Howard Marks does not test for a bubble by looking at the price. He tests for it by listening to the room. A real one, in his account, is a period in which investors are not asking whether prices are too high, because the question does not occur to them. On that test, the current argument about artificial intelligence fails to qualify.
Marks is co-chairman of Oaktree Capital Management and has been writing memos to clients since the 1990s, a record that includes warnings before the 2000 collapse and purchases during 2008. His point this time is narrow and often misread. When the financial press, the sell side and the dinner party are all debating whether AI is a bubble, the risk awareness that precedes a top is still present. Bubbles end when that awareness disappears, not when it becomes loud.
He does not deny the behaviour that produces bubbles. Fear of missing out is visible, and he says so. Money is moving into companies on a story about future demand, and some of those companies will not exist in a form anyone recognises. But that fear is a poor timing device. It can persist for quarters, and the people who act on it early are often rewarded long enough to convince themselves they were analytical.
His second test is valuation, and there the answer is less comfortable. Forward price-earnings ratios on the large indices are elevated but not at the extremes of 2000, when the companies leading the market had no earnings to discount and the multiples were computed on hope. The cyclically adjusted measure, which smooths ten years of profit, sits near the top of its historical range. Both statements are true, and they point in different directions.
A forward multiple reflects the earnings of the largest companies, which today are among the most profitable businesses ever assembled. The smoothed measure reflects a decade of profit that includes a financial crisis and a pandemic, and it has signalled that American equities were expensive for most of the past ten years, during which they compounded handsomely. A measure that has been wrong for a decade is not useless. It is measuring something other than the next twelve months.
The part of his argument that bears most directly on this market is what ends a cycle. Bubbles rarely die of high valuations. They die when the financing stops. The 2000 collapse was preceded by a repricing of telecom debt, and the 2007 collapse by a repricing of mortgage credit. In both cases the equity market kept its faith for months after the lenders had lost theirs, and in both cases the lenders were right.
That is why the credit market deserves more attention than the price of the stocks. The AI buildout is being financed through structures that separate the borrower from the parent: project companies, special purpose vehicles, leases, and vendor arrangements in which the supplier of chips helps fund the buyer of chips. These structures work while capital is available and the equipment holds value. They become difficult when both assumptions are tested at the same time, because the entity that owns the hardware has no other business to fall back on.
Marks has made a related point about being right versus being early. An investor who sold in 1997 was correct about the internet and wrong about the next three years, and most investors are judged on the next three years. That asymmetry is why bubbles inflate past the point where the analysis says they should, and why the people who call them are usually ignored until the call is no longer useful.
The practical conclusion he draws is not a sell recommendation, and reading it as one misses the argument. Not mature is not the same as safe. A market can be 20 percent overvalued, stay there for a year, and deliver a poor return without ever producing the crash the word bubble implies. The absence of a crash is not evidence that the price was reasonable. It is evidence that the buyers had time.
For anyone holding AI exposure, the useful question is not whether this is a bubble. It is what would make the financing stop, and how quickly that would travel. A rise in the cost of high-yield debt for the second tier of computing companies, a supplier that stops extending credit to its customers, a bank that declines to renew a construction loan, a lease disclosure that reveals the residual value assumed for equipment that is already a generation old: any of these could end the argument without settling it.
The awkward feature of his framework is that it gives comfort to both sides. To the bulls, it says the top is not here, because everyone is still worried. To the bears, it says the top will not be identifiable when it arrives, because the thing that identifies it is the disappearance of exactly the worry now on display. Both can hold the same view and act differently.
His answer to the question of what to do when a price cannot be judged has been consistent: he does not predict, he prepares. Size positions so that being wrong is survivable, avoid structures that require refinancing at a price you do not control, and accept that the last year of a bubble produces most of the return and most of the damage.
The debate about AI will continue, loudly, which on his own logic is the strongest argument that the end is not imminent. What the debate cannot tell you is what the end will look like. That will be decided in a credit market, by people who are not arguing about whether the technology is real.
