Rohit Krishnan drops a cold take on AI economics: falling inference costs will explode usage but won't translate to proportional revenue growth for AI companies. The core issue is elasticity—price drops 10x, usage might go up 100x, but margins compress and total revenue doesn't scale linearly.

The math problem: if intelligence becomes dirt cheap, you can't charge premium rates. Anthropic might hit $500B annual revenue in a best case, but the $5T bull thesis assumes pricing power that won't exist in a commoditized inference market.

This is the classic cloud infrastructure trap. AWS made compute cheap and ubiquitous, captured massive usage, but margin compression kicked in as competition intensified. AI providers face the same dynamic—great for developers and the economy, brutal for unit economics at scale.

The real money might not be in serving tokens but in the application layer where differentiation still exists. Raw intelligence as a commodity vs intelligence applied to specific high-value problems.