Ken Griffin explains "AI crash" in detail: computing power is similar to aviation, three major traps—"marginal pricing", "depreciation period" and "corporate budgets"

On October 21st, during the Citadel Annual Global Markets Future Summit, Griffin stated in an interview with Bloomberg reporter Dani Burger that the market is shifting from the "thematic buying spree" of the past two years to true fundamental research. He said: "What I think is happening in this market is that people are really starting to do their homework."

In the past two years, the market logic was simple and direct: a trillion-dollar data center construction boom began, leading to buying in at every key node along the industry chain—Nvidia, high-bandwidth memory DRAM, super-multilayer ceramic capacitors (MLCC), as well as suppliers from South Korea, Taiwan, Japan, and cloud giants—virtually guaranteeing profits. Now, this logic is being re-examined. Recently, AI trading was also hit hard, and the semiconductor sector has seen its longest losing streak since the financial crisis.

Ken Griffin warns: GPUs are highly similar to airlines, with extremely low marginal operating costs—once computing power supply outpaces demand, spot leasing and API call prices will collapse instantly, repeating the "dark fiber" bust of the 2000 telecom bubble.

Trap One: Marginal Pricing Collapse—The "Airline Moment" for GPUs

Griffin drew a direct analogy with the airline industry.

"GPUs are very similar to airlines, with extremely low marginal operating costs. If there's an empty seat on a flight, even selling it for $20 is still profitable," he said. "It is precisely this feature that means, once there is overcapacity among airlines, pricing goes into free fall."

He then extended this logic to the compute market: "If we reach a point where GPU supply exceeds demand, marginal pricing will essentially go into free fall."

Griffin referenced the 2000 telecom bubble as a historical comparison: that year, massive amounts of optical fiber were laid but never lit due to insufficient demand, becoming "dark fiber," and telecom prices plummeted. His question: Is the compute market heading down the same path?

"This would completely upend all the investment assumptions people rely on today. That's the first issue."

Trap Two: Depreciation Model Breakdown—"Useful Life" Overestimated

Griffin’s second question goes straight to the valuation models of investment institutions.

"Some people are still able to collect high rents using GPUs that are five or six years old. The likelihood of this situation lasting another five or six years is very low," he said. "As a result, more and more people are worried: I can no longer assign the same capitalization rate or income model to the seventh, eighth, or ninth year of a chip as I did twelve months ago."

In other words, if investment institutions continue to discount depreciation using high asset residual values for 7 to 9 years, their valuation models will face systemic repricing pressure.

Trap Three: Tightening Enterprise Budgets—The Era of ROI Calculations Has Arrived

Griffin pointed out that the logic of enterprise AI spending has fundamentally shifted.

"America’s most astute companies are becoming very clear-minded—knowing which businesses require paying top dollar for OpenAI and Anthropic’s state-of-the-art models, and which can run on open-source models at extremely low cost."

He added: "If you look at those companies truly making good use of AI internally, they are rapidly becoming very effective at maximizing the return on each dollar spent on AI. I think the entire corporate sector is looking at bills from Anthropic and OpenAI and then saying: Okay, we need to carefully consider how to ensure we’re really gaining productivity from this."

A $750 Billion Head-On Collision with Enterprise Belt-Tightening

This trend of tightening enterprise budgets is directly at odds with the expansion plans of AI foundation model companies.

According to The Wall Street Journal, Anthropic disclosed in its S-1 filing that it plans to spend $750 billion on compute capacity in the future and that spending is still increasing.

Reporter Dani Burger asked Griffin directly: Is there any tension between these two things?

Griffin’s answer was brief and direct: "Yes. You’ve already answered the question with your question."

He then noted that the key issue is whether Anthropic can make its models "essential"—for example, in life sciences, if its models can truly help cure cancer or other deadly diseases, then every pharmaceutical company would pay any price for the technology. "But right now, we simply don’t know the answer."

Bubble Question: History Repeats, But Outcomes May Differ

On the increasingly heated topic of an "AI bubble" in the market, Griffin offered a historically informed answer.

"At this moment, there is almost certainly a localized bubble. Any fundamental economic transformation usually comes with a bubble," he said. "I just wish I’d held onto my Amazon stock back then."

His view is that what’s left after bubbles burst is often the technical infrastructure that changes the world, and companies that survive the bubble will become the market leaders of the coming decades.

But he also emphasized that this does not mean the current pricing risks can be ignored—especially in the three dimensions of marginal pricing, depreciation assumptions, and enterprise budgets, the market’s repricing has only just begun.