Private AI · Video Production
You keep one clip in eight. You are paying for all eight.
Every per-second API bills you for the takes you delete. Own the render pipeline and iteration becomes free, which changes how your creatives work rather than just what you spend. We benchmark the open-weight models on your shots, train them on your brand assets, and build the queue that keeps the hardware busy.
/// run the numbers
What a delivered clip really costs you.
Not the rate card price. The price including every take you rendered and threw away to get the one you shipped.
Your output
Clips that actually reach a client or a channel.
How many takes you generate to keep one. Be honest, most studios are between 6 and 12.
Seconds of finished output per clip.
Default is LTX's published 4K rate of $0.30 per second. Set it to whatever you actually pay Runway, Veo, Kling or LTX.
Hardware you would buy
1x RTX PRO 6000 Blackwell · 96 GB
The real entry point. 96 GB is the first tier that clears Wan 2.2's 80 GB bar, so this is where you stop choosing models around your hardware.
True cost per delivered clip, on an API
$24.00
You render 3,200 seconds a month to deliver 400, so you are paying for 2,800 seconds of video nobody will ever watch. On your own hardware that same clip costs $4.50 in running cost.
Monthly API spend
$960
At the rate and volume you set.
Up front, owned
$11k
Studio tier hardware. Indicative.
Breaks even
Month 15
Then running cost only.
Can that hardware keep up?
2% busy
That volume needs about 13 GPU hours a month against roughly 700 available. That machine would sit idle almost all month. The Pilot tier at $4,000 handles this volume with room to spare, and the difference buys a lot of API credit. We would rather sell you the smaller box.
Over 36 months
$17k kept over three years, and iteration stops being a line item your producers have to police.
How to read this. Hardware prices and throughput are indicative figures for deployments of this shape, not a quote, and throughput in particular swings by roughly ten times depending on which model you land on. This compares hardware against API fees only. Our build fee is separate and quoted once we have seen your shot list. We benchmark on your actual footage before anyone buys a GPU.
/// the models, september 2026
What you can actually self-host today.
Vendor minimums assume aggressive offloading and often describe the smallest model in the family. Where secondary sources disagreed by a factor of three, we went to the official repository.
| Model | Params | Floor | Production | Output | Speed | Licence |
|---|---|---|---|---|---|---|
| LTX-2.5Lightricks | 22B | 16 GB | 48 to 96 GB for 4K with headroom | 720p, 1080p, native 4K, up to 20s, 24/25 fps | 10s clip in 6.8s on two GB200 superchips. 23.7s through the vendor API at 1080p. | LTX-2.x Community, free under $10M ARR |
| Wan 2.2 T2V-A14BAlibaba | 27B MoE, 14B active | 80 GB | 80 GB, H100 or a 96 GB workstation card | 480p, 720p | Roughly 10 to 12 minutes per 5s 720p clip on an H100. | Apache 2.0 |
| Wan 2.2 TI2V-5BAlibaba | 5B dense | 24 GB | 24 to 32 GB | 720p at 1280x704 | 5s 720p clip in under 9 minutes on a single consumer card. | Apache 2.0 |
| HunyuanVideo 1.5Tencent | 8.3B | 14 GB | 24 to 48 GB | 480p and 720p native, 1080p by super-resolution | Step-distilled 480p image-to-video in about 75 seconds on an RTX 4090. | Tencent community terms |
| HunyuanVideo (original)Tencent | 13B | 60 GB | 100 to 120 GB at 1080p, so an H200 | 720p, 1080p | Roughly 12 to 18 minutes per 5s 720p clip. | Tencent Community Licence |
The licence question nobody asks until it is too late
Only the Wan family is plainly permissive under Apache 2.0. LTX-2.5, the strongest of these and the only one with synchronised audio, is free for organisations under $10M annual recurring revenue measured across all affiliates and subsidiaries, and negotiated above that. Tencent ships HunyuanVideo under its own community terms rather than a standard open-source licence. If you are an agency, your revenue and your client contracts decide which of these you can legally build a service on. We check that before recommending a stack, because unwinding it later means retraining every LoRA.
/// what to buy
Four tiers. One real decision.
VRAM decides which models you can run. Card count decides how much you can ship. Almost everyone gets this backwards and buys one expensive card when they needed three cheaper ones.
Utilisation is the whole game
Owned hardware only beats an API when the cards are busy. A render queue that batches overnight is what turns a capital purchase into a lower cost per delivered second, which is why we build the queue and not just the pipeline.
Pilot
1x RTX 4090 or 5090
24 to 32 GB · about $60/mo to run
Runs: Wan 2.2 TI2V-5B, HunyuanVideo 1.5, LTX-2.5 at lower resolutions
One artist proving the workflow. Clears the licence question and the quality question before you spend real money. Not a production line.
Studio
1x RTX PRO 6000 Blackwell
96 GB · about $180/mo to run
Runs: Everything on the list, including Wan 2.2 14B and LTX-2.5 at 4K
The real entry point. 96 GB is the first tier that clears Wan 2.2's 80 GB bar, so this is where you stop choosing models around your hardware.
