Reference
AI compute economics glossary
Plain-English definitions of the terms that actually decide what AI compute costs — the metrics, constraints and trade-offs behind $/M-token. 21 terms.
The price of running a model, per one million tokens.
How compute is procured — on-demand, reserved, or owned.
The volume where self-hosting beats paying a hosted API per token.
Reusing a processed prompt prefix, billed at a discount.
Mintok’s 1–8 scale of how capable a model is — separate from its price.
The maximum tokens a model can consider at once.
The dollar cost of a single agent or task invocation.
The emissions/energy twin of $/M-token.
Grams of CO₂ emitted per kWh of grid electricity.
The fast on-package memory on an AI accelerator — capacity and cost driver.
The fraction of an accelerator’s peak memory bandwidth actually used.
The fraction of an accelerator’s peak compute actually used.
An architecture where only a subset of parameters activate per token.
The two phases of LLM inference, with different bottlenecks.
Datacenter overhead multiplier — total power ÷ IT power.
Running a model at lower numeric precision to cut memory and cost.
The itemized component list and cost of an AI server rack.
A model’s scale bucket — Flagship, Balanced, Fast, or Nano.
The all-in lifetime cost of running compute, not just the sticker price.
Steady-state generation speed during decode.
How long from request to the first output token.
Definitions maintained by Mintok. See the methodology for how these feed the sizing and cost models.