Renting GPUs without a usage review

Renting a modern accelerator from a large provider increasingly means a corporate account, an identity check and an acceptable-use policy that reaches into what you are permitted to train or serve. For plenty of legitimate work — research on models the vendor would rather not host, fine-tuning on a corpus you cannot describe to a third party, or simply not wanting a procurement process — that is a poor fit.

Why the gating exists

Some of it is genuine risk management and some is liability theatre. Either way the effect is the same: access to an H100 now routinely requires a conversation about what you intend to do with it, and a record of who you are attached to the answer. Very few providers will rent one against a prepaid balance and no name.

What we do instead

• You get the whole physical card. No MIG partition, no time-slicing, full VRAM, and nvidia-smi reports what the spec sheet promised.
• No dataset inspection, no model gating and no use-case questionnaire. The Acceptable Use Policy covers crime, not capability.
• Hourly billing from the prepaid balance. Destroy the instance and the meter stops that minute.
• No identity anywhere in the chain — an email, a password and a coin balance.

Sizing it

1. 01Inference on open-weight modelsVRAM is the whole question. A 24 GB 4090 handles most models up to roughly thirteen billion parameters quantised; 80 GB on an A100 or H100 is what you need above that.
2. 02Fine-tuningBudget two to three times the memory you would need for inference on the same model. LoRA and similar methods cut this substantially and often bring a job back within reach of a 4090.
3. 03Throughput-bound workH100 over A100 where the budget allows — the gap on transformer workloads is much larger than the specification difference suggests.
4. 04Rendering and simulationA 4090 or 5090 is usually better value than a data-centre card. The clocks are higher and the memory ceiling rarely binds.

Practical notes

Images ship CUDA-ready with recent cuDNN, PyTorch and TensorFlow, or you can start from bare Debian and build the stack yourself. Checkpoint to a storage volume rather than to local scratch, so an instance you destroy does not take a week of training with it. And take the hourly billing seriously: a card left running over a forgotten weekend is the most common surprise on any GPU bill anywhere.

Published by NoDMCAVPS, an offshore host that files automated DMCA notices instead of forwarding them. What we still remove is listed in the acceptable use policy.