​
Provider economics

What drives provider credits today, and the assumptions behind the illustrative token-launch scenarios referenced by the Litepaper. Last reviewed: 23 September 2026.
Earlier versions of the Litepaper quoted per-Mac annual figures and hardware-payback lines from tokenomics scenario modeling. Those figures were removed on 23 September 2026. The scenario inputs below are kept for context. They are illustrative and depend on a planned token launch and an assumed token price. They are not live marketplace payouts, not projected earnings, not investment returns, and not guarantees.
Illustrative, not guaranteed. What a Mac actually earns depends on buyer demand, uptime, routing tier, which models it can fit, thermal limits, receipt quality, and which regime applies: marketplace credits (live) or token emission (planned). Nothing on this page is a forecast or a recommendation to buy hardware or tokens.

​
Two income regimes

An off-chain reward ledger for verified work is kept in pre-beta as a preview. It cannot be withdrawn and is not marketplace income. See Economics → reward ledger. If you put both regimes in one planning spreadsheet, label which regime each row belongs to.

​
Live marketplace economics (today)

Sources: Payments & payouts and Metering & billing.
Rate arithmetic, not a forecast: from the live rate card, 1 million completion tokens on qwen3-32b is 0.22gross,and0.22 gross, and 0.198 of that is provider credit. The same volume on qwen3-8b is 0.027grossand0.027 gross and 0.0243 provider credit. How many tokens your Mac actually serves is the unknown, and it depends on buyer demand. What moves your credits:
  • Traffic routed to your Mac. This depends on buyer demand, pool competition, uptime, and routing tier. Provisional providers route at reduced weight; trusted and pinned route at full weight.
  • The model you serve. Rates differ by model row on the rate card, by up to about 80× on completion tokens.
  • Settlement. Only rows with a valid, on-time receipt are payable; see Verified model settlement.
  • Buyer pricing in pre-beta. Public buyers use a free daily quota, so today’s demand is not paid demand.
The live rail has no public fixed “$X/year per Mac” figure. Read your own served volume and earnings in the Malibu app.

​
Hardware reality checks

  • Unified memory limits which rows a Mac can serve. The signed autotune feed lists a minimum RAM for each row. Examples: meta-llama/llama-3.2-3b-instruct needs 4 GB, qwen3-8b 12 GB, openai/gpt-oss-20b 24 GB, qwen2.5-coder-32b-instruct and qwen3-32b 48 GB, z-ai/glm-4.5-air 80 GB, and openai/gpt-oss-120b 90 GB. Higher-rate rows route only when a Mac that fits them is online.
  • Sustained load is not burst peak. Sustained throughput is lower than a short benchmark peak. Apply a duty-cycle discount for 24/7 planning. See Benchmarks & methodology.
  • Interactive use. The mining design targets yielding to buyer requests within about 50 ms. That is a v1 design target, not a shipped measurement.

​
Sensitivity checklist

Before you treat any annual number as a forecast, stress-test it against:
  1. Uptime. Lid-close sleep drops the WebSocket; see Payments → why your Mac might drop.
  2. Pool competition. More providers share the same demand unless buyer volume grows with them.
  3. Paid demand. Public buyer usage is unpaid in pre-beta.
  4. Token price. Emission income is highly sensitive to it. Marketplace credits are priced in USD on the rate card.
  5. Thermal limits and duty cycle. Sustained load is lower than a burst benchmark peak.
  6. Receipt quality. Missing or invalid receipts pay zero; see Verified model settlement.