Key Takeaways
- • Citadel CEO Ken Griffin warned that a glut of AI chips could send GPU pricing into freefall, comparing GPUs to airline seats that still earn money when sold cheaply.
- • Griffin said bubbles almost certainly exist within AI but stopped short of predicting a collapse, noting that companies emerging on the other side of bubbles lead markets for decades.
- • Griffin cited the telecom fiber build-out around 2000 as a precedent in which excess capacity drove communication prices down.
- • Wall Street firm BTIG expects a prolonged chip slump that could run into 2027 after the Philadelphia Semiconductor Index (SOX) fell 11.4% in a week, its largest drop since April 2025.
- • Yardeni Research President Ed Yardeni forecasts the S&P 500 at 8,400 by mid-2027 and argues the 1999-2000 bubble comparison no longer fits.
Summary
Citadel CEO Ken Griffin said a glut of artificial intelligence chips could send pricing into freefall, according to remarks he made to Bloomberg's Dani Burger at Citadel's annual Future of Global Markets Conference in New York, as reported by BeInCrypto. Griffin compared graphics processing units (GPUs) to airline seats, noting that GPUs have a very low marginal cost to run. He argued that an unfilled seat sold for as little as $20 still earns money, so fares collapse once capacity exceeds demand. Applying that logic, he said that if GPU supply outstrips demand, pricing could drop into freefall, upending the investment assumptions behind today's AI buildout.
Griffin said bubbles almost certainly exist within AI, though he stopped short of predicting a collapse, and added that companies that emerge on the other side of bubbles lead markets for decades. He pointed to the telecom build-out around 2000, when companies laid more fiber than demand required and communication prices fell, as a historical parallel.
On recent market moves, Griffin described the AI-linked selloff as investors doing their research and said hedge fund deleveraging should be the 10th explanation anyone reaches for. He also questioned the useful life of a GPU: chips five or six years old still earn rates he called extraordinary, but he doubted that would hold for another five or six years. He added that corporate buyers are sorting which tasks justify premium OpenAI and Anthropic models and which can run on cheaper open-source alternatives, a dynamic he said strains against the labs' spending.
The bearish view is not universal. BTIG now expects a prolonged chip slump that could run into 2027, after the Philadelphia Semiconductor Index fell 11.4% in a week, its largest drop since April 2025. Yardeni Research President Ed Yardeni, by contrast, forecasts the S&P 500 at 8,400 by mid-2027 and says the 1999-2000 bubble comparison no longer fits. Whether AI revenue catches up with the capacity being built will determine how much of that spending earns a return.
Griffin said bubbles almost certainly exist within AI, though he stopped short of predicting a collapse, and added that companies that emerge on the other side of bubbles lead markets for decades. He pointed to the telecom build-out around 2000, when companies laid more fiber than demand required and communication prices fell, as a historical parallel.
On recent market moves, Griffin described the AI-linked selloff as investors doing their research and said hedge fund deleveraging should be the 10th explanation anyone reaches for. He also questioned the useful life of a GPU: chips five or six years old still earn rates he called extraordinary, but he doubted that would hold for another five or six years. He added that corporate buyers are sorting which tasks justify premium OpenAI and Anthropic models and which can run on cheaper open-source alternatives, a dynamic he said strains against the labs' spending.
The bearish view is not universal. BTIG now expects a prolonged chip slump that could run into 2027, after the Philadelphia Semiconductor Index fell 11.4% in a week, its largest drop since April 2025. Yardeni Research President Ed Yardeni, by contrast, forecasts the S&P 500 at 8,400 by mid-2027 and says the 1999-2000 bubble comparison no longer fits. Whether AI revenue catches up with the capacity being built will determine how much of that spending earns a return.
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.