The marketplace rents real GPU hardware by the second. This explains what you are choosing between and what to check before you start the clock. ## On-demand and interruptible On-demand instances run until you stop them. Interruptible instances cost less and can be reclaimed when someone bids higher for the same hardware. The choice follows from whether your job can be interrupted safely. Training that checkpoints regularly is a good fit for interruptible — losing ten minutes since the last checkpoint is an annoyance. Anything holding state that cannot be reconstructed, or serving live traffic, should be on-demand. ## Reading a listing Each machine shows its GPU type and count, VRAM, CPU and RAM, network speed, region, and an hourly rate per GPU. Two fields deserve attention. **Verification status.** A verified machine has had its self-reported hardware checked against what its host agent actually reports — GPU model and count compared directly, not taken on trust. **Reliability rating.** Derived from the machine's health-check history: the proportion of checks it has passed. A fast machine with poor reliability will cost you more in restarted work than it saves per hour, and the DLPerf-style score shown on listings already factors reliability in for exactly that reason. ## Sizing before you rent Training needs substantially more memory than inference on the same model — optimizer state and gradients commonly triple the requirement over weights alone. This is the single most common reason a rental runs out of memory ten minutes into paid time. Work out weights, add optimizer state and gradients if training, add activation memory for your batch size, then compare against the listing. Then add margin for setup, data transfer, and at least one failed attempt, because billing runs while you debug. ## Getting in Once an instance is running you get SSH access, and JupyterLab where the image provides it. Instances are reachable over TLS through the platform, so you do not need to open ports or set up tunnels yourself. Data transfer is worth planning. Pulling a large dataset onto a fresh instance is billable time before any work starts, and doing it twice because the first attempt failed is a common way to double the cost of a short job. ## Tearing down Billing stops when the instance stops, so stop it. The most expensive mistake on any rental marketplace is an instance left running after the work finished — an idle GPU bills exactly like a busy one. Today's inventory is small and honest about it: check the marketplace for what is actually online right now rather than assuming a given card is available. Capacity is being added.

Renting GPUs on the Inferix marketplace: a practical guide