Interview

Why The AI Bubble Will Burst: The Most Logical Case — With Paul Kedrosky

Source attribution: Original source

Interview details

Publication date: Aug 15, 2026
Duration: 1h 4m

Speakers

Paul KedroskyVenture Capitalist, skvcap.com

Viewpoints in this interview

AI spending has reached an historically unusual scale.

Kedrosky argues that AI infrastructure investment has surpassed most previous major capital-expenditure cycles when measured against GDP, fixed investment, and contributions to GDP growth.

Paul Kedrosky AI capital spending Source at 2:22

Data center expansion increasingly depends on external financing.

Kedrosky says the financing mix has shifted away from relying primarily on technology companies' own cash flows, with more than half of data center funding coming from external financing by the second quarter of 2026.

Paul Kedrosky Data center financing Source at 3:47

The compressed pace of AI spending increases investment risk.

Kedrosky argues that AI infrastructure spending is occurring faster than previous infrastructure buildouts, leaving capital markets less time to reassess projects and expected returns between investment rounds.

Paul Kedrosky AI capital spending Source at 4:44

Investors evaluate data centers against comparable project-finance returns.

Kedrosky says external capital providers generally evaluate data center projects using market-based returns comparable with commercial real estate and other project-finance opportunities rather than requiring revenues proportional to an enormous theoretical AI market.

Paul Kedrosky Data center economics Source at 6:43

Data centers require continuing capital unlike conventional real estate.

Kedrosky argues that the real-estate analogy breaks down because data centers require recurring hardware, cooling, and infrastructure investment, creating a duration mismatch and continually diluting returns.

Paul Kedrosky Data center economics Source at 10:37

Falling token prices undermine data center revenue economics.

Kedrosky characterizes tokens as a rapidly deflating output of data centers and argues that persistent price declines make it difficult to support fixed investor return expectations.

Paul Kedrosky Token price deflation Source at 12:53

GPU longevity varies substantially with how chips are used.

Kedrosky says GPUs used intensively for training can have different failure characteristics from inference-focused chips, making broad assumptions about a uniform GPU lifespan misleading.

Paul Kedrosky GPU lifecycle Source at 17:21

Model convergence is shifting competition toward price.

Kedrosky argues that frontier models continue improving but are becoming harder for ordinary users to distinguish, making price an increasingly important competitive differentiator.

Paul Kedrosky AI model convergence Source at 19:18

Usage growth may not offset compounding token price declines.

Kedrosky rejects a simple Jevons-paradox argument, contending that the usage growth required to compensate for repeated steep token price declines would be extraordinarily large and therefore difficult to treat as the most probable outcome.

Paul Kedrosky Token price deflation Source at 20:03

Frontier labs are moving upward to seek higher returns.

Kedrosky argues that frontier AI companies are expanding into higher-value applications because the underlying token business is becoming less attractive, which could create conflicts with companies building on their models.

Paul Kedrosky Frontier model strategy Source at 21:53

Transformative technology does not guarantee attractive investment returns.

Kedrosky says he considers AI extraordinarily transformative while distinguishing technological importance from whether the scale and economics of current investment are financially justified.

Paul Kedrosky AI technology value Source at 24:30

Selling tokens weakens claims that models can absorb entire industries.

Kedrosky argues that if frontier models could reliably capture the enormous economic value sometimes attributed to them, their developers would have stronger incentives to deploy that capability directly rather than primarily sell access to the models.

Paul Kedrosky Frontier model strategy Source at 28:04

AGI expectations should not make AI spending effectively unpriceable.

Kedrosky rejects using the possibility of AGI as an unquantifiable option that can rationalize essentially unlimited spending when conventional economic justification is weak.

Paul Kedrosky AGI investment thesis Source at 38:37

AI combines several forces historically associated with major bubbles.

Kedrosky says the current AI boom combines technology enthusiasm, real estate, loose credit, and government policy, with geopolitical competition adding further pressure to continue investing.

Paul Kedrosky AI bubble drivers Source at 45:27

Many different catalysts could halt the AI investment cycle.

Kedrosky says no single trigger is necessary for an AI investment reversal because higher capital hurdle rates, geopolitical fragmentation, government involvement, and other factors could independently weaken project economics.

Paul Kedrosky AI bubble catalysts Source at 48:23

Utility-like capital intensity could force lower technology valuations.

Kedrosky argues that hyperscalers exhibiting utility-like capital requirements should not necessarily retain the valuation multiples associated with less capital-intensive technology businesses.

Paul Kedrosky AI company valuations Source at 50:25

Investors may challenge the value of massive pre-training expenditures.

Kedrosky predicts growing pressure to reduce large pre-training runs if investors conclude that more model improvement can be obtained through post-training, quantization, harnesses, and related techniques.

Paul Kedrosky AI training economics Source at 53:37

AI-related debt could spread broadly through financial institutions.

Kedrosky argues that hyperscaler-related debt can migrate through asset managers, insurers, banks, and other institutions, making the eventual distribution of AI-related credit exposure difficult to see clearly.

Paul Kedrosky Financial system risk Source at 55:54

Heavy capital concentration makes AI unattractive for his venture strategy.

Kedrosky says his venture investing generally avoids crowded AI sectors because many well-funded competitors chasing the same opportunities can structurally reduce expected returns, despite his belief in the technology itself.

Paul Kedrosky AI investing Source at 56:43