OpenAI and Anthropic are each running at roughly $30B annual revenue, equal to 0.1% of US GDP apiece. Add cloud and adjacent AI services and the sector has gone from near zero to 0.25-0.5% of GDP in a few years. If both labs hit $100B by end of 2025 as widely speculated, AI will represent approximately 1% of US GDP run rate by end of 2026. The original piece digs into whether this growth gets mismeasured the way internet productivity was in the 2000s, and what that measurement failure could mean for regulatory responses that only see the job losses and miss the gains.
Memory supply from Hynix, Samsung, and Micron is now a binding constraint on compute buildouts for at least the next two years. Gil argues this creates an artificial asymptote on model capability gains, with no single lab able to break significantly ahead until 2028 at the earliest. That constraint effectively locks in the current LLM oligopoly. A separate dynamic is worth reading in full: Meta's aggressive researcher compensation triggered what Gil calls a distributed IPO across the top labs, making a few dozen to a few hundred researchers post-economic all at once, with the behavioral distortions that typically follow a liquidity event.
Tokens are functioning as a unit of economic denomination in the same way bandwidth did in the late 1990s. Cursor subsidizes inference as a user acquisition strategy. Allbirds, a shoe company, just raised a convertible note to build a GPU farm. The piece does not resolve where this goes. It raises the right questions: what happens to compute depreciation cycles when new supply is constrained, whether algorithmic breakthroughs contained inside one lab could still produce a runaway leader, and whether the compute ceiling extends well past 2028. Worth reading for the framing, not the conclusions.
[READ ORIGINAL →]