Plain English with Derek Thompson

Episode: BEST OF: This Is How the AI Bubble Could Burst
Date: January 27, 2026
Host: Derek Thompson
Guest: Paul Kedrosky (Investor, SK Ventures/MIT Center for the Digital Economy)

Overview

In this "best of" episode, Derek Thompson revisits his urgent and deeply-researched conversation with tech investor and analyst Paul Kedrosky. Together, they dissect the unprecedented scale of AI infrastructure spending in the US—currently amounting to $300–400 billion annually—and draw historical parallels with past technology bubbles like railroads, telecom, and broadband. Most strikingly, they warn about the fragile economics underpinning this boom: aging hardware, capital misallocation, energy strains, and highly opaque investment vehicles that could set the stage for an economic crisis. The episode balances cautionary analysis with optimism for AI’s actual transformative potential, especially in less-hyped, deep business applications.

Key Discussion Points & Insights

Setting the Stage: The Scope and Scale of AI Spending

• Historical Context:
  • Derek opens by stressing that AI is not a future technology; it is already the single most important economic driver of the moment, with spending comparable to, or exceeding, prior infrastructure booms.
  • “American tech companies will spend about 300 to $400 billion on artificial intelligence... more in nominal dollars than any group of companies has ever spent to do just about anything.” (Derek, 04:31)
• GDP & Markets:
  • Half of US GDP growth and over half of recent stock market gains are tied to AI-related investment.

Paul Kedrosky’s Thesis: What Makes this AI Boom Different?

• Concentration of Capital:
  • Spending is "incredibly concentrated"—not just in corporations but geographically (e.g., Northern Virginia).
• Unprecedented Nature:
  • Both historically huge and unique because AI infrastructure (like GPUs) has a much shorter useful life than railroads or fiber optic cable.
    • “The lifespan of a GPU is on the order of two and a half to three and a half years. This is nothing like the spending that's being done on railroads... or fiber.” (Kedrosky, 13:49)
  • Capital is going mostly to rapidly depreciating assets that must be replaced every 2–3 years, unlike century-old railroad tracks or decades-old telecom cables.
    • "[GPUs are] closer to bananas than anything... They're closer to bananas than steel." (Kedrosky, 16:02)
• Perverse Market Incentives:
  • Market rewards spending for spending’s sake, likening it to Bitcoin hoarding by companies.
  • “[The market] is rewarding you for doing this, even though it makes no economic sense to spend at this level, because there’s no way I can recoup the value…” (Kedrosky, 16:24)

How AI Spending Warps the Broader Economy

• Capital Redirection:
  • Echoes the “death star” effect of telecom in the '90s: capital is being sucked from traditional manufacturing (and potential onshoring) into AI, raising borrowing costs elsewhere and possibly hollowing out the rest of the economy.
    • "The same phenomenon… If I’m a small manufacturer… the hurdle rate just got a lot higher… because they're comparing me to this other part of the economy that will accept giant amounts of money..." (Kedrosky, 20:19)
• Investment Dynamics:
  • Large firms and investors prefer writing huge checks for data centers over many small ones for small manufacturers, adding another layer of inertia.

AI, Energy, and Community Tensions

• Rising Power Demands:
  • Data centers are driving up electricity prices and straining grids; utilities prefer large, reliable buyers (data centers) over many small residential ones.
    • "[Data centers] are seen as essentially... a very clear leverage on raising energy prices..." (Thompson, 27:01)
• Community and Political Pushback:
  • Early signs of NIMBY (“Not In My Backyard”) resistance; offshoring of data centers may increase as local communities push back.
    • "At night they would go outside their houses and they hear [a] hum and it's like, I didn’t sign up for…” (Kedrosky, 28:14)

The Mechanics of the Bubble: Opaque Financing and Risk

• Hyperscalers & Off-Balance-Sheet Moves:
  • Big tech is using Special Purpose Vehicles (SPVs) and other structures to keep AI spending off their books, making tracking real risk much harder.
    • “So what we’re seeing... is these SPVs... where I have a stake in it as Meta, some giant private debt provider... I don't own it, right?” (Kedrosky, 33:56)
    • “For me, that's the factor to watch. And it's just begun accelerating.” (Kedrosky, 34:41)
• Who’s Exposed When It Blows?
  • Not just hyperscalers and their private equity partners, but also:
    • Construction/industrial suppliers (e.g., air conditioning manufacturers)
    • REITs (Real Estate Investment Trusts): "Somewhere between 10 and 22% is already directly data center related… if you’re a conservative investor... you’re already in [the AI bubble]…” (Kedrosky, 39:28)
    • Insurance: Private equity has bought insurance firms to use their premiums as capital for data centers—creating asset-liability duration mismatch reminiscent of the 2008 financial crisis.
      • "Private equity and private credit have purchased insurance companies... the premiums get reinvested...and that's created a new source of risk." (Kedrosky, 48:24)

