It's going to take a lot of artificial intelligence to solve the problems of AI, says MIT fellow and venture capitalist Paul Kedrosky

Model convergence. Irresponsible debt financing. A discounting of the boom-bust cycle. MIT expert Paul Kedrosky thinks the AI boom shares many similarities with previous bubbles in technology.

The ambitious capital-expenditure plans and expensive market valuations of AI companies are fundamentally at odds with the revenue growth these companies can generate and the actual data-center utilization currently being witnessed.

This warning, one of several recent reality checks on AI, comes from a venture capitalist, investor and research fellow at the MIT Initiative on the Digital Economy. Paul Kedrosky was interviewed Friday on the financial podcast the Meb Faber Show, and he offered some troubling observations and provocative predictions about the development of AI technologies.

Paul Kedrosky.

His first concern is that humans aren't very good at anticipating what he calls "the exponential function," when technologies are adopted and then scaled rapidly, "moving in ways that are both unexpected and more consequential than people realize."

Comparing the advent of AI with previous revolutions, like the building of a railroad network, he said it, too, is generating its own bubble - inflated by loose credit, some kind of policy stimulus, a compelling technological narrative and, usually, some kind of real-estate component.

For Kedrosky, then, the behavioral aspects of manias and bubbles don't change: Investors chase things at the wrong time and then panic. The same phenomena that we saw in the global financial crisis "are all in place now, again." AI, he observed, "is such a good example of it."

An extension of this fad is the massive increase in debt financing. Most of the investment into data centers has shifted from internal cash flows to external debt financing, using a variety of instruments like asset-backed securities and private credit, that now accounts for somewhere between 15% and 18% of investment-grade credit. This is competing directly for capital with sovereigns that are already under pressure for funding because of fiscal largesse.

Goldman Sachs now estimates 2026 capex among hyperscalers will exceed $1 trillion.

Kedrosky is troubled that fixed-income investors are piling money into data-center projects backed by hyperscalers without really understanding what is happening in those data centers and whether they are sufficiently utilized. This utilization issue is a focus of Kedrosky's. In some data centers, he is hearing anecdotally, he said, the usage of graphics processing units is only around 35% or 40% because some companies are "hoarding colossal amounts" in case they come to a point at which they don't have enough chips to provide the computing capacity. Looking forward, Kedrosky thinks this will cause huge problems with overcapacity, he said.

The price of the tokens by which corporates pay frontier labs for AI usage has been falling dramatically for months now. For Kedrosky, token pricing represents "the first hyperdeflationary commodity in the history of modern economies" because, whenever they come into contact with other parts of the economy or not, "they do huge damage to it," causing price deflation. With these falling token prices, Kedrosky calculates that frontier model companies must achieve 400% annual unit growth "just to stand still financially," he said.

Token-expenditure index declines may reflect a shift to cheaper models.

The conversation also touched on the distorting impact that the mega-IPOs of SpaceX (SPCX) and eventually Anthropic and OpenAI are likely to have on capital markets, as fund managers have to reduce holdings in established companies to generate liquidity for new issues.

On top of that, Kedrosky is alarmed, he said, at the model convergence and diminishing returns of large language models' performance. Moreover, AI is shortening chip-design cycles and, sooner or later, this will lead to a return of the boom-and-bust cycle of the semiconductor industry, he said.

-Jules Rimmer