What a ComfyUI Pipeline Actually Costs a Post House

Everyone quotes the GPU price and calls it the budget. The GPU is the smallest surprise. Here's what a production ComfyUI pipeline actually costs a post house — and where the money really goes.

The hardware is the honest part

A realistic local inference box is a high-VRAM NVIDIA card (a 5090-class GPU is the practical sweet spot today), a fast CPU, 64–96 GB of RAM, and NVMe that doesn't bottleneck model loads. Add a colour/post workstation — often a Mac Studio — for finishing. That's real money, but it's a one-time, knowable number.

Cloud is the opposite: near-zero capex, and a bill you can't predict. For steady production volume, local hardware is usually cheaper per render and keeps assets on-premise.

Model licensing is the quiet line item

Open weights don't automatically mean free for commercial use. Licence terms differ by model and version, and some restrict commercial output or require attribution. Budget time — or a legal review — to map the licences of every model in your graph, and pin versions so a silent update can't change your obligations.

Artist time is the biggest hidden cost

A ComfyUI graph is not a product. Turning a working node graph into something a team can run reliably takes documentation, sensible defaults, error handling and naming that matches your delivery specs. Expect the "make it usable" phase to cost more than the "make it work" phase.

The build is 20% of the cost. Making that build survive a Monday deadline is the other 80%.

Retraining is recurring, not one-off

New models ship constantly and no single model is best at everything — you'll end up combining several. That means periodic re-evaluation, not a build-once-and-forget system. If you train LoRAs for a look or a character, add the training and curation time too.

What this means for your budget

• Hardware: one-time, knowable — the easy part to quote.
• Licensing: low dollar cost, real review time — don't skip it.
• Artist/engineer time: the dominant cost — budget for usability, not just a demo.
• Retraining: recurring — plan an ongoing maintenance line.

Get those four right and the pipeline pays for itself. Get them wrong and you build something impressive that nobody uses. If you want the numbers for your specific volume, that's what the pipeline discovery is for.