same CBO outlook expects this year's deficit to reach $1.9 trillion, or 5.8% of GDP, rising to 6.7% by 2036, with rising net interest costs driving much of that increase. More borrowing means the market has to absorb more Treasury debt. If demand does not keep pace with that supply, investors can demand a higher yield to hold it.

A Fed staff paper from May 2026 (opens in a new tab), updated on 31 August, attempts to measure how much government debt itself pushes rates higher. It estimates that for every additional percentage point of debt relative to GDP that markets expect, the long-run interest rate that neither stimulates nor restrains the economy rises by roughly 0.01 to 0.02 percentage points. The term premium on a 10-year bond rises by another 0.02 to 0.03 percentage points.

The CBO's projected increase from 101% of GDP this year to 120% by 2036 is 19 percentage points. Applying the paper's estimates mechanically would imply roughly 0.2 to 0.4 percentage points on the long-run rate and another 0.4 to 0.6 percentage points on the 10-year term premium. That is an illustration of the paper's coefficients, not a forecast. Actual yields will depend on much more than the debt path. But the arithmetic shows how small effects per percentage point can become meaningful over a decade.

That is consistent with Waller's point that fiscal concerns have been pushing yields higher. Markets can reprice long-term bonds as soon as they expect heavier Treasury supply, without waiting for a ratings decision.

The buyer base changed too

Supply is only half the story. Who holds long-term bonds has changed too.

For much of the postwar period, a large share of private-sector retirement saving sat in traditional defined-benefit pensions. Employers promised a retirement benefit, while professional managers decided how much of the portfolio went into bonds, stocks, and other assets. Workers generally did not choose the allocation themselves. That system began to give way after Congress created Section 401(k) in the Revenue Act of 1978 (opens in a new tab), as retirement saving shifted toward defined-contribution plans that workers increasingly directed themselves.

The Labor Department's Form 5500 history makes the shift concrete. In 1975, private-sector defined-benefit plans had about 27.2 million active participants (opens in a new tab), compared with 11.2 million in defined-contribution plans. Defined-benefit plans remained ahead until 1984. By 2023, defined-contribution plans had 96.4 million active participants, versus 11.1 million in defined-benefit plans. Congressional Research Service (opens in a new tab) shows the same institutional shift in 2021: 65% of private-sector workers had access to a defined-contribution plan and 51% participated, compared with 15% access and 11% participation for defined-benefit plans.

Household balance sheets show the broader change in retirement wealth as well. Following the method in a St. Louis Fed FRED Blog note (opens in a new tab), pension entitlements (including IRA and 401(k) claims) and direct stock holdings (corporate equities plus mutual-fund shares) can each be expressed as a share of household net worth. In 1960, stocks were about 18.3% of net worth and pensions about 14.7%. By 1970 pensions had edged ahead, 18.5% against 17.1% for stocks. The gap widened through the inflation decade: in 1975 pensions were about 20.2% of net worth while stocks had fallen to 9.6%. Pensions stayed larger for decades, peaking near 27.8% in 2010. Stocks caught back up around 2018 (24.9% pensions vs 23.8% stocks) and then pulled ahead. By year-end 2025 the split was about 18.9% pensions and 32.6% stocks. A rising stock share of net worth is not proof that fewer dollars are going into bonds. Household net worth grew, stock prices rose, and retirement money shifted into accounts that lean toward equities. The chart is about composition of household wealth, not a count of Treasury buyers.

Inside 401(k)s, the allocation is tilted toward equities. The EBRI/ICI database (opens in a new tab) puts equities, including the equity share of balanced funds and company stock, at about 75% of participants' assets at year-end 2023, up from 68% at year-end 2007. Bond funds represented a much smaller share.

Treasuries and bonds are still available inside these accounts, and target-date funds gradually add more fixed income as workers approach retirement. But the allocation is now largely determined by workers and default fund designs rather than by pension managers matching assets to a promised liability.

That matters when the Treasury needs to place more long-term debt. The traditional private-sector pension system created natural demand for long-duration bonds because funds had long-term liabilities to match. The defined-contribution system does not create the same automatic bid. It leaves more of the allocation to households and the funds they choose or default into.

