Every bit of AI jargon used across this collection, explained simply — no prior knowledge assumed. If a word in one of the sheets tripped you up, it's almost certainly defined here. Skim a section, or jump straight to what you need.
The little chunk of text an AI model reads and writes. Roughly ¾ of a word — "tokenization" might be 3 tokens. Everything (price, speed, energy) is counted in tokens.≈ syllables for a machine.
Input = the text you send (your question + any context). Output = the text the model writes back. On paid plans, output usually costs several times more than input.
Whatever you type or send to the AI — a question, an instruction, a document to summarize. The model responds to the prompt.
The maximum amount of text (in tokens) a model can consider at once — your prompt plus its reply plus the conversation so far. Long chats and big attachments fill it up.≈ the model's short-term memory.
The big AI model behind chatbots like Claude, ChatGPT, and Gemini. Trained on huge amounts of text to predict and generate language.
A smaller, faster, cheaper model. Modern small models match the big flagships of ~18 months ago on most everyday tasks — which is why "right-sizing" saves so much.
The model's internal adjustable values — a rough proxy for its size and capability. "qwen3:8b" means 8 billion parameters. More isn't always better for a given task.
A provider's most capable (and usually priciest) model — e.g. Claude Opus, GPT-5.5, Gemini Pro. Smaller siblings trade some smarts for speed and cost.
The act of running a trained model to get an answer. Every chat reply is one inference. This is what you pay for and what draws power day-to-day.Seen in: AIBox energy, demand vs. supply
The one-time, very expensive process of creating a model from data — done once by the provider. After that, the model is used over and over via inference.
Models whose files are published so anyone can download, run, or self-host them (e.g. Meta's Llama, DeepSeek). The opposite of closed, API-only models.Seen in: pricing sheet, AIBox
The component that chops text into tokens. Different models count differently, so the same sentence can be more or fewer tokens depending on the model.
A subscription like Claude Pro or ChatGPT Plus (~$20/mo). You pay a fixed price and get an allowance of usage — you never see tokens, just limits.≈ an all-you-can-eat buffet.
The developer doorway into a model. Instead of a chat app, software connects directly and is billed per token used.≈ a utility meter.
How API pricing is quoted: dollars per million tokens. E.g. "$5 / MTok input" = $5 for every million tokens you send.
A ceiling on how fast you can send requests (tokens or requests per minute) — not a bill, just a pace limit so the service stays fair and stable.
If you send the same context repeatedly, the provider can cache it and charge up to 90% less on those repeated parts. A big cost saver for apps.
Submit a pile of work to run later (within hours) instead of instantly, for roughly half price. Great for non-urgent jobs.
Choosing the smallest model that still does the job. Using a giant model for a trivial task wastes money and power; matching model to task is the single best optimization.
Consumer plans are fixed — the bill can't surprise you. API spend is variable — it scales with how much you use, which is ideal for automation.
How fast something uses energy right now. A 60 W idle box is drawing 60 watts continuously. Like speed (mph), it's an instant reading, not a total.
A watt-hour is one watt for one hour. A kilowatt-hour (kWh) is 1,000 of those — the unit on your electric bill. Power × time = energy.Watts = speed; watt-hours = distance traveled.
One billion kWh — country-scale energy. Used to describe how much electricity all data centers consume (~1,000 TWh/yr globally).
A measure of climate impact: grams of carbon dioxide (or its warming equivalent) released. The carbon "cost" of the energy used.
How much carbon your local electricity carries, in gCO₂ per kWh. Solar/nuclear ≈ low; coal/gas ≈ high. The same query is "greener" on a cleaner grid.
Power Usage Effectiveness — extra energy a data center spends on cooling and overhead beyond the computers themselves. PUE 1.09 = 9% overhead (very efficient).
Emissions from manufacturing the hardware, before it's ever switched on. Reusing old hardware makes this ~0 — a real green advantage.
What a device draws while doing nothing but staying on. For a lightly-used always-on box, idle dominates the whole footprint.Seen in: AIBox energy
Full: charge each query its share of the 24/7 idle power (treats the box as a dedicated appliance). Marginal: count only the extra power used while actually working (treats the box as "on anyway"). The per-query number can swing wildly between them.
Water some data centers evaporate to cool their chips, measured in mL per query. An air-cooled home box uses ~0.
The specialized chip that runs AI fast. Originally built for video-game graphics, now the workhorse of AI inference and training.
Storing a model's values at lower precision so it runs faster and cheaper, with little quality loss. A common efficiency trick.
Training a compact model to imitate a large one — capturing most of the smarts at a fraction of the cost to run.
A model design where only the relevant slice of the network activates per query, so a "big" model runs at a small model's cost.
A small model drafts several tokens ahead and the big model checks them in bulk — producing answers faster for the same quality.
The observed trend that, for equal quality, AI inference gets roughly 10× cheaper every year — faster than computing got cheap in the PC era.Seen in: the optimistic case
When something gets cheaper/more efficient, people use much more of it — so total consumption can rise even as each use gets greener. Why demand keeps climbing.
A way to plot wildly different sizes on one chart: each step multiplies instead of adds. Lets "1 million" and "1 quadrillion" share a graph. Watch for it — bars aren't to linear scale.Seen in: who burns the tokens
"×330" means 330 times the starting value — a measure of growth rate, not absolute size. Two ×-figures compare how fast things grew, not how big they are.
Measured = read from real instruments or logs. Estimated = a reasonable default where no measurement exists. The sheets mark which is which so you know how much to trust a number.Seen in: AIBox energy
M = million, B = billion, T = trillion, Q = quadrillion (1,000 trillion). "3.2 Q tokens/month" = 3,200 trillion tokens a month.
About — This guide covers the vocabulary used across the five sheets in the collection. Definitions favor clarity over precision — for exact figures, methods, and sources, see the individual sheets. Missing a term? It's a living page; it can grow.