The DGX Spark is a seriously impressive box. It has 128GB of unified CPU and GPU memory in one package, and a GB10 Grace Blackwell chip that handles models which would need several server-class GPUs on a traditional setup. And it sits on a desk.
In a lot of the setups we have seen, it sits on one person’s desk and does one person’s work.
The hardware can serve a whole team at once, so the hardware is not what holds it back. Linux is. Getting a Spark set up properly for a real AI workload takes command-line skills that most researchers, developers, and creative people do not have and should not need. So the Spark turns into a personal workstation by default, because sharing it takes too much setup.
What the Linux barrier looks like
When someone new unboxes a DGX Spark, here is the list in front of them. They configure network settings from the command line, install the CUDA toolkit and the NVIDIA container runtime, pull and configure container images for their workload, set up storage paths and mount points, manage environment variables for each application, and sort out conflicts when two workloads want different library versions.
Every one of those steps has documentation. The documentation assumes you already know Linux. For example, a researcher who wants to run AlphaFold on a Spark to look at protein structures should not have to learn the difference between a volume mount and a bind mount first.
So the Spark gets used by the one person on the team who is comfortable in a terminal, usually at a fraction of what it can do, and everyone else works around the hardware instead of on it.
What changes with a browser
Orion runs on the Spark and shows a catalog of workloads in a browser. You open the URL, log in, and see cards for the applications you have access to: Jupyter notebook, VS Code IDE, Open WebUI and Ollama for local models, and any custom research environments your IT admin has set up.
You click a card. The workload launches in about sixty seconds and you get a full-screen browser interface to the application. When you are done, you close it and the resources go back to the pool.
You never touch a command line or configure an environment, and workloads do not fight each other, because each one runs in its own isolated container with its own dependencies. It feels closer to opening a web app than to running a Linux server.
The Spark’s 128GB of unified memory becomes a shared pool instead of a fixed allocation for one workload. Several people can have workloads running at the same time. The scheduler places each workload on available resources and enforces the limits set in its template. No single workload can take the whole device unless an administrator sets it up that way.
When one Spark is not enough
When a team outgrows a single Spark, Orion clusters several of them into one unified compute plane. A second Spark joins the cluster in minutes. The catalog stays the same and workloads schedule across both devices. Users never see the cluster layout. They just see their workloads running.
The same goes for mixed hardware. A Spark next to older workstations, Intel NUC nodes, or cloud instances all shows up as one pool the scheduler can use. A workload that needs the Spark’s Blackwell GPU lands there. A workload that only needs CPU and memory can land on any node with room. The scheduler handles placement and the user clicks a card.
Running the model locally
The Spark’s unified memory also makes it practical to run a large language model locally, so the whole team gets an AI assistant with no per-request bill and no data leaving the infrastructure. We wrote up how we run a 35-billion parameter model on one Spark separately.
What this looks like for different teams
For a research team, the Spark becomes a shared compute resource where anyone can launch their environment without going through IT. Shared storage with the right file permissions means everyone works on the same data. The local model helps with analysis and writing without the data leaving the institution.
For a development team, the Spark hosts shared dev environments, CI agents, code review automation, and the shared inference endpoint, all on one device. That is a full AI-assisted development setup on hardware that fits under a desk.
For a creative team, image generation and GPU-accelerated creative tools run as shared workloads. Nobody waits for one person to finish with the hardware before the next person starts. The Spark serves everyone at the same time.
What utilization looks like in practice
A DGX Spark running as a single-user Linux workstation is busy while its owner is working and idle the rest of the time. That means nights, weekends, meetings, and any work that does not need the GPU.
Run the same Spark with Orion, several users, and some automated agent workloads, and that idle time fills up. CI agents run after hours. Automated analysis jobs run overnight. Hardware that was covering one person’s working day now covers the team around the clock.
The cost math is simple, up to a point. A capable API-based AI service charges per request, every month, and that adds up fast when a team uses it all day. Because the bill grows with every request, teams get hesitant about using AI as freely as they should. Once you own a Spark, local inference costs you power plus your Orion subscription, and neither goes up when someone sends another request. For teams that lean on AI all day, that tips toward local compute faster than most people expect.
The API side does win in one case. A Spark is still one box with 128GB of memory. If your team needs the very largest models, or a lot of them at once, you will outgrow one device. That is when clustering several Sparks, as described above, starts to matter.
See how Orion turns a DGX Spark into a team platform. Get 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.
