"Running H3 needs at least 24GB of VRAM" — that line scares a lot of people off. But you don't have to buy that VRAM; you can rent it. This article runs the numbers: at what usage level buying a card pays off, and when renting is almost always the better deal.
| Item | Local (own GPU) | Cloud (rented GPU; YouYun Zhisuan 2026-09 prices as example) |
|---|---|---|
| Upfront cost | ~¥10,000-class for a 24GB GPU; more for 32GB-class | ¥0 |
| Running cost | Electricity ~¥0.4-0.6/hour (whole-system load) | ~¥2.15/hour for a 4090; ~¥3.32/hour for a 5090; no charge while powered off; ¥0.15/hour to keep your data in GPU-less mode |
| Model files | You supply the 42.5GB+ download and SSD space | Platform's public model library, no download (some platforms ship a prebuilt H3 image for one-click deployment) |
| Environment upkeep | You maintain CUDA, dependencies, and drivers yourself | Preinstalled image, ready to use out of the box |
| Privacy | Footage never leaves your machine | Footage is uploaded to the cloud (evaluate carefully for sensitive projects) |
| Available GPUs | Whatever you bought | Switch between 4090 24G / 4090 48G / H20 96G on demand |
Take the 4090 24G (comfortably runs H3 at 4-8 steps, 540p-768p): renting costs ¥2.15/hour, so a purchased card at ¥15,000 equals about 7,000 hours of rental time. In other words:
| Your situation | Recommendation |
|---|---|
| Just starting out / under 20 hours a month | Go straight to the cloud: spend a few dozen yuan over a month or two and figure out the model and workflow first |
| Already have a 12-16GB GPU, want the occasional 768p run | Low-end local + cloud for the rest: roll at 540p locally, rent a 4090 for final renders |
| Have a 16GB+ GPU, use it 10-20 hours a week | Local first, moving to the cloud for 15-second clips or high resolutions |
| Studios running 40+ hours a week | Local first, cloud for peaks: run batch jobs locally, rent cloud capacity when racing a deadline |
| Sensitive footage (client contracts, unreleased products) | Local first, or pick a cloud option with private storage |
On a 4090, a 768p, 5-second H3 clip at 4-step distillation takes roughly 2-4 minutes to sample and decode (depending on the acceleration setup and resolution; you can use the WhichH3 calculator to estimate per-clip time for your target configuration first). At ¥2.15/hour, that's about ¥0.1-0.2 per clip — cheaper than most closed-source APIs, and cheap enough to reroll over and over without wincing.