Artificial intelligence requires an extraordinary amount of physical infrastructure.

And all of it costs money.

Estimates cited by the Federal Reserve Bank of Dallas put the required investment in AI data centers at somewhere between $3 trillion and $5 trillion over the next several years.

Until recently, much of that spending could be financed from the enormous profits generated by America’s largest technology companies.

But the scale is becoming too large for that alone.

So the AI boom is increasingly coming to the bond market.

And that creates an unusual problem.

The United States government, which is already dealing with pressure at the long end of the bond market and has recently expanded its long-term Treasury buyback program, needs that same money.

The Treasury must continually sell enormous quantities of bonds to finance federal deficits and refinance existing debt.

AI companies are now competing for some of the same long-term capital.

The Dallas Fed estimates AI-related investment-grade bond issuance could create as much as $360 billion of new 10-year-equivalent duration this year.

That would equal roughly one-eighth of the duration being supplied by the United States Treasury.

There are only so many investors willing to lock up money for decades.

And insurance companies are especially important in that equation.

An insurance company may know that it will have liabilities stretching decades into the future. So it buys long-term assets that can help match those obligations.

Treasury bonds have traditionally been particularly useful for that purpose.

But now imagine an insurer looking at two choices.

It can lend money to the United States government for 30 years.

Or it can buy a corporate bond paying a higher yield from a major technology company financing the AI boom.

Suddenly, the Treasury has more competition.

And that could push long-term borrowing costs higher.

But there is another side of this story.

What happens if all these AI investments don’t make enough money?

That question may ultimately matter even more.

AI, like many other important emerging technologies in the past, may be an extraordinary technology while still producing disappointing investment returns.

Railroads, airlines, and the internet transformed the world.

But none of them prevented investors from losing enormous amounts of money financing companies that ultimately failed.

AI could present a similar problem.

Companies are spending extraordinary sums building the infrastructure necessary to produce artificial intelligence.

But the price of intelligence itself is under enormous downward pressure.

Stanford University’s AI Index finds that the leading models are becoming increasingly similar in performance.

There is also growing competition from open-weight models.

And some are developed by organizations with very different economic objectives from American technology companies.

DeepSeek founder Liang Wenfeng recently said his company is prioritizing the pursuit of artificial general intelligence over maximizing profits and expects its leading models to remain open-source.

American policymakers can restrict foreign AI products inside the United States.

But they cannot eliminate global competition.

That matters because a completed data center isn’t automatically a profitable data center.

And this is where the worst-case scenario becomes interesting.

Imagine that AI revenue grows, but not enough to justify the investment.

Future investors will demand larger premiums before financing new infrastructure.

Meanwhile the United States government may have to offer higher yields to finance its enormous deficits.

Higher yields would increase federal interest expense.

That could require even more borrowing.

And more borrowing would create even more bonds for the market to absorb.

That is the dangerous feedback loop.

It is also important to understand that none of this requires AI to fail. Quite the opposite.

AI could become one of the most important technologies ever created.

The risk is that AI succeeds while the investments made to build it fail financially.

Ultimately, no one knows where this is all going.

But it is worth spending some time connecting the dots on the financing risks connected to AI infrastructure.

The United States is attempting one of the largest private infrastructure buildouts in modern history at the same time its government has an enormous continuing appetite for borrowed money.

Both are competing for capital.

And eventually, somebody has to pay it back.