AI Buildout Economics
2 analysis reports
A constraint-led AI buildout screen inspired by Gavin Baker's watts-and-wafers framework. The model ranks 33 AI infrastructure and Mag 7 stocks using physical scarcity exposure, revenue growth and scale, price momentum, quality, valuation, and risk. The 9 July 2026 run puts TSM, NVDA, MU, ASML, KLAC, LRCX, AMAT, CRDO, AVGO, VRT, and GOOGL at the top of the research queue.
A bottom-up model of one 1GW GB200-class datacenter: ~$38-41B upfront capex and ~$7.1B/year to run; renting GPUs at $3.30/GPU-hour earns a 42.8% operating margin; the lab on top needs ~$31B/year of inference revenue per gigawatt at a 60% gross margin to justify the compute. The crux is the net revenue index — throughput per watt (H100 900k to B200 2.8M tok/sec/MW, 3.1x) rising faster than token price falls means revenue per gigawatt expands ~56% even as prices halve. All figures tie out to the author's sheet; the ARK $8.5B vs $7.14B annualized-cost divergence is flagged explicitly.