Provider economics

Assumptions behind earnings tables in the Litepaper. Last reviewed: July 2026.
The Litepaper quotes annual earnings ranges (e.g. Mac Studio Ultra 10,00010,000–13,000/year at MALIBU=MALIBU = 2). Those numbers are illustrative scenarios from tokenomics modeling — not live marketplace payouts and not guarantees.
Illustrative, not guaranteed. Actual earnings depend on buyer demand, uptime, model tier, thermal limits, token price, and whether you are on marketplace credits (live) vs token emission (planned v0).

Two income regimes

Never mix the two in one planning spreadsheet without labeling the regime.

Live marketplace economics (today)

Source: Payments & payouts, Metering & billing.
Earnings scale with:
  • Requests routed to your Mac (pool size, tier, uptime)
  • Model served (rate card per model)
  • Fault rate (invalid/missing receipts → zero credits)
There is no public fixed “$X/year per Mac tier” on the live rail — compute from your served volume using /providers/{id}/earnings.

Token-scenario tables (Litepaper)

The Litepaper Year-5 base-case table assumes: These projections require token launch, emission curve, and buyer demand not guaranteed today.

Hardware reality checks

From Phase 1 mining research (mining-prototype-phase1.md):
  • Fanless Air may sustain ~76% of peak hashrate over 30 minutes — apply a duty-cycle discount for 24/7 planning.
  • Mining at full throttle can be hostile to daily interactive use; cooperative scheduling yields to buyer requests (~50 ms at v1 design target).
  • Unified memory caps which catalog models a tier can warm; serving gpt-oss-120b requires Studio Ultra-class RAM.

Sensitivity checklist

Before treating any annual number as a forecast, stress-test:
  1. Uptime — lid-close sleep drops WebSocket; see Payments → provisional vs pinned.
  2. Pool competition — more providers → shared demand unless buyer volume grows proportionally.
  3. Token price — emission income is highly sensitive; marketplace USDC income is not (today).
  4. Thermal / duty cycle — sustained load ≠ burst benchmark peak.
  5. Receipt quality — faults zero credits; see Verified model settlement.