Official Comfy Cloud compute and growth docket#

Status: Decision log Scope: Official Comfy Cloud, distinct from independent ComfyUI hosting companies and the local open-source application.

2026-10-04 — Upstream distribution and managed execution#

Takeaway#

Comfy Cloud's clearest disclosed advantages are upstream ComfyUI distribution and eliminating local setup. Its serving GPU class is advertised, but the inspected pages do not disclose whether those GPUs are purchased, leased or rented through another operator.

Compute and procurement#

Published product configuration, read 2026-10-04: Cloud advertises Blackwell RTX PRO 6000 hardware with 96 GB VRAM, preloaded models, supported custom nodes and portable workflows. Users are billed for active generation rather than idle editor time. These describe the customer product, not the supplier contract or ownership of servers. Cloud product page.

The pricing page also lists a separate Comfy API deployment product with GPU-specific hourly-equivalent rates. That table must not be mistaken for the Cloud editor's credit consumption or the operator's GPU procurement cost. Closed-model Partner Nodes execute on partner infrastructure; their existence does not identify who supplies Cloud's open-model GPU fleet. Pricing and deployment products, MCP execution boundaries.

Acquisition and monetization#

Current published offer: five free GPU runs without a card lead into Standard, Creator and Pro subscriptions. Monthly list prices shown are $20, $35 and $100 respectively; yearly equivalents are discounted and are not monthly checkout prices. Included credits, top-ups and higher-tier capabilities monetize ongoing usage. This snapshot supersedes older trial-credit descriptions for comparison purposes, without rewriting those historical notes. Pricing.

Distribution mechanism: official Cloud offers a managed entry point for people already discovering ComfyUI models, nodes and workflows, reducing hardware/install friction while preserving a path to local use. This is an inference from the product's position, not a measured funnel. The product page documents the setup/convenience offer, not conversion rates. Cloud product.

Agent access: the official MCP connects either local ComfyUI or Cloud to agent clients. Cloud generation requires a subscription; local open-source generation uses the user's machine, while partner-model calls consume credits. This adds an access surface and a paid Cloud entry point, not evidence that all local users pay Comfy. Official MCP.

Reported scale and financing#

Company announcement, 2026-04-24: Comfy reported four million ComfyUI users and announced a $30 million financing, bringing cumulative funding to $47 million. It describes investment in Cloud infrastructure, local experience and ecosystem support. The users refer to the broader ComfyUI community—not Cloud subscribers or Cloud MAU—and financing is not revenue. No Cloud ARR or free-to-paid conversion was established by this source. Official funding announcement.

Implication and next evidence#

Inference: upstream distribution, compatibility and convenience may matter as much as low hardware cost. Owning our GPUs alone does not counter an official product's built-in audience. Our useful comparison is successful, reliable workflow execution and paid-user latency, not GPU name or the size of the open-source community.

Unresolved: ownership/lease split, named open-model infrastructure supplier, procurement terms, utilization, custom-node coverage for our users, Cloud-specific active customers/revenue and retention. The pricing page presents differing API concurrency descriptions across sections; do not publish one as a resolved limit without confirmation. A matched workflow test remains undone.

Evidence trail#

Sources were read on 2026-10-04 local time; prices are advertised snapshots, not a tested checkout. See the source manifest, earlier product profile and coverage index.

2026-10-04 — New deployment product, live region catalog and Runpod comparison#

Product boundary and growth update#

Official launch, 2026-09-30: Comfy API is a separate deployment product, not merely the curated Cloud editor exposed over HTTP. Builds package models, private custom nodes and dependencies; immutable releases serve through autoscaling endpoints. Paid-plan access does not make deployment usage part of the editor subscription allowance. Earlier restrictions on user-installed Cloud nodes must not be generalized to this new product. Launch announcement, Cloud API quickstart.

