Data current through August 7, 2026.
The AI spending boom has an interest-rate problem.
The biggest technology companies are pouring hundreds of billions of dollars into data centers, chips, power, cooling, and network capacity. At the same time, the U.S. government is borrowing heavily to finance deficits and refinance existing debt.
Those two facts are often turned into a simple story: Washington and Big Tech are fighting over the same pile of money.
That is too simple.
The real connection runs through the price of capital. U.S. Treasury yields help set the starting interest rate for much of the financial system. If those yields stay high, AI projects have to clear a higher bar. If investors also become more worried about AI debt, the bar rises again.
The AI boom does not need a bond-market crash to run into trouble. It only needs capital to stay expensive long enough for the weaker projects to stop making financial sense.
First, what does “crowding out” actually mean?
When the federal government spends more than it collects in taxes, it has to borrow the difference. It does that by selling Treasury bills, notes, and bonds.
In August 2026, the Treasury projected $739 billion of privately held net marketable borrowing for the July-through-September quarter. That is new net borrowing, not the same thing as saying every Treasury security that matures represents fresh demand for cash.
That distinction matters.
A Treasury bill may mature and be replaced by another Treasury bill. That is a rollover. The government is refinancing existing debt. Net borrowing measures how much additional marketable debt the government expects private investors to absorb after those movements are accounted for.
“Crowding out” is the idea that heavy government borrowing can make financing harder or more expensive for private borrowers.
But there is no fixed bucket of dollars sitting in a room.
The capital pool can expand or shrink. Foreign investors can buy Treasuries. Banks can change their balance sheets. Pension funds can shift portfolios. The Federal Reserve can tighten or loosen financial conditions. Higher government spending can also support economic growth, which may encourage private investment.
So crowding out is better understood as pressure, not a one-for-one swap.
Research cited in the project material finds that larger government debt flows can displace some corporate borrowing through financial intermediaries. But the size of the effect depends on demand for Treasuries, Fed policy, foreign capital, inflation expectations, and investor appetite.
For AI, the question is simple: does government borrowing help keep the general cost of money high at the same time the AI buildout needs more outside financing?
Treasury yields set the starting price
This is where bonds connect directly to AI.
Treasury yields are often treated as the risk-free benchmark in dollar markets. That does not mean U.S. government debt has literally zero risk. It means investors use Treasury rates as the baseline price for lending money, then demand extra return for taking additional risk elsewhere.
A company does not usually borrow at the same rate as the U.S. government. Investors demand extra interest because a company or project has a greater chance of running into trouble.
That extra interest is called a credit spread.
A credit spread is simply the additional return investors demand for lending to a company instead of lending to the U.S. Treasury for a similar period.
Think of it like this:
Treasury yield + credit spread = approximate corporate borrowing cost
Suppose a Treasury bond yields 4.6%.
Now suppose investors want another 2.25 percentage points to lend to a data-center project because that project carries more risk.
The borrowing cost is roughly:
4.6% + 2.25% = 6.85%
That extra 2.25 percentage points is the credit spread.
The spread can change even when Treasury yields do not. If investors become nervous about the borrower, the spread widens. If they feel safer, it narrows.
This means AI financing can get hit from two directions at the same time.
Treasury yields can rise because inflation stays high, government borrowing remains heavy, or investors demand more compensation for holding long-term bonds.
Then credit spreads can rise because investors become less confident that a data center, AI lab, or heavily leveraged borrower will produce enough cash to repay its debt.
A project that once borrowed at 5% can suddenly face 7% or 8% financing without anything physically changing about the building.
That can completely change the economics.
Meta’s Hyperion data-center financing gives us a real example of this structure. Meta and Blue Owl created a separate vehicle to help finance a project with development costs around $27 billion. The debt was tied to long-term lease commitments from Meta rather than being funded entirely as ordinary Meta corporate debt.
This is one reason AI financing is becoming harder to understand from a quick look at Big Tech balance sheets. Some of the borrowing sits in bonds, some in leases, some in project vehicles, and some in private markets.
