Skip to content
Pool stats中文Sign inStart mining
Menu

Peryn GPU Clouds agreement

Last updated: 30 October 2026. Version 1.3.

This agreement is between Peryn, the operator of the Peryn service ("Peryn", "we"), and the company named on your GPU Clouds account ("you"). You accept it in the GPU Clouds portal (app.peryn.ai/cloud) for that company. Our terms of service also apply. Where they differ on the subjects below, this agreement applies.

1. Definitions

  • Software: Peryn's agent (pc-agent), its installers and Helm chart, the mining software it runs, and every update of them that we publish, in the signed, compiled form we distribute.
  • Pool: the Peryn mining pool for the Pearl network.
  • Machines: the GPU servers you control on which you install the Software with your account's fleet keys.
  • Mined PRL: the PRL the Pool credits to you for work your Machines submitted, before the Fee.
  • Fee: the Duplex fee in section 6.

2. Licence

2.1 We grant you a non-exclusive, non-transferable licence, without the right to sublicense, while this agreement runs, to install and run the Software on your Machines, only to mine for the Pool under your account.

2.2 Your employees and contractors may install and operate the Software for you. You are responsible for them.

2.3 The Software contains open-source components, each under its own licence. Nothing here limits your rights under those licences for those components.

3. What you may not do

You may not, and may not let anyone else:

  • (a) copy, modify or create works based on the Software, except for the copies needed to install and run it;
  • (b) decompile or reverse engineer the Software, except where the law allows it regardless of this agreement;
  • (c) distribute, sell, rent or otherwise make the Software available to anyone else;
  • (d) use the Software to mine for anyone other than the Pool, or for any other purpose than section 2.1;
  • (e) remove, bypass or weaken its signature checks, update mechanism, fleet keys or other security controls, or run a modified or unsigned version.

You keep your fleet keys confidential and revoke in the portal, at once, a key you believe has leaked.

4. Your Machines come first

4.1 Your own workloads come first. Depending on how you install it, the Software runs below your workloads in your scheduler (Kubernetes, Slurm), gives a GPU back when your own process starts on it, or, in Duplex mode, mines only while your own inference server on the GPU is idle and pauses as soon as it is busy. It follows the rules you set in the portal: time windows, how many GPUs at once, and in Duplex mode the most GPU memory the miner may use.

4.2 We never raise a GPU's power limit. In Duplex mode, while mining on a GPU, the Peryn agent may lower that GPU's power limit, never above the limit you set. It restores your limit within about 1 ms, before your own work runs on that GPU, and never overrides a limit you or your tooling changed. After any stop, crash or restart of the agent, your limit is restored before mining resumes. You can turn this off with PC_DUPLEX_POWER=off. Outside Duplex mode we never change a GPU's power limit. On supported GPU models the mining software may set the GPU's memory clock while it mines and restores it when it stops; in Duplex mode the agent never changes your GPU clocks. We change no other GPU setting unless you turn that on in the portal, and then only while the Software holds that GPU.

4.3 You remain responsible for your Machines: power, cooling, hardware, network and their security, and for your agreements with your own customers. Mining uses your GPUs and electricity; you decide how many GPUs, when and how much.

5. What leaves your Machines

5.1 Only machine metrics and mining results leave your Machines. Your customers' data never does. The Software mines only on numbers it generates itself. It does not read, copy or send your customers' requests, prompts or answers, your models, model weights or files, your KV cache, or your traffic, and it does not sit in the path of your traffic. None of it reaches Peryn: not in mining shares, not in machine metrics, not in logs. The Software uploads no logs.

5.2 Machine metrics. The Software sends these to Peryn's servers, signed with each Machine's own key, over TLS, on outbound connections only (you open no inbound port):

  • (a) Enrolment and identity: the Machine's public key and the public keys it makes for its GPUs, the fleet key it was installed with, the install channel (for example Kubernetes, Slurm or Linux), the Software's version, and a sequence counter of its reports.
  • (b) The Machine's label: the name you set in PC_MACHINE_LABEL, or else one we generate. On the Slurm idle install only, the default is the node's short host name; set PC_MACHINE_LABEL if the host name says anything you do not want to share. The label also names the Machine's GPUs in the portal. At the Pool, each GPU's worker name is an identifier we assign ("gc-" and 16 characters of the GPU's id with us), never the label.
  • (c) Hardware inventory: CPU model and number of cores; memory; disk size and free space; operating system, kernel and architecture; runtime type and Docker version; for each GPU its model, architecture, compute capability, memory size, power limits, maximum and current clocks, PCIe generation and width (maximum and current), PCI bus id, interconnect and fabric state, and a one-way hash of its id (never the id itself); the driver and CUDA versions; and codes and short messages for driver, GPU or fabric problems.
  • (d) How our work runs on each GPU: its state (downloading, starting, running, paused, failed), download progress, restarts, exit and error codes, the number of our miner's own proof checks that passed or failed, a summary of the GPU's clock and power while our work ran on it, the memory-clock state, and whether its current miner keeps mining while new software downloads; when our work held the GPU and why it stopped (for example a hand-back, a time window, an update), how quickly a GPU was handed back, and how long a GPU sat idle; in Duplex mode, whether mining is paused and the number and total time of its pauses (never the reason or the process that caused one), and the GPU's power-limit state while mining: full or lowered, why (for example turned off with PC_DUPLEX_POWER, or the limit cannot be changed), and when the agent restored your limit.
  • (e) About the Software: its update results, a pause's source and the reason your operator gave for it, and whether it needs a download credential.
  • (f) Network metadata: the public IP address of each connection, which our servers see when the Software reports, downloads the mining software and its updates, and mines. Our web servers' access logs keep the address for at most 14 days; after that we keep only the country it comes from and a keyed pseudonym of its network (to tell clients apart). Some of our signed software is downloaded from the public hosts we publish it on (such as GitHub's container registry, ghcr.io) or from an image mirror: the one set in PC_IMAGE_MIRROR, which the installer of a new Linux or Slurm machine sets to the third-party mirror ghcr.nju.edu.cn when GitHub's registry is slow from your network (--mirror none turns it off; a Duplex install never sets it by itself). That mirror sees the Machine's IP address and which of our public images and archives it fetches. These hosts act under their own terms; every download is checked against its signature or digest. To choose, that installer (not a Duplex one) contacts ghcr.io and the mirror once to time them.
  • (g) Connection choices (from version 0.5.13): which of our pool endpoints the Machine's miners use, that endpoint's region and the Machine's last median connection time to it, and the image mirror in use, if any. To pick the fastest endpoint, the Software opens three connections to each of our pool endpoints every 30 minutes (sooner after a failed one); they carry no data.

