GPU Test Privacy Notice

Who we are

Strike DFS, Inc. (Liquid Compute) runs the GPU test and decides how the information described here is used. A GPU test checks NVIDIA GPU nodes you administer: our servers connect to them over SSH with access you grant, run hardware diagnostics and performance tests, and report the results against reviewed thresholds. You can start a test on our website without an account, or from an AI agent through our MCP server, which requires signing in to a Liquid Compute account.

What the test collects

Node access details: each node's public IP address, SSH port and user name, and its SSH host key. A test started through the MCP server is a signed-in test: we also record the company and the user account that started it. A test started on our website needs no name or email address; it is tied to your browser by a session token, of which we store only a one-way digest.

Machine information read from each node: hostname; operating system, kernel and architecture; CPU model and core counts; memory, local storage and network port details; system vendor and product name; checks of the test's prerequisites, such as Python version, passwordless sudo, free disk space and the NVIDIA tools and services the test uses; and, for each GPU, its model, UUID, serial number, VBIOS, driver and CUDA versions, memory, PCIe link, power limit, ECC, MIG and error-counter state, interconnect topology, GPU error (XID) entries from the kernel log, and the processes using it (process name, ID and GPU memory). We also record the names, but not the values, of NCCL and UCX environment variables set on the node.

Test results: the raw output of the diagnostics and benchmarks (NVIDIA DCGM, CUTLASS GEMM, nvbandwidth and NCCL), GPU telemetry sampled under load (temperature, power, utilization, clock throttling and error counts), records of the test files we copied to each node and removed, and the evaluated report. We also keep a timeline of the test, including when you confirmed that you are authorized to give us SSH access and that the nodes are reserved for the test. The test does not collect the data, models or code of your workloads.

How we use it

We use this information to run your test and produce your report; to operate, support and secure the service, including diagnosing failed tests and preventing abuse; and to improve our tests and their thresholds. We may combine results from many tests into de-identified, aggregated statistics about GPU health across the market, such as failure rates by GPU model, driver version or hosting provider, and publish them. Published statistics do not identify you, your company or your machines. If you ran a signed-in test, we may contact you through your Liquid Compute account about your tests, or to ask whether you would like to supply compute on Liquid Compute. We do not sell this information or share it for cross-context behavioral advertising. The GPU test page uses cookieless product analytics from PostHog's EU-hosted service to count page views and test starts; those events do not include your node details.

Our reasons for using it

We run a test because you ask us to, and only after you confirm that you are authorized to give us SSH access to the nodes and, before any load starts, that the nodes are reserved for the test and you are authorized to run it. We use test information to operate, secure and improve the service, to produce de-identified statistics, and to contact signed-in users because these are our legitimate business interests, limited as this notice describes. You can object to any of these uses, or ask us to stop contacting you, using the contact below.

Who receives it

Authorized Liquid Compute personnel can see test records to run, support and secure the service. Service providers process the information on our behalf: Vercel hosts and protects the website and processes limited technical information, such as IP address and browser, request, and security signals; our infrastructure providers run the servers that connect to your nodes and store test records and outputs; and PostHog receives the analytics events described above. On a signed-in test, other users on your company's Liquid Compute account can see and manage the company's tests. When you create a report link, anyone with the link can open that copy of the report, which numbers your nodes and leaves out their addresses and SSH details. We do not otherwise share test information, except where the law requires us to.

How long we keep it

We keep test records, including node access details, machine information, raw test outputs, reports and the test timeline, until they are deleted at your request; they do not currently expire automatically. Raw outputs are kept in private, encrypted storage. The time limits you see in the product limit access, not storage: the browser session that sets up and runs a test works for 24 hours, your report stays open to you for 30 days after you create the test, and a report link stops working 30 days after it is created, or sooner if you revoke or replace it. After deletion, limited residual copies may remain temporarily in provider backup systems.

Your choices

Remove our line from authorized_keys on a node to end our access to it. On a test started on our website, Revoke link on the report page turns off its report link; a new link, from the website or your agent, replaces the previous one. Email privacy@liquidcompute.com to ask a privacy question; to request access to, correction of, or deletion of your test information; to object to a use described here; to turn off a report link; or to ask us to stop contacting you. To find a test run without an account, we may ask for details that identify it, such as the node addresses and when it ran. We may need to verify your identity and may limit a request where applicable law permits or requires it.

Changes and contact

We will update this notice and its effective date if our GPU test practices materially change. Contact Strike DFS, Inc. (Liquid Compute) at privacy@liquidcompute.com.