Plain English with Derek Thompson Plain English BEST OF: This Is How the AI Bubble Could Burst
54 snips
Jan 27, 2026 Paul Kedrosky, investor and writer known for VC work and tech-economics analysis. He explains why AI data-center CapEx could form a historic bubble. He breaks down GPU-driven costs, short asset lifespans, opaque SPV financing, and grid and private-credit risks. He flags warning signs and sketches how AI could still yield practical, revenue-generating uses.
AI Snips
Chapters
Transcript
Episode notes
AI Spending Is Economically Massive
- AI infrastructure spending in 2025 is unprecedented and large enough to materially affect GDP growth.
- The spending is extremely concentrated in chips, a few firms, and small geographies like Northern Virginia.
GPUs Are Short-Lived, Not Steel
- GPUs make up roughly 50–60% of a data center's cost and they depreciate rapidly over ~2.5–3.5 years.
- That rapid depreciation makes AI CapEx fundamentally different from long-lived infrastructure like rail or fiber.
Market Rewards Unsustainable CapEx
- Markets currently reward hyperscalers for huge AI CapEx even when it makes little economic sense to spend that much.
- Companies then try to present the spending as less risky by emphasizing the long-lived 'shell' instead of short-lived GPUs.