The OECD's 2026 Global Debt Report (opens in a new tab) makes the duration point explicit across markets: the shift from defined-benefit to defined-contribution schemes has been one of the forces structurally reducing institutional demand for long-term bonds. The shift can also move allocations away from government debt and toward higher-yielding corporate bonds. In the OECD's pension survey, debt securities remained near 47% of pension-fund assets, with central government bonds making up just under 60% of those debt holdings. That is still a large bond portfolio. What it does not recreate is the old private-sector defined-benefit machinery that bought long-duration bonds to match promised retirement payments. Banks still hold Treasuries in part because they count as high-quality liquid assets (opens in a new tab) under liquidity rules, while insurers still need fixed income to match their own liabilities. Those buyers are real. They are simply not interchangeable with a pension board managing a multi-decade liability.

Washington does not need every 401(k) dollar to move into Treasuries for this to matter. Nor does a smaller share of one type of investor mean total Treasury demand has collapsed. Banks, insurers, foreign buyers, mutual funds, and remaining pension funds still buy government debt, and the Treasury market itself is much larger than it was when defined-benefit plans dominated private retirement saving. The narrower point is that supply has risen while the composition of the buyer base has changed.

That matters at the margin. If the next buyer is less willing to pay today's price for a fixed stream of payments, the bond's price has to fall. Because those payments are fixed, a lower price means a higher yield. That is how the market clears. The question is whether the marginal buyer will absorb the additional Treasury supply at the same yield as before. If not, the adjustment comes through a higher cost of capital.

The AI problem is the price

Waller also pointed to competition for capital from AI infrastructure. After the late-April earnings season, Yahoo Finance (opens in a new tab) added up the 2026 capital spending plans of Microsoft, Amazon, Meta, and Alphabet and got close to $725 billion, using the high end of each company's guidance: Microsoft at $190 billion, Alphabet at $180 to $190 billion, Meta at $125 to $145 billion, and Amazon at around $200 billion. The money is going into physical infrastructure: data centers, the chips inside them, and the power needed to run them.

That is large enough to matter as another source of demand for capital alongside Treasury borrowing. Some of the spending is financed internally and some through borrowing, but the economic problem is the same: more capital has to be committed today in the hope of generating returns later.

Real technology and expensive capital have lived side by side before. The internet was real, and it remade commerce. That did not stop the Nasdaq (opens in a new tab) from rising 86% in 1999 alone, peaking at 5,048 on 10 March 2000, then falling about 77% by October 2002 as cash-strapped startups collapsed. Alongside the equity mania, telecom firms laid so much fiber that by September 2002 TeleGeography estimated only 2.7% of installed fiber nationwide was actually in use (opens in a new tab), with much of the rest sitting unused as "dark fiber." The network still mattered later. What broke first was the assumption that near-term cash flows would cover what had been paid to build it.

The 2026 question is what a 4.8% long-term interest rate does to an AI buildout this large. Higher rates increase the discount applied to far-off profits, and they raise the return each new data center needs to earn just to justify the spending. AI can be real and valuable while the financing behind it still becomes harder to make work.

How the buildout gets financed

This leaves two sides of the same market. Washington needs to issue more long-term debt. Companies building AI infrastructure need to raise and deploy more long-term capital. The Treasury market tells you what investors require to supply that capital. The AI financing market tells you what happens when that capital gets expensive, and when the buildout itself is increasingly leveraged against future contracts.

The $725 billion figure is the spending plan. A growing share of the AI buildout is being financed outside ordinary corporate bonds, through separate vehicles that borrow against data centers, leases, and contracted cash flows.

Meta shows the hyperscaler version at scale. In October 2025 it announced a joint venture (opens in a new tab) with funds managed by Blue Owl Capital to develop and own the Hyperion data-center campus in Richland Parish, Louisiana. Blue Owl funds own 80% of the venture. Meta keeps 20%. The parties committed to fund their shares of roughly $27 billion in development costs for buildings plus long-lived power, cooling, and connectivity. Meta leases all of the completed facilities back from the venture on an initial four-year term with extension options, and it provided a capped residual-value guarantee for the first 16 years of operations if a lease is not renewed or is terminated under certain conditions. Meta also said a portion of Blue Owl's capital would be funded by debt sold to PIMCO and other bond investors through a private offering. S&P Global Ratings (opens in a new tab) assigned a preliminary A+ rating to about $27.3 billion of senior secured debt at Beignet Investor LLC, the Blue Owl holding company formed to own that 80% stake. The big loan sits in the financing vehicle. Meta's exposure shows up through the lease, the guarantee, and its minority stake.

