AI bonds hit $200 billion — it was not the government crowding out markets
US long-term Treasury yields reached a 19-year high in August 2026, and government deficits are only part of the reason. Large technology companies borrowed roughly $200 billion this year to build AI data centres — about a quarter of net Treasury issuance by Nomura's estimate — inverting the textbook crowding-out effect so that corporations are crowding out the government
The three lines
- Scale — big tech borrowed roughly $200bn in 2026, near 25% of net US Treasury issuance
- Price — Alphabet issued 30-year paper at 6.4%; Meta's data centre bonds cleared above 7.5%
- Effect — corporate and mortgage supply added about 0.3pp to the 10-year yield (BofA)
Key questions
- What does an inverted crowding-out effect mean?
- The textbook version says that when a government borrows heavily it absorbs available capital, leaving less for private borrowers and pushing rates up — the government crowds out companies. What is observed in the United States in 2026 runs the other way. Highly rated AI companies are issuing bonds on terms attractive enough that investors sell Treasuries to buy them. Companies are crowding out the government.
- How much have AI companies actually borrowed?
- The answer depends on what is being counted. Large technology company borrowing in 2026 is estimated at roughly $200 billion (Nomura). Taking Amazon, Meta, Nvidia, Oracle and Alphabet together, corporate bond funding rose from $39.7 billion in 2024 to $144.7 billion in 2025 and $218.3 billion so far in 2026. Counting all AI-linked new bonds, issuance reached $270 billion by early July 2026 — close to double the full-year 2025 figure.
- Does this affect borrowing costs outside the US?
- Indirectly. US long-term Treasury yields function as a reference rate for long-term borrowing worldwide, so a rising 30-year yield tends to pull other sovereign long bonds and any product priced off them, including fixed-rate mortgages, in the same direction with a lag. The transmission is not one-for-one — domestic policy rates, bank funding costs and credit spreads all intervene. This page did not confirm the size of that pass-through.
Economics textbooks contain an entry called crowding out. A government borrows heavily, capital is drawn into government paper, competition for what remains pushes rates up, and the party pushed aside is private business.
In the United States in August 2026, that arrow is pointing the other way.
1. The numbers underneath the headline
The US 30-year Treasury yield reached 5.31% on August 17, its highest since 2007. The 10-year stood at 4.724%.
The standard explanation pairs "fiscal concerns" with "large corporate bond issuance by AI companies." Opening the second half gives this:
| Measure | Amount |
|---|---|
| US investment-grade corporate issuance, 2026 | ~$1.5 trillion (+36% year on year) |
| Large technology company borrowing (Nomura estimate) | ~$200 billion |
| Above as a share of net Treasury issuance | ~25% |
| Increase in net corporate bond supply (Barclays) | $474 billion |
| AI-linked new bonds through early July 2026 | $270 billion |
| Reference — full-year 2025 AI-linked bonds | roughly half the above |
The last two lines carry the speed. Seven months of issuance approached double the preceding twelve.
Narrowing to five companies — Amazon, Meta, Nvidia, Oracle and Alphabet — corporate bond funding went from $39.7 billion in 2024 to $144.7 billion in 2025 to $218.3 billion so far in 2026. A 5.5-fold increase in two years.
2. Where the arrow flips
What matters is not the volume but who is now offering the higher price.
| Issuer | 30-year cost of funds |
|---|---|
| US government (30-year Treasury) | 5.31% |
| Alphabet (30-year corporate) | 6.4% |
| Meta (data centre bonds) | above 7.5% |
Alphabet is paying 1.15 percentage points over the government. When an issuer of near-top credit quality offers that premium, the arithmetic changes for a pension fund or insurer with 30-year liabilities to match. There is now a reason to sell Treasuries and move.
Selling Treasuries pushes their price down, and a falling bond price is a rising yield — the mechanism this page set out on August 18 in "What a 30-year government bond is."
So corporations crowd out the government. Crowding out, inverted.
3. By how much
Estimates differ in size but not in direction.
| Institution | Estimate |
|---|---|
| Bank of America | Corporate + mortgage bond supply added ~0.3pp to the 10-year yield in 2026 |
| Nomura | Big tech borrowing equals ~25% of net Treasury issuance |
| Dallas Fed | $300bn of AI-linked issuance implies up to $360bn of new duration supply — about one-eighth of Treasury supply |
| Goldman Sachs | Five hyperscalers: $250bn this year, projected $400bn next year |
The Dallas Fed's phrase "duration supply" refers not to the cash raised but to the total interest-rate risk the market must absorb. Longer maturities impose more of it, so $300 billion of issuance converts into $360 billion of burden.
Two cautions. Bank of America's 0.3 percentage points is a combined figure for corporate and mortgage supply; the corporate-only share is not published. And no official decomposition separates the fiscal-deficit contribution from the AI-borrowing contribution. What the evidence supports is "AI borrowing accounts for a substantial share" — no more precise than that.
4. Why the loop matters
Higher yields come back around to the borrowers.
- Higher rates raise the cost of the next issue. Meta clearing above 7.5% is that signal.
- Higher rates raise the discount rate on equities. Companies whose profits sit furthest out fall hardest — the US semiconductor gauge dropped 5.5% on August 18.
- A depressed share price makes equity funding harder, which sends the company back to the bond market.
Debt raised for the buildout simultaneously raises the cost of the next round and depresses the equity that would otherwise substitute for it. This is the structure behind the recurring question of whether AI infrastructure spending is peaking faster than expected.
5. What remains open
- Decomposition. No settled figure divides the yield rise between deficits and AI borrowing.
- Counting bases. The $200 billion (Nomura), $218.3 billion (five companies) and $270 billion (all AI-linked bonds) in this article count different things. Adding or comparing them directly produces error.
- Cross-border transmission. How much of the US 30-year move reaches Korean long-term rates and fixed-rate mortgages was not established.
The instrument itself is explained in "What a corporate bond is," and the equity mechanism in "Why rising bond yields push stocks down."
Sources
- Newspim — Big tech's AI borrowing snags US long-term rates: the inverted crowding-out effect
- Financial News — Governments and AI firms both borrowing: US long rates at a 19-year high
- Global Economic — Meta's AI data centre borrowing cost jumps to 7.5%
- Financial News — AI rush drives a corporate bond boom as the line around 'safe' bonds erodes
- Newspim — 30-year US Treasury yield jumps on AI borrowing and fiscal concerns; dollar weakens