Analyzes the shifting economics of AI inference driven by agent harnesses and decentralized labs.

Inference is the New Oil

10x cheaper inference cost BUT usage increases by 30x

The rise of AI agent harnesses (OpenClaw, Hermes) drastically increased the inference usage — billions of tokens burnt everyday (and not stopping anytime soon)

Anthropic banned the usage of harnesses with Claude subscription, saving compute/inference for enterprise clients + for training Mythos (the next frontier model)

Open-weight labs experience much higher demand due to cheap/efficient inference, optimized for long running workflows

Decentralized AI labs found initial PMF from offering token subsidy-driven inference infra + private AI moat. Token alignment & unique perpetual credits further lock-in users

The economics of AI across Closed AI labs, Open-weight labs, and DeAI labs are changing fast. 

"Gone are the days when players compete on intelligence alone"

Today, it's all about feeding a massive fleet of token-hungry agents while anchoring a customer base that can’t afford to leave

Read the Full piece on Substack

[Link in bio]