How the AI Bubble Will Pop (www.derekthompson.org)

AI’s infrastructure build-out is being cast as a classic speculative bubble: companies are pouring unprecedented capital into data centers and chips—tech firms are projected to spend ~$400B this year and potentially >$500B annually by 2026–27—while end-user market receipts remain tiny (U.S. consumers spend ~$12B/year on AI services). Analysts highlight striking warning signs: blockbuster seed rounds for pre-product startups (e.g., a $2B seed at a $10B valuation), momentum-driven equity flows detached from fundamentals, falling enterprise AI usage in some firms, and financial sleights-of-hand (SPVs and accounting shifts) that downplay real infrastructure costs. The technical and economic implications are acute. For hyperscalers, GPUs make up roughly 60% of data‑center costs, with cooling/energy the other large component—so chip supply and electricity are systemic choke points, concentrating capital and geography (Northern Virginia and a few other hubs). The allocation mirrors past infrastructure manias (telecom in the 1990s): huge centralized investments can siphon capital from manufacturing and other sectors, raise hurdle rates, and create fragility tied to a handful of chipmakers (e.g., Nvidia). Expect political pushback (NIMBY, environmental concerns), offshoring of data centers to lower-cost regions, and a painful market correction if revenue growth fails to justify the ongoing "Apollo-scale" spending cadence.