The AI cloud providers use a related structure, closer to the equipment. CoreWeave's second-quarter report to the SEC (opens in a new tab) describes financing facilities that can be drawn over time through separate companies set up to isolate the financed assets from the rest of the business. Those companies hold the servers and the related customer contracts. The loans are secured by the equipment and the payments those customers have agreed to make. By 30 June, CoreWeave had already drawn more than $12 billion across its earlier facilities, with additional borrowing capacity still available.

That structure turns future contracted revenue into borrowing capacity today. The loans work as long as customers keep paying, the machines keep earning, and the pledged assets keep enough value to carry the debt.

CoreWeave has been scaling that structure through the capital markets all year. In March it closed an $8.5 billion loan (opens in a new tab) backed by high-performance computing gear and an associated customer contract. Lenders can only seize those pledged assets rather than the whole company, and the loan was the first of its kind to receive investment-grade credit ratings. In May it followed with a $3.1 billion loan sold to a wide group of investors (opens in a new tab), opening the same kind of paper to a broader market. In August it closed another $2.6 billion facility (opens in a new tab) whose roughly five-year life sits beyond customer contracts that average about three years, which means lenders are betting those customers will renew.

The supplier side is getting closer too. NVIDIA invested $2 billion (opens in a new tab) in CoreWeave stock in January, while CoreWeave's platform runs on NVIDIA gear. The Financial Times Lex column (opens in a new tab) has described NVIDIA's wider pattern of equity stakes, credit guarantees, and related support for chip customers as a modern form of supplier financing: helping buyers afford more of the seller's own products. The live question is who holds the financing risk if the contracted computing demand fails to arrive.

A 4.8% long rate bites harder in that world. Higher long-term rates pressure valuations of far-off profits, and they raise the hurdle on every debt-financed data center that needs tomorrow's leases or contracts to service today's loans.

Cheap funding can disappear abroad

Japan has been a major source of cheap global funding. For years Japanese yields sat near zero, which gave Japanese investors a reason to buy higher-yielding bonds abroad. If those yields rise enough, that incentive fades. Japanese money can stay home, demand for foreign bonds can fall, and long yields elsewhere can rise even if the Fed does nothing.

The leveraged version is the yen carry trade: borrow cheaply in yen, buy a higher-yielding asset elsewhere, and hope the yen stays quiet. When Japanese borrowing costs rise or the yen strengthens, anyone who used that funding to hold Korean stocks, U.S. equities, credit, or other risk assets has a reason to sell. The mechanism is real. How large it is on any given week, and whether it timed any particular crash, is much harder to prove from the public sources in this piece.

That plumbing is no longer free. In June the Bank of Japan raised its short-term policy rate to 1% from 0.75% (opens in a new tab), the highest in about three decades.

The FRED monthly series only runs through June 2026, when the yield averaged 2.67%. The move continued after that. On 1 September, Reuters (opens in a new tab) had Japan's 10-year at 3% for the first time since 1996, and noted that higher Japanese yields may keep Japanese money at home.

Korea shows how the pressure travels

Korea is where expensive capital showed up in an equity market that was already a concentrated AI bet. When U.S. long-term yields rise, investors generally demand higher returns from riskier assets elsewhere. Foreign money has less reason to stay, and positions built on cheap funding get harder to carry.

South Korea's main stock index, the KOSPI, closed September 2025 at 3,424.60. By June 2026 the monthly close was 8,476.48, more than double in nine months, per Yahoo Finance (opens in a new tab). Reuters (opens in a new tab) put the June peak at 9,114.55. Two stocks did most of the lifting, Samsung Electronics and SK Hynix, both of which make the memory chips that go into AI servers. Reuters put the pair at more than half of the KOSPI's weighting.

Leverage turned that concentration into a household-finance accident. In late May, after regulators cleared the products to keep retail money at home, Korea listed single-stock leveraged ETFs (opens in a new tab) designed to deliver twice the daily move in names including Samsung and SK Hynix. CNBC (opens in a new tab) reported that by late July Korean retail investors had put a net 14 trillion won, about $9.7 billion, into those products. The Korea Financial Investment Association (opens in a new tab), cited by Reuters, put domestic margin loan balances at a record 38.63 trillion won on 24 June.

When the chip stocks broke, the leveraged products broke faster. CNBC, using LSEG data, said the KODEX SK Hynix single-stock leverage ETF fell more than 80% from its 23 June peak, and the Samsung equivalent nearly 75% from its 3 June peak, while the KOSPI itself was down almost 35% over the prior month. Citibank (opens in a new tab), as reported by Chosun Daily, estimated Korean individuals lost about $38.7 billion on leveraged products in the correction. Regulators halted (opens in a new tab) new listings of those products in mid-July. On 28 July, Yonhap (opens in a new tab) had the KOSPI down 10.84% to 6,023.66, with foreigners selling a net 4.97 trillion won while Korean retail bought 4.33 trillion.

