Key Takeaways
- • Ken Griffin, founder of Citadel, warned at the firm's annual Global Markets Future Summit on October 21 that AI computing power investments face three traps: collapsing marginal pricing, overstated GPU depreciation periods, and tightening corporate AI budgets.
- • Griffin likened GPU economics to airlines, noting extremely low marginal operating costs mean that once supply exceeds demand, spot leasing and API prices could fall sharply, drawing a parallel to dark fiber after the 2000 telecom bubble.
- • Griffin said the market is in a 7-to-9-year depreciation assumption problem, since current high rents for five- to six-year-old GPUs are unlikely to persist, undermining residual-value assumptions for years seven through nine of a chip's life.
- • Corporate buyers are shifting from trial-and-error AI spending to ROI-focused cost minimization, a direct tension with Anthropic's disclosed plan to spend $750 billion on computing power, per The Wall Street Journal citing its S-1 filing.
- • Griffin acknowledged that localized AI bubbles almost certainly exist but argued that infrastructure remaining after a bubble often reshapes the economy, and that repricing along these three dimensions has only just begun.
Summary
Citadel founder Ken Griffin used his appearance at the firm's annual Global Markets Future Summit on October 21 to tell Bloomberg reporter Dani Burger that the market is moving away from the theme-driven buying of the past two years toward fundamental research. He said, "I believe what this market is experiencing is that people are finally starting to do their homework." Over the prior two years, buying across the AI supply chain, including NVIDIA, high-bandwidth memory (DRAM), Murata Manufacturing's multilayer ceramic capacitors, suppliers in South Korea, Taiwan, China and Japan, and cloud computing giants, was treated as nearly guaranteed profit. That logic is now being re-examined, with semiconductors experiencing their longest consecutive decline since the financial crisis.
Griffin identified three risks. The first is marginal pricing. He compared GPUs to airlines, which have extremely low marginal operating costs, and warned that once computing supply exceeds demand, spot leasing and API call prices could collapse quickly. He cited the dark fiber glut of the 2000 telecom bubble as a historical parallel. The second is depreciation. Griffin noted that some entities still charge high rents for GPUs five to six years old, a situation he doubted would persist, and said investors can no longer apply the same capitalization rates or revenue assumptions to years seven through nine of a chip's life that they did a year earlier. Institutional valuation models relying on seven- to nine-year horizons therefore face pressure for reassessment. The third is corporate budgets. Griffin said leading U.S. companies are deciding which business units need premium models from OpenAI and Anthropic and which can run on open-source alternatives, and that firms are reviewing AI bills to ensure genuine productivity gains.
Griffin also addressed Anthropic's plans. According to The Wall Street Journal, Anthropic disclosed in its S-1 filing that it plans to spend $750 billion on computing power, with spending continuing to rise. Asked whether this conflicts with corporate cost-cutting, Griffin replied, "Yes. You have already answered the question by asking it." He said the key question is whether Anthropic's models can become essential utilities, citing life sciences as an example, but added, "But we do not know the answer yet."
On the bubble debate, Griffin said localized bubbles almost certainly exist, since fundamental economic transformations are typically accompanied by them, and added, "I only wish I had held on to my Amazon shares." He argued that the infrastructure left after a bubble bursts often changes the world and that the survivors become long-term market leaders. He cautioned, however, that current pricing risks in marginal pricing, depreciation assumptions and corporate budgets have only begun to be repriced.
Griffin identified three risks. The first is marginal pricing. He compared GPUs to airlines, which have extremely low marginal operating costs, and warned that once computing supply exceeds demand, spot leasing and API call prices could collapse quickly. He cited the dark fiber glut of the 2000 telecom bubble as a historical parallel. The second is depreciation. Griffin noted that some entities still charge high rents for GPUs five to six years old, a situation he doubted would persist, and said investors can no longer apply the same capitalization rates or revenue assumptions to years seven through nine of a chip's life that they did a year earlier. Institutional valuation models relying on seven- to nine-year horizons therefore face pressure for reassessment. The third is corporate budgets. Griffin said leading U.S. companies are deciding which business units need premium models from OpenAI and Anthropic and which can run on open-source alternatives, and that firms are reviewing AI bills to ensure genuine productivity gains.
Griffin also addressed Anthropic's plans. According to The Wall Street Journal, Anthropic disclosed in its S-1 filing that it plans to spend $750 billion on computing power, with spending continuing to rise. Asked whether this conflicts with corporate cost-cutting, Griffin replied, "Yes. You have already answered the question by asking it." He said the key question is whether Anthropic's models can become essential utilities, citing life sciences as an example, but added, "But we do not know the answer yet."
On the bubble debate, Griffin said localized bubbles almost certainly exist, since fundamental economic transformations are typically accompanied by them, and added, "I only wish I had held on to my Amazon shares." He argued that the infrastructure left after a bubble bursts often changes the world and that the survivors become long-term market leaders. He cautioned, however, that current pricing risks in marginal pricing, depreciation assumptions and corporate budgets have only begun to be repriced.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.