Before deploying a team AI workspace, confirm the team workflow, fresh DGX Spark environment, Ubuntu 24.04 ARM64, Docker and KVM, power and unified-memory behavior, network and offline preparation, data boundaries, individual accounts, performance, and licenses. Team AI Box provides browser access to local AI, team files, messaging, wireless casting, and virtual devices.
1. Confirm the Team’s Actual Use Cases
A shared AI workspace is most useful when employees need to process files, transcribe meetings, translate documents, review AI results, or use virtual devices. Team AI Box supports local AI conversations, image generation, OCR and document processing, speech transcription, translation, team file management, direct and group messaging, wireless casting, virtual Android phones, and virtual Windows and Linux computers.
Start with one workflow rather than enabling every feature at once:
Meeting recording
↓
Speech transcription
↓
Human review
↓
AI-assisted meeting notes
↓
Approved file in the team drive
Other suitable starting points include document processing, translation review, customer-support material, technical demonstrations, and virtual device testing. Review the Team AI Box capabilities before deployment.
OpenShell Serves a Different Layer
Teams may also encounter NVIDIA OpenShell when evaluating DGX Spark. OpenShell is designed for isolated, per-developer GPU sandboxes and agent-development workflows. Team AI Box addresses a different layer: a browser workspace for human collaboration around documents, messaging, transcription, translation, and virtual devices. The two approaches can coexist, but this checklist focuses on the human team workspace.
2. Confirm the DGX Spark and System Environment
Team AI Box is intended to run on a supported NVIDIA DGX Spark, not on an ordinary office laptop. The official Team AI Box deployment guide lists the following requirements:
- A fresh supported NVIDIA DGX Spark
- Ubuntu 24.04 ARM64
- Docker Engine and Docker Compose v2
- NVIDIA drivers and container support
- KVM
- Git
- Python 3.10 or later
- Several hundred gigabytes of temporary free space
Run the installation directly on the DGX Spark:
git clone https://github.com/weerocksllc/teamaibox.git
cd teamaibox
sudo bash install/from-source.sh
Cloning the repository alone does not complete the installation. The installer also downloads dependencies and models, builds runtime images, creates the installation package, and installs the workspace. A fresh installation is recommended; do not run it over unverified business data or an existing production environment.
3. Check DGX Spark-Specific Platform Behavior
The generic requirements are necessary but not sufficient. Confirm these DGX Spark behaviors before opening the workspace to a team.
Use the Supplied Power Adapter
Use the supplied NVIDIA power adapter. A different adapter can reduce performance, prevent boot, or cause unexpected shutdowns. If performance is unexpectedly low after setup, check the power connection and perform a cold power cycle according to NVIDIA’s hardware guidance before troubleshooting the application stack.
Do Not Treat nvidia-smi Memory Output as Dedicated VRAM
DGX Spark uses unified memory. On this platform, nvidia-smi may show “Memory-Usage: Not Supported” even though process information is available. This does not by itself indicate a driver failure. Use system memory tools such as free -h and the DGX Dashboard when planning capacity.
Leave Memory Headroom
Because CPU and GPU workloads share the same physical memory pool, an oversized model or concurrent workload can affect the whole appliance. Leave headroom for the operating system, Team AI Box services, virtual devices, and other users. Test memory-heavy workflows before promising a particular level of concurrency, and keep a recovery plan that does not depend only on SSH.
Refresh Docker Group Membership
If installation or preflight checks use Docker as a non-root user, confirm that the deploying account has Docker access. After adding the account to the docker group, start a new login session or run newgrp docker so the current shell receives the updated group membership.
Allow Time for First Boot and Updates
The first setup may include system and firmware updates. Use a reliable wired connection when possible, and do not interrupt a process that is downloading or rebooting simply because it takes longer than an ordinary application installation.
Check ARM64 Compatibility
DGX Spark is an ARM64 platform. Team AI Box manages its own installation path, but any additional packages, containers, drivers, or developer tools should be checked for ARM64 support before deployment.
4. Confirm Internet Access, Caching, and Offline Plans
Team AI Box is local-first, but initial source preparation requires internet access to download dependencies, models, and build inputs. Before deployment, confirm that the Spark can reach the required sources and has enough bandwidth and temporary storage.
You can run the preflight check with:
sudo bash install/from-source.sh --check
If the process is interrupted, preserve the build workspace and rerun the same command so completed stages and verified downloads can be reused. A cloned repository is not a complete offline installation package. If the target must remain disconnected, prepare the complete package on a compatible connected machine first.
