Rentals · what we learned · from our own benchmark runs · Updated October 2026
Every number on GPU Battle comes from a GPU we rented by the hour, on Vast.ai, RunPod and Modal. That's more than 2,500 benchmark runs across hundreds of rentals, and plenty of them went wrong. These are the mistakes that cost us money or gave us bad numbers, what they look like, and how to avoid them.
1. Shopping by the hourly price. The hourly rate isn't the bill; the job is. Qwen3 8B runs at about the same speed on an RTX 5090 (244 tok/s) as on an H100 (244), but the 5090 rents for $0.39 an hour against $2.14, so a million tokens costs $0.44 instead of $2.43. Divide the rate by the speed for your model before you compare.
2. Not checking the power limit. On consumer hosts, power caps are the norm, not the exception. Rented RTX 3090s we checked were capped at 230 to 315W of their 350W-plus stock limit, 4090s at 320 to 420W of 450W, and one 5090 at 450 of 575W. A capped card is a slower card at the same price. Run nvidia-smi and look at the power limit in the first minute, and leave if it's well under stock.
3. Ignoring the CPU, RAM and disk. A cheap RTX 3060 we rented sat on a 2016 Xeon that took 30 minutes to compile llama.cpp, all of it billed. We now refuse hosts with fewer than 16 CPU cores, 32GB of RAM, 800Mb/s of internet or 500MB/s of disk. The GPU is only as fast as the machine feeding it.
4. Paying for downloads. A 40 to 60GB model takes a while to download, and on most platforms you pay the GPU rate the whole time. On Vast.ai the host also sets its own bandwidth price. For our image benchmarks on popular cards, the download was most of the roughly 40-cent cost per job. Download once to cheap storage (Modal lets you do it on a CPU machine first), or pick hosts with fast, cheap bandwidth.
5. Trusting the first number. One RTX A5000 host passed every check and still ran Qwen3 8B at 11 tok/s instead of about 120. A hot container read 12 to 29% low. Before you trust a host, run one small known model and compare against a reference like our tables; bad hosts show up in the first minute.
6. Picking a card with a tiny pool. Price doesn't predict whether a rental works; the number of healthy hosts does. When we last checked, Vast.ai had about 30 healthy RTX 4090 hosts and 29 RTX 5090s, but about 10 RTX 3090s and only 3 RTX 3060s, all on old Xeons. With a deep pool you can drop a bad host in a minute and try the next for about a cent.
7. Renting a card the model doesn't fit. A model that doesn't fit in VRAM doesn't run slowly, it doesn't run, and you pay for the setup either way. Check the model's measured VRAM first; every model page here lists it, and the giant models are strict: Qwen3 235B needs about 160GB, and GLM-5.3-Flash needs the B300's 288GB.
8. Forgetting load time on giant models. On a B300, loading a 150 to 250GB model took us 2.5 to 4.5 minutes before the first token, at about $7 an hour. Batch your work so you load once, and don't restart the container between jobs.
9. Leaving it running. The most expensive GPU is the one you forgot. We destroy every instance the moment a job finishes; if you're renting by the hour, set a timer or use a platform that shuts down when the job ends.
Before you rent: a one-minute checklist
| Check | What to look for | Why |
|---|---|---|
| Cost per job | rate ÷ measured speed for your model | the cheapest hour is often not the cheapest job |
| Power limit | nvidia-smi: limit near the card's stock TDP | capped cards run slower at the same price |
| Host machine | 16+ cores, 32GB+ RAM, fast disk and internet | slow hosts bill you for waiting |
| Bandwidth price | per-GB charges on marketplace hosts | big model downloads add up |
| VRAM | the model's measured peak vs the card | too small means it won't load at all |
| First result | one small known model vs a reference number | bad hosts show up in a minute |
| Shutdown | destroy, don't just stop | stopped instances can still bill storage |
Price the job, not the hour; check the power limit, the host machine and the bandwidth price before you start; test one small model against a known number; and destroy the instance when you're done. On Qwen3 8B a rented RTX 5090 costs $0.44 per million tokens against $2.43 on an H100, so the right card for your model matters more than the platform.
Everything here comes from the rentals behind GPU Battle's benchmarks on Vast.ai, RunPod and Modal in August to October 2026: host checks we log before every run (power limit, CPU cores, RAM, bandwidth, disk), runs we threw out as bad hosts, and the cheapest hourly rates we track. Cost-per-token figures use our measured Qwen3 8B speeds and October 2026 rates.