5.3 Mining results. The mining software connects to the Pool and sends the shares your GPUs found (proofs of the mining work, made only from numbers our software generated and the Pool's mining job), with your PRL payout address, each GPU's worker name, the mining software's name and version, and the hashrate figures the mining protocol carries.

5.4 Read on your Machines, never sent. To decide when to pause in Duplex mode, the Software looks, on the Machine itself, at the processes that have a GPU open: their names, which GPUs they use and how much CPU and GPU time they use, and it reads their environment variables only to recognise our own processes. None of this is sent to Peryn or stored; the reason for a pause stays in the Machine's local log.

5.5 Signed software only. The Software runs only mining software signed by Peryn, checks every download against its signature or digest, and updates itself only from our signed releases. The fleet keys of a GPU Clouds account are for mining only: no Peryn customer traffic is sent to your Machines.

5.6 How we use it. We are the controller of these metrics and use them only to run the service: assigning and pausing work, applying your rules, paying, your statements and portal, preventing abuse, and support. They describe machines, not people. Personal data can appear only in labels and pause reasons your operators type; please keep it out of them. The names and emails of your account's members are kept under our privacy policy.

5.7 Changes. We will not change the Software so that it sends your customers' data or your models to us. If we add a new kind of data to this section, we tell you 30 days in advance.

5.8 Your part. You run only our signed, unmodified Software, and you do not set it up to mine on your customers' data or with your models. Because Peryn does not process your customers' data, Peryn is not your processor for it.

6. The Fee and what you receive

6.1 The Duplex fee is 2% of the Mined PRL your Machines mine, and there is no other fee. It replaces, on that PRL, the pool fee in our terms of service (section 12.4); it is never added to it. It applies to every Machine enrolled in your account, in Duplex mode or not, including what they mine to the PRL payout address you set in the portal. Other rigs mining to that address pay the pool fee for every miner. It applies even where any of our other terms would give your other mining a lower rate; that other mining keeps its own rate. The Fee is taken from each credit when the Pool credits it to you. It is never invoiced separately. There is no setup or sign-up fee.

6.2 Standard payouts. Your Machines' Mined PRL is credited under the Pool's standard method (PPLNS), not under predictable payouts.

6.3 Payouts. Mined PRL less the Fee is paid to your PRL payout address under the Pool's published payout rules (minimum amount, schedule, block maturity, sanctions screening), as for every miner of the Pool. You change your payout address in the portal with your passkey; the change takes effect within a few minutes, once we have emailed the account's owner about it.

6.4 Statements. Each month the portal gives you a statement: Mined PRL, the Fee and what you were credited, per day and per GPU, the payouts sent, and the Fee's value in USD at each day's PRL/USD fixing.

6.5 Changes to the Fee. We may change the Fee with 30 days' notice in writing. A change applies only to PRL credited after it takes effect. You may end this agreement before then (section 8).

6.6 No promise of earnings. What your Machines mine depends on the Pearl network's difficulty and rewards, the PRL price, your hardware and how long your GPUs are idle. We make no promise about the amount or value of Mined PRL.

6.7 Serving through Peryn is separate. If you also serve requests from Peryn's own customers, what you earn for that follows our partner terms, and the Fee does not apply to it.

7. Updates and pauses

7.1 The Software runs only software signed by us and updates itself from our signed releases. You do not block security updates.

7.2 We may pause the Software on your Machines at once if you break section 3 or 5.8, to protect the Pool, the Pearl network, other users or your Machines from a security risk, or where the law requires it. We tell you why as soon as we can. Mined PRL already credited to you is not affected.

8. Term and ending

8.1 This agreement starts when you accept it in the portal and runs until either party ends it with 30 days' notice in writing (in the portal or by email).

8.2 Either party may end it at once if the other seriously breaches it and does not put it right within 14 days of notice, or becomes insolvent.

8.3 When it ends, we revoke your fleet keys and the Software stops mining; you uninstall it from your Machines within 14 days. Mined PRL credited before the end, less the Fee, is paid out under section 6.3. Sections 3, 5, 6.3 and 9 continue to apply.

9. Ownership, liability and law

9.1 The Software and all rights in it remain ours or our licensors'. You own your Machines, your data and your customer relationships.

9.2 The Software is provided as is. Liability, governing law and disputes follow our terms of service (sections 15 and 17): Dutch law, the courts of Amsterdam.

9.3 You may not transfer this agreement without our written consent. We may transfer it to a company that takes over the Pool. Notices go to the contacts on your GPU Clouds account and to legal@peryn.ai.