AI’s Economics Don’t Make Sense [Ad Free]

Source: Where’s Your Ed At Date: 2026-04-28 URL: https://www.wheresyoured.at/ais-economics-dont-make-sense-ad-free/

Summary

Ed Zitron’s detailed analysis argues that AI subscription products are structurally loss-making: Microsoft lost $20+/user/month on Copilot while charging $10, and Anthropic’s token burn rates meant users were consuming $8–$13.50 of compute per $1 of subscription revenue. At infrastructure scale, 100MW data centers yield only ~16.7% gross margin after depreciation, and Oracle’s Stargate Abilene commitment requires OpenAI to generate $852B in revenue and funding through 2030.

Implications

• Feeds the token economics thread. The fixed-fee/variable-cost mismatch is the central fragility: when subsidies end and billing shifts to consumption, user defection becomes the most likely outcome. The Anthropic $150–250/month Claude Code figure Zitron surfaces suggests even developer tooling is being cross-subsidized.
• The infrastructure commitment overhang is the most concrete risk. Oracle’s $115B+ debt tied to OpenAI demand is the systemic lever — not the AI companies’ own balance sheets but their cloud partners’ debt exposure. A demand miss doesn’t just hurt the lab; it strains a major cloud provider.
• Counter-signal to the Anthropic 80x growth report. Growth at that speed can mask per-unit economics that don’t improve with scale if the underlying compute cost curves don’t compress faster than pricing pressure.