Is There an A.I. Bubble? A Debate. | The Ezra Klein Show

1. Kedrosky links AI capex to historic overbuilds

Timestamp: 00:01:58 to 00:07:20

Paul Kedrosky argues that every major general-purpose technology buildout in Western history has ended in vast overcapacity and financial crisis. With trillions in AI credit spreading across insurers and 401ks while 50% of US net worth sits in equities, a prisoner's dilemma forces tech giants to overinvest.

Significance (High): Frames the AI boom not as a question of technological utility, but as a systemic financial fragility where retail investors are heavily overleveraged to a single capital expenditure trade.

Sources in support: Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.)

Sources against: Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Neutral sources: Ezra Klein (Host, The Ezra Klein Show)

2. Azhar cites strong GPU demand and fast returns

Timestamp: 00:07:20 to 00:12:45

Azeem Azhar contends that the AI sector is currently in a productive boom rather than a bubble because computing capacity remains vastly undersupplied. Hyperscalers and neoclouds are recouping their GPU investments within 18 to 24 months, and the share of revenue coming from non-lab enterprise customers is steadily growing.

Significance (High): Provides a concrete empirical counter-thesis to bubble warnings by showing that physical data center assets are generating rapid cash returns from broadening enterprise adoption.

Sources in support: Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Sources against: Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.)

Neutral sources: Ezra Klein (Host, The Ezra Klein Show)

3. Plummeting token prices and coder skew, per Paul

Timestamp: 00:12:45 to 00:16:36

Paul Kedrosky argues that massive token volume does not equal financial value because performance-adjusted token prices are falling roughly 13-fold annually. Furthermore, software developers account for roughly 80% of frontier token revenue, whereas mainstream users need compressive tasks that run fine on cheap, primitive models like Meta's Muse.

Significance (High): Exposes a critical flaw in frontier lab revenue projections if mainstream consumers and businesses refuse to pay premium prices for leading-edge models.

Sources in support: Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.)

Sources against: Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Neutral sources: Ezra Klein (Host, The Ezra Klein Show)

4. Separating infrastructure from model lab survival

Timestamp: 00:16:36 to 00:20:15

Azeem Azhar insists on separating the long-term viability of frontier labs like OpenAI and Anthropic from the trillion-dollar buildout of GPU data centers. Even if enterprises shift to cheaper open-weight models and hurt frontier lab margins, that usage still monetizes the underlying hyperscaler compute infrastructure.

Significance (Medium): Clarifies that a business model failure at a headline AI startup does not automatically render the physical data center buildout unprofitable.

Sources in support: Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Sources against: Ezra Klein (Host, The Ezra Klein Show), Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.)

5. Monster IPOs as insider exits and customer risk

Timestamp: 00:20:15 to 00:31:36

While Azeem Azhar defines a bubble burst as a 50% crash in investment and asset prices, Paul Kedrosky points to historic Shiller CAPE multiples and warns that the upcoming OpenAI and Anthropic IPOs are massive insider exit events. Kedrosky also highlights extreme customer concentration, noting Meta alone accounted for roughly 10% of Anthropic's 2025 revenue.

Significance (High): Alerts public market investors that upcoming trillion-dollar AI listings may serve as liquidity exits for insiders while masking heavy reliance on a handful of tech buyers.

Sources in support: Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.)

Sources against: Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Neutral sources: Ezra Klein (Host, The Ezra Klein Show)

6. Why Apple abstained from the frontier model race

Timestamp: 00:31:49 to 00:36:22

Ezra Klein notes that Apple spends roughly $15 billion in annual capex compared to $100 billion to $200 billion at rival hyperscalers, yet its stock has outperformed. Azeem Azhar attributes Apple's restraint to a culture of strict product control, whereas Paul Kedrosky argues Apple fumbled its AI efforts and accidentally benefited as a free rider.

Significance (Medium): Demonstrates that abstaining from massive infrastructure spending and instead building complementary services on top of commoditized AI can reward shareholders.

Sources in support: Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.), Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Neutral sources: Ezra Klein (Host, The Ezra Klein Show)

7. Hyperscaler debt issuance and housing crowd-out

Timestamp: 00:36:22 to 00:44:14

Paul Kedrosky refutes the idea that pristine tech balance sheets protect the economy, noting that tech is now the largest US issuer of investment-grade bonds, utilizing off-balance-sheet special purpose vehicles and pushing up 10-year Treasury yields. Azeem Azhar acknowledges that external financing is approaching his 30% risk threshold and that data centers are crowding out housing and skilled trades.

