Interactive AI token cost and carbon map

In the recent report Moving AI inference can cut carbon at little extra cost, Uptime Intelligence examines how the location of AI inference affects its cost and carbon footprint.

The report shows that, for the same infrastructure, changing the location of the inference workload affects carbon per token far more than it affects cost per token. Across 40 countries and regions, carbon per token varies by several hundred times, while cost varies by about 15%, because hardware capital accounts for most of the cost of a token and power for only a small share. Variation of the grid emission factor — which measures the amount of greenhouse gas emissions released per kWh —in large countries can be as significant as the variation between them. In practice, many operators have limited choice: interactive inference needs to be close to users, and sovereignty requirements may keep data and processing within national borders. Workloads that are less sensitive to latency, however, can move to locations with cleaner, cheaper power.