Cato Dezorra
The $50 Trillion Problem:
In the 1987 movie Spaceballs (a parody of Star Wars), the villain “Dark Helmet,” in pursuit of the film’s hero, commands the enemy ship to go to its maximum speed, many times the speed of light – a level called “Ludicrous Speed.” It’s so stupidly fast that the stars outside go beyond looking visually stretched by speed, to a point where the light looks like “plaid.” (Fun fact: that’s where the name of Tesla’s fastest model, the Model S Plaid comes from.)
Dating myself with that reference (Spaceballs came out when I was five, so I loved it) but “Ludicrous speed” is the only way to describe the far-beyond-irrational level of capital expenditure spend the AI/data center buildout is now reaching. It is getting to a level so massive, it dwarfs any other stock market bubble in history – and it projected to keep on going.
To put all this into perspective, a new study just published by financial services firm PricewaterhouseCoopers projects that the world will spend an almost incomprehensible $31.6 trillion on data-center infrastructure between now and 2050. And that’s merely the firm’s central scenario. If AI adoption develops faster than expected, PwC says the number could approach $50 trillion.
The United States alone is projected to capture $15.1 trillion. Annual worldwide data-center capital expenditures supposedly rise from roughly $800 billion in 2026 to $1.1 trillion in 2030 and $1.8 trillion by 2050.
For perspective, the entire U.S. economy currently produces a GDP of a little over $30 trillion annually – and remember, that is the claimed value of every single good sold and service performed in the entire economy for an entire year (including all the circular spending, fraud, waste, and government overspending included to make that figure sound better than it is. AI spending alone is projected to be an enormous part of that, all by itself.)
PwC argues that this investment cycle will dwarf the railroads, electrification and internet buildouts because those were largely one-time infrastructure projects. AI infrastructure is different: GPUs, servers and networking equipment become obsolete and must continually be replaced. PwC makes a good point here – it assumes major equipment refreshes every four to six years, producing a gigantic recurring capital cycle lasting decades.
But there is a rather important question buried underneath these breathtaking forecasts:
Where exactly is all this money going to come from?
Two gigantic fraud & waste machines, chasing the same money
For the first phase of the AI boom, this question wasn’t particularly pressing. Microsoft, Amazon, Alphabet and Meta generated extraordinary amounts of cash and possessed fortress-like balance sheets. They could and did finance enormous capital expenditures internally.
But starting in late 2025, the numbers began becoming too large even for them.
Estimates for AI-related capital expenditures through 2030 now run roughly $5–7 trillion, and borrowing is increasingly becoming part of the financing equation. U.S. hyperscalers have burned through all their cash and are now aggressively issuing bonds; and having burned through most American demand for these high-yield bonds, their latest project has been raising over €40 billion from European investors as they hunt globally for pools of capital large enough to finance the buildout.
American Big Tech now represents nearly 10% of new Euro-denominated nonfinancial corporate bond issuance. European policymakers are already beginning to worry that American AI companies could crowd European companies out of their own capital markets (well, welcome to the club mates.)
And that is where the AI story collides with something much larger. The United States government is hunting for people who want to buy its bonds too – and this is critical because it needs to sustain that demand at lower interest rates in order to keep the cash machine flowing for both itself and for the AI monstrosity.
For perspective, the Congressional Budget Office projects a $1.9 trillion federal deficit this year, rising to $3.1 trillion annually by 2036. In spite of all Trump’s bloviating about “draining the swamp” and reigning in reckless federal spending, under him the U.S. government has been more reckless than ever, with larger deficits and more overall debt added under him than under any president in US history.
Under its current-law baseline, Washington will need to borrow another approximately $26 trillion between the end of 2025 and 2036, pushing debt held by the public to roughly $56 trillion. Gross federal debt reaches approximately $64 trillion by that time (as if the already staggering $40 trillion wasn’t enough…on the other hand, once you hit 40, what’s another 26?)
Interest expense itself will become one of the primary forces driving those deficits higher, as debt needs to continually be refinanced, and more money needs to be borrowed to pay the interest payments on the debt that already exists.
And yes, the US government is that irrational of a spender; if it were a household, it would have maxed out all its credit cards, but simply opened up new credit cards to make the minimum payment on the old credit cards, and when those are maxed out it will just do the same thing again – no end in sight.
For this reason, publicly held debt is projected to rise from roughly 101% of GDP today to 120% by 2036—the highest level in American history. And importantly, that is not counting any significant interest-rate increases, which would be the only way to tackle the inevitable inflation that would result from all this borrowing/spending.
Now, put these two forecasts next to each other.
Washington needs roughly $26 trillion of additional borrowing over the coming decade.
