AI video workstation for marketing: which NVIDIA RTX PRO Blackwell for Wan 2.2 and LTX-2.5

Which workstation for AI video, by job

For AI video in marketing, a workstation with one NVIDIA RTX PRO 6000 Blackwell Workstation Edition or Max-Q (96 GB) is the size that holds the large open models, Wan 2.2 A14B, LTX-2.5 and the avatar models, in 16-bit precision. An RTX PRO 5000 Blackwell with 72 GB holds them in compressed formats, and a 24 or 32 GB RTX PRO 4000 or 4500 suits drafts with the small Wan model, upscaling, voices and stills. A second card lets two jobs run at once; when several people generate at the same time, a server serves them better.

This comparison is for freelancers, small studios and in-house marketing teams that want to buy one machine for product clips, social ads, presenter videos and upscaling, and for the IT person who looks after it. It sets four card tiers side by side, explains what the software setup covers, and covers main memory, storage, power and noise. The workflows themselves are in our guides to product videos from photos, local AI avatar videos, a local alternative to Runway and Kling and AI video upscaling, and our guide to video generation GPU requirements lists the memory each model states for team servers.

One term matters for most of this article. VRAM is the memory on the graphics card. A video model's files must sit there while it works; what does not fit is moved back and forth from the computer's main memory (offloading), which works but makes every clip slower.

Four RTX PRO tiers for video side by side

card memory, bandwidth and power from NVIDIA's datasheets; model memory from the Wan model cards (24 GB for TI2V-5B, 80 GB for A14B, both with offloading) and from file sizes on Hugging Face (Comfy-Org Wan 2.2 repackaged, Lightricks/LTX-2.5), read 30 October 2026, totals our arithmetic.

Why memory decides the tier. Wan2.2-TI2V-5B, the small Wan model, generates 720p at 24 frames per second and runs on "a GPU with at least 24GB VRAM" with offloading, its card says; ComfyUI's tutorial even fits it on 8 GB with its own offloading. The large Wan2.2 A14B models state "at least 80GB VRAM" for their 720p command. ComfyUI's FP8 files of the A14B image-to-video model are 14.3 GB per expert, and with the FP8 text encoder (6.7 GB) and the VAE they come to about 35.6 GB. LTX-2.5's 22-billion-parameter transformer is 42 GB in 16-bit, 21.5 GB in INT8 and 18.7 GB in NVFP4, and its Gemma 4 12B text encoder adds 26.3 GB, or 15.4 GB in INT8. Our avatar video guide adds up the avatar chains: about 28.5 GB for InfiniteTalk in FP8 and 49.6 GB in 16-bit. Working memory for the frames comes on top of every figure and grows with resolution and length.

By our usable-memory rule, 90 per cent of the memory the driver reports less 3 GiB per card, a 48 GB RTX PRO 5000 offers about 40.0 GiB (43 GB), so it holds the FP8 Wan set or the NVFP4 LTX set with about 7 to 9 GB to spare; the 72 GB card offers 61.5 GiB and holds either with more room. The 96 GB card is the first that meets Wan's own 80 GB figure and holds LTX-2.5 in 16-bit (about 68 GB of files). Of the 95.6 GiB the driver reports on an RTX PRO 6000, 83.0 GiB are usable by the same rule, which keeps about 19 GiB beside LTX-2.5's 16-bit files for frames and decoding. Send us the clips you plan and we name the card that holds them.

Why the card's compute sets the time. Video diffusion is mostly arithmetic: our Wan 2.2 server article counts about 5.2 × 10^17 operations for one five-second 720p clip with the A14B model at Wan's default 40 steps. NVIDIA's AI TOPS figures show the spread between the tiers: 1,178 for the RTX PRO 4000 (datasheet), 1,617 for the RTX PRO 4500, 2,064 for the RTX PRO 5000, 3,511 for the Max-Q and 4,000 for the RTX PRO 6000 Workstation Edition. No publisher states minutes per clip for Wan 2.2, LTX-2.5 or the avatar models on any RTX PRO card, and we found no independent measurement. The one published video-AI test on these cards we found is Puget Systems' review of June 2025 (updated July 2025): in Topaz Video AI 6.2 enhancement at 1920 × 1080, the RTX PRO 6000 Blackwell Workstation Edition was 30 per cent faster than the RTX 6000 Ada. The test enhanced existing footage and included no generation.

The software: interfaces, models and licences

A workstation with only an operating system and drivers cannot make a video yet. It also needs the interfaces, the models in the formats that suit the card, and ready workflows.

GitHub releases and README of Comfy-Org/ComfyUI; Hugging Face model cards and licence files of Wan-AI, Lightricks, MeiGen-AI, meituan-longcat, ResembleAI; ComfyUI SeedVR2 tutorial; read 30 October 2026.

ComfyUI is the main interface. Its README lists Wan 2.2, LTX-Video 2 and 2.3 and other video models, describes asynchronous queueing and a local API, and ships workflow templates; Lightricks states that official LTX-2.5 templates ship in ComfyUI. App Mode, which NVIDIA's blog of 10 March 2026 calls App View, "presents workflows in a simplified interface": a marketer enters a prompt, picks an image and starts the job, while the node view stays available for the person who builds the workflows.

