Citadel founder Ken Griffin said artificial intelligence is "almost certainly" in a bubble, but he stopped short of predicting a crash and argued that the companies that emerge from one will lead markets for decades.

Griffin, founder and chief executive of hedge fund Citadel and market maker Citadel Securities, made the remarks before a New York audience on October 20 at the Future of Global Markets Conference, an annual event hosted by Citadel Securities. "Any radical transformation of our economy is generally intertwined with a bubble," he said.

He drew parallels to the dot-com era and said he wished had held on to his Amazon stock. Companies that survive a speculative episode "are the market leaders for decades to come," Griffin said, adding that the current froth was no reason to sell.

Nvidia shares closed at $216.47 on Oct. 20, hours before the talk, then rose 2.5% to $221.97 the following day. On October 21, the Invesco QQQ Trust (QQQ) increased 1.1% to $725.20.

From weeks to two or three hours

Citadel's own trading desks show the upside. AI agents—programs that carry out multistep jobs without a human overseeing each step—have trimmed some of the firm's research from six to eight weeks to two or three hours. It is precisely this kind of measured productivity gain that helps explain why institutions keep funding the technology even as voices like Griffin's flag bubble conditions.

Griffin's view of AI flipped in 2026. He was dismissive of the technology at Davos in January, then said he was "unsettled" by its capabilities at Stanford's business school in May. AI is opening a strong stretch for entrepreneurs, he added, and already threatens white-collar jobs.

GPUs compared to airplane seats

Griffin compared GPUs, the chips used to train and run AI models, to airplane seats: because selling one more of either costs little, prices can crash when supply outstrips demand. He pointed to the telecom boom around 2000, when companies laid far more fiber-optic cable than anyone needed, driving down call and data rates. In that episode, much of the surplus fiber was later absorbed by growing internet traffic, even as many of the carriers that financed it went under—history's version of the survivor dynamic Griffin described.

Chips that are five or six years old still rent for what he called extraordinary rates, but he said he did not expect them to command that much for another five or six years.

Uneven value from premium AI models

Buyers are becoming more selective, Griffin said, determining which jobs require premium models from OpenAI and Anthropic and which cheaper open-source models can handle. "Corporate America has been very uneven in its ability to extract value from its AI models," he said, describing finance teams reviewing their lab bills and asking where the productivity gains are.

As Cryptopolitan reported, Google launched Gemini 4 Argon on October 13, priced at an introductory rate of $2 per million input tokens and $10 per million output tokens. Anthropic's Claude Opus 5.5 costs $4 and $20, respectively, while OpenAI's GPT-6 Astra costs $10 and $50. Meta subsequently released Muse Spark Open, an open-weight model with 816 billion parameters that developers can download and use for free under limited commercial terms. That spread—flagship output prices differing by a factor of five, with a free open-weight option beneath them—is the concrete cost backdrop for the triage Griffin described among corporate buyers.

Nvidia sells the GPUs Griffin described. The company has $279 billion in supply commitments, mostly for memory for its Vera Rubin systems, and reports third-quarter results on November 17 with guidance of $108.0 billion. Those results will offer a near-term reading on whether demand is keeping pace with the commitments behind the buildout Griffin was assessing.

This article was first published by Cryptopolitan.