Podcast Summary

Plain English with Derek Thompson

Episode: "Yes, AI Is a Bubble. There Is No Question."
Date: March 17, 2026
Host: Derek Thompson
Guest: Paul Kedrosky

Episode Theme

Derek Thompson revisits the provocative question: Is AI in a financial bubble? Joined by investor and writer Paul Kedrosky—whose initial "AI bubble" thesis he had previously adopted—Derek explores bubble dynamics, technological infrastructure cycles, the latest shifts in the AI market, and the implications for workers, companies, and investors. The conversation uses historical analogies (like railroads and fiber optics) to examine whether current AI investment is sustainable—or headed for a reckoning.

Major Discussion Points & Insights

1. Defining the AI Bubble: Theory and Historical Parallels

• Carlotta Perez’s Framework: Derek introduces Perez’s idea that technological revolutions follow a pattern of over-exuberance (overbuilding, overinvesting), a crash, then a productive golden age. (00:47)

• AI Investment Numbers: The sector is seeing blockbuster investment—approximately $700 billion/year on AI chips, data centers, and related infrastructure. Thompson likens this to “one Manhattan Project every three to four weeks” (03:30).

• Bubbles ≠ Pointless Tech: Both Thompson and Kedrosky stress that calling something a bubble doesn’t mean the underlying tech is useless. Railroads, radio, fiber optics, and the internet were all bubbles with world-changing impact. (04:10, 09:36)

  “Railroads were a very good idea… But that didn’t prevent people from overfunding startup… railroads. … [Half] of those track miles were eventually abandoned. Does that mean that railroads were a bad idea? No, we just wildly overbuilt…”
  — Paul Kedrosky (12:44)

• Rational Bubbles:
  Kedrosky introduces the concept of a "rational bubble"—where every participant's incentives lead to collective overbuilding, even though individually each decision seems reasonable. All necessary ingredients (loose credit, tech fervor, government policy, real estate) are present at once for AI, making the moment historically fragile. (11:34)

2. Recent Changes & Evidence of a Bubble

• Stock Market Reversal:

  • The “MAG7” tech stocks (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, Tesla) had dominated returns, but in early 2026, they all turned negative year-to-date, even as the S&P 500 hovered near flat.

  • This signals a change in how markets value AI-related CapEx—initially market cap soared with AI investment, but now such spending actually drags down valuations, a “gestalt shift” in sentiment. (18:47)

    "Two years ago, if I added a dollar of AI CapEx, the market rewarded me with two dollars of market cap... About six months ago... it became neutral... then... it reversed completely... The market started taking away market cap."
    — Paul Kedrosky (18:47)

• Persistent Infrastructure Spend:
  Despite market punishment, tech giants continue to ramp up AI infrastructure spending. Game theory: “If I’m the first one to stop spending, I lose the chance to be a future consolidator.” (22:51)

• Rising CapEx Consumes Free Cash Flow:
  These investments are now eating larger slices of even the biggest tech companies' free cash flow—at the expense of activities like share buybacks that offset stock-based compensation.

  • More spending is financed by debt, especially private credit, now under pressure. (25:33)

3. Winners, Losers, and Market Flows in the AI Wave

• SaaS-pocalypse:

  • The new “commodity” of tokens (units of AI compute) is in a deflationary spiral; their price drops 70-90% per year, unlike standard commodities.

  • SaaS (software as a service) companies that rely on selling digital info products are being walloped, as tokens lower barriers to entry and eat into margins and moats. (31:25)

    "We're having the emergence for the first time in 100 years of a new industrial commodity... tokens are in a deflationary spiral… The SaaS apocalypse is the market... saying, I see now all the people this could hurt, because it reduces moats, lowers barriers to entry, and changes the economics of their business."
    — Paul Kedrosky (31:25)

• Value Flowing to Energy (& Infrastructure):

  • Energy stocks outperform as AI’s voracious compute appetite drives demand for more power, data centers, and supporting infrastructure.
  • Hyperscalers, unable to rely solely on the grid, are building private “behind the meter” natural gas plants. (35:43)

4. Misunderstandings and Benchmark Debates in AI Progress

• Productivity Illusions:

  • Recent US productivity spikes are not due to direct AI efficiency gains, but mostly to parabolic CapEx—which boosts GDP math without real output growth. (42:25)

    “The reason why productivity is rising has nothing to do with AI. … as I increase [AI investment], if hours worked stay constant, productivity increases, but doesn’t mean we’re at all more productive. It just means we’re spending a lot of money on AI.”
    — Paul Kedrosky (42:25)

• The Orchestration Layer is Key:

  • Recent “aha” moments in AI utility come less from model breakthroughs and more from improved orchestration layers (e.g., Anthropic’s Claude Code; OpenAI’s Codex).
  • These “nannies” organize the sprawling, sometimes bratty behavior of AI models, making them usable for complex, multi-step knowledge work. (44:40)

• Benchmark Disputes:

  • Debate over whether models are still improving quickly:
    • Kedrosky cites composite benchmarks (e.g., Epoch) showing year-over-year model improvement slowing (from 12%+ to 2-3%), even as cost and training times soar.
    • Derek points out some other benchmarks show progress has accelerated recently. (46:00–50:29)

5. Is There an Escape from Bubble Logic?

• The 'Token Consumption' Argument:

  • Derek floats a “vibes-based” hypothesis: if millions of knowledge workers use AI “agents” all day, token demand (and thus compute spend and revenue) will soar, potentially shrinking the gap between investment and revenue.

