Podcast Summary: This Is How the AI Bubble Could Burst

Plain English with Derek Thompson | The Ringer | September 23, 2025

Episode Overview

Derek Thompson explores the possibility of a major economic bubble forming around AI infrastructure spending. Joined by investor and writer Paul Kedrosky, the conversation dives deep into how unprecedented capital is being poured into AI tech, the historical parallels with previous tech-driven booms and busts, why AI infrastructure is inherently different, and what could trigger a dramatic reversal—what a “burst” would look like. The discussion maintains a balanced tone, warning about bubble dynamics while recognizing AI's long-term transformative potential.

Key Discussion Points & Insights

1. The Unprecedented Scale of AI Spending

• Annual Spending: US tech firms are deploying $300–400 billion per year on AI infrastructure—surpassing any other technological buildout in history. (02:02)
  • Derek: “These companies are not anywhere close to earning back that $400 billion... That’s why you’re starting to hear some people wonder if the AI buildout is turning into the mother of all economic bubbles.”
• Comparison to Historic Booms: Previous bubbles included railroads (19th century), radio, automobiles, aviation, the telegraph, and broadband. Many had massive crashes. AI spending is comparable if not even more concentrated and aggressive. (03:50–04:20)

2. Physics of the AI CapEx Boom

• Where the Money Goes:
  • About 60% of a data center’s cost is GPUs (graphics chips).
  • The rest: energy (power), cooling systems, and construction. (09:29)
  • The lifespan of GPUs is short (~2.5–3.5 years), making them more “like bananas than railroads.” (14:58)
    • Kedrosky: “They’re closer to bananas than anything else.” (14:56)
• Historical Difference: Rail lines and fiber optics were durable assets; if demand lagged, value could be realized years later. GPUs depreciate rapidly—if demand is delayed, the assets are nearly worthless. (11:36–14:08)

3. Economic and Labor Market Effects

• AI Creating a Gravity Well: AI isn’t just driving tech—it’s warping the entire American economy. Tech jobs are shifting toward AI, other sectors like manufacturing/construction are declining as capital pours into data centers. (17:04–20:31)
  • Thompson: “AI is like this star that is pulling in all these resources gravitationally from throughout the economy.” (17:39)
• Onshoring “Paradox”: Attempts to revive US manufacturing (e.g., tariffs on Chinese goods) are being undermined because capital prefers big, concentrated AI investments, leaving small manufacturers in the lurch. (20:31–22:48)

4. Energy: A Brewing Political Flashpoint

• Energy Inflation: Data centers’ voracious power needs are pushing up grid demand and energy prices, with utilities preferring large, reliable corporate buyers. (23:27)
• Community Pushback: Residents in rural/exurban areas find themselves surrounded by humming data centers, leading to not-in-my-backyard (NIMBY) reactions and possible political opposition. (26:13)
  • Kedrosky predicts more data centers will be offshored and/or face local resistance soon. (26:13–27:46)

5. Signs of a Bubble, and What Might Pop It

• Financial Engineering:
  • Big tech firms (Meta, Google, Microsoft) are increasingly using opaque off-balance-sheet vehicles (SPVs) to finance data centers, reminiscent of CDOs and other opaque instruments of 2008 fame. (29:57–33:10)
    • Kedrosky: “What I’m watching is how they’re moving the financing off balance sheet... For me that’s a reflection of, I don’t want the credit-rating agencies to look at what I’m spending.” (31:07)
• Who’s Exposed:
  • Not just hyperscalers and PE firms—construction, air conditioning, REITs (real estate income trusts), and by extension conservative retirees/investors are all tied to the AI boom. (37:40–42:47)
    • “If you’re a conservative investor with a REIT in your portfolio... Go have a look inside your REIT, see what’s actually in there. You’re soaking in it.” (38:48)
• “No Place to Hide”: Because AI data centers now account for up to 20% of major REITs’ assets, even cautious investors are tied to the bubble—often without knowing it. (41:17)
• Classic Bubble Tells: Growing use of SPVs, lengthening supply chains for supporting industries (air conditioning, interconnects), and increasing opacity in financial reporting. (42:54)
  • “You never know if you’ve spent enough on capital till you’ve spent way too much.” (29:57)
• Insurance and the Temporal Mismatch: Private equity now owns insurance firms, funneling policyholder premiums into volatile investments (data centers) that don’t match the long-term liabilities of insurance—echoing the classic “borrow short, lend long” structure that exploded in 2008. (46:54)

