Plugging an AI Supercomputer into IntelliJ: Our DGX Spark Story
A small, heavy box arrived at the desk.
Inside: an NVIDIA DGX Spark — a desktop AI box that looks like a PC, and behaves like a tiny datacenter. The obvious question for us was:
How do we make this thing feel invisible from inside IntelliJ?
This post is about the path we took to connect DGX Spark and IntelliJ in a way that:
- keeps everything local and offline-friendly,
- keeps security teams relaxed,
- and keeps developers in flow, not in network manuals.
There is also a tiny hardware trick involved! And the shiny user experience we are fighting for!
I am going to demo the device, the magic cable (JetCable) at AWS re:Invent, come visit JetBrains booth for the real demo!
Why Local AI Matters for IDE Workflows
Cloud AI is amazing — until the first security review. And of course, there are many of us who are happy to use the frontier models (e.g. Anthropic, OpenAI, Google, xAI, …). There is a way to many regulated businesses, where it’s not yet possible or profitable.
Many teams live under constraints like:
- strict compliance rules
- code and data that must never leave the corporate network
- air-gapped or tightly segmented infrastructure
- flat usage plans (you own the hardware, you use it 100% of the time)
For those teams, “send code to an external model endpoint” is not an option.
NVIDIA DGX Spark gives a different model:
- AI hardware is on the desk or on-prem
- the box can be configured with no direct internet access
- all prompts, code, embeddings, and models stay inside the company perimeter
From the local development perspective, we want:
- AI assistance and heavy compute for code
- minimal configuration
- and strong guarantees that the bits never leave the building
The integration story started from a simple goal:
Treat DGX Spark as a local development tool, not “another cloud”.
Step Zero: “Just Put It on the Network”
The first obvious attempt followed the official playbook:
- Plug DGX Spark into the corporate network (Wi-Fi or Ethernet).
- Finish OS setup, accounts, security.
- Connect from dev machines over the LAN.
- Install agents, tools, and talk to the device via IP.
On paper, this is fine.
In real corporate networks, we hit the usual obstacles:
- VLANs and firewalls.
- Ports that work in one environment and disappear in another.
- VPNs that break discovery.
- “Networking changes” with lead times measured in sprints.
- Broken flow (instead of trying, you’re configuring and rebuilding for hours)
We want a different experience:
- We unbox NVIDIA DGX Spark.
- We connect one cable.
- IntelliJ quietly detects a device and offers to use it for AI and heavy tasks.
No manuals, no “Which IP did DHCP assign this time?”, no ticket to “open port 12345”. So we decided to remove the corporate network from the diagram.
Attempt #1: USB Gadget Magic
The natural idea for a direct link is USB-C. Modern devices can expose, over USB:
- a network interface
- a serial/COM port
- or a custom vendor protocol
The experiment:
- make DGX Spark pretend its USB-C is a network adapter;
- let it run a small DHCP server, assign an IP to the dev machine;
- expose a small API / SSH / custom protocol over this private point-to-point link.
On Linux, gadget drivers make this possible. Early tests were promising. Then we remembered a detail: this needs to work on Windows, macOS, and Linux developer machines.
That implies:
- custom drivers
- code signing and OS-specific packaging
- keeping up with random OS updates
- debugging strange USB behaviour on a random laptop the night before a release
Technically, possible and feasible. As a product, not something we want to base everything on. We stepped back and asked a simpler question:
What is the most boring piece of hardware that all OSes already understand?
Answer: a regular network adapter.
The Final Trick: Two NICs and a Tiny Cable
The design we ended up with is almost disappointingly simple:
Use two standard Ethernet adapters connected by a short cable, one on the NVIDIA DGX Spark side, one on the developer machine. Both go into USB-C.
Diagram version:
[IntelliJ laptop]
|
| USB-C
|
v
[USB-Ethernet #1] == short cable == [USB-Ethernet #2]
^
|
USB-C
|
v
[NVIDIA DGX Spark]
And this setup is working greatly! We make the NVIDIA DGX Spark run as DHCP and DNS for the network interface, so it means it will assign the developer machine the specific IP address. The software (say an IntelliJ plugin) will know about the port, or it will even look up it via DNS or either way.
Current State
We are running vLLM on the box, JetBrains AI Assistant can connect and use it, we’re experimenting to let JetBrains Junie use the box as the primary model too. So far it’s the demo phase to show what we can do and what’s coming next!
From the product standpoint, we take an assumption that there will be more powerful LLM models with open weights and the hardware to run the inference as fast as possible. We are experimenting with the “tool” product to make software development much more pleasurable and still more productive.
Stay tuned!
Related Work
There are many areas were we definitely need to look in at, for example
- NVIDIA Sync tool
- SSH tunneling to connect to the NVIDIA Spark
- Using JetBrains cloud service to activate the device and let it with offline afterward
- Automatic configuration and pre-population of the software and AI model images
- Running AI Assistant, Junie, and JetBrains Mellum on the solution
- Making the solution distribute licenses
- Running local RAG’s in the device or near-by
- Making the cable smarter to support us with all of that
The magic cable itself appears to have much of the potential, I’m looking forward to sharing more with you. The cable sources are available on GitHub. We are going to post more, stay tuned, or reach out in DMs for more details.
New World & AWS re:Invent 2025
There are so many ideas that we have and found working on this project, stay tuned, and let me know (via LinkedIn) what you think or would expect from us. I will update.
I’ll be deming the solution, the JetCable at AWS re:Invent this year, stop by at JetBrains booth, and let’s talk about LocalAI. See ya! Drop me a line via LinkedIn or Twitter to stay in touch.
Here is the demo from AWS re:Invent 2025: