Cost to produce a PRL
Three kinds of operator produce PRL, and a coin costs each of them something different. This page prices all three against the market, shows the assumptions behind them, and works through the arithmetic line by line.
Who is producing, and on what assumptions
Pick an operator, or set the inputs yourself. Everything below recalculates: the chart, the tables and the working.
All assumptions
Market price against cost to produce
The PRL price and what a coin costs each kind of producer. Where the price sits relative to a line, that producer is above or below water. Log scale by default: the cheapest and dearest producers are more than three times apart.
Unit cost by producer
Cost of goods sold per PRL, broken out by cost type. Read across: what each producer pays for electricity, hardware and forgone inference, and what that works out to per coin. Each row says which card it is priced on, because the right card is not the same for all three.
Return on the fleet, by use
The same hardware put to three uses. Capital is charged in full to all three, so the comparison is of what the fleet earns, not of how it is financed.
Working
Every headline figure above, derived step by step. Each line shows the inputs and the operation, so the result can be checked with a calculator.
Contention between the two workloads
The one input that decides the answer, and the only one nobody has measured. Derived here from published GPU behaviour rather than assumed.
Hourly revenue per GPU, by use
What a single card earns per hour mining, against what the same hour sells for as inference. Mining moves with the PRL price and the size of the network; the inference line is today's market rate.
Sensitivity to inference throughput
How the fleet's daily profit changes with the share of inference throughput that survives alongside mining. The break-even crossing is the decision point.
Hardware
Throughput per card, power draw and purchase price. Sources disagree by up to two times on consumer throughput, so the range column shows the spread rather than implying a precision we do not have.
Electricity
What miners actually pay against posted industrial tariffs — roughly a five-fold gap. The two are kept apart because averaging them would distort every figure above.
GPU rental rates
What an hour of this hardware sells for. Two independent routes to the same number: where the rental market and token revenue agree, both are probably right.
Token prices — open-weight models only
What an operator can charge per million tokens sets what a GPU-hour is worth. Only models you can host yourself are listed: a proprietary API price is a product price, not a compute cost.
Where to next
Related pages that answer the questions this one raises.