The Benefits of Bubbles (stratechery.com)

The piece argues that while current AI spending looks like a bubble — citing OpenAI’s roughly $1.4 trillion in deals versus about $13 billion of reported revenue and widespread capex booms across hyperscalers — that bubble may nonetheless be productive. Drawing on Carlota Perez’s model (bubbles fund an “installation” phase) and Hobart & Huber’s idea of “inflection bubbles,” the author says speculative mania coordinates many actors to build both physical and cognitive capacity: infrastructure, protocols, talent and complementary services. Historical parallels include the dot‑com era’s overbuilt fiber and concurrent innovations (e.g., XMLHttpRequest, Linux/x86 stacks, Google’s commodity-scale architecture) that enabled decades of growth long after many firms failed. For AI/ML the stakes are mixed. A lot of capital has flowed into GPUs (accelerating Nvidia to ~ $5T valuation), but chips are short‑lived assets (typical hyperscaler depreciation ~5 years), so GPU spending risks being transient. More durable outcomes are plausible: new fabs (TSMC, Samsung building in the U.S.) and renewed domestic semiconductor capacity, plus government support for firms like Intel, could create long‑lived supply chains and compute capacity. Bottom line: an AI bubble will probably overshoot and correct, but — if investment focuses on lasting manufacturing, shared platforms and talent — it can produce the infrastructure and coordination that unlock a genuine “deployment” period for AI.