Citadel CEO Ken Griffin warned that an oversupply of chips could weigh on prices across the AI industry if GPU capacity grows faster than demand. The risk could challenge investment assumptions behind the current buildout of AI infrastructure, though Griffin did not predict a broad market collapse.
Griffin compares GPUs with airline seats
Speaking at Citadel’s annual “Future of Global Markets” conference, Griffin compared GPUs with airline seats, where the marginal cost of operating additional capacity is relatively low. An airline may still sell a seat for $20 if the fare covers some of its costs. But when capacity substantially exceeds demand, prices can remain under pressure.
A similar dynamic could emerge in the GPU market, he said. Once chips are installed, the additional cost of running computing resources may be relatively low. If a large wave of new chips comes online without a matching increase in customer demand, suppliers could cut prices to keep their capacity in use. That could alter expectations for GPU profitability and valuations.
Telecom fiber buildout offers a historical parallel
Griffin pointed to the telecom industry’s fiber-optic buildout around 2000, when companies laid more network capacity than the market needed. After the expansion, communications prices fell, and excess capacity put pressure on business models built around expectations of strong demand. He cited that episode as a historical comparison for investment in AI infrastructure.
Data centers, cloud providers and businesses are now increasing their AI computing capacity, with GPU purchases a major part of that expansion. If new capacity grows faster than demand from model training, inference and other applications, prices for chip rentals, computing services and related equipment could also come under pressure. For investors, the question is not only whether AI demand will keep growing, but whether the market can absorb new supply quickly enough.
Griffin sees a bubble, not an imminent collapse
Griffin said there is almost certainly a bubble element in AI, but he did not say a market crash is imminent. He also noted that some companies could emerge from a market shakeout and lead the industry for decades.
The distinction matters because long-term value in AI may not move in step with short-term asset prices. Changes in supply and demand could affect chip prices, revenue from computing rentals and valuations of related companies, even as businesses with strong technology, customers and operations remain competitive after an industry adjustment. The useful life and scale of GPUs, alongside the pace of actual demand, remain central to assessing returns on AI investment.