tokn.supply is a decentralized inference provider. Independent operators run open AI models on machines they already own; you reach them through one OpenAI-compatible API at the market price, and every prompt is sealed to the machine that answers it, so no one in between can read it. We are in private beta, and we are building this to last.
Why tokn exists
Spending on AI inference is growing faster than almost anything else in software, and nearly all of it flows to a handful of data-center operators. We think the people who already own capable hardware should be able to sell into that demand, earn a return that grows as their machines do, and do it without anyone reading their customers' prompts. tokn is that market: a professional, permanent way to turn spare compute into tokens, and tokens into income.
For buyers one API
If you already use an OpenAI-style API, the only change is the base URL:
# Same request you already send. Different host.
curl https://api.tokn.supply/v1/chat/completions \
-H "Authorization: Bearer $TOKN_KEY" \
-d '{"model":"meta/llama-3.2-1b-instruct","messages":[{"role":"user","content":"hello"}]}'
The OpenAI SDKs work unchanged: chat with streaming, and embeddings.
GET /v1/models lists what is available to your key.
Credits are prepaid on your account page and drawn down per token as you go. No subscription, no minimum, no expiring credits, and your card details never touch our servers. If we are at capacity you are told at once and pay nothing; a request that produces no tokens costs nothing.
Buyer accounts are by invitation while the beta is private. Join the list and tell us what you want to run; you get an API key and starter credit, no card and no sales call.
For providers your hardware, earning
You install the tokn client on a machine with a capable GPU. It fetches and verifies the exact model builds your hardware serves well, opens sealed requests that only it can read, and answers them. You are paid per token your machine serves, at the posted market price, and you can stop at any time.
A Linux command-line client is available now and runs as a system service. Windows and macOS apps, with command-line clients for both, follow as the beta widens to more providers.
We are onboarding providers in small groups so that every node is measured before it carries paying traffic. Join the list with the hardware you have; consumer GPUs with 8 GB of memory or more are exactly what the current catalogue is built for.
Models per million tokens
We serve the models the large marketplaces underserve and that people actually use: small, fast instruction models and embedding models with steady demand. Each one is chosen so that it pays the provider who serves it and is worth buying at the market price, and each is pinned to exact, verified weights.
Every model we list stays served. As the catalogue grows we add to it, and when something has to change we say so well ahead of time. A provider who buys hardware for tokn and a buyer who builds on it should both be able to count on that.
Privacy by construction
When independent people run the machines, privacy cannot be a policy. So your request is encrypted at our edge to the single machine that will answer it. The network that carries it holds no key, and nothing anywhere in the system stores prompt or completion text. The provider's machine can answer your request; nobody, including us, can read it in transit or afterwards.
The full mechanism, what is guaranteed, and where its honest limits are, is on the technology page.
Join the beta
Whether you want to use the API or supply compute, this is the door. We invite in small groups, and we write to everyone on the list.
You are on the list. We will write to you at that address.
That did not look like an email address. Please try again.
We could not take that just now. Please try again in a minute, or write to hello@tokn.supply.
Prefer email? hello@tokn.supply reaches the same people.
Private beta. Two models, a Linux client, and a network served today from hardware we operate, opening to independent providers through the beta. Access on both sides is by invitation. What we publish here is what runs; the details are on the technology page.