Howard Marks built Oaktree Capital on a simple idea: the most important fact in any investment is the price of money. For months, Wall Street has been waiting for that price to fall. Marks is now telling his clients to stop waiting.
In a memo dated September 22, the co-founder of the world’s best-known distressed-debt investor argued that the AI data-center buildout, projected to absorb more than $5 trillion worldwide through 2030, will compete directly with the U.S. Treasury for the same pool of savings. That contest, he wrote, argues against a decline in interest rates anytime soon.
The memo, titled “Shall We Repeal the Laws of Economics – Part III,” lays out the arithmetic. The federal government must refinance maturing debt and issue roughly $2 trillion in new securities each year. The technology industry, meanwhile, has committed to enormous spending on chips, servers, power and buildings. The two demands arrive at the same moment.
Marks cited a McKinsey & Company projection that more than $5 trillion will be spent through 2030 on data centers directly tied to AI. JPMorgan analysts have reached a similar figure for the full infrastructure bill. Either way, the numbers dwarf what the biggest spenders can cover from cash on hand.
“The need to fund the massive federal deficits comes on top of the routine need for capital that accompanies the growth of the U.S. economy,” Marks wrote, “and to that is added the multi-trillion-dollar investment in AI.” The five companies leading the buildout – Microsoft, Alphabet, Amazon, Meta and Oracle – held roughly $350 billion in cash among them at the end of the third quarter, a fraction of the total bill, he has noted.
The gap between those balances and the scale of the commitments must be filled somewhere, and that somewhere is increasingly the bond market. Oracle, Meta and Alphabet have all issued 30-year bonds to finance AI investment in recent months. In some cases the yield those bonds offer over comparable Treasurys is less than a percentage point, a thin premium for lending money for three decades against technology no one has proven can be repaid over that horizon.
Marks’s argument rests on the oldest rule in economics. “The simplest rule of economics is that increased demand for something causes its price to rise,” he wrote. When the something is capital, its price is the interest rate.
“It’s entirely understandable, therefore, that this growing demand for capital should put upward pressure on the price of money: interest rates,” he added. The government’s financing needs do not pause while technology companies raise capital, and the two draw on the same investors.
The pressure Marks describes is not evenly distributed. Short-term rates respond to what the Federal Reserve does; long-term rates respond to the supply of and demand for capital over decades. A wall of 30-year issuance from technology companies, layered on top of record Treasury borrowing, is exactly the kind of force that pushes long-term yields higher even if the central bank cuts its own rate. Economists call that wedge the term premium, and Marks’s memo is, in effect, an argument that it will stay wide.
The warning carries weight because Marks sits on the lending side of the market. Oaktree manages one of the largest credit portfolios in the world, and its parent, Brookfield, is raising a $10 billion fund for AI infrastructure with commitments from sovereign wealth funds and Nvidia. The firm is not a detached critic. It is one of the parties bidding for the same capital.
Marks framed the tension as a collision of two enormous appetites. Federal deficits that once looked cyclical have become structural. The AI buildout is the largest private investment program since the spread of the internet, and unlike that earlier era, much of it is being financed with debt rather than equity.
He has been careful not to call the AI boom a bubble. He has written that the technology is real and the demand is genuine, and that the danger is less a collapse than a mispricing of capital. But the financing structures remind him of earlier manias. In the late 1990s, the telecom buildout was funded heavily with debt, and when the glut arrived, the losses fell first on lenders. He has argued that the right way to own a winner-takes-most revolution is through equity, not debt, and that borrowing to build data centers is the riskiest way to hold the same bet.
The practical consequence, he argued, is that bond investors will demand higher yields to absorb the supply. Companies paying more to borrow face a higher bar before their data-center investments clear a profit. The arithmetic tightens for everyone at once.
The question Marks leaves open is how long the pressure lasts. He has cautioned before that every transformative technology draws more capital than it can productively absorb, and that the reckoning arrives only after the overbuild. Whether the AI boom ends in a glut of data centers, a shortage of electricity, or neither, the financing has already begun to reshape the yield curve.
For now, his conclusion is deliberately narrow. He is not predicting a crash. He is predicting that money will stay expensive, and that the people building the future on borrowed cash are betting they can outrun the cost of it.
