# Token Economics of Inference A **token** is the unit a language model reads and writes, roughly three-quarters of an English word. Models are priced per million input and output tokens, so tokens are the unit of both usage and revenue. **[[Inference]]** is running a trained model to produce tokens. Every query, every coding-agent step, every document summary is inference. It is now the larger and faster-growing share of AI compute ([[Inference is Eating AI Compute]]). ## Where the money goes The value chain for one token: 1. **Model owner.** Earns a margin if the model is proprietary and only available through its API. 2. **Server.** Whoever runs the GPUs that generate the token: a frontier lab, a hyperscaler, a [[Neoclouds|neocloud]], an inference API company. 3. **Hardware.** Nvidia and other chip vendors, paid upfront. With a closed model, the owner captures layers 1 and often 2. With [[Open Weights]], layer 1 drops to zero: anyone can serve the model, providers compete on price, and the margin shifts to whoever serves most efficiently ([[Systems and Compiler Depth]]) and keeps GPUs busiest ([[AI Inference Infrastructure]]). ## Workloads vs revenue A Chinese open model running on an American cloud generates little or no revenue for the Chinese lab. What the lab gains is the workload: the usage, the feedback, the ecosystem of fine-tunes built on it. What frontier labs lose is pricing power at the API layer. ## Capacity is the other constraint Serving capacity is finite. If demand for a new model outruns available GPUs, providers queue or pause signups. This is the practical meaning of "capacity binds on both sides". Related: [[Hyperscalers]], [[Tokenomics and its evolution]], [[The Cross-Pacific AI Chessboard]]