Citadel founder Ken Griffin said the AI boom almost certainly has pockets of excess. But, he argued, a bubble does not mean the technology itself is doomed.
In a conversation with Bloomberg reporter Dani Burger at Citadel Securities’ annual Future of Global Markets conference, Griffin ran through artificial intelligence, GPU supply and demand, corporate AI spending, and the latest market volatility. His basic point was pretty direct: the real issue is not whether AI matters. It is whether today’s AI infrastructure valuations lean on assumptions about future demand and pricing that are just too rosy.
Griffin compares GPUs to airline seats
Talking about the recent sell-off in semiconductor stocks, Griffin said this is not just the market suddenly giving up on AI. What investors are doing now, he said, is taking a harder look at profits across the whole supply chain. For a while, the AI trade looked easy enough: if big tech companies were going to pour hundreds of billions, or even trillions, of dollars into data centers, then investors could buy cloud providers, chipmakers, and suppliers of parts tied to Taiwan, South Korea, and Japan.
Now? The questions have gotten sharper. Which companies really own the choke points in the supply chain? How much longer can current gross margins hold up? And will demand still justify today’s pricing two or three years from now?
Griffin pointed to DRAM and multilayer ceramic capacitors, or MLCCs, and said those segments do have supply constraints right now. That scarcity, he said, has helped produce very high profits. Still, investors are pushing on a different question now: how long does that scarcity actually last?
For GPUs, Griffin reached for an airline comparison. "A GPU is like an airline, where the marginal cost of operating is very low." He said that once an airline has already decided to run a flight, selling one more seat does not cost much. Even a $20 ticket can be better than letting the seat go empty. If the market winds up with too much capacity, airlines can cut fares hard to fill those seats. Prices can drop fast.
He said GPU computing services could run into the same dynamic. If GPU supply climbs above real computing demand, data center operators may start cutting prices just to keep utilization up. That would put pressure on investment cases built around unusually high rental rates.
Griffin also brought up the telecom business during the 2000 internet bubble. Companies laid down huge amounts of fiber, but supply eventually got far ahead of demand. The result: large volumes of unused dark fiber and a steep drop in communications-capacity pricing. Investors in AI infrastructure, he said, need to think seriously about a similar risk.
Older GPUs may not keep earning at current levels
Griffin also raised another issue beyond oversupply: the economic life of GPUs. He said some GPUs already in use for five or six years are still able to earn meaningful rental prices in the market.
But that may not last forever. As newer GPU generations arrive, efficiency gets better, and AI workloads change, it is still an open question whether older hardware will keep supporting high pricing. Griffin said investors should not just assume a GPU’s revenue in year seven, eight, or even nine will look like it does now.
He said this matters especially for companies borrowing heavily to build AI data centers. If the window in which equipment can throw off strong cash flow ends up being shorter than expected, or if compute rental pricing falls, then payback periods could stretch out and hit both project valuations and financing capacity. Even if demand for AI compute keeps growing, that still does not guarantee every company investing in GPU infrastructure will preserve today’s profit levels.
Companies are rechecking AI bills as open-source models gain ground
Another thing worrying Griffin is return on investment from the customer side. He said US companies are not all getting the same economic value from AI, and some of the more advanced users are already measuring the cost-benefit tradeoff across different models much more carefully.
In real terms, that could mean paying for premium models from OpenAI or Anthropic only when the task truly needs top-end reasoning. Simpler work could shift to far cheaper open-source models. So yes, overall AI usage might rise. But, he said, that does not mean companies will keep paying premium prices for the most expensive models.
Griffin said companies are going back through the bills they pay to OpenAI and Anthropic and asking whether those costs are really turning into productivity gains. Dani Burger also mentioned market chatter that Anthropic, while preparing listing documents, had laid out future spending that could reach $750 billion. Griffin said there may be a real tension here: AI companies keep increasing capital spending, while enterprise customers try to push model-usage costs lower.
He added that the ending still depends on how much value the models can create. If AI eventually helps pharmaceutical companies crack cancer treatment or other major medical problems, businesses may be willing to pay enormous sums for that capability. So the future pricing power of AI models, he said, is still highly uncertain.
A bubble does not erase the value of the technology
Griffin did not wave away the bubble argument. He said, "There is almost certainly some bubble at this moment, and any technological revolution that fundamentally changes the economy is usually accompanied by a bubble."
He pointed back to the 1999-2000 internet boom. Plenty of internet companies then had unrealistic business models and valuations. And yet the internet still became basic infrastructure for modern society.
Griffin even said he still wishes he had not sold Amazon stock back then. In his view, big technological shifts like railroads and the internet both passed through stretches of excessive capital investment. But after the bubble cleared, the companies with real competitive strength often came out as industry leaders for decades.
AI has compressed some research work from weeks to hours
Even while warning about excess in AI investing, Griffin said the past year of progress has changed his own view in a fundamental way. He recalled that some high-end white-collar work at Citadel, work that once took PhD- and master’s-level researchers four to six weeks, can now in some cases be finished in a matter of hours with AI tools while producing the same answer.
He said seeing that firsthand was deeply discouraging at first. Why? Because it also suggested white-collar jobs could come under pressure and long-standing corporate moats could erode fast. At the same time, he said AI is creating openings for a new generation of founders that would have been hard to picture before.
Griffin said that if startups can get access to specialized capabilities at very low cost, capabilities that only large companies could previously afford, the result could be a golden era for entrepreneurship. For Citadel itself, though, he said human researchers are still at the center of generating alpha.
He compared Citadel’s fundamental investing work to journalism. The key, he said, is understanding a company’s business model, market size, unit economics, and management incentives, and then building the right investment narrative. AI can help collect and process parts of that information set. But researchers still have to put those inputs together and turn them into an investment judgment. Trade execution, on the other hand, is already one of the most automated functions inside Citadel.