milvey

Rent a GPU pod with ComfyUI already running.

6 cards from 24 to 141 GB of VRAM. Metered per second, shown in credits per minute. Stop the pod and GPU billing stops.

See GPU prices Join the list
Crowned hedgehog on an ornate throne (description; prompt not published)

6 GPU types, from 4 to 21 cr per minute.

GPUVRAMGood forRegionsAvailabilityPrice
RTX 409024 GBFast drafts, image models, small videoSG · TYOIn stock4 cr/min
RTX 509032 GBBigger batches, quicker draftsSGLimited5 cr/min
L40S48 GB14B video at 480p, steady throughputSG · FRAIn stock5 cr/min
RTX PRO 600096 GB14B video at 720p, LoRA trainingSG · IADLimited8 cr/min
H10080 GBHeavy pipelines, fast long rendersIADLow stock17 cr/min
H200141 GBHeaviest pipelines and long rendersIADLow stock21 cr/min

Sample rates and availability. Metered per second, shown in credits per minute.

150 free credits run an RTX 4090 for 37 min.

on an RTX 4090 at 4 cr/min
37 min
on an RTX 4090 at 4 cr/minsample
on an L40S at 5 cr/min
30 min
on an L40S at 5 cr/minsample
on an RTX PRO 6000 at 8 cr/min
18 min
on an RTX PRO 6000 at 8 cr/minsample
on an H200 at 21 cr/min
7 min
on an H200 at 21 cr/minsample

Wan 2.2 14B · 720p needs about 40 GB, so it fits on the L40S and larger cards.

The matrix compares VRAM with what common workloads need. A benchmark of your own workflow gives the real figure.

GPUFLUX · 1MP
needs 16 GB		Wan 2.2 14B · 480p
needs 24 GB		Wan 2.2 14B · 720p
needs 40 GB		LoRA training
needs 80 GB
RTX 4090 24 GBfitsfitsout of memoryout of memory
RTX 5090 32 GBfitsfitsout of memoryout of memory
L40S 48 GBfitsfitsfitsout of memory
RTX PRO 6000 96 GBfitsfitsfitsfits
H100 80 GBfitsfitsfitsfits
H200 141 GBfitsfitsfitsfits

Sample VRAM needs.

A new pod is ready to open in the browser in about 11 seconds.

connect · pod_7a2
# ComfyUI in the browser (port 8188)
https://pod-7a2.milvey.example

# Shell (port 22), key-based
ssh root@pod-7a2.milvey.example

# Jupyter (port 8888)
https://pod-7a2.milvey.example:8888
pod log · first 11 seconds
[00:00:01] pulling image milvey/comfyui:cu128 (cached)
[00:00:03] mounting volume vol_brand (120 GB)
[00:00:05] starting ComfyUI on :8188
[00:00:09] loaded 4 custom node packs
[00:00:11] ready · https://pod-7a2.milvey.example

Example pod. Your pod gets its own private URL.

A stopped pod keeps its disk, and the disk is still billed.

Stop lora-train-brand?

  • GPU billing stops now.
  • The 120 GB disk is kept and still billed while stopped.
  • Start it again later with the same files.
CancelStop pod

Terminate lora-train-brand?

  • All billing stops now.
  • The 120 GB disk is deleted and can’t be recovered.
  • Download brand_lora_v4_step1800.safetensors first.
CancelTerminate and delete disk

Example dialogs. Idle pods can stop on their own after a timeout you set per pod.

Tell me which GPU you need and for how long.

Join the list and reply with the card and the workload. I’ll tell you whether it fits and roughly what it costs in credits.

Email when signups open · Unsubscribe any time