AI is crashing US Treasuries — and even if it succeeds, it will be hit with heavy taxes

On September 24, US long-dated bonds collapsed across the board. The 30-year Treasury yield rose to 5.48%, a 20-year high; the 10-year broke above 5.2%, the highest since the financial crisis. Long-end yields have risen almost in a straight line over the past two months. Treasury Secretary Bessent twice stepped up buyback interventions, each time buying only one day of respite before the market knocked prices back down the next day.

Too-large deficits, too-high oil prices, the Fed raising rates again — these are all real underlying forces, but none of them are new variables in this round. What truly makes this round different is that a competitor has emerged on the demand side that doesn't care about price at all.

AI is crowding the Treasury out of the market

To rush-build data centers and computing power, tech giants are issuing ultra-long-dated bonds at a record pace, and almost without regard to price.

As of August, the five hyperscale cloud giants had issued nearly $230 billion in bonds this year — more than double all of last year — with a large share concentrated in 20-, 30-, and even 40-year maturities.

These deals would be front-page news in any year.

Meta is only 22 years old, yet last October it issued a bond locked in for 40 years — by the time it matures, Zuckerberg will be 80. That single deal was $30 billion, setting a record for the largest single corporate bond issuance. This April, Meta added another $25 billion.

Alphabet even issued 100-year sterling bonds. Amazon raised $54 billion in a single deal in March. Oracle is now rated only two notches above junk, with negative free cash flow, yet it keeps issuing bond after bond.

Why don't they care about interest rates? Bessent himself marveled in an August interview that these companies borrow without even looking at rates, because they are convinced AI's returns will be so high that today's interest costs will be trivial.

In other words, the brake that would normally automatically curb the impulse to borrow when rates rise has failed for this AI demand.

And the money willing to lock up "long money" is only so much: pension funds, insurers, long-duration bond funds — the long end rests on these buyers, and their money is finite. Money that flows into tech companies' long bonds cannot flow into Treasuries.

Alphabet's 30-year corporate bond yield sits at 6.4%. If Treasuries want to win that money back, they can only push their own yields up.

This is effectively a "reverse crowding out": in the past, government bond issuance crowded out private investment; now it's the AI giants' turn to crowd the government out of the bond market.

Nomura estimates that tech giants alone have borrowed roughly $200 billion in long-term funds, a quarter of the Treasury's full-year issuance of medium- and long-term debt. Investment-grade bond funds have also been trimming Treasuries and adding corporate bonds this year, and foreign private money buying US corporate bonds has now exceeded its purchases of US Treasuries — something that was previously hard to imagine.

A massive, price-insensitive private borrower permanently stationed at the long end is itself a rewriting of the rules of the game.

Bessent's buybacks can't even reach here — each is only a few billion dollars, a drop in the bucket in a Treasury market of over $30 trillion, and it can only adjust the maturity structure of Treasuries themselves, not the pace of tech giants' bond issuance.

There is also a self-reinforcing loop hidden here: tech giants' bond issuance pushes up long-end rates, and higher rates in turn raise the funding costs of AI projects themselves — the extra premium AI giants pay when issuing new bonds has surged from just over 2 basis points last year to 12 basis points.

AI will only make the US government poorer

If it were merely competing with the Treasury for money, the problem would still be solvable, because if AI really succeeds in making the economy grow rapidly, tax revenue will eventually catch up, the government's debt burden will ease, and yields will naturally fall.

But the problem is that the AI boom likely will not expand the government's tax base. Productivity and GDP may surge, yet the government's ability to service debt could actually deteriorate.

This deterioration shows up in two ways: the long term and the near term.

In the long term, US individual income taxes and payroll taxes together account for more than three-quarters of federal revenue — meaning the government's fiscal lifeline is tied to the payroll — while AI's economic effect is precisely to compress the value and income of labor. Companies do more with fewer people; profits rise, but total wages fall, and the government collects less and less from income and payroll taxes. If AI were to reduce labor's share of GDP by 10 percentage points, the finances of most developed countries would be in serious trouble.

In the near term, the "Big and Beautiful Act" passed in 2025 has sharply cut US corporate taxes. The act restored the policy of full expensing of capital expenditures in the year they are made. The intent was to stimulate investment, but for tech giants that were already going to pour money into data centers at any cost, this is not an incentive but a windfall. Microsoft disclosed in July that its current tax payable was only $2.5 billion, compared with $14.1 billion a year earlier — an 80% drop — while profits were surging. That is the magic of depreciation. The Congressional Budget Office's September data confirms this: in the first 11 months of this fiscal year, corporate income tax revenue fell by a quarter year over year, a decline of nearly $100 billion.

Revenue shrinking, spending swelling, tax revenue falling off a cliff. The deficit can only continue to be financed by issuing debt. In the first 11 months of this fiscal year, the federal deficit has already piled up to $2 trillion, total national debt just crossed the $40 trillion mark in August, and interest alone now eats up about $1 trillion a year.

And under the current system, expecting AI to ultimately save the day by expanding the existing tax base is simply unrealistic.

The more successful AI is, the more it cannot escape heavy taxes

The more successful AI is, the more thoroughly it replaces labor, and the less tax the US government collects from the payroll.

At the same time, government spending will not shrink as a result — it will actually swell — because more people will need retraining and welfare to sustain their lives.

As the government's fiscal gap widens, money can only be found in the one place that is still expanding — the profits AI creates.