Ken Griffin, the billionaire founder of Citadel, has articulated a thesis gaining traction among sophisticated investors: the current artificial intelligence boom contains many characteristics of a speculative bubble, yet the structural winners will dominate markets for years to come. This nuanced view—bullish on outcomes, skeptical of valuations—reflects the tension defining today's AI market. Griffin's comparison of graphics processing units to airline seats offers an unusually clear framework for understanding why commodity oversupply can destroy economics without eliminating the underlying technology's transformative potential.
The GPU shortage that began during the pandemic created artificial scarcity, driving prices to levels that bore little relation to manufacturing costs or long-term demand fundamentals. As supply chains normalized and manufacturers ramped production, the market faced an inevitable correction: excess capacity seeking buyers. Griffin's airline analogy is instructive here. Empty seats on a flight represent sunk infrastructure costs that cannot be recovered; airlines cut fares to marginal cost rather than leave planes grounded. Similarly, once GPU fabs are built and operational, producers face enormous pressure to maintain utilization rates, which compresses margins regardless of earlier optimism about pricing power. This dynamic suggests current AI infrastructure spending forecasts—many projecting hundreds of billions annually—may prove overstated once the productive capacity actually materializes.
What distinguishes Griffin's position from pure skepticism is his recognition that bubbles and transformative technology adoption are not mutually exclusive. Companies will inevitably overspend on compute resources, training models with questionable ROI, and building redundant infrastructure. But this wasteful investment also establishes which AI platforms and applications generate genuine value, which business models prove defensible, and which organizations can convert AI capabilities into competitive advantages. The survivors of this culling will possess entrenched network effects, data advantages, and proven monetization mechanics—precisely the characteristics that support sustained dominance across decades rather than single-cycle outperformance.
For investors and builders, Griffin's framework suggests a two-tier strategy: skepticism toward near-term valuations and near-term AI projects lacking clear unit economics, yet conviction in platforms demonstrating sticky adoption and defensible moats. The bubble in AI spending does not invalidate artificial intelligence itself; it merely prices in unrealistic assumptions about deployment speeds and return timelines. When the contraction comes—as it almost certainly will—the question becomes not whether AI matters, but which organizations survive the transition from hype to productive utility.