Glossary
The language of inference procurement, explained.
5 min read·User manual
Purchasing terms
| Term | Meaning |
|---|---|
| Agreement | The accepted model, quantity, delivery terms, rate card, and supplier commitment |
| Buying approval | A broker’s recorded decision that an organization may accept quotes and firm offers. An account can research and prepare requests before this approval exists |
| Firm offer | Provider terms backed by capacity, available for eligible acceptance until expiry |
| RFQ | Request for quote: a buyer’s requirements for suppliers to price |
| Quote | An expiring provider response with prices and terms |
| Spot | Immediate service delivery after valid acceptance and prerequisites |
| Future delivery | Service consumption during an agreed future window |
| Amendment | An accepted revision, such as a value-preserving future model replacement |
| Mandate | Recorded permission for a broker to act for a buyer, with an expiry and defined rights |
| Revocation | Withdrawal of a specific authorization; historical agreements and records remain preserved |
Models and delivery
| Term | Meaning |
|---|---|
| Model identifier | The exact source name/version used to identify a model |
| Training house | The organization associated with a model’s creation or publication |
| Provider | A supplier serving a model through an endpoint |
| Provider association | A proposed or verified arrangement linking a model and onboarded provider, with delivery type and optional endpoint/buyer scope |
| Verification status | A broker’s pending, verified, or rejected decision on a provider association; distinct from capacity reservation |
| Endpoint | A configured route to a supplier’s inference service |
| Reference catalogue | Dated public metadata and prices used for research, separate from executable supply |
| Context window | The amount of input and generated context the model can handle within its limits |
| Modality | A type of content, such as text, image, audio, or video |
| Parameter count | A published architecture property, distinct from request-generation settings |
| Generation parameter | A request control such as temperature or a reasoning setting |
| Capacity pool | Shared, time-bound supplier capacity reserved against accepted obligations |
| Provider allocation | A portion of an agreement with its own supplier, model, allowance, prices, and delivery terms |
| Partial quote | A provider quote for part of an RFQ, combined with other selected quotes to cover the full request |
| Composite acceptance | One acceptance that reserves all selected provider allocations together and creates one complete agreement |
| Throughput | The permitted delivery rate, such as tokens per minute; the RFQ’s numeric minimum is entered in Required tokens per minute |
| Tokenizer | A model-specific method for splitting content into tokens |
| Concurrency | The number of requests allowed to run at the same time |
Prices and usage
| Term | Meaning |
|---|---|
| Token | A unit used by a model to represent content; tokenization depends on the model |
| BT | One billion tokens; a wholesale allowance is expressed in input-equivalent BT |
| 1 BT | One billion input-equivalent tokens, the minimum for a new agreement |
| Input-equivalent | Consumption expressed relative to the accepted input base rate |
| Allowance | The purchased input-equivalent quantity available for consumption under an agreement |
| Rate card | Accepted prices for input, output, cache, and other supported billable classes |
| Multiplier | A usage class’s price divided by the applicable nonzero input base price |
| Cache read | Reuse of eligible cached input under provider-specific conditions |
| Cache write | Creation of an eligible cached entry, with its own price and retention conditions |
| Batch | An explicitly supported execution mode with its own lifecycle and delivery deadline |
| Reservation | Allowance held before dispatch or while a result needs reconciliation |
| Reconciliation | Matching validated supplier usage to the reserved allowance and accepted rates |
| Exception | A reconciliation state in which the outcome of admitted work is uncertain and awaits evidence-based resolution |
| Evidence resolution | An audited broker decision to settle authoritative usage or release after supplier-confirmed no execution |
| Ledger | Durable records of admitted work, usage, charges, and exceptions |
| Idempotency | Replaying the same identified operation without creating a duplicate acceptance |
Access terms
A buyer API key lets an application authenticate for the buyer’s permitted agreements. A provider credential authenticates the server to a supplier and remains on the server. They are different secrets and are not interchangeable.
An organization is the ownership boundary for business records. A role permits categories of actions; a mandate grants a broker specific authority for a named buyer.
Need assistance with this workflow?
Contact the wholesale desk with your organization and record reference.