Datacentre
1x H100 80 GB or H200 141 GB
80 to 141 GB · about $420/mo to run
Runs: All models at maximum quality, HunyuanVideo at 1080p
Buy this for the quality ceiling, not for throughput. If you want more clips per day, more of the Studio tier is better value per rendered second.
Production line
4x RTX PRO 6000 Blackwell
384 GB total · about $700/mo to run
Runs: Four concurrent jobs, or one job sharded across cards
These models scale across cards with no NVLink needed, so four cards is close to four times the output. This is what an agency shipping daily actually buys.
/// what we build
The hardware is the easy part. This is the work.
Anyone can install ComfyUI. Getting consistent, on-brand output from it at volume, with a queue that keeps a $46,000 machine busy, is a different job.
Model Selection On Your Shots
We benchmark the open-weight shortlist against your actual reference frames and brand look, not a public leaderboard. Motion, faces, text legibility, colour, whatever your work turns on.
LoRA Training On Your Assets
Your products, your talent, your house style trained into the model so output lands on brand without twenty prompt attempts. This is the part that removes most of the iteration waste.
ComfyUI Pipelines Your Team Can Drive
Node graphs built for your workflow and locked down so an artist can render without touching Python. Versioned, so a look that worked in March still works in September.
A Render Queue, Not A Person Waiting
Jobs submitted, batched overnight and picked up in the morning. Utilisation is what makes hardware cheaper than an API, and a queue is what gets you utilisation.
Review, Versioning And Delivery
Generated takes land somewhere your team and your client can actually review them, with the prompt and seed attached so a good result can be reproduced rather than rediscovered.
/// is this for you?
Great for some studios. Wrong for others.
This fits if
- ✓You generate video every week, not for the occasional pitch
- ✓Your per-second API bill has become a real line item
- ✓You iterate heavily and throw most takes away
- ✓Clients demand their footage and brand assets stay in-house
- ✓You want the same look reproducible six months later
This does not fit if
- ✗You make a handful of clips a month, where APIs are cheaper
- ✗You need the absolute frontier look on one-shot text prompts
- ✗Nobody on the team will own the pipeline after handover
- ✗You need it working next week with no benchmarking
- ✗Your revenue puts you over a licence threshold you cannot negotiate
/// questions
Before you buy a GPU.
How much does a private AI video pipeline cost?+
A scoped quote needs your reference shots, output requirements, expected volume and existing hardware. Benchmarking, model setup, asset preparation, LoRA training and the render queue affect the engineering cost. The calculator on this page estimates hardware and running costs; it is not a quote for our implementation services.
How long does it take to set up a private AI video pipeline?+
The schedule depends on benchmark results, hardware availability, the readiness of your assets and the amount of custom training and pipeline work. Include time for review, reproducibility checks and team handover. Clip rendering time is separate from the time needed to implement the pipeline.
Is open-weight video actually competitive with Veo or Sora?+
For controlled shots where you supply the reference, the input image and the brand assets, the gap is much smaller than the demo reels suggest, and LTX-2.5 generates synchronised audio in the same pass. For one-shot text-to-video on an open-ended prompt, the closed frontier models are still ahead. The honest answer depends on how much control you have over the input, so we benchmark on your shots before you buy anything.
What GPU do we actually need?+
96 GB is the first tier that clears Wan 2.2's 80 GB requirement, which is why an RTX PRO 6000 Blackwell is the realistic entry point rather than a 4090. You can start on 24 to 32 GB with the smaller models to prove the workflow, and plenty of studios do. Beyond that, adding cards buys throughput rather than capability, because these models parallelise across GPUs without needing NVLink.
Can we use these models commercially?+
Depends on the model, and it is the question most people skip. The Wan family is Apache 2.0, which is plainly permissive. LTX-2.5 is free for organisations under $10M annual recurring revenue and negotiated above it, measured across affiliates. Tencent ships HunyuanVideo under its own community terms. We check the licence against your revenue and your client contracts before recommending a stack.
How long does a clip take to render locally?+
Anywhere from about 75 seconds to 18 minutes for five seconds of output, depending entirely on which model and resolution. LTX-2.5's headline 6.8 seconds was measured on two GB200 superchips, not a workstation. Model choice moves throughput by roughly ten times, which is why we benchmark before specifying hardware.
What happens when a better model comes out next month?+
That is the strongest argument for owning the pipeline. Open weights mean you swap the model and keep the LoRAs, the node graphs and the queue. We test the new model against your reference set and only recommend the move when it scores better on your work.
Do we need to hire someone to run it?+
No, but somebody on your side should own it. We build the pipelines so an artist drives them without touching code, and we train your team on the handover. If you would rather we kept operating it, that is a separate retainer.
Handling sensitive records rather than making video? See private AI for regulated data →
Bring us three shots you had to redo.
We will tell you which open model handles them, what hardware it needs, and whether owning the pipeline actually beats your current API bill.