  • Kedrosky rebuts: Coding is weirdly unique; code generation is deterministic and expansionary (generates lots of new tokens), while most white-collar work is compressive (summarizing, reducing info). So you can’t extrapolate coding agent demand to all knowledge workers. (54:43)

    "The idea that by bringing AI to white collar workers, I can duplicate the rampant token consumption… is a misunderstanding of the nature of this commodity... White collar work is compressive, not expansive."
    — Paul Kedrosky (54:43)

• Jevons Paradox in AI?:

  • If tokens get super cheap, will new, token-hungry use-cases emerge? Kedrosky: that’s speculative and doesn’t rebut current bubble logic. (57:30–58:55)

6. Chips: Are They Bananas or Steel?

• Debate on AI Hardware (Nvidia):
  • Kedrosky previously argued that GPUs have fast obsolescence (“closer to bananas than steel”). Recent evidence shows chips like Nvidia’s H100s retaining surprising value.
  • He responds: It depends on how chips are used; those not used for frontier model training last longer. But Nvidia faces new low-cost competition in inference (companies like Talas); artificial scarcity can’t last. Over-ordering by AI firms could lead to sharp corrections. (61:02–65:31)

7. How the Bubble Might Pop—and Who Wins After

• Bubble Pop Mechanisms:

  • Credit market stress: private credit firms exposed to SaaS and data centers could be forced to “gate” redemptions.

  • Sharp repricing at suppliers like Nvidia if orders turn out to be egregiously inflated.

  • Token “dumping”—as profits evaporate, token prices may collapse, further eating into SaaS and tech margins.

  • Software engineering jobs (2–3 million in US) most keenly threatened by advanced AI/automation. (65:31–67:24)

    "I think it just is a sharp repricing of companies... because their order book turns out to be 2 and 3x times overordering... we've seen this over and over again."
    — Paul Kedrosky (65:31)

• Winners After the Bust:

  • Not obvious that new tech winners will emerge within AI itself (unlike Amazon and Netflix after the dotcom bust).
  • Tech megacaps are becoming like utilities (high CapEx, low/no growth), so value shifts out.
  • "HALO" (Heavy Asset, Low Obsolescence) firms—manufacturers, grid suppliers, waste management—are outperforming and may absorb more value as tech becomes less profitable. (68:52)

Notable Quotes & Memorable Moments

• On the duality of bubbles:

  "Sometimes bad people, in the process of crashing the US economy over and over again and ruining people's lives, nonetheless build technologies that in the long run we can't imagine modern life without."
  — Derek Thompson (16:52)

• On the insanity of current spending:

  "It's one Manhattan Project every three to four weeks. It's one Apollo Program every five months. This is an insane amount of money for the private sector to put into anything." — Derek Thompson (03:30)

• On market logic, the point-of-no-return for Big Tech:

  "If I'm the first one to announce that I'm no longer spending on AI Capex, what's going to happen to me in the market? The market's going to say, well, you just folded up … so you have no choice. Even though it has negative value, the consequences of actually withdrawing from the game are even worse."
  — Paul Kedrosky (22:51)

• Paul’s tongue-in-cheek rename for the AI mega-stocks:

  "Just because Mag7 was coined after an old Western, the Magnificent Seven, I flipped it and I said, okay, they're now the Hateful Eight."
  — Paul Kedrosky (20:25)

• On “tokens” as a new class of commodity:

  "The new industrial commodity that's emerging… is this thing called tokens... tokens are in a deflationary spiral. … That's not like copper… This is a very unusual commodity."
  — Paul Kedrosky (31:25)

• Derek’s summary of the eat-yourself era:

  “Software eats the world. Software becomes the world. Software eats itself.” — Derek Thompson (70:22)

Timestamps for Key Segments

• [00:47] – Tech bubbles in historical perspective: Perez, railroads, internet
• [09:36] – Kedrosky lays out his “AI is a bubble” core argument
• [12:44] – Railroad analogy, repeat patterns
• [18:47] – Stock market's shift in attitude to AI CapEx
• [22:51] – Spending continues anyway: game theory in action
• [25:33] – Debt and private credit now fueling CapEx amid cash flow pinch
• [31:25] – SaaS-pocalypse and “tokens as commodity”
• [35:43] – Energy’s boom, infrastructure tailwinds
• [42:25] – Productivity illusions and macro misunderstanding
• [44:40] – Orchestration layers and the true drivers of recent AI gains
• [46:00–50:29] – The benchmark debate: Are models still improving?
• [54:43] – White-collar AI usage won’t mirror coders
• [61:02] – Are GPUs “bananas or steel”? The Nvidia debate
• [65:31] – How the bubble could pop: overordering & market corrections
• [68:52] – Who wins in the long run? The “HALO” thesis

Overall Tone and Takeaways

• Analytical, conversational, wryly skeptical—balanced between historical perspective and appreciation that bubbles can drive lasting societal change despite near-term carnage.
• Strong focus on historical analogy, empirical trends, and wariness against hype.
• Acknowledges uncertainty and “nobody knows anything” as a mantra, yet lays out clear, data-driven reasons for skepticism about indefinite AI expansion.

For Listeners Who Haven’t Tuned In

This episode is a lively deep-dive into not just whether AI is a bubble—but how, why, and what that means. Expect an engaging blend of economic theory, business history, tech skepticism, market dynamics, and speculation about future winners and losers in the age of industrial AI. It’s accessible to laypeople but richly detailed for industry watchers.