Computation connects algorithms to processors, memory and electricity. Different stages of an AI system can face different bottlenecks. Training a model and responding to many users are related activities with distinct resource demands.

Ask how much useful work is delivered for the resources consumed. Efficiency, availability and total demand should be considered together. A claim about scaling needs explicit assumptions about hardware, software and the infrastructure that supports them.

Sources & further reading

Links point to original reporting or primary research.

  1. International Energy Agency · Energy and AI (2025)
  2. Vaswani et al. · Attention Is All You Need (2017)