Cost to produce a PRL

Pearl's reference miner is built on vLLM and runs alongside inference on the same card — mining saturates compute at 95–100% while using under 30% of memory bandwidth, which is exactly the capacity an LLM serving decode leaves idle. So the real comparison is three-way: inference alone, mining alone, or both at once. The question that decides it is how much inference throughput you give up to mine, and whether the PRL is worth more than what you lost.

Who is mining?

Four operator archetypes, each with the hardware, power price and accounting that actually describes them. Pick one, then adjust below.

Fleet and costs

One set of assumptions drives every figure on this page. All of it re-runs against live chain state and the stored price.

What running both actually costs

The two retention figures are the model's weakest inputs, so they are derived rather than guessed. Published characterisations of LLM serving put prefill at 92% tensor-core utilisation and decode at 28% — so a decode-heavy GPU leaves most of its compute idle, and that idle compute is what a miner can take.

The same coin, costed four ways

None of these is the right answer on its own — they answer different questions, and an operator needs more than one of them. The spread between them is the honest uncertainty about what a PRL actually costs.

Show the working

Every headline number above, with the arithmetic that produced it — check it with a calculator rather than taking it on trust.

Cost to produce one PRL — who can still afford it

The same coin, two cost bases, against the price they both sell into. When the price line sits below the standalone miner's cost but above the AI operator's, only operators with inference revenue can mine at a profit.

Mining only vs AI only vs blended

Same fleet, same capital charge, three strategies. Capital is charged in full to all three — the question is what the fleet should do, not what a marginal coin costs.

Revenue per GPU-hour: mining vs inference

What one card earns per hour doing each job. Mining moves with price and network size; the inference line is what that GPU-hour sells for on the neocloud market today.

How much inference can you afford to lose?

Mining adds revenue; the throughput it costs inference subtracts it. The crossing point is the only number an operator actually needs.

GPU economics

Throughput figures disagree between sources — the range column is the honest spread, not a precision claim.

Power

What miners actually pay versus posted industrial tariffs — a ~5× gap. Averaging the two would wreck the model, so they are kept separate.

What a GPU-hour sells for

Two independent routes to the same number. Where they agree, both are probably right.

Token prices

The sell side. What an operator can charge per million tokens sets what a GPU-hour is worth.

Where to next

Related pages that answer the questions this one raises.