6. How and When Might the Bubble Burst?

• Trigger: When rental income for GPUs/data center space declines below sustainable levels—Kedrosky estimates this crunch could arrive ~2–2.5 years out (i.e., 2027–2028). (53:44)
  • “It’s hard to imagine, absent something changing materially... that it doesn’t happen even faster than that.” (54:08)
• Cascading Effects: Pullback would ripple from chipmakers (Nvidia, etc.) through REITs, insurance capital, private credit ETFs, and the broader market.
• Timing & Politics: The projected crunch aligns with the 2028 US Presidential Election cycle, raising the risk of tech woes crossing over into political chaos. (55:01)

7. Counterarguments and Bullish Possibilities

• High Margins—For Now: Some argue that even with declining rental rates, data centers’ current margins are robust. If demand stays high, firms might absorb years of high CapEx. (50:00–53:07)
• Potential for Breakthroughs: If AI applications at scale deliver new multi-billion-dollar revenue streams (e.g., “the Google for everything”), returns may catch up with spending. But so far, usage and adoption have lagged CapEx. (48:50–53:07)
• Subsidized by Investors: The current model is sustained by venture and private capital, not actual commercial demand.

8. Where AI’s Promise Still Lies

• Transformative Potential: The fact that AI is in a classic infrastructure bubble does not mean it lacks significance—historical bubbles often precede genuinely world-changing technologies.
• Most Interesting Use Cases: While consumer-facing chatbots get headlines, the most transformative AI applications are “mundane”— automating the “grammar of business,” connecting suppliers and making markets more efficient by handling messy, boring tasks humans do today. (56:01)
  • Kedrosky: “People don’t want chat added to everything... the interesting stuff is under the hood, messy, boring. The language of business talking to each other... and that’s the stuff that this is tremendous at.” (57:19)

Notable Quotes & Moments

• On the Nature of AI CapEx:

  “They’re closer to bananas than steel.”
  — Paul Kedrosky (14:58)

• On Bubble Logic:

  “You never know if you’ve spent enough on capital till you’ve spent way too much.”
  — Paul Kedrosky (29:57)

• On Financial Engineering Risks:

  “What I’m watching is how they’re moving the financing off balance sheet... For me that’s a reflection of, I don’t want the credit-rating agencies to look at what I’m spending.”
  — Paul Kedrosky (31:07)

• On Hidden Risk in Retirement Portfolios:

  “If you’re a conservative investor with a REIT in your portfolio... Go have a look inside your REIT, see what’s actually in there. You’re soaking in it.”
  — Paul Kedrosky (38:48)

• On Policy Error and Misreading the Economy:

  “Now we’re like that barking dog in terms of understanding the drivers of economic growth right now. We think it’s because of tariffs... it’s being driven by completely different things.”
  — Paul Kedrosky (45:14)

• On Dead-End AI Uses:

  “We got sucked into the idea of [AI that] sounds so much like us... That’s kind of a dead end... the interesting stuff is all deep under the hood, messy, boring.”
  — Paul Kedrosky (57:19)

Timestamps for Crucial Segments

• The AI Spend in Historical Context: 02:02–06:50
• Where the Money Goes & Why GPUs Are Not Railroads: 09:00–14:08
• AI and Energy, NIMBYism, and Political Fallout: 22:48–27:46
• Opaque Financing and Bubble Symptoms: 29:57–33:10; 34:15–36:13
• Who’s Exposed to the Crash (REITs, retirees): 37:40–42:47
• Classic Bubble Headlines and Underlying Causes: 42:54–45:14
• Insurance, Private Equity, and Timing Mismatch: 46:54–48:50
• How and When It Might Burst: 53:07–55:01
• The Real Promise of AI: 56:01–58:31

Tone & Takeaways

Despite dire warnings, both Thompson and Kedrosky believe that AI, like the railroad or internet before it, will still transform the world—even if a spectacular bubble and painful fallout come first. The real risks, they argue, are not only financial but political and social, as AI’s gravitational pull distorts American economic realities and populates everyday investment vehicles with hidden, highly correlated risks.

For listeners:
This episode provides an unvarnished, historically informed look at today’s AI boom, drawing clear lines from economic theory to present-day policy and portfolio risk. The conversation’s plain language, memorable analogies, and deep technical grounding make this essential listening—or reading—for anyone interested in technology, markets, or the future of the US economy.