5. Define Data and External Model-Service Boundaries
Local AI tasks can run on the DGX Spark while external model services remain optional. Before deployment, decide:
- Which files must remain on the appliance
- Which tasks may use external model services
- Whether customer or confidential files may be sent externally
- Which users may configure external connections
- Where backups should be stored
- How logs and task results should be retained
Local-first does not automatically mean permanently offline. Initial installation, model downloads, optional integrations, and external model services can affect the data boundary.
Use a simple classification scheme: public data for testing, internal data for the team workspace, customer data requiring approval, and sensitive or regulated data requiring additional review. These are team governance decisions; they should not be interpreted as a claim that Team AI Box automatically provides a complete approval, audit, or backup system.
6. Create Accounts and Test Access
Do not use one administrator login for every team member. After installation:
- Sign in with the initial administrator account.
- Change the initial password immediately.
- Create ordinary user accounts.
- Decide which people should use each application and file area.
- Test both allowed and restricted access.
- Verify the appliance is not exposed to the public internet with default credentials.
Team members should connect to the installed appliance through its current LAN HTTPS address:
https://<BOX_LAN_IP>/
Verify that the locally generated certificate belongs to the intended device. Separate administrators from ordinary members, and assign reviewers or owners for sensitive files and external-service decisions. The exact access boundaries should be confirmed against the controls available in the installed workspace.
7. Test Performance, Concurrency, and Licenses
Hardware specifications alone cannot guarantee a fixed number of simultaneous users. Performance depends on the selected model, model size, context length, file size, task type, number of concurrent users, and whether virtual devices run alongside AI tasks.
Before inviting the full team, test:
- Two or more users signing in at the same time
- Text and file-processing tasks running concurrently
- OCR, transcription, or translation
- Saving and reusing files through the team drive
- First model startup and later request times
- Failure, retry, and recovery behavior
- Reachability after a memory-heavy task
Virtual Windows and Linux computers require an authorized compatible ARM64 ISO and any required operating-system license. Virtual Android applications and Google services have their own compatibility and licensing conditions. Third-party code, model weights, NVIDIA components, firmware, Windows, Android applications, and user-provided installation media may retain their own terms. Team AI Box is an independent project by Weerocks LLC and is not an NVIDIA product, certification, or endorsement.
Deployment Checklist
Before inviting the team to use the workspace, confirm that:
- [ ] The first workflow and users are defined.
- [ ] Ubuntu 24.04 ARM64 and the documented DGX Spark requirements are confirmed.
- [ ] Docker, KVM, Git, Python, and NVIDIA support are ready.
- [ ] The supplied power adapter is in use.
- [ ] The team understands unified-memory behavior and leaves headroom.
- [ ] Docker group membership has been refreshed in a new session.
- [ ] First-boot and firmware updates have completed.
- [ ] Several hundred gigabytes of temporary storage are available.
- [ ] Internet access or a complete offline package is available.
- [ ] External model-service rules are documented.
- [ ] Individual accounts and the initial password change are complete.
- [ ] The LAN HTTPS address and certificate have been verified.
- [ ] Realistic concurrency and recovery tests have passed.
- [ ] ARM64 installation media and operating-system licenses are available where required.
- [ ] Users understand that AI-generated results require human review.
Frequently Asked Questions
Does a team AI workspace require cloud services?
No. Team AI Box is designed to run on a local NVIDIA DGX Spark. Local AI tasks can run on the appliance, while external model services remain an optional configuration choice.
Do team members need to install AI software on their own computers?
Usually not. Team members can access the browser-based workspace from their computers and use the available AI applications, team files, messaging, and virtual devices.
Can Team AI Box be installed completely offline?
The target appliance can use an offline package, but initial source preparation must be completed on a compatible connected machine. Cloning the repository alone does not create a complete offline installation.
Does Team AI Box include Windows?
No. Virtual Windows and Linux computers require an authorized compatible ARM64 ISO and any required operating-system license. They do not replace the DGX Spark host operating system.
Can team size determine concurrency?
No. Concurrency depends on the models, workload types, context length, and hardware conditions. Test actual business workflows before promising a specific number of simultaneous users.
Why can nvidia-smi show “Memory-Usage: Not Supported”?
DGX Spark uses unified memory rather than dedicated VRAM. Use free -h or the DGX Dashboard to understand system memory availability.
Conclusion
A reliable DGX Spark deployment starts before the first AI conversation. Confirm the workflow, hardware, platform-specific behavior, network, data boundary, accounts, recovery plan, performance, and licenses before the workspace goes live.
Team AI Box turns one NVIDIA DGX Spark into a browser-accessible workspace for local AI, team files, messaging, wireless casting, and virtual devices. Start with one practical workflow, validate access and performance, then expand as the team learns which applications and policies it needs.
Read the Team AI Box deployment guide, review the capabilities, or follow How to Share One DGX Spark with Your Team.