— output tok/s at saturation
Mining only · net/day
—
—
Serving only · net/day
—
—
Hybrid · net/day
—
serve first, mine the idle time
Break-even demand
—
serving beats mining above this
Electricity at full load, all three modes: —/day · —/month
Annual return on hardware
—
—
Net $/day vs. serving demand
Serving
Mining (flat)
Hybrid
Where these numbers come from
- PRL/hour: anchored on our own 24-hour RTX 5090 run — 0.289 PRL/h at 361.5 TH/s, network at 51.49 EH/s, after a 2% fee (2026-09-30 → 10-01) — scaled by your card's hashrate and today's network hashrate.
- Hashrate & power, measured: H100 743 TH/s @ 698 W · RTX 5090 350 @ 500 W · RTX 4090 310 (rated draw) · RTX 3090 67 @ 220 W (2026-10-04). H200 = H100-class, an estimate we have not verified.
- Network hashrate & PRL price: read live when the page loads; both inputs are editable because PRL is volatile.
- Serving throughput, measured on one RTX 5090: Qwen3-Coder-30B 9,024 tok/s (2026-10-03) · Qwen2.5-7B 7,385 (2026-10-01) · Qwen3.8-27B 1,051 (2026-10-01). Whole-card totals under full load; one request on the flagship runs about 86 tok/s (2026-10-07). Only these three wear the green measured badge — everything else is an estimate, and our one cross-GPU check of that method ran 1.6× optimistic, so treat estimates as an upper band.
- Model prices: live from /v1/models; if the fetch fails, a 2026-10-04 snapshot is used and labelled.
- Your share: 75% of what the node bills, paid in USDC — the pool's published terms. Inference tier in private beta.
- Hybrid: what our installer runs — serve first, mine the idle time, at the card's full mining rate. Not available on 24 GB cards (not enough memory headroom); the calculator says so.
- Not included: rent, cooling, resale value, prompt-token revenue, or any guarantee of serving demand.
A model, not a promise. Not financial advice.