For years, we kept saying we needed something cheaper than a workstation and faster to get into than a render farm.
A way to throw a heavy GPU job at the cloud with a single API call, have it run in seconds, and come back with the output. No render queue. No artist sitting around watching a workstation think.
Just submit, run, done.
So we built it.
Meet Spark Fuse.
Everything that doesn’t fit a render job
Fuse has all the goodness of our infinitely scalable SmartCompute render farms, but for everything that does not fit neatly into a Maya, Houdini, Karma, Arnold, Redshift, C4D, After Effects, or Nuke render.
Simulations, cache-baking, neural texture work, ML inference and training, asset processing, LoRA training, heavy ComfyUI workflows. Anything that wants serious GPU horsepower for ten minutes or ten hours and needs it within a couple of seconds.
If you have used RunPod or Modal, it is a bit like those, but cheaper, instant availablility, with more GPU options, aligned to TPN security requirements, and built directly into Spark Cloud Studio.
That last part matters. Fuse speaks the same language as the rest of Spark. Same login. Same billing. One stack. Outputs stream straight to your ShareSync storage, so the job does not feel like it has disappeared into another system somewhere
Two flavours
-
InstantCompute from $0.49/hr for an NVIDIA T4 GPU up to $3.15/hr for NVIDIA RTX PRO 6000 Blackwell GPUs. Sub-2-second cold start from a warm pool. Feels local.
-
SmartCompute from $0.29/hr for NVIDIA T4 GPU up to $1.19/hr for NVIDIA RTX PRO 6000 Blackwell GPUs. 60 to 180 second Spot launch, up to 70% cheaper, for jobs where you can tolerate the slower start.
Why we’re excited
- Outputs land in ShareSync as the job runs. No copying files around, no babysitting an S3 bucket or a URL. Your /Compute Jobs/ folder fills up while you are working on something else.
- Live log streaming. Submit a job and watch your stdout and stderr roll into your terminal in real time. Training loss, frame counts, whatever you are writing out, it is right there.
- Any GPU you actually want. NVIDIA T4 GPU for quick inference, NVIDIA RTX PRO 6000 Blackwell GPU with 96GB of VRAM for the heavy lifting, NVIDIA A100, and NVIDIA H100 for the truly monstrous workloads.
- It is quick. We tested it on a stack of Houdini sims, including a super heavy FLIP setup and a 512³ smoke sim with full raytraced volumetric shadows, emission, and self-shadowing. A 180-frame demo loop, sim included, in about 20 seconds. Faster than most artists’ local cards can manage, with the EXRs available in Nuke on a Spark workstation a few seconds later.
Spark Fuse is in private early access right now.
We are starting with the kinds of jobs that usually fall awkwardly between a workstation and a render farm: LoRA training, neural rendering, volumetric sims, ML inference at scale, heavy procedural work, and anything else where iteration speed matters more than watching another queue crawl along.
That is the space Spark Fuse is built for.
Not another machine to sit inside.
Another way to get the heavy work moving.