Significance (High): Reveals how AI capital demands are actively distorting sovereign bond yields and diverting construction labor away from domestic housing.

Sources in support: Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.), Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Neutral sources: Ezra Klein (Host, The Ezra Klein Show)

8. Enterprise AI adoption and repricing human labor

Timestamp: 00:44:20 to 00:50:16

Azeem Azhar explains that outside Silicon Valley, traditional enterprises are adopting AI gradually over a 5-to-10-year horizon, primarily increasing revenue per employee and slowing hiring while creating higher-paid verification bottlenecks. Paul Kedrosky counters that AI is repricing cognitive labor faster than any previous technology, threatening to destabilize workers long before new job categories emerge.

Significance (High): Highlights the tension between gradual corporate integration and the rapid downward repricing of white-collar and service-sector wages.

Sources in support: Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Sources against: Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.)

Neutral sources: Ezra Klein (Host, The Ezra Klein Show)

9. Public backlash and post-crisis AI regulation

Timestamp: 00:50:16 to 00:57:01

Ezra Klein and Paul Kedrosky argue that aggressive labor replacement or a major AI security failure could trigger sudden, strict regulation akin to the FDA or an atomic energy board, with Kedrosky estimating a 50% to 60% cumulative probability of a triggering crisis. Azeem Azhar views local data center protests as healthy democratic feedback and notes that New Mexico data center subsidies have reached $4 million per job created.

Significance (Medium): Identifies political and regulatory backlash as a major underpriced risk to AI valuations, fueled by lopsided local tax subsidies and public unease.

Sources in support: Ezra Klein (Host, The Ezra Klein Show), Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.), Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

10. AI agent spam and the loss of social friction

Timestamp: 00:57:01 to 01:01:57

Ezra Klein warns that autonomous AI agents like Muse flooding restaurants with calls and managers with automated reports could make large swaths of the economy unpleasant. Azeem Azhar notes that while only 2.5% of Americans pay for AI subscriptions, free tools push screening costs onto society, and Paul Kedrosky compares the removal of societal friction to a denial-of-service attack on modernity.

Significance (Medium): Illustrates how consumer-facing AI agents externalize nuisance costs onto businesses and institutions, potentially sparking widespread refusal to engage with automated systems.

Sources in support: Ezra Klein (Host, The Ezra Klein Show), Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.), Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

11. Benchmarks that would change both guests' minds

Timestamp: 01:01:57 to 01:04:03

Asked what evidence would reverse their positions, Azeem Azhar states he would worry if US enterprise AI spending growth halved for two consecutive quarters. Conversely, Paul Kedrosky says he would abandon his bearish stance if macroeconomic data showed US productivity growth, labor participation, and GDP growth rising significantly above post-World War II trends.

Significance (Medium): Establishes clear, falsifiable macroeconomic and corporate spending benchmarks for investors tracking the health of the AI boom.

Sources in support: Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.), Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Neutral sources: Ezra Klein (Host, The Ezra Klein Show)

12. Book recommendations on technological upheaval

Timestamp: 01:04:03 to 01:06:38

To close the episode, Paul Kedrosky recommends George Eliot's Middlemarch, Emma Griffin's Liberty's Dawn, and Forrest McDonald's biography of utility tycoon Samuel Insull to illustrate how ordinary workers and empires experience technological revolutions. Azeem Azhar recommends James Burke's The Day the Universe Changed, Ivan Turgenev's Fathers and Sons, and Voltaire's Candide.

Significance (Low): Offers listeners historical and literary frameworks for understanding how societies absorb the psychological and financial shocks of new general-purpose technologies.

Sources in support: Paul Kedrosky (Partner at SK Ventures and Fellow at M.I.T.), Azeem Azhar (Technology Entrepreneur and Author of Exponential View)

Neutral sources: Ezra Klein (Host, The Ezra Klein Show)

Kedrosky's Venture Capital Role (Low severity)

Type: Financial

Paul Kedrosky is a partner at the venture capital firm SK Ventures, operating directly within the technology investment ecosystem.

Significance: Active venture investors hold portfolio exposures and valuation sensitivities that can shape their public critiques of competing frontier tech capital allocation.