The problem this actually solves
The graph is the easy part of a ComfyUI setup to share. What does not travel is everything underneath it: the custom node packs, the exact model files, the LoRAs, the dependency versions that happen to work together. A new artist joining a project has traditionally had to reproduce that environment locally before they can open the workflow the rest of the team is using.
Custom Builds attack that directly. One person packages the setup once, starting from a workflow or a Comfy Desktop snapshot, and the builder collects the nodes and models it needs and helps resolve dependency conflicts. Anyone in the workspace can then pick that Build in the Comfy Cloud editor and start creating, in the same canvas, with the same tools. When the team adds a LoRA, it can update the Build and roll the change out. Developers can run the same setup through the API.
That is a real change in what Comfy Cloud offers: teams can package the supported components of their own setup and use the resulting Build in the cloud editor and through the API, instead of rebuilding the environment artist by artist.
Where the meter runs
The pricing structure deserves a slower read than the launch framing gives it.
GPU time is billed by the second, which sounds precise and fair, and the rates are stated plainly: $1.75 an hour at the bottom of the range, $11.23 an hour at the top, depending on the GPU. Comfy is also upfront that merely opening or editing the canvas does not start a GPU.
The sentence to hold onto is the other one. Charges apply "while workers are running, including during a keep-alive window or when workers are kept warm." In plain terms, the meter does not stop the moment your render finishes. A worker kept warm for the next job, which is exactly what you want during an active production day, is billable time. How long the keep-alive window runs, and how aggressively workers stay warm under load, will matter to real-world cost as much as the headline hourly rate. Comfy's post does not put a number on the window.
There is a second, quieter line item: your models live on network storage, billed monthly. For a team whose Build carries a shelf of large video models and LoRAs, that is a standing charge that exists whether or not anyone renders anything that month.
None of this makes the offer bad. Per-second billing is more honest than rounding to the hour. It does mean the comparison to weigh is not "cloud rate versus GPU price." It is the paid plan, plus worker runtime including warm time, plus monthly storage, set against your own local costs.
The limits in the announcement itself
Two qualifications are in Comfy's own text, and both matter for the teams most likely to want this.
First, a Build packages your team's supported nodes. Comfy links to a supported custom nodes list and concedes that complex packs can take some troubleshooting. The setups with the most exotic node stacks, which are often the ones teams most want to standardize, are the ones most likely to hit that boundary. Check the list before assuming your pipeline qualifies.
Second, collaboration is tiered. Custom Builds are on every paid plan, but inviting teammates into the workspace is a Team plan feature, and concurrency starts at two simultaneous jobs per Build on Standard and Creator. A small studio can package its setup on a lower tier; running a whole team on it at once is what the upper tier is for.
Sources
- [1] Comfy (Eric Sun), "Your custom ComfyUI is now in Comfy Cloud," blog.comfy.org, November 18, 2026Read source
- [2] Comfy, Custom Builds documentation (referenced from the announcement)Read source