Wan 2.2 on a GPU VPS: how much VRAM do you need for video generation with ComfyUI

Generating video with AI models locally remains impractical below a certain VRAM threshold: a still-image model will run even on a laptop, but a 14-billion-parameter video model won't. For anyone managing infrastructure and weighing up whether renting a GPU makes sense, the difference between "technically possible" and "usable" is measured in hours per generated clip.\n\nWan 2.2, the open-weight video generation model from Wan AI released under the Apache 2.0 licence, is now natively supported in ComfyUI via predefined workflow templates, with no need to build the graph by hand. The technical detail that matters when sizing a server: the hybrid Wan2.2-TI2V-5B model runs with roughly 8 GB of VRAM thanks to ComfyUI's native offloading, while the two 14B expert models (one for high noise, one for low noise, using a mixture-of-experts architecture) require considerably more memory. Wan's reference implementation calls for 80 GB or more of VRAM for A14B inference on a single card, a threshold that rules out every consumer GPU and a good chunk of the mid-range rental market.\n\nFor installation you'll need: an Ubuntu image with NVIDIA drivers and the CUDA toolkit, root or sudo SSH access, and enough disk space for the weights (the 5B model weighs roughly 10 GB, the 14B text-to-video pair in fp8 comes to around 29 GB, and the image-to-video pair in fp16 reaches about 57 GB). The practical route is: clone the ComfyUI repository, create a Python virtual environment, install a build of PyTorch matched to the card's CUDA version (cu130 or higher is needed on NVIDIA 20-series cards and above), download the weights from the Comfy-Org/Wan_2.2_ComfyUI_Repackaged repository on Hugging Face, then start the server bound to 127.0.0.1 and reach it via an SSH tunnel: ComfyUI has no authentication of its own, so exposing it on 0.0.0.0 means opening up a remote code execution endpoint to anyone who finds the port.\n\nContabo offers a GPU VPS with a dedicated NVIDIA RTX PRO 6000 card with 96 GB of GDDR7, 18 vCPUs, 96 GB of RAM and 900 GB of NVMe storage, at a fixed monthly fee that, according to the official price list, starts at €999 per month plus VAT ($1,199, £969), with a 15% discount on 12-month contracts and 20% on 24-month contracts. With 96 GB of VRAM, the card covers every Wan 2.2 variant, including the 14B models, without needing to split the workload across multiple GPUs. Available locations for this product are in Europe and the central United States.\n\nAll Wan 2.2 weights are distributed under the Apache 2.0 licence, which permits commercial use provided the copyright notice and licence text are retained.