Developer acquisition, announced 2026-10-01: an October 5–19 app-building challenge offers credits to the first 100 entrants, cash/hardware prizes and distribution through an open-source feature-app catalog. Runpod and NVIDIA participate as sponsors/judges. This establishes a named ecosystem relationship with Runpod, not a compute-supply contract. The competition had not begun at retrieval; acquisition and conversion outcomes remain unknown. Official challenge announcement.

Live infrastructure evidence#

Observed with existing account access, 2026-10-03 16:48 UTC / October 4 local: the official deployment catalog returned a datacenter-level US-NE-1 option with RTX PRO 6000 Server Edition, 96 GB VRAM, $4.54 per worker-hour and low availability. Country/global groups also listed H100 SXM, H200 and B200. This was a catalog query, not proof of capacity successfully allocated or the placement of curated Cloud jobs. Endpoint: https://platformapi.comfy.org/deploy/v1/compute-catalog?levels=all; the route is supported by pinned official CLI source.

Primary-source corroboration: Runpod documents US-NE-1 with S3 endpoint https://s3api-us-ne-1.runpod.io/, and its status service records maintenance for that region. The founder supplied the storage-documentation lead. This exact region-code correspondence, the public partnership and the pricing pattern below make Runpod a strong supplier-integration hypothesis. They do not establish physical server ownership, exact building/city, contractual terms, exclusivity or Cloud-editor placement. Runpod storage regions, Runpod region maintenance.

Advertised retail pricing, not supplier costs#

Full 96 GB GPUs only; exclude 24/48 GB MIG slices, RTX 6000 Ada and RTX A6000. All rates below are USD/hour equivalents, retrieved October 4 local. No reservation, negotiated rate or regional checkout was tested.

Sources: Runpod pricing, page dated 2026-09-27, Comfy API pricing; Comfy's live catalog corroborated its rates. Formula: (Comfy / Runpod − 1) × 100. Every listed Comfy serverless-equivalent rate is Runpod's displayed rate × 1.30 rounded to cents. That consistency supports the integration hypothesis, not a measured margin.

Founder comparison preference: Pod rental is the main economic benchmark for an always-warm, customizable ComfyUI runtime. On compute rates alone, an always-on $2.09/hour Pod crosses over when Comfy's billed worker time reaches 2.09 / 4.54 = 46.0% of the same period. This is a hypothetical retail-cost comparison, not actual job utilization, matched performance or proof Comfy pays the public Pod price. Setup, maintenance, availability and service differences remain unpriced.

Billing boundaries: Comfy charges warm minimum workers throughout their lifetime; flex billing includes startup/loading, execution and currently 30 seconds of idle time. A 50 GB ephemeral container disk is included; staged model storage is separately billed and persists until the last deployment using it in the region is deleted. Runpod also bills initialization and idle time, with a documented default 5-second idle timeout and separate storage. Thus equal generation duration does not imply equal billed duration. Comfy deployment guide, Runpod billing.

Callback experiment and limits#

Observed: Cloud returned 3,768 node definitions, but the inspected image-loader interfaces exposed filenames rather than an obvious public-URL input. Its system_stats response contained an empty devices list and zero RAM values; it is not worker GPU evidence.

A temporary public endpoint served only a synthetic 2-by-2 PNG. Our control request succeeded. JSON URL import to POST /api/assets returned HTTP 400 and directed callers to /api/assets/download; one corrected request there returned HTTP 400 because external-source metadata could not be retrieved. No Comfy callback arrived by the 17:00 UTC cutoff. The control caller's address is our network, not Comfy's worker. No asset ID, job or deployment resulted from those requests; restrictions were not bypassed.

An asset-service callback would identify that service's egress, not necessarily a GPU worker. A private custom-node worker test was subsequently approved with zero minimum workers, at most one worker, no models, a 10-minute test deadline and a $1 target budget. Its execution/result is pending in this snapshot. Neither a catalog region nor a callback NAT address establishes hardware title.

Evidence trail and remaining questions#

The continuation manifest records public captures and authenticated observation checksums. Raw account-specific catalogs, headers, credentials and identifiers are not committed or uploaded. Privacy text names generic cloud-hosting categories; the trust-center download was only an HTML shell, so no readable supplier list was established.

