Ken Griffin: GPU glut could repeat telecom's fiber collapse

1. Three emotions at once — ▶ 1:05

   Griffin says high-level white-collar work that took PhDs and Masters students 4 to 6 weeks can now be produced by pressing a button and waiting a few hours. His reaction combined alarm about white-collar jobs, dismay as a business owner at competitive moats being "filled in by bulldozers", and excitement about a golden era for young entrepreneurs. He resolves the conflict with a historical argument: free-market transitions hurt incumbents but leave lives better, as 2026 compares with 1926. That is a pattern from past transitions, not evidence about this one.

2. Reskilling and pedigree-blind startups — ▶ 2:14

   Workers in the sweet spot of what AI does well, such as English-to-German translators, have in his view probably already lost their jobs, and he says the US needs reskilling "in a profound way" but sees talk without action. For incumbents, collapsing moats force faster innovation, customer value creation and cost reduction, which he says benefits consumers. Startups, he adds, hire hard workers without perfect pedigrees, giving them a new career lane. His prediction of unimagined new products is asserted rather than evidenced.

3. Research remains human — ▶ 5:07

   Griffin describes Citadel's core business as research, "almost like journalism": understanding business models, TAM, economics and management incentives, then assembling the narrative and deciding what multiple to apply. AI sheds light on bits of this, but he still believes the vast majority of the firm's alpha comes from its people's creativity and analysis. Trading, the means of monetizing research, has become the most automated and technologically forward part, because efficient execution lets managers run bigger positions and extract more from their ideas.

4. GPUs as airline seats — ▶ 8:24

   Like an unfilled airline seat sold for $20, a GPU has low marginal cost, so overcapacity pushes pricing into freefall. Griffin cites the telecom buildout around 2000, when more fiber was built than demand required and some was never lit. If GPU provisioning exceeds demand, he says, marginal pricing collapses and the investment assumptions behind today's buildout unravel. A second concern is useful life: 5-6-year-old GPUs earn extraordinary prices now, but he doubts that persists, and years 7-9 of a chip cannot carry the same capitalization rate. He frames both as open questions.

5. Uneven corporate returns on AI — ▶ 10:31

   Corporate America has been uneven at extracting value from AI. Sophisticated companies are deciding which parts of the business justify very high prices for frontier models from OpenAI and Anthropic and which can run on open-source models at a fraction of the cost. The interviewer cites a reported $750 billion of spending ahead at Anthropic against corporate cost-cutting. Griffin agrees the tension is real. In his view it depends on whether models become must-haves, for example by unlocking cures, so that customers pay almost any price. Nobody yet knows the answer.

6. Bubbles within the transformation — ▶ 13:27

   There are almost certainly bubbles within this moment, Griffin says, because radical economic transformations are generally intertwined with them. He recalls the metrics used to value 1999-2000 internet companies, some with fanciful business models, yet notes the internet is now woven into daily life. Railroads are another example. What remains after a bubble is often game-changing, and companies on the other side lead markets for decades. He does not say which layers of AI resemble the survivors and which the casualties.

7. Deleveraging as answer ten — ▶ 15:30

   Asked whether the recent sell-off was hedge fund deleveraging, Griffin says he has seen the industry "leveraged to zero" at least 10,000 times in the past decade, and calls the explanation a way of washing your hands of accountability. Reaching for an excuse, he argues, stops learning. He wants deleveraging to be answer ten, with answers one through nine asking what other investors, such as Capital Group, Fidelity or Wellington, understand that Citadel does not. Prices are changing for reasons, and the job is to find them.

8. Anatomy of a portfolio rescue — ▶ 17:34

   Griffin says Situational Awareness needed to raise cash and protect a substantial year-to-date gain, and Citadel bought a complicated portfolio in under 24 hours, then spent several days managing its risk. His thesis: when a manager looks distressed, fast-money accounts front-run expected selling and turn "a possibility into a crystallized reality", while a buyer of the portfolio typically sees a sharp reversal once confidence returns. He says the fund was caught in three mid-sized storms, including chatter about foundation model growth rates and pension rebalancing, with too much leverage at peak. This is Citadel's account, as the counterparty.

9. Robotics instead of tariffs — ▶ 27:41

   Griffin calls the tariff battles with allies a disaster that has been inflationary and severed supply chains. He argues the US will not industrialize efficiently that way, and that most people do not want factory jobs. His alternative is robotics-run factories, the reason he gave $3 billion to Carnegie Mellon University, which he says leads the world in robotics and has about 100 robotics companies around its campus. He also expects Republicans to almost certainly lose one or both houses in the midterms, and faults the president's explanation of the Iran war.

10. Growth as the fiscal path — ▶ 35:05

    Griffin says pandemic spending approached what the US spent to win World War II as a share of GDP, leaving no fiscal space to renew tax cuts. With what he calls a $40 trillion federal deficit, crowding out is back, visible in mortgage rates and the 30-year yield. His priorities, in order: growth, education, a Simpson-Bowles-style up-or-down vote on entitlement reform, and high-skilled immigration. Growing out is "the only direction", he says. If Washington ignores the bond market, he expects higher inflation, a weaker dollar and a falling standard of living. He does not quantify the growth required.

11. Physical proximity as strategy — ▶ 47:31

    Citadel's center of gravity, Griffin says, is somewhere over the Atlantic between London, Miami and New York. To keep cohesion, it keeps each vertical's teams in one location, with whiteboards and writable walls in offices and boardrooms. His claim is that nothing beats person-to-person interaction, particularly for developing younger talent through immersion with a boss who is also a mentor and with peers. He offers it as conviction rather than data, and as the counterweight to a geographically dispersed firm.

