Working vocabulary.

A capacity future is a forward agreement to take delivery of compute capacity — typically expressed as GPU-hours of a named hardware class — in a specified future quarter. It is structurally analogous to a natural-gas or aluminium future: physical-or-cash settled, hedgeable, and traded against a forward curve.

Capacity futures matter because hyperscaler capex planning already operates on multi-year horizons. A standardised contract simply makes the implicit forward market explicit, and lets parties without their own balance-sheet exposure participate.

Data exclusion is the right to keep specified data out of training corpora. It is the negative-space counterpart to output rights: rather than licensing what comes out, it constrains what may go in.

A working data-exclusion market needs three things that do not yet robustly exist: a fingerprinting standard precise enough to identify included content after the fact, an audit regime that can verify compliance, and a settlement mechanism for breaches. Each is non-trivial. The legal framework is arriving faster than the technical one.

A FLOP-hour measures sustained training work the way a kilowatt-hour measures sustained electrical work. One FLOP-hour is the compute delivered by one floating-point operation per second, run for one hour — 3,600 operations.

In practice, contracts quote PFLOP-hours or EFLOP-hours, and qualify them with a precision (FP8, BF16, FP32) and an interconnect grade. A training-FLOP commodity is only meaningful once both are fixed, because otherwise two “equivalent” FLOP-hours can produce wildly different training throughput.

A grade is the contract-grade quality classification that lets a commodity behave as a commodity. Crude oil has Brent and WTI; wheat has US No. 2 Hard Red Winter; electricity has firm vs. interruptible. Without a grade, a “barrel of oil” is meaningless as a unit of trade.

For AI, grading is the open problem. Provisional schemes split inference along latency, context window, and evaluation-suite floors: T1-FRONTIER (frontier capability, strict latency, high price floor), T2-PRODUCTION (steady-state production load), T3-BULK (batch, relaxed latency, lowest floor). The challenge is that capability moves faster than any standards body has historically moved.

A token is the elementary unit a language model produces — most often a sub-word fragment of two to four characters in English text. Inference markets adopt the token as their natural unit of trade because it is quantifiable to many decimal places, additive across requests, and broadly comparable across providers once a quality grade is fixed.

The catch is that “one token” only behaves like a commodity once you also specify which model class produced it, at what latency, with what context window, against what evaluation suite. That bundle is what a contract on inference tokens has to nail down.

The interconnect grade captures the network fabric that joins accelerators inside a training cluster — bandwidth per link, link latency, topology (fat-tree, dragonfly, torus), and collective primitives supported.

It belongs in any training-FLOP contract because two clusters with identical raw FLOP totals can deliver radically different throughput on the same model: a poor interconnect strands FLOPs in synchronisation overhead. Without an interconnect grade, “one PFLOP-hour” is a misleading number.

A load curve is the time-series of demand against capacity. In electricity it is the canonical operational object: dispatch, pricing, and capacity planning all hang off its shape.

Inference load is already showing recognisable curve features — pronounced diurnal peaks (US working hours dominate), weekly cycles, correlated bursts driven by news cycles, and growing nighttime troughs that are absorbing batch jobs. Whether AI ends up with electricity-style peakers (hardware that exists only to cover the top of the curve) is one of the more interesting open questions.

Output rights are the licensing layer over generated content: who may use a model’s output, for what, where, and for how long. They are already a meaningful unit of trade in segments where the output is the product (stock imagery, voice, code), and they sit on familiar intellectual-property infrastructure.

The interesting question is whether output rights remain a downstream licensing concern or whether they become a primary contractual unit on which compute is procured — i.e. you buy not “tokens” but “the right to ship N tokens of T2-PRODUCTION output for purpose P”.

A settlement window is the interval inside which delivery must occur for a contract to be fulfilled. Power markets settle in five-minute or hour-ahead windows; commodities settle in business-day or monthly windows.

For inference tokens, the settlement window is what makes “latency” a contractual property rather than a marketing one. A T1-FRONTIER token contract might require delivery within a 200ms window; a T3-BULK contract might allow a 24-hour window. The window collapses two intuitions into one: how soon and how reliably.

The tick size is the smallest price increment a contract can move in. It is the resolution of the market: a tighter tick lets prices discover finer signal; a wider tick reduces noise and friction.

For an inference-token contract priced in cents per million tokens, a tick of $0.01 / Mtok means the order book can express moves of one cent on the M-token. Picking a tick is one of the most under-discussed design choices in any new market — too tight and the book churns on noise, too wide and the contract loses informational fidelity.