Still unresolved: a worker-level callback and provider corroboration; exact physical placement; purchased versus leased GPUs; supplier rates/commitments; Cloud-specific growth, retention and revenue; matched workflow latency and reliability. The evidence narrows the supplier question without converting retail-price arithmetic into an ownership or profit claim.

2026-10-04 — Approved private probe stopped at the account-credit gate#

Observed outcome: the supplied key could read Builder base-image/target catalogs and resolve a model-free workflow. One private node ZIP (1,676 bytes) was uploaded through the official blob API. Build creation then returned HTTP 402, PAYMENT_REQUIRED, reason: no_credit_balance. The response says an active subscription is required and legacy-plan subscribers also need a positive balance. No Build ID, release, deployment or GPU job was created. The key being active did not establish billable deployment access.

The synthetic custom node was limited to reporting CUDA device name/VRAM, two explicitly named non-secret provider hints, and one GET to our fixed public callback. It never ran remotely. Its source and package remain temporary local artifacts; no credentials were included. The blob upload destination was Google Cloud Storage, evidence of artifact storage only—not the GPU provider.

Cleanup: both accountless tunnels and callback listeners were stopped. A deletion request addressed only our newly created blob; the server returned HTTP 404. The inspected official CLI provides blob listing, but no blob-delete operation. Therefore removal of the uploaded 1.7 KB blob is unconfirmed, not successful cleanup; retention/garbage collection must not be assumed. No existing customer resource was modified. No deployment was created to trigger GPU-worker/model-staging billing.

Next gate: the user may supply a usable existing credit balance or waive the worker test; we will not purchase credits or upgrade a plan. Region/catalog, partnership and retail-rate evidence remain valid; no worker egress or physical location was measured. Other competitor research can proceed independently. The continuation manifest records the failed Build request and cleanup limit.

2026-10-04 — Correction: credits exist; execution is restricted by Free tier#

The earlier next-gate wording was too broad. Authenticated billing diagnostics confirmed positive prepaid/effective funds and has_funds: true, while the same key reported subscription_tier: FREE. An active subscription-status row is not proof of a paid plan. Personal balance amounts and account identifiers are omitted from this docket. A repeat of the private Build request returned the same HTTP 402 and no_credit_balance reason, despite those positive funds; that error must not be presented as evidence of an empty account.

Independent execution check: a valid, model-free EmptyImage → SaveImage workflow, producing only a synthetic 2-by-2 image if accepted, was submitted once through POST https://cloud.comfy.org/api/prompt. It returned HTTP 403, FREE_TIER_NOT_ALLOWED, with the message that API-key authentication is unavailable for free-tier accounts. No job ID or output was created and no worker ran. This confirms that the current key cannot execute even the regular Cloud workflow route; it is not merely a private-deployment restriction. The official Cloud API overview and Cloud quickstart explicitly require a paid subscription. Partner Node prepaid credit access is a separate execution boundary and does not identify Comfy's GPU fleet.

Revised conclusion: existing credits are sufficient in amount for a tiny probe, but do not establish entitlement to these execution products. An existing paid-plan account would be needed to resume; no purchase, upgrade or restriction bypass is authorized. Read-only region/pricing evidence remains usable, and the supplier hypothesis remains Runpod with worker placement unmeasured. This is a documented verification limit, not a finding that Comfy has no funds or no APIs. The updated manifest records the diagnostic captures.

2026-10-04 — Current investigation handoff#

The public-source, authenticated read-only and bounded execution-diagnostic pass is finished. Earlier “pending” probe/credit-gate entries are historical observations superseded by the Free-tier check above. No successful GPU execution, worker callback or physical-server location was obtained. Current credits do not remove the execution entitlement gate. Resuming that specific experiment needs existing paid-plan access; it is not a reason to buy additional credits or silently upgrade. Temporary callback services are stopped; deletion of the uploaded probe blob remains unconfirmed. The cross-competitor synthesis retains the Pod-rental comparison and separates evidence from supplier inference.