Big Tech still has cash. AI is consuming much more of it.
The phrase “Big Tech has become debt dependent” goes too far.
Microsoft, Amazon, Alphabet, and Meta still run huge businesses that generate large amounts of operating cash.
But the direction has changed.
AI infrastructure requires physical investment on a scale that software companies were not built around. (See also: who will pay for the AI buildout and why AI data centers may be riskier than they look.) Data centers need land, transformers, power contracts, cooling systems, networking equipment, backup generation, and enormous amounts of chips.
The first research pass found capex rising very quickly across the major hyperscalers. Alphabet reported $44.9 billion of property and equipment spending in Q2 2026. Meta reported roughly $31 billion of capex and finance-lease principal payments. Amazon’s spending was also running at an extremely high level.
The important point is not that these companies are suddenly broke.
It is that AI is eating a much larger share of the cash their core businesses produce.
That pushes the next dollar of expansion toward other financing channels:
- Corporate bonds
- Long-term leases
- Special-purpose vehicles
- Private credit
- Joint ventures
- Strategic investors
Morgan Stanley estimated that AI-related issuers had sold about $236 billion of debt globally through May 2026 and projected the full-year total could approach $570 billion. That estimate includes hyperscalers, chip companies, and data-center developers. It is a forecast, not a final audited number.
The bond market has absorbed that issuance so far.
That matters. We are seeing financing pressure, not a financing shutdown.
Higher rates hit AI twice
The first hit is easy to understand: debt gets more expensive.
The second hit happens in the stock market.
Investors value a company partly by estimating the cash it may produce in the future and then asking what those future dollars are worth today.
That calculation uses a discount rate. You do not need a finance degree to understand the basic idea.
Imagine someone promises to give you $100 ten years from now.
If your required return is 5%, that future $100 is worth about $61 today.
If your required return rises to 10%, the same future $100 is worth only about $39 today.
The future payment did not change.
The price you are willing to pay for it today did.
This matters most for companies whose valuations depend heavily on profits expected years from now. That is why rising rates can hit expensive growth stocks harder than mature businesses that already produce steady cash.
We saw a version of this in 2022. The Fed raised rates rapidly — a reminder that rate hikes and quantitative tightening work differently, Treasury yields moved higher, and technology valuations fell hard. Slower post-pandemic growth and changing investor sentiment also mattered, so rates were not the only cause. But the basic discount-rate mechanism worked exactly as finance theory says it should.
AI now faces both sides of that pressure.
Higher yields can make the infrastructure more expensive to build.
And higher discount rates can reduce what investors are willing to pay for the future profits the infrastructure is supposed to create.
The weak links matter more than the biggest tech companies
A lot of AI-bubble arguments focus on Microsoft, Meta, Amazon, or Alphabet.
Those may not be the first places to look for a break.
The weaker points are likely to sit farther out in the financing chain.
Think about a data-center developer with high debt, a private AI company burning cash, or a project that depends on one or two large customers.
Those borrowers do not have Microsoft’s balance sheet.
Private credit matters here.
Private credit is lending done outside the traditional public bond market and ordinary bank-loan market. Large investment firms raise money from investors and lend it directly to companies.
The Bank for International Settlements reported that private-credit lending to software companies had grown above $500 billion by the end of 2025. Business development companies alone had around $115 billion of software exposure.
That is real exposure.
But the BIS did not say a systemic AI credit crisis had already started. In fact, it noted that low leverage and secured lending may limit some spillovers. That is an important check on the bearish story.
So the risk is not “private credit is already broken.”
The risk is that a weaker revenue outlook could hit a financing structure that was built when money was cheaper and investors were more forgiving.
The circular financing problem is real, but it needs careful language
AI financing also contains relationships where the investor, supplier, and customer can sit inside the same small network.
Amazon and Anthropic are a good example.