The technology industry simultaneously envisions $5–7 trillion of AI capital expenditures by 2030, followed by an infrastructure cycle eventually totaling $31.6 trillion—and perhaps $50 trillion—globally.
These aren’t independent numbers. They are competing claims on the same capital.
Someone has to buy all this debt
There isn’t some magical warehouse containing unlimited cheap money.
Every Treasury bond, corporate bond, data-center loan and private-credit facility ultimately represents a claim on somebody’s capital. (Granted, the U.S. Treasury and the Federal Reserve can both just “print money out of thin air” to buy the U.S.’s own debt – but it can only do this so much, and that causes more inflation.)
Pension funds, insurers, banks, sovereign wealth funds, households and foreign investors have finite balance sheets. When Washington issues trillions of dollars of Treasuries while Amazon, Alphabet, Microsoft, Oracle and the rest of the AI ecosystem simultaneously demand trillions for data centers, chips and power infrastructure, those borrowers increasingly compete for investors.
We’re already seeing it. And that is just a tiny preview of the larger problem.
If the supply of bonds rises faster than investor demand, bond prices fall and yields rise. Higher Treasury yields then increase Washington’s interest expense, requiring still more borrowing merely to service the existing debt.
But those same higher yields then raise the cost of financing AI infrastructure, throwing all current projections completely out of whack. Suddenly a data center whose economics looked attractive when capital cost 4% looks considerably less attractive at 7%.
This creates an extraordinarily dangerous feedback loop: More government borrowing → higher yields → higher government interest expense → more government borrowing.
At the same time:
More AI borrowing → higher corporate yields → worse data-center economics → greater required AI profitability.
Both machines depend upon capital remaining sufficiently cheap. Yet their combined demand for capital pushes in precisely the opposite direction.
And money isn’t even the only constraint
Even if PwC’s $31.6 trillion somehow materializes, there remains the awkward problem of constructing everything it supposedly purchases. PwC itself identifies electricity availability as the foremost constraint determining where data-center investment can occur.
Semiconductor disruptions alone could reduce projected global investment by nearly 20%. Data centers also require enormous quantities of transformers, turbines, cooling equipment, transmission infrastructure, copper, construction labor and increasingly scarce electrical-generation capacity.
So consider what these forecasts actually require us to believe.
- We need tens of trillions of dollars of capital for all this stuff.
- We need enough investors willing to finance it, at economically tolerable interest rates.
- We need sufficient electricity generation.
- We need transmission infrastructure.
- We need continually expanding semiconductor production.
And after spending all that money, we still need AI to produce enough actual economic value to generate an adequate return on perhaps the largest capital-investment cycle civilization has ever attempted. And overwhelmingly, it’s already failing to do that, as many individual companies and independent analysts have reported.
S&P is already warning that hyperscaler spending is growing faster than previously anticipated, financing structures are becoming increasingly opaque, and measurable returns remain uncertain. The six largest U.S. hyperscalers alone are now expected to spend more than $7 trillion through 2030, potentially putting their credit ratings at risk if the promised earnings don’t materialize (though, they have plenty of tricks to move that toxic debt “off-balance sheet” and onto shell companies they make up; they’ve already been doing that, as I’ve previously reported…and this is while things supposedly look financially terrific!)
That’s a fairly important caveat to the assumption that capital simply “exists.”
Eventually somebody loses (and it’s always regular people)
This leaves policymakers with an increasingly unpleasant problem. They can allow market forces to operate normally, and yields may have to rise enough to attract sufficient capital. But that threatens both sides of the equation.
Higher rates worsen Washington’s debt spiral, while simultaneously undermining the economics supporting trillions of dollars of AI investment.
Alternatively, the Federal Reserve can eventually step back in and create enough liquidity to suppress borrowing costs. But then the adjustment doesn’t disappear. It simply moves somewhere else—into inflation, asset prices and the value of the currency. (My prediction: this is what they will do, by categorizing AI spending as “national defense,” and just dumping the resulting inflation onto the people, who can’t do anything about it anyway.)
That is why the real story behind these increasingly absurd AI projections isn’t whether the world will actually spend $31.6 trillion or $50 trillion building data centers.
It is that those projections are being made alongside equally absurd projections for sovereign borrowing, apparently under the assumption that both can simultaneously obtain effectively unlimited amounts of affordable capital. They can’t. It’s not possible.
Eventually, the U.S. will have to make a choice between funding the biggest ever Ponzi scheme on earth – the U.S. government; or the biggest ever financial bubble in history – AI.
Either way, the math simply ain’t mathing. Either outcome will be very ugly.
*Cato Dezorra is a former military/security professional turned independent commentator.
Source: https://catodezorra.substack.com/p/the-50-trillion-problem-there-isnt