Ken Griffin's central claim is that AI's capability is no longer the uncertain variable. The uncertain variable is who captures its economics. Griffin, who runs Citadel, accepts that models now compress weeks of skilled white-collar work into hours. What he questions is whether the money flowing into compute, and the prices charged for frontier models, can survive once buyers and investors start doing the arithmetic. That distinction between a technology that works and a business structure that pays organizes almost everything he says about markets. It also explains how someone who describes going home depressed one Friday can describe himself, minutes later, as optimistic.

The capability evidence is first-hand. Griffin says work that took PhDs and Masters students 4 to 6 weeks now returns the same answer after a button press and a few hours of waiting. His reaction had three parts: fear for white-collar employment, dismay as a business owner at competitive moats "filled in by bulldozers", and excitement about new company formation. He resolves them with a historical argument that free-market transitions hurt incumbents and workers but leave lives better, as the gap between 1926 and 2026 suggests. That is a pattern, not proof. His concrete prescription, reskilling at scale, comes with the complaint that Washington talks but does not act.

His sharpest analytical move concerns supply. GPUs, like airline seats, cost little to run at the margin, so a seat sold for $20 still adds money. That is why airlines struggle: once capacity exceeds demand, prices fall freely. Griffin points to the telecom buildout around 2000, when more fiber was built than demand required, some of it never lit, and the price of telecommunications went straight down. If GPU provisioning exceeds demand, he says, marginal pricing would go into freefall and upend today's investment assumptions. He presents this as a question mark rather than a forecast, but it is the question on which much of the buildout rests.

A second issue is how long a GPU keeps earning. Some operators are getting extraordinary prices on chips 5 or 6 years old, but Griffin doubts that persists for another 5 or 6 years. Investors, he says, cannot give year 7, 8 or 9 of a chip the same capitalization rate or revenue model as one used 12 months ago. Add the high margins in DRAM and other areas of real supply constraint, which investors now ask how long will last, and the market's mood shifts. For two years the thematic trade was to buy everything: hyperscalers, chipmakers, obscure suppliers in Korea, Taiwan and Japan. Now, he says, people are becoming more discerning, which he reads as investors finally doing research.

On the demand side Griffin is least certain, and says so. Corporate America has been uneven at extracting value from AI. The most sophisticated companies are sorting which parts of their business need state-of-the-art models from OpenAI and Anthropic at very high prices and which can run on open-source models at a fraction of the cost. The interviewer cites a reported $750 billion of spending ahead at Anthropic, set against corporate cost-cutting. Griffin agrees there is tension. Its resolution, he says, depends on whether models become must-haves, for instance by unlocking cures, so that customers will pay almost any price. Nobody knows that yet.

Griffin does not deny a bubble. He says there are almost certainly bubbles inside this moment, and that radical economic transformations are generally intertwined with them. His reference points are the 1999-2000 internet and the railroads. What remains after a bubble is often game-changing, he argues, and the companies on the other side lead for decades. As history this is persuasive, but as guidance it is thin, because it does not say which layers of AI will be which. Read alongside his telecom analogy, it suggests an editorial inference: a technology can transform the economy while the pricing of an overbuilt layer collapses. Both stories can be true at once.

The same discipline governs how he reads market turmoil. Asked whether the recent sell-off was hedge fund deleveraging, he says he has seen the industry "leveraged to zero" at least 10,000 times, and calls the explanation a way to wash your hands of accountability. He wants it to be answer ten, with the first nine asking what other investors understand that he does not. Citadel's core, he says, is research, with trading as the way to monetize it, and the vast majority of its alpha still comes from people. Even as AI does PhD-level work, he holds that the human-assembled narrative is the source of edge. He offers no mechanism for why that will persist.

The Situational Awareness episode is his own example of the principle. By his account, the fund needed to raise cash and protect a substantial year-to-date gain, and Citadel bought a complicated portfolio in under 24 hours, then spent several days managing its risk. Fast-money accounts, he says, front-run expected selling and turn "a possibility into a crystallized reality", while a buyer of the portfolio usually sees a sharp reversal once confidence returns. He adds that the fund was caught in three mid-sized storms, including chatter about foundation model growth rates and pension rebalancing, with too much leverage at peak. This is the counterparty's telling; the fund's perspective is absent.

Griffin extends the growth-first logic to policy. Tariffs, he says, have been inflationary and have severed supply chains, and he prefers robotics-run factories, the reason for his $3 billion gift to Carnegie Mellon. On the fiscal side he says pandemic spending approached World War II levels as a share of GDP, leaving no room to renew tax cuts, and with what he calls a $40 trillion federal deficit crowding out is back. His priorities are growth, education, an up-or-down vote on entitlement reform and high-skilled immigration. Growing out is "the only direction", though he never says how much growth would be enough, or how quickly AI must deliver it. The interviewer also notes that Republicans raised spending too, and he agrees.

The thread connecting the compute glut, the chip-life question and the fiscal hope is the same: how much of AI's value shows up as pricing power for those who own the models and chips, and how much flows to customers as lower costs. Griffin's optimism for consumers and his worry about overcapacity both imply the second. His own list of risks, though, does not square easily with a reported $750 billion spending commitment at a single lab unless models become indispensable. That outcome is possible, but he offers no evidence for it, and his growth-based fiscal plan leans on the same assumption without testing it.