From a Customer Question to $0.80 per Minute of Generated AI Video.

It started with a regular customer meeting.

We had scheduled one of our discovery calls. The customer was looking for ways to improve their Governance, Risk, and Compliance processes with our GRC automation service.

Nothing unusual for us. This is exactly the kind of conversation we have with companies trying to reduce manual work, improve internal processes, and understand what can realistically be automated.

But somewhere in the middle of the conversation, we went completely off-topic.

The customer started talking about their marketing pipeline.

They were producing videos using AI, including dubbing content into different languages.

Nothing to do with GRC.

Then came a question I wasn’t expecting:

“By the way, if you do local AI, can you make this run locally for us?”

Interesting question.

We work with local models. Gemma, MedGemma, Qwen, Mistral — these are some of the model families we deal with, mostly for text-related tasks.

The Next Morning, We Started Looking for Answers

The next morning, my team and I started researching what was actually available.

At first, the task sounded relatively simple.

Find an open-weight video generation model, deploy it on locally, and see whether we could produce something comparable to the commercial services the customer was already using.

But characters had to remain recognizable across different scenes. Voice and sound were important parts of the production process.

That shortened our list considerably.

We looked at Wan 2.2, LTX-2.3, LTX-2.5, HunyuanVideo, and MiniMax H3.

But we weren’t looking for a model that could simply animate a picture.

We needed something capable of working with several types of references while producing synchronized audio and keeping characters reasonably consistent.

During the research, MiniMax H3 — and particularly MiniMax H3 Ref2VA — caught our attention.

Two and a half hours to generate five seconds video!

Our initial environment was an Apple M2 Pro with 32 GB.

We wanted to understand what was possible with hardware we already had before moving to something more expensive.

And it worked.

We generated a 5.17 seconds video.

Generation time:

2 hours, 30 minutes, and 16 seconds.

Eight denoising steps alone consumed around 142 minutes. Video decoding added another 7.7 minutes. Peak GPU memory usage reached 20.06 GB.

So yes, the pipeline worked.

But clearly this wasn’t something we could recommend for a customer’s AI video production process.

The model wasn’t the immediate problem.

Hardware was.

We Needed GPU Power

Then came another question.

Hourly GPU rental was interesting. Persistent storage was useful.

But something else attracted my attention…

ComfyUI Is Like n8n, But for Models

By this point, we already had the video generation pipeline running on Mac.

We didn’t need ComfyUI to prove the model could run.

What I liked was the interface.

If you’ve worked with n8n, you’ll recognize the basic concept.

In n8n, you connect nodes to automate processes between applications and services.

ComfyUI uses a similar visual approach, but for generative models.

You connect nodes for loading models, adding image, video, and audio references, configuring generation parameters, and producing the final output.

You can see the entire pipeline in front of you.

You can change individual steps without constantly modifying scripts.

For me, that was much more convenient than doing everything through GitHub.

So for the proof of concept, the direction became clear:

Not Every RTX 5090 Performs Like an RTX 5090

Then things became more interesting.

We started experimenting with RunPod Community Cloud.

The idea sounded good: rent an RTX 5090 at Community pricing, generate the videos, terminate the GPU when finished.

But we noticed something unexpected.

Some RTX 5090 hosts had a GPU power limit of only 400 W.

And the performance difference was enormous.

At first, it looked like Community Cloud might simply be too unpredictable.

Then we found another Community RTX 5090 with a 536 W power limit.

The actual problem was:

How do we find a good GPU before spending time and money downloading models and attempting a generation?

Find a GPU

Normally, you launch the GPU, connect to it, download the models, and start generating.

But if the machine turns out to be slow or unreliable, you’ve already wasted time.

Our pruned model set was approximately 72 GB.

Downloading 72 GB before discovering that the GPU is useless is not a great workflow.

So we added GPU discovery to our deployment script.

The logic is relatively simple.

Our script can:

1. 1Request an RTX 5090 Community Cloud pod with 120 GB of disk space.
2. 2Wait for SSH and verify CUDA availability through PyTorch or ComfyUI.
3. 3Check the GPU power limit.
4. 4Sample idle GPU utilization.
5. 5Reject hosts reporting less than 500 W or averaging more than 20% GPU utilization before our workload starts.
6. 6Terminate the rejected pod, wait two minutes, and try again.

A cheaper GPU per hour doesn’t necessarily produce cheaper video.

Once the pipeline was stable enough to test repeatedly, we started measuring generation time and infrastructure cost.

For consistency, we used a five-second workload with 20 sampling steps at 736 × 1312 resolution.

We compared three configurations.

The result was interesting.

When we found a fast Community host, the estimated infrastructure cost was approximately half the Secure Cloud cost.

But on slower Community machines, the projected cost increased to roughly $3.20–$4.90 per generated video minute.

In our measurements, Community Cloud only remained cheaper than Secure Cloud when the host completed a sampling step in approximately 31 seconds or less.

That is exactly why our GPU discovery script became useful.

These numbers include estimated storage charges but exclude startup time, idle resources, failed generations, editing, and other production overhead.

And What About Startup Time?

Cost wasn’t the only thing that mattered.

We also wanted to understand what happens when somebody actually wants to start working.

The difference wasn’t only cost.

It was predictability.

Secure Cloud or Community Cloud?

At this point, we had two practical deployment options.

There is no reason to pretend that one option is universally better.

Let the customer decide whether cost or predictability matters more.

And there is another important distinction.

Both of these options still use rented GPU infrastructure.

They provide more control over the execution environment and the models we run, but they are not the same as running everything on physical hardware inside the customer’s own infrastructure.

Does any of this actually make financial sense?

Commercial video services already exist.

You subscribe, upload your references, generate your video, and pay for the credits.

The difference is substantial.

Even compared with the cheapest Seedance option Team plan, our Secure Cloud full-model configuration came out around 72% cheaper.

And Community configuration reached approximately $0.80 per generated minute.

That’s around nine times cheaper than the Seedance Team figure in our comparison.

What Started as a GRC Conversation

We started with GRC and ended up building a video generation pipeline.

Not exactly what we had planned.

But that’s what I like about working directly with customers. Sometimes the most interesting projects start with a question that has nothing to do with the original meeting.

We had to investigate, experiment, fail a few times, and figure things out.

Now we have a working proof of concept, actual performance measurements, and a much better understanding of the costs.