📊 Model inputs — these are assumptions, not measurements
Every figure below is an input we chose, not a number this pool has observed.
We have not yet run a GPU long enough to measure any of them. Treat the
output of this calculator as a model, and check each input against your own
costs before you rent anything.
- PRL mining yield: assumed ~$0.50 per H100-hour. Not measured.
Depends on live difficulty and PRL price, both volatile.
- H100 operating cost: assumed ~$2.00/hr. Sanity-checked against
spot rentals on 2026-08-10, where H100 NVL ranged $2.12–$3.33/hr.
- Inference gross: assumed ~$6.00 per H100-hour. Not measured, and
not derived. It presumes both a model mix and a level of paid customer
traffic that the inference tier does not have today.
- Total surplus: ~$4.50/hr is simply the three figures above combined.
It inherits every assumption in them and is not an observed margin.
- Consumer GPUs (3090/4090/5090): the combined mine+inference kernel is
compiled for sm90 (H100/H200) only, so the "Inference gross" assumption above
does not apply to them and this calculator does not model any inference income
for them. They can separately register as a GPU serving node
and earn a share of inference revenue when a model fits their VRAM (or a
multi-GPU rig serves a bigger one) — that path exists, but we have no measured
rate for it yet, so it is left out of this calculator rather than assumed.
- Enterprise GPUs (H100/H200) are the only hardware the combined
mine+inference tier modeled above supports.
What we have actually measured, on an H100 NVL on 2026-08-10: the Pearl
vLLM miner served 6,958 inference requests at 445 tokens/sec aggregate with the
mining kernel active on all 80 layers. We have not yet measured PRL earned
per GPU-hour — at current network difficulty a single card solo-mining would wait
months for one block, so that number requires the pool share layer we are still
building. We will publish it here when it is measured, and not before.