The first version of Juno ran on four Raspberry Pis and my wife’s old gaming PC. It ran game servers before it ran anything else.
Before that I spent about sixteen years in visual effects. I started as a compositor at Digital Domain, taught myself Python to build tools for the artists around me, and got pulled into pipeline work and then infrastructure at Cinesite and other studios. When I pitched the idea for Juno at a studio, the answer was that it was ten years out. So I built it at home.
Building it on hardware that small was a bit ridiculous. It was also the constraint that forced us to build something worth building. When you have almost no hardware, there is nothing spare to give. Every CPU cycle has to count, and a GPU has to go to whichever workload needs it right then.
That is where Orion came from, trying to make something work on hardware that had no business running it.
What VFX teaches you
The VFX industry has been solving compute problems that the rest of the world is only now running into.
Render workloads are bursty by nature. A production has nothing to render for months, then needs to render everything at once, then has nothing again. Static infrastructure for that pattern is either massively over-provisioned in the quiet months or nowhere near enough during crunch. Neither works when the delivery date is fixed.
Artists need very different amounts of compute depending on what they are doing. A layout artist needs a responsive workstation but not much GPU. A particle simulation artist needs GPU capacity that would embarrass a small data center, for exactly as long as the sim takes to bake. The infrastructure has to serve both without parking expensive hardware on one of them forever.
And the people using it are artists, not infrastructure engineers. They should not have to think about where their compute comes from any more than they think about how electricity gets to the wall. The system has to be good enough to disappear.
I learned all of this inside studios before Juno was a company, which turned out to be a useful way to start.
The scheduling problem
The first thing we figured out is that scheduling is where most of the waste hides.
Most infrastructure treats scheduling as solved. You have resources and jobs, and you match them. Kubernetes does that out of the box, and so does any cluster manager. The trouble is that simple matching leaves a huge amount of capacity on the table.
When you give one GPU to one artist and another GPU to another artist, both GPUs are idle most of the time. Interactive work follows people. The artist is on a call, reading notes, or in a review, and the GPU waits. Meanwhile another artist is trying to bake a simulation and cannot get a GPU because every one of them is “allocated.”
So we built scheduling that treats the GPUs as a shared pool. Essentially, the artist gets their GPU when they need it, and when they do not, that capacity goes to the job that does. We do not slice the GPUs ourselves, NVIDIA does. We decide what runs where. The artists do not notice, and the utilization numbers change a lot.
The portability problem
The second thing VFX taught us is that production environments move around all the time.
A production might start on-prem because the studio has local hardware. It might burst to cloud during crunch when local capacity runs out. It might run in a remote location because a key supervisor is there. The environment has to follow the work.
That is where container-native workstations came from. If the workstation is a container instead of a physical machine, it runs wherever there is capacity. The artist connects to the same environment in the studio, at home, or on location. The project state, the application stack, and the storage mounts all travel with the workstation.
We built it because VFX productions needed it. Later we found that researchers running computational experiments need the same thing. They want to pick up exactly where they left off no matter which cluster runs their job. Teaching labs want workstations that behave the same for students on campus or off.
What “works on anything” means in practice
Starting on a gaming PC and a handful of Pis gave us software that runs on whatever hardware is around, which is not what we expected.
Orion runs on NVIDIA, AMD and Intel GPUs and on CPU-only nodes. It runs on bare metal, VMs, and cloud instances, needs no outside connection to work, and scales from a single node up to whatever size the customer needs. When we moved it off the Pis and onto AWS, it took about two hours, because it was just Kubernetes. Funny enough, some of those Pis still run our test jobs.
That range comes straight from building something that had to run on four Raspberry Pis and an old gaming PC. When you cannot assume hardware, you build for any hardware. That turns out to be exactly what enterprise customers need too, especially the ones with a mix of on-prem hardware, cloud instances, and edge nodes that are all a little different.
Why this matters if you are not in VFX
The pressures VFX production has lived with for decades are showing up in other industries now, pushed by AI workloads and the broader move to GPU-heavy compute.
Research organizations deal with bursty demand from experiments that need GPU capacity for days and then nothing for weeks. Life sciences teams have the same utilization problem render farms always had. A team running a disconnected site needs portable, isolated compute that runs on whatever hardware is in the room.
The patterns match even when the specifics differ, and the software that handles them is the same software that started on a gaming PC in my house.
Working outside VFX was never the plan. We were trying to get artists out of the queue, mostly because we thought it was a cool problem to solve, and the rest of the world turned out to have the same problem.
If you want to see what that looks like for your environment, we are happy to show you. Book a demo
WRITTEN BY
Alex Hatfield
Alex co-founded Juno to fix how enterprise compute gets done. He leads product vision and customer strategy, and spends most of his time working directly with infrastructure and research teams pushing the limits of what their hardware can do.
