What a worker earns.

Three numbers decide a payout: how fast your machine generates, how much of the day it is given work, and the price per token of the model it serves. The first is measurable on your own hardware. The second is not knowable in advance. The third is published per model and shown in the app.

Arithmetic

One formula. Every input is either measured on your machine or published for the model you serve.

# price is in USDC per token, so divide micro-USDC by 1,000,000
USDC per day = tokens/sec × 86,400 × utilisation × price × 0.80

# the same formula in the units the app shows
USDC per day = tokens/sec × 86,400 × utilisation
             × price_per_million_tokens / 1,000,000 × 0.80

The 0.80 is the worker share. The remaining 20% is the protocol fee.

Tokens per day

Generation rate against the share of the day the network has work for the machine. Pick the row nearest your measured rate and the column nearest a day you think is realistic.

tokens/sec10%25%50%100%
1086,400216,000432,000864,000
25216,000540,0001,080,0002,160,000
50432,0001,080,0002,160,0004,320,000
100864,0002,160,0004,320,0008,640,000
2502,160,0005,400,00010,800,00021,600,000
5004,320,00010,800,00021,600,00043,200,000
10008,640,00021,600,00043,200,00086,400,000

A day is 86,400 seconds. 100% is a machine that is generating for every one of them, which is the upper bound and not a forecast.

Price per token

The price belongs to the model, not to the machine serving it. Whoever publishes a model to the catalogue sets its price when they upload it, and can change it later. Every worker serving that model is paid the same price. A worker cannot set a price of their own.

The price is quoted in micro-USDC per token. One micro-USDC per token is one US dollar per million tokens, which is the easier unit to reason about.

At one dollar per million tokens, a machine generating 100 tokens per second for a full day produces 8,640,000 tokens. That is 8.64 USDC before the split, and 6.91 USDC to the worker. For any other price, multiply that figure by the price in dollars per million.

A first payout to a new wallet needs at least 2 USDC, because the protocol creates the USDC token account and pays its rent. At the rate above the floor is cleared inside a day. Anything below it carries into the next batch.

Your throughput

Speed depends on the model, its quantisation, the card and whether the weights fit in VRAM. A figure quoted for the same model on different hardware does not transfer.

$ autarcho-node probe --model /models/<file>.gguf --tokens 128
model=/models/<file>.gguf tokens_requested=128 elapsed_ms=... bytes=...
output="..."

probe runs inside the worker image. Divide the token count by elapsed_ms / 1000 for tokens per second, once the weights are loaded. The first run of a model includes the load and is not the rate.

The same command prints one line per device before its result, in the index order GGML_VK_VISIBLE_DEVICES takes. A GPU is not always faster than the CPU.

Limits

  • Utilisation depends on the jobs the network routes to you. There is no guaranteed share of a day, and the column you pick is an assumption.
  • The price is per model and can change, in either direction, after you install.
  • Electricity, hardware and tax are your own costs, and they are not in the formula.
  • No earnings are guaranteed.