Provider Terms

You may supply inference only as a business. By applying, you confirm that you act for the purposes of your trade, business or profession, and not as a consumer.

Give us complete and accurate information in your application and keep it up to date. We may approve, reject, ask for changes to, or suspend any application or account at our discretion. You may receive live requests only after we approve you and you go live. While in staging, you receive only test requests.

Request content (prompts, inputs, files, outputs and related metadata) is confidential information of the Buyer. You must:

• use request content only to serve the request it belongs to;
• not read, inspect, log, copy, sell or share request content, or use it to train, evaluate or improve any model, except as your declaration for that offering expressly allows;
• not keep request content for longer than your declared retention period. If you declare that you store no content, delete it as soon as the response is complete, except for data kept only in memory or cache to honor cache terms you quoted, and delete that when those terms end;
• serve each offering only from the country you declared for it;
• protect request content with appropriate technical and organizational security measures, including encryption in transit and access controls; and
• tell us within 48 hours of becoming aware of any unauthorized access to, or use or disclosure of, request content, and help us respond to it.

You are responsible for anyone you use to run your offerings, including hosting providers and staff, as if their acts were your own. Where request content contains personal data, you process it only on documented instructions, which these Provider Terms and the request are, and you will sign any data processing terms we reasonably require.

Every declaration for an offering must be accurate and current: the model, weight precision, engine, context window, capabilities and data handling. Serve each awarded request with exactly the model weights and precision you declared. Do not substitute another model, a distilled, fine-tuned or further quantized version, or a different system prompt, and do not change parameters the request sets, unless the request allows it. Update a declaration before the change it describes takes effect.

You must have all the rights and licenses needed to serve each model, and you must follow its license.

A quote is a binding offer. When we award it, you must serve that request on the replica named in the quote, at the quoted prices and terms. An award is not a further chance to decline. Decline when you quote if you cannot honor the terms. If we quote for you from your saved prices, set your concurrency limit so that you can serve every request we award.

You may refuse to serve a request only where serving it would break the law. If you refuse, tell us promptly and give the reason.

Report usage accurately. The input, cached input and output token counts you report must match the tokens you actually processed and returned. Do not inflate usage, misreport cached tokens, or charge a Buyer for a request outside the Service. We charge Buyers and pay you at the awarded prices for the usage we accept for each request. We may review, correct or reject usage that is missing, invalid or implausible.

If you quote cache terms, you commit to them for their whole term, and you cannot cancel them. Follow-up requests within the cache window are billed at your cached input rate for at least the floor described in the documentation, whether or not your engine still holds the cache. Keep enough cache capacity to honor every cache term you quote, and do not evict cached data in a way designed to avoid them.

Start serving each awarded request promptly. Stream tokens as soon as they are ready. Do not delay, abandon or slow down an awarded request. Serve it with at least the speed you declared or quoted for it. We measure every offering’s failures and speed, and use them to rank quotes and in our decisions under Section 10. We do not guarantee any volume of requests.

We may send test, canary and verification requests to your endpoints at any time, and compare your outputs with reference outputs, to check that you comply with these Provider Terms. We do not pay for these requests. You must cooperate with reasonable requests for information about how you serve your offerings.

You may use the quoter, installer and Provider SDK only to take part in the Service. Do not modify them to misreport load, usage or capabilities. Keep your quoter credentials secret, and keep your quoter up to date when we tell you an update is required.

You earn the awarded prices for the usage we accept on each request you serve. We currently charge Buyers a service fee and take no commission from your earnings. We may introduce a Provider fee with at least 30 days’ notice.

Before your first payout, you must complete identity checks and give us valid payee, bank and tax details, and a signed supplier agreement where we ask for one. Earnings mature after a holding period, currently seven days, and we pay matured earnings above the minimum payout amount, currently US$50, on the schedule in your supplier agreement or payout settings. You are responsible for your own taxes. You do not earn from requests sent by a workspace you belong to.

We may hold, offset or reverse earnings for a request that breached these Provider Terms, that involves usage under review, or that a Buyer disputes or is refunded for because of your act or omission, and where we reasonably suspect fraud. We will tell you which earnings we are holding and why.

If you breach these Provider Terms, we may do any of the following:

• lower the ranking of your quotes, or stop sending you requests;
• suspend any offering, quoter or your account;
• refund the affected Buyers and recover those amounts from your earnings, or hold or cancel the earnings for the affected requests;
• if you commit fraud, inflate usage or breach Section 2, withhold all of your unpaid earnings, not only those for the affected requests;
• end your participation as a Provider; and
• report illegal conduct to the authorities.

Repeated failed requests, misrepresenting a model or precision, inflating usage, and breaching Section 2 are material breaches. These remedies are in addition to any other rights we have.

Section 11 of the Terms of Service limits each party’s liability. Its limits do not apply to your breach of Section 2 or to your obligations under this Section.

In addition to Section 12 of the Terms of Service, you will defend and indemnify Canopy X and the affected Buyers against third-party claims, and the losses and reasonable costs that follow, arising from your breach of Section 2 or Section 3, from your infrastructure, or from serving a model you did not have the right to serve.

We may update these Provider Terms in the way Section 8 of the Terms of Service describes, and the date at the top of this page shows when they last changed. We may ask you to accept a new version before you supply inference under it. If you do not agree to a change, you may also end these Provider Terms by telling us within 14 days of the notice.

You may stop quoting at any time. Before you take an offering offline, serve every request already awarded to it and honor every cache term it quoted until that term ends. We pay your remaining matured earnings after final reconciliation, less any amounts we may hold or recover under Section 9 or Section 10.

Sections 2, 9, 10, 11 and 13 survive the end of these Provider Terms, and Section 2 continues to apply to any request content you still hold.