Wan2GP, a one-window interface for the same models, has its own licence (WanGP Community License 2.0) and is not part of the software we set up.

The licences of the models matter more for marketing than for most uses, because client and ad work is commercial. Wan 2.2, InfiniteTalk and SeedVR2 are under Apache 2.0, LongCat-Video-Avatar-1.5's weights and Chatterbox under MIT. The LTX-2.x Community License requires "Entities with annual revenues of at least $10,000,000" to obtain a paid licence, counting "all subsidiaries, affiliates, and companies under common Control" together. LTX-2.5 is not part of the setup we deliver; a company that meets its licence terms installs it under that licence. We do not install HunyuanVideo 1.5 or MiniMax H3, whose licences exclude the EU. Presenter and voice clips that copy a real person's face or voice need that person's consent, and Article 50 of the AI Act requires deployers to disclose deep fakes from 2 August 2026; our avatar video guide covers both, and your legal counsel confirms what applies to your clips.

Updates. ComfyUI states a weekly release cycle, and new models arrive with new workflows. Decide within your team who updates the installation, and re-run one reference workflow after each update before client work continues.

You can order the workstation alone, the setup of ComfyUI, Wan 2.2, the avatar and voice models and SeedVR2 alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Compressed FP8 and NVFP4 models on 48 and 72 GB cards

FP8 and NVFP4 store each weight in 8 or about 4 bits instead of 16, so the files shrink: one Wan 2.2 expert from 28.6 to 14.3 GB, the LTX-2.5 transformer from 42 to 18.7 GB. NVIDIA reported in its blog of 10 March 2026 that NVFP4 in ComfyUI makes generation "2.5x faster and VRAM is reduced by 60%", and FP8 "1.7x faster and VRAM is reduced by 40%". Those figures were measured on GeForce RTX 50 Series cards, with an RTX 5090 in the chart; NVIDIA published no such figure for RTX PRO. The RTX PRO Blackwell cards have the same FP4 Tensor Cores, and LTX-2.5's card lists its NVFP4 file with Blackwell GPUs.

This is what lets a 48 or 72 GB card run models whose 16-bit files need 96 GB. Compressed weights can change the picture slightly, so compare a few clips in both formats before you switch client work to the smaller files.

Main memory, storage, power and noise

Main memory. Offloading uses the computer's main memory. By the arithmetic in our Wan 2.2 server article, each A14B job in 16-bit keeps about 69 GB there. Our workstation configurations follow NVIDIA's rule of at least twice the GPU memory, so 192 GB next to one 96 GB card and 384 GB next to two; our guide to how much RAM a GPU server needs explains the rule.

Storage. The models in this article take a few hundred GB on disk: the 16-bit A14B image-to-video experts alone are 57.2 GB, LTX-2.5's 16-bit transformer and text encoder 68.3 GB, the LongCat avatar repository 74.9 GB. Plan fast NVMe storage for the models separately from the storage for projects and finished clips, and keep prompt, seed and model version with every clip.

Power and heat. Video generation keeps the card at full load for minutes per clip, so plan with the full figure. The single-card tiers draw 145 to 600 W on the card alone. Our article on two RTX PRO 6000 Workstation Edition estimates about 1.8 kW at the wall for such a tower, 7.8 A on 230 V, within one 16 A circuit planned at 80 per cent, and notes that the flow-through coolers release their heat inside the case; two such towers do not fit one circuit. Two Max-Q come to about 1.1 kW and four to about 1.8 to 2 kW, as our Max-Q power guide estimates, and all of it becomes heat in the room.

Three or four cards. NVIDIA states up to four Max-Q in one workstation. Puget Systems (July 2025) found that directly adjacent Max-Q cards have little room to draw air and recommends a slot gap; Exxact (updated March 2026) reports that all four adjacent cards throttled without its custom airflow and held full 300 W at up to 85 °C with it. Three or four Max-Q therefore belong only in a chassis built for them, as our guide to the four-card Max-Q workstation describes, and two Workstation Edition cards are the realistic maximum for that edition.

Noise. Full load for minutes at a time keeps the fans fast for long stretches. Place the workstation where they do not disturb calls, or in a side room with the team working in ComfyUI through the browser.

When a workstation stops being enough

A video job holds a card for minutes, so the number of people who generate at the same moment sets the card count more than the daily volume. One person works well on one card. With two or three people submitting at once, each job waits for the one before it, and a second card helps. Beyond that, a server with one job per card is the better tool: four RTX PRO 6000 Server Edition cards run four clips at once and draw about 3.4 kW by our estimate, which needs a server room with a 32 A or three-phase feed. The Server Edition is also the only RTX PRO 6000 edition with vGPU, for artists who work remotely.

Our guides to the AI server for creative agencies, to a GPU server or workstations for a team and to the Wan 2.2 video server cover that step.

We supply RTX PRO workstations from one RTX PRO 4000 to two RTX PRO 6000 Workstation Edition or four Max-Q, and servers for larger teams; tell us which videos you make, how many people generate at once and where the machine will stand, and we propose the tier that fits.

FAQ