The Training Tax
Security Boulevard, Monday, July 20th, 2026
AI training costs are small next to inference expenses, which compound with user growth and dominate spending.
While training a frontier model costs roughly $100 million, serving it at scale costs hundreds of millions monthly.
At ChatGPT-level usage with 100 million daily users, inference spending overtakes training cost in nine days and reaches 97.7% of total expense within a year. Price cuts and efficiency improvements matter far less than the rate of user growth.
The conclusion is that success in AI deployment depends on controlling inference cost, making serving efficiency the real competitive moat rather than training budget.