AI Is Too Expensive

Source: Where’s Your Ed At Date: 2026-05-19 URL: https://www.wheresyoured.at/ai-is-too-expensive/

Summary

Zitron argues LLM infrastructure is economically unviable across three levels: hyperscalers ($800B+ invested, needing $3T AI revenue to break even), AI labs (Anthropic spending $3 per $1 inference revenue; both OpenAI and Anthropic burning billions annually), and enterprises (Stripe burning $94K/day in tokens; ServiceNow worried about annual Claude Enterprise budgets being exhausted five months in). Cites Anthropic March 2026 affidavit: $5B lifetime revenue against $10B spent. Anthropic gross margins dropped to 40%, OpenAI to 33%. Predicts token budget collapse when the first executive cuts spending. Published on I/O day.

Implications

• Feeds the token economics thread directly: Zitron’s $94K/day Stripe figure and ServiceNow budget exhaustion are the most concrete enterprise consumption data points yet. These are the demand-side numbers the thread has been missing.
• Counter-narrative timing: publishing the economics indictment on the day Google announces $5B more in TPU infrastructure is pointed. The supply-side bet ($800B+) and the demand-side skepticism run in the same news cycle.
• Feeds the bear case fractures thread: this piece is more data-grounded than the “AI’s Circular Psychosis” piece that drew Piper’s critique. If Zitron can hold this tone, the bear case re-consolidates around economics rather than fraud allegations.