Amazon has committed large amounts of capital to Anthropic. Anthropic, in turn, has committed to spending heavily on AWS infrastructure.
That does not prove fake demand or bad accounting.
But it does make the flow of money harder to interpret.
If a cloud company invests in an AI lab and that lab uses part of the capital to buy cloud services from the same company, some reported demand is connected to the financing relationship itself.
The same question appears around chip suppliers, AI labs, data-center developers, and strategic investors.
The question is not, “Is the revenue fake?”
The better question is, “How much demand would still exist if the financing loop became harder to fund?”
That is the number investors should care about.
Telecom already gave us a warning
The late-1990s telecom and fiber boom is a better comparison than saying AI is simply another dot-com bubble.
Telecom companies spent heavily building fiber networks because internet traffic was clearly going to grow.
They were right about the technology.
They were wrong about the timing, financing, and amount of capacity the market could absorb immediately.
Telecom capital spending reached enormous levels around 2000. Then spending collapsed. Defaults surged. Companies failed. Huge amounts of fiber sat unused for years.
But much of that fiber later became useful infrastructure for broadband, streaming, cloud computing, and the modern internet.
That is the lesson.
A technology can be economically important and still produce terrible investment returns for the people who finance too much capacity too early.
AI could follow a similar path.
The data centers may be useful.
The chips may be useful.
The power infrastructure may be needed for decades.
None of that guarantees every project deserves to be financed at today’s price.
But this is not 2000
There is also a major problem with pushing the telecom comparison too far.
Many of the biggest AI spenders today are profitable, investment-grade companies with established customers and enormous cash flows.
The current hyperscalers are very different from the highly leveraged or unprofitable telecom companies that failed during the earlier buildout.
And the debt market is still open.
Large AI-related bond deals have cleared. Meta completed its Hyperion financing. AI-linked debt issuance has accelerated sharply rather than freezing.
That tells us investors still believe they are being paid enough to supply capital.
The AI boom can survive high rates if revenue grows fast enough to cover the higher cost of money.
It can also survive if hyperscalers slow spending before their balance sheets become stressed, if compute demand remains strong, or if falling chip and inference costs improve the economics.
Higher rates create pressure.
They do not guarantee failure.
What would tell us the cycle is actually breaking?
This is the part I would watch instead of trying to predict a crash date.
1. AI credit spreads begin widening.
If Treasury yields stay high and investors start demanding much larger spreads from AI borrowers, financing costs rise from both directions.
2. Major debt deals struggle.
A large hyperscaler, data-center developer, or project vehicle has to pay far more than expected, shrink an offering, delay it, or cancel it.
3. Hyperscalers cut capex guidance.
A real reduction in planned data-center spending would matter more than another bearish analyst interview.
4. Data-center projects get delayed or canceled.
That would show financing pressure moving into the physical economy.
5. Private-credit losses rise.
Watch non-accruals, restructurings, payment-in-kind interest, and loan markdowns in funds with meaningful technology or data-center exposure.
6. AI customers reduce compute commitments.
This would be especially important. It would show the problem moving from financing into actual demand.
7. Earnings forecasts come down.
If analysts reduce future profit estimates at the same time discount rates remain high, technology valuations get squeezed from both sides.
That sequence would be much stronger evidence than Treasury yields crossing one dramatic number.
The Macro Board Watch view
AI capital spending is a contested force.
During the buildout, it can be inflationary. It creates demand for power, construction, chips, workers, land, and financing.
But if the financing cycle breaks, the effect can flip.
Capex gets cut. Projects stop. Credit tightens. Technology stocks fall. (For the mechanics of a disorderly selloff, see what actually causes a stock market crash.) Suppliers lose orders. Investors become more cautious.
That turns an inflationary investment boom into a deflationary financial shock.
The bond market sits near the center of that switch.
For now, the evidence says AI financing is getting more expensive and more complex, but capital is still available. The signal changes when spreads widen, deals start failing, capex gets cut, and demand weakens at the same time.
That is the signal to watch.