Keep your customers, add a second revenue
Serve your own traffic on Pearl’s mining-enabled vLLM; it also earns PRL. No rerouting, no reselling. One integration, 1% pool fee. Talk to us.
Your traffic · PRL through our pool
Same GPU. Same work. Two revenues.
Serve AI as you do today. The same computation also earns network rewards through our pool.
Partnership
Two ways to work with us.
Add our paid demand
Want inference jobs from our API on idle capacity, paid in USD per token? Talk to us.
Our traffic · USD per token + PRL
Ask us for the miner image, then create your first node in the miner portal.
What you run
One container. Outbound only.
-
One container
vLLM with Pearl’s miner and our node software, on your GPU host.
-
Outbound only
It connects out to our network; no open ports, no inbound access.
-
Your hardware, your rules
You choose which GPUs join and when they leave.
-
Measured on NVIDIA H100 SXM
-
H200 runs the same software
-
Running other GPUs? Ask us.
Estimate
Estimate your earnings.
Pick a GPU and a load. Ranges come from our measured H100 SXM runs and today’s Pearl network difficulty.
8 × H100 SXM (80 GB)
Tokens served
4.58B–14.9Btokens / day
137B–448B tokens per 30 days
- Chat5.15B–12.8B
- Long prompts6.11B–14.9B
- Long answers4.58B–11.6B
USD from our API jobs (at the partner rate)
60.4–181USD / day
1,811–5,428 USD per 30 days
Demand is routed by agreement today. Talk to us.
- Chat60.4–150
- Long prompts62.2–152
- Long answers71.5–181
PRL from the same work
9.14–31.6PRL / day
274–947 PRL per 30 days
- Chat9.14–27.2
- Long prompts13.7–31.6
- Long answersNot measured
Long answers: mining is not measured on this shape, so it has no PRL figure.
Estimate. Your results depend on load and network difficulty.
How we estimate
Throughput and mining rates are measured on an H100 SXM under three request shapes, three runs each, warm-up discarded. With mining on, a Pearl GPU serves about three quarters (73–78%) of the tokens per second of a standard vLLM server on the same card. H100 NVL and H200 figures are scaled from the H100 SXM measurement. Under steady traffic it mines about half as fast as when it runs warm-up batches alone. PRL is the expected value at today’s network difficulty and block reward; rewards arrive per block, so a single day can be higher or lower. Full method: docs.
Payouts
Earn in USD and PRL.
Payouts begin at launch. Minimum 10 PRL. Batches are prepared daily at 02:00 UTC and sent after review.
Verification
We check the work.
Privacy
Prompts stay in memory.
Our node software never logs or stores prompts or answers. The pool briefly checks a small slice of intermediate model values from mined requests in memory, then discards them.
Before and after
Every AI data centre can earn twice.
Contact
Tell us about your fleet.
We reply within one business day.