Signs and Symptoms of a Bursting Bubble

• What to Watch:
  • Hyperscalers increasingly stepping away from direct investment, leaving more to SPVs and private credit.
  • Missed earnings by industrial suppliers to data centers (signaling slowdown).
  • Data center rental rates aligning with or falling below cost due to GPU price compression.
• Cascade Effect and Opacity:
  • “There’s nowhere to run... It's a complex system, but there is a single point of failure... a couple of semiconductor stocks who are highly leveraged to everything that's going on and yet have metastasized across each of these pieces…” (Kedrosky, 42:48)
• Potential for Systemic Crisis:
  • Conditions align: over-leveraged sector, opaque funding, systemic lender exposure, narrative of “this time it’s different.”
    • “It absolutely does. It has all the pieces.” (Kedrosky, 48:24)

Counterarguments and Possible Ways the Bubble Doesn’t Burst

• Strong Free Cash Flow:
  • Tech giants have immense cash reserves and could theoretically weather years of heavy spending and shrinking margins.
  • Breakthroughs: If AI quickly proves vastly more profitable in business applications, leading to large returns, crisis may be averted.
• Where Kedrosky Might Be Wrong:
  • If rental margins on GPUs and data centers stay extremely high and stay subsidized longer than he expects; "The way I'd be wrong is that margin doesn't continue to decline. That even though it declines, it doesn't decline back to the point where it's no longer economically viable..." (Kedrosky, 53:57)

When Does It Go Sideways? How Soon?

• Timeline:
  • Projecting current trends, alignment of data center costs and rental revenues could become unsustainable in about 2–2.5 years.
  • “[A] naive projection... would put you into about two and a half years out where they're no longer earning a risk adjusted return commensurate with the cost of doing business as a data center..." (Kedrosky, 55:14)
  • Coincides, ironically, with the next US presidential election cycle (2028).

Beyond the Hype: Where AI is Actually Transformative

• Real Value Lies in the Mundane:
  • The biggest future impact is not in chatbots, but in deep business process automation and interoperability—language translation between business systems, data harmonization, etc.
  • “The most interesting applications… are at a deeper level. Think about, I’m a small manufacturer and I’m trying to bring on a bunch of new suppliers. All of them have different systems…” (Kedrosky, 57:32)
  • Automation and flexibility in the “language of business” can lower barriers, increase competition, and benefit the wider economy.

Notable Quotes & Memorable Moments

• "The lifespan of a GPU is on the order of two and a half to three and a half years. This is nothing like the spending that's being done on railroads... or fiber."
  — Paul Kedrosky, 13:49

• "They're closer to bananas than steel."
  — Paul Kedrosky, 16:04 (on how quickly GPUs lose value)

• "Huge capital is being sucked to one narrow part of the economy, just like we did in the 90s with telecom... the exact same thing is happening now."
  — Paul Kedrosky, 20:19

• "It's not like people want chat added to everything... The interesting stuff is all deep under the hood."
  — Paul Kedrosky, 59:03

• "It's a complex system, but there is a single point of failure... a couple of semiconductor stocks who are highly leveraged to everything..."
  — Paul Kedrosky, 42:48

• "It has all the pieces..."
  — Paul Kedrosky, 48:24 (on the ingredients for a financial crisis being present)

Timestamps for Important Segments

• 04:30 — Derek’s introduction on the scale of AI spending and bubble analogies
• 08:05 — Paul’s thesis: GDP growth, capital pools, and “crazy math”
• 12:43 — Fundamental differences between AI infrastructure and prior tech buildouts
• 16:02 — “Closer to bananas than steel” and implications for recouping investment
• 19:09 — How AI spending distorts labor and capital markets (the “death star” analogy)
• 23:54 — Big investors' preference for megadeals over many smaller bets
• 24:33 — AI’s impact on energy costs and early NIMBY tensions
• 31:27 — The mechanics of SPVs and off-balance-sheet financing
• 39:11 — Who loses if the bubble pops? Exposure across REITs and traditional portfolios
• 44:25 — What bubble warning signs would look like
• 48:24 — How insurance and private equity are intertwined in the risk
• 53:57 — Kedrosky’s best-case scenario for bubble skeptics
• 55:13 — When the math stops making sense: a 2–2.5 year window
• 57:32 — True potential of AI: deep process/business communication, not chatbots

Conclusion

Thompson and Kedrosky’s exchange is a sweeping analysis of both the risks and rewards of the current AI gold rush. By highlighting similarities and differences between previous industrial bubbles and today’s AI megaprojects, they make a strong case that current capital flows are both historic and potentially precarious. The episode invites policymakers, investors, and the broader public to recognize how deeply the AI buildout underpins current economic growth—and how a correction could ripple throughout not just tech, but the entire financial system. At the same time, Kedrosky’s optimism for AI’s “boring,” infrastructural applications suggests that, as with railroads and broadband, enduring value may outlast any crash.