The next AI bailout could break the bond market

The next AI crisis may not start with an AI company going bankrupt.

It may start when investors realize how much financing has been built around infrastructure that has not produced the cash to pay for itself yet.

The numbers are getting ridiculous.

Microsoft, Amazon, Alphabet and Meta are expected to spend roughly $630 billion on AI infrastructure and data centers in 2026 alone. Include the wider group of cloud and infrastructure providers and the figure rises to about $811 billion.

And that is just one year.

Oxford Economics estimates cumulative AI investment could reach roughly $3.8 trillion from 2024 through 2028.

Someone has to finance that.

And increasingly, it isn’t all coming from cash flow.

Oracle raised $43 billion of debt in fiscal 2026 and expects roughly another $40 billion of debt and equity financing in fiscal 2027. As of August 31, Oracle had $125.3 billion of borrowings and another $288 billion of mostly data-center leases that had not yet started.

Then there is Meta.

Its enormous Hyperion data center project in Louisiana has absorbed more than $50 billion of investment.

About $27 billion of bonds were issued through a separate financing vehicle.

Meta owns only 20% of the project, leaving much of the debt off Meta’s own balance sheet.

That structure is not unique.

Morgan Stanley estimates roughly $3 trillion of financing and leasing structures tied to AI infrastructure are sitting off the balance sheets of Nvidia, Broadcom and the hyperscalers.

This is where the story gets strange.

The AI companies don’t necessarily have to carry all the debt themselves.

The debt can sit inside special-purpose vehicles.

The data center can sit somewhere else.

The equipment can be financed by somebody else.

The customer signs a long-term computing contract.

The bank finances the infrastructure.

The chip company sells the GPUs.

Everybody gets to call something an asset.

Until somebody wants the cash.

Oracle’s Project Jupiter is already showing what happens when the financing gets uncomfortable.

About $18 billion of loans tied to the New Mexico data center were recently quoted at 89 to 91 cents on the dollar. Banks have reportedly struggled to distribute the debt to other investors.

Then Oracle issued a force-majeure notice because of uncertainty surrounding the project’s power supply and construction schedule.

Oracle says the project remains on track.

Fine.

But the debt is already trading at a discount.

That is the market putting a price on the risk before anyone has declared the project a failure.

And there is another problem hiding inside the AI buildout.

The machines don’t last forever.

AI infrastructure has to be continually upgraded as new chips become available. The faster Nvidia and its competitors improve performance, the faster yesterday’s equipment can become economically obsolete.

So the industry has a strange capital cycle.

Borrow billions.

Build the data center.

Buy the GPUs.

Wait for customers to generate enough revenue to cover the investment.

Then start spending again on the next generation before the first generation has necessarily produced its full return.

That means the AI boom isn’t just one enormous investment.

It could become a series of enormous refinancing events.

And that brings us to the Treasury market.

The U.S. government itself needs staggering amounts of refinancing.

The Government Accountability Office says Treasury needed to refinance $9.7 trillion of maturing securities in fiscal 2026.

Interest expense on debt held by the public was already about $1 trillion in fiscal 2025.

So now you have two gigantic borrowers competing for capital.

Washington needs buyers for Treasury securities.

The AI industry needs buyers for corporate bonds, project debt, private credit and infrastructure financing.

When AI debt starts looking riskier, investors demand higher returns.

When Treasury yields rise, the AI industry’s cost of capital rises too.

And when AI financing gets more expensive, some of those enormous future projects stop looking as attractive.

That is the feedback loop.

AI infrastructure requires financing.

Higher yields make financing more expensive.

More expensive financing reduces the return on new data centers.

Lower returns make lenders more cautious.

More cautious lenders demand even higher yields.

And suddenly the industry’s biggest problem isn’t whether people want AI.

It is whether somebody is willing to finance the next $4 billion data center.

This is why a government rescue would be so dangerous.

Not because a bailout is currently happening.

There is no evidence of that.

The problem is what happens if the industry becomes too financially interconnected to unwind cleanly.

If banks are exposed.

If pension funds own the bonds.

If infrastructure funds own the projects.

If cloud companies have long-term commitments.

If utilities have built generation for the new demand.

If chipmakers have expanded production around the assumption that the buildout continues.

At that point, allowing everything to clear through bankruptcy could create losses across multiple layers simultaneously.

The political pressure to refinance, guarantee or support the system would become enormous.

And any rescue would eventually run into the Treasury market.

The government would be trying to stabilize a private-sector debt problem while already refinancing trillions of dollars of its own debt.

That’s the part that makes this different from a normal tech bubble.

A normal bubble can deflate.

The capital disappears.

The bad projects die.

The survivors buy the assets cheaply.

But what happens when the bubble is built on infrastructure that the financial system has already borrowed against?

Then somebody owns the debt even after the dream disappears.

That’s why I don’t think oil is necessarily the thing to watch.

Oil can spike and fall.

The AI infrastructure cannot be switched off that easily.

The buildings remain.

The power contracts remain.

The leases remain.

The bonds remain.

The debt remains.

And if the next generation of AI spending requires another trillion dollars while the bond market is demanding higher yields, the industry eventually has to prove something Wall Street has mostly taken on faith:

That the cash arriving later will be large enough to justify all the money borrowed today.

If that proof fails, the problem won’t stay inside the technology sector.

It will move into credit.

And if credit gets pushed onto the government balance sheet, it can move again.

Straight into the Treasury market.

Not financial advice.