The same channel showed up again this week. On 2 September, New York hit 4.818% on the 10-year and the S&P 500 closed up 0.46% (opens in a new tab), while Chosun Daily (opens in a new tab) had the KOSPI down 3.99% to 6,562.72. The sources in this piece stop short of showing that yen funding caused the July crash. They do show a concentrated AI equity market moving when U.S. long rates jump, with leverage sitting on top of the same two chip names.

The U.S. already owns a concentrated AI bet

The Magnificent 7, Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla, accounted for roughly one-third of the S&P 500's market cap (opens in a new tab) at the end of 2024 and still held that share through 2025, per Fidelity. Motley Fool (opens in a new tab), using August 2026 data, put them at about 33.9% of the index. That is less extreme than Korea's two-stock majority, and still enough that a broad U.S. index is no longer an average of 500 names.

Americans mostly meet that weight through retirement plumbing rather than twice-levered single-stock bets. ICI (opens in a new tab) put 401(k) assets at $9.9 trillion at the end of March 2026. Anyone in a market-cap S&P 500 or total-market fund already owns a large slice of those seven names. FINRA (opens in a new tab) still put U.S. margin debit balances near $1.42 trillion in July 2026, and MacroMicro (opens in a new tab) estimates leveraged ETFs at about 7.1% of Korea's ETF net assets against 1.2% in the United States. Korea put 2x daily products on the two stocks that already made up more than half the benchmark. A U.S. shock would be less dramatic on the surface, but it could spread through millions of accounts that thought they simply owned "the market."

What has not broken

In the United States, companies with weak credit ratings pay more to borrow than the government does. That gap is the junk-bond spread, the market's price for the risk those companies default. On 3 September the ICE BofA U.S. High Yield spread (opens in a new tab) was 2.65 percentage points, tight enough that investors are barely charging extra for risk.

In August 2007, the Fed's own briefing materials (opens in a new tab) already noted that these spreads had started widening in June, months before stocks fell apart in 2008. Credit tends to move first. This week U.S. junk bonds did not confirm the Treasury selloff, which is the best argument that the story is still the price of money. Korea's July crash already showed that a concentrated equity market can break while U.S. junk spreads stay calm. The 2.65% spread is the number I'd want to see move before calling this a broader American corporate-funding problem.

The place that argument can't be settled in public is private credit: loans made by investment funds directly to companies rather than through banks or public bond markets. The Financial Stability Board (opens in a new tab), the international body that watches for system-wide risk, put the sector at $1.5 trillion to $2 trillion in May 2026 and said it has never been tested in a severe downturn at this size. Public junk bonds are calm. The loans nobody has to price every day are the ones that can't prove the calm is real.

What to watch

The 10-year Treasury yield, inflation, and the junk-bond spread tell you whether this is still a story about expensive capital or something worse.

If the 10-year stays around 4.8%, inflation slowly comes down, and credit spreads stay tight, the system is absorbing the cost: companies can still refinance, AI projects can still earn enough to justify the build, and Treasury auctions still clear without a jump in the required yield.

The concerning mix is long-term yields stuck near 4.8% or higher, inflation that will not return to 2%, and the junk-bond spread moving away from 2.65%. Pressure would then be showing up in private borrowing costs, not only in government debt. Weaker firms would find it harder to roll their loans, and highly financed AI infrastructure would find it harder to roll its own.

Japan is a separate pressure point. If Japanese yields keep rising, sending yen overseas gets less attractive, and global funding can tighten without another Fed hike. Watch whether that shows up first in foreign demand for bonds, in risk assets, or in the financing behind AI infrastructure. The size and timing of that channel remain hard to prove from the sources in this piece.

The more useful early signal may arrive before the broad junk market moves. A CoreWeave-style facility can get harder to refinance, a data-center lease can be worth less, a customer can delay a renewal, and a lender can demand more protection against the same pledged gear. None of that needs a recession or a bank failure. It only needs the price of capital to be too high for some of the buildout.

There is enough money to fund Washington and AI. The open question is what return that money will demand to keep doing it.

For now, the bond market is asking for more. The rest of the financial system has not yet decided how much more it is willing to pay.

References

Categories

Tags