Dark Tokens

Dark tokens are the source’s label for AI usage that is economically real but not visible as revenue at Anthropic, OpenAI, or another closed model lab. In More Trillion Dollar IPOs, Anthropic $3T, Zuck’s Price War, China Ends Open Source?, Trump Accounts, Chamath argues that open-source or self-hosted model calls may show up as compute demand on Nvidia, neocloud, or infrastructure balance sheets rather than as AI-lab revenue.

The concept qualifies simple duopoly claims. If revenue is concentrated in a few frontier labs, those labs may still dominate monetization; but if large volumes of open-model inference are hidden inside customer compute, cloud usage, routers, or internal platforms, revenue share can overstate closed-lab usage share.

Key Claims

• Dark tokens make AI Revenue Legibility harder because usage may be visible only through indirect infrastructure demand.
• Open weights, self-hosting, and routing can shift economic value from model APIs to hosting, chips, power, memory, and orchestration.
• The concept does not prove closed labs are weak; it says revenue concentration and usage concentration are different measurements.
• Dark tokens can strengthen Closed Model API Moat Pressure when customers can get acceptable results from cheaper or locally deployed models.

Connections

• Open Source AI Models, Model Routing Cost Control, AI Inference Cost Structure, and Closed Model API Moat Pressure - model-choice and economics branch.
• Nvidia, neoclouds, OpenRouter, and MaaS Infrastructure - infrastructure layers where dark-token demand may appear.
• Anthropic, OpenAI, and AI IPO Valuation - companies and valuation frames affected by usage/revenue interpretation.