The Economics of AI: Token Costs and the Real Price of Intelligence

Artificial intelligence is sold as magic. The invoices are not. Every prompt, every retrieved document, every agent loop, and every retry converts directly into a metered quantity of tokens, and tokens are now the dominant variable component in the cost of goods sold for software companies that have embedded generative features into their products. ICONIQ's July 2026 data places the average gross margin on an AI product at a projected 53% for 2026, up from 45% in 2025 and on track for 59% in 2027, still far below the 80% benchmark that defined mature software-as-a-service, and public companies disclosing model-driven cost pressure are reporting margins ten to seventeen points below their pre-AI baselines. Inference accounts for roughly 23% of total AI product cost at the scaling stage, second only to talent at 26%.

Research library