Autarcho

Live network · USDC payouts enabled

AI by the token, not by the month.

Autarcho is an inference network. Independent operators serve open GGUF models from their own hardware, the gateway meters every token, and you pay only for what you generate. No subscription. No lock-in.

GGUF only Catalogue shown in the app Outbound-only workers USDC on Solana
01 — Use

For people who want an answer, not another subscription.

Open a chat, ask, close the tab. Credit is spent per token and the price is shown before you send. Small models, no monthly fee.

  • Metered per token, priced per model
  • No subscription and no lock-in
  • Top up with USDC on Solana
Open the chat
How using it works →
02 — Earn

For people with a GPU and a power bill.

Serve GGUF models on hardware you already own. The operator keeps 80% of every job, paid in USDC on Solana. Outbound-only: no inbound ports, no public IP, no account.

  • 80% of job value to the operator
  • Payout address only — never your private key
  • x86_64 Linux · Docker · 16 GB RAM
Run a worker
Read the full guide →

01 — The network

Three parties, one job.

There is no central model host. A prompt is routed to whichever operator is online and serving that model, and both sides of the exchange are measured.

01

You send a prompt.

The chat client asks the gateway which models are available and what each one costs. Credit is reserved against the job's cap before it starts.

02

The gateway routes and meters.

It picks an operator that is online for that model, counts tokens as they stream, and records a receipt for the work.

03

An operator answers and gets paid.

Their GPU generates the response. Verified work is batched and settled to their wallet in USDC on Solana.

02 — Straight answers

What this is, and what it isn't.

Most networks in this space are sold with adjectives. This page is written so that you can decide before you spend anything.

It is

  • Model-agnostic. There is no fixed menu of models baked into the client. The gateway registers GGUF models with a price each, and the app lists what is registered and being served.
  • Metered. You buy credit and spend it per token. There is no monthly reset and no subscription to cancel.
  • A real market. Independent operators serve the work and keep 80% of what it is worth.
  • Verifiable settlement. Every payout is a USDC transfer on Solana mainnet, visible on-chain.

It isn't

  • A frontier-model replacement. Today's catalogue is small models: long documents and hard reasoning are not what they are for. Use a bigger model for that.
  • Confidential. Operators run your prompt on their own machines. Do not send anything you would not put in a stranger's logs.
  • Open source. The software is proprietary. Operators get a narrow contract instead: outbound-only, no private key, no inbound ports.
  • A guaranteed income. Earnings depend on demand and the models you serve. Nothing is promised.

03 — Supply

An idle GPU is a cost with no revenue.

A worker joins the network in a few commands, pulls a model the first time someone requests it, and serves it from then on. No inbound ports, no public IP, no allowlist. Payment is 80% of job value in USDC.

See the requirements Worker guide
Host
x86_64 Linux · Docker Engine + Compose v2
Memory and disk
16 GB RAM · 20+ GB free for the model cache
Network
Outbound TCP 443 and 4001. No inbound ports, no public IP, no DNS.
GPU
Optional. CPU is fully supported; CUDA 75 (Turing) is the validated GPU path.

04 — Money

No subscription. No monthly fee.

Two ways to look at the same number: what you spend, and where it goes.

You pay per token

Metered

Credit is spent as tokens are generated. The price is set per model and shown in the app before you send a message.

The operator keeps

80 / 20

Eighty percent of every job goes to the operator who served it. Twenty percent funds the gateway and the protocol.

Nothing resets

No monthly fee

There is no allowance that resets and no subscription to cancel. What you buy stays on your account until you spend it.

Early access: per-model prices are visible in the app before you send. There is no headline rate to quote here, because cost depends on the model and the number of tokens you generate.

05 — Questions

The ones that decide it.

Do I need crypto to use the chat?
Today, yes. Credit is bought with USDC on Solana. Card payments are not available yet. Better to say that here than at checkout.
What do operators earn?
Eighty percent of the value of each job they serve, in USDC, batched from verified receipts and settled on-chain. There is no fixed payout schedule and no guaranteed amount. A wallet's first payout needs at least 2 USDC to cover creating its USDC token account; anything below that is carried into the next batch rather than lost.
Can an operator read my prompts?
Yes. Your prompt runs on their machine. Autarcho does not claim confidentiality and there is no end-to-end encryption in this release. Treat prompts as visible to the operator serving them.
Is the software open source?
No. The worker and gateway are proprietary. What operators can verify is the contract — EULA version and hash recorded at install, pinned image tags, a pinned gateway PeerId — and the money, because payouts are USDC transfers on Solana.
What hardware does a worker need?
x86_64 Linux with Docker Engine and Compose v2, 16 GB RAM, and 20+ GB of free disk. A GPU is optional: CPU workers are fully supported. CUDA 75 is the validated GPU path in 0.1.0; Ampere and Ada images are not published yet.
How do I know I am talking to the real gateway?
The installer pins the gateway by PeerId and refuses a bootstrap response for a different peer. You can cross-check the published PeerId yourself against /bootstrap.json.

Pick a side of the network.

Open the chat Run a worker