Back to Home

Comfy Cloud Runs Your Own Nodes, Then Bills You for It

ComfyUI has spent two years winning an argument it never fully intended to win: that the best way to make images and video with AI is to own the graph. Node by node, pack by pack, the community built things no closed product would ship. The catch was always the machine. Your workflow lived on your box, with your forty custom nodes and your 30 GB of models, and handing it to a colleague collapsed into a README full of pip errors.

On November 18, Comfy closed that gap. Custom Builds now package your nodes, models, LoRAs and Python dependencies into a named environment that runs in Comfy Cloud, on GPUs you pick from a menu that tops out at a 180 GB B200. Your team opens the same canvas with the same nodes and gets the same result without installing anything. It is, technically, the most important thing the company has shipped since it open-sourced the app.

It is also the moment the project's economics stopped being a footnote.

What actually shipped

The mechanics are unglamorous. You start from a workflow or a Comfy Desktop snapshot. The builder reads the graph, finds the custom node packs and models it needs, and helps you resolve dependency conflicts. You publish the build and it becomes a numbered version. The serverless build usually finishes in 10 to 15 minutes. Then anyone in the workspace selects it in the editor, and the Run menu flips between stock Comfy Cloud and your environment.

Private and internal node packs are supported, which matters more than the marketing copy suggests. Studios that maintain proprietary nodes for their own pipeline can now push them into a managed cloud without exposing them to the ecosystem.

Two cheaper escape hatches are worth knowing before anyone spends a weekend on a build. Creating, publishing, and keeping a Build costs nothing. If your workflow only needs a LoRA or a single model, and stock Comfy Cloud already has the loader node, you can upload a .safetensors file of 2 GB or less to the Model Library and pick it in lora_name. That needs a paid plan and a deployment created after the Model Library existed.

The bill is the story

Here is where my enthusiasm gets a governor. Custom Builds are included on every paid plan, but the word "included" covers the packaging, not the compute. A deployment bills GPU time by the second while its workers are on, plus network storage for as long as the deployment exists. The published rates:

  • RTX PRO 6000 MIG, 48 GB: $1.75 per hour, at roughly half the speed of the full card
  • RTX PRO 6000, 96 GB: $3.49 per hour
  • H100 SXM, 80 GB: $6.23 per hour
  • H200 SXM, 141 GB: $7.71 per hour
  • B200, 180 GB: $11.23 per hour

Storage runs $0.20 per GB per month, so a 20 GB model library costs about $4 a month before a single frame renders. Keep-alive defaults to three minutes after the last job, and that window is billed at the GPU rate. Comfy is upfront about this. What it means in practice is less charming.

The workflows ComfyUI users actually build are the worst possible fit for per-second billing. They are long, they branch, they load four checkpoints nobody else uses, and they iterate for hours with a human staring at the canvas. On a 96 GB RTX PRO 6000 that is $3.49 an hour, or roughly $28 for an eight-hour session where a good fraction of the time is spent thinking, not sampling. On a B200 the same afternoon costs $90.

The trap in the keep-alive

Every cloud GPU product has this shape, and all of them rely on the same weakness: nobody deletes a deployment at 2 a.m. A build with a warm worker and a three-minute keep-alive, left up over a weekend, quietly bills all weekend. The company is not hiding it. But "billed by the second" and "billed for the hours you forgot" are two different sentences, and launch posts tend to say the first one out loud.

Why this still matters

None of that makes the launch a mistake. It makes it a trade. Local ComfyUI asks you for a GPU, a weekend of maintenance, and an acceptance that your pipeline dies the day a node author abandons a repo. Custom Builds asks you for money and hands back a numbered, reproducible environment that a new hire can open on day one. For a studio billing clients the second deal is obviously correct, and it is what every closed competitor has sold for two years.

The open ecosystem just proved it can match that deal without giving up the graph. That is the real headline, and it is a better one than any benchmark score. The claim that open tools are toys for hobbyists with big desktop cards is now dead, and it died on a pricing page rather than a demo reel.

What I want to see next is whether the build system stays honest over time. Versioned builds are a reproducibility win, right up until a platform update quietly changes a base image and your identical build renders differently. Whether Comfy publishes those changes and pins the runtime will decide if Custom Builds become infrastructure or another subscription that slowly drifts away from the thing you tested.

For now, my advice is unromantic. If you are a team of more than one person, and your workflows depend on node packs that stock Comfy Cloud does not carry, this is the first version of the cloud product worth paying for. If you are a solo artist with a working local rig, nothing here changes your life, and $4 a month for models already sitting on your drive is not a bargain. The graph was always the point. The bill is the part you now have to read.

Comments

No comments yet. Be the first to share your thoughts!