Can One DGX Spark Support an AI Team?
The Real Limits of a Shared DGX Spark Team Workstation
When teams consider buying an NVIDIA DGX Spark, the important question is rarely whether a model can run.
The real question is:
Can one DGX Spark support several people using local AI at the same time?
The short answer is yes—but only when the team treats DGX Spark as a shared local AI workstation, not as an infinitely scalable cloud server.
DGX Spark can support local model execution, document processing, image generation, speech transcription, AI application development, and virtual device testing. With Team AI Box, team members can access these capabilities through a browser-based workspace that also includes team files, messaging, and virtual devices.
However, the number of people a single system can support depends on the model, context length, simultaneous requests, virtual devices, available unified memory, runtime, and the response time the team considers acceptable.
This article focuses on capacity planning: concurrent users, workload mix, response speed, memory headroom, and when to add more hardware. For the account, file-sharing, messaging, and browser-access setup, see How to Share One DGX Spark With Your Team.
DGX Spark is an AI workstation, not a managed data center
NVIDIA positions DGX Spark as a desktop AI system for developers, researchers, data scientists, and teams that need local AI computing.
The official specifications include a 20-core Arm CPU, up to 128 GB of unified memory, and support for AI models up to 200 billion parameters under the stated conditions. NVIDIA also documents multi-Spark configurations for larger models and workloads. See the official DGX Spark hardware specifications.
These specifications describe the hardware. They do not guarantee:
- A fixed number of users
- A fixed number of concurrent requests
- A specific tokens-per-second result
- A guaranteed response time
- Per-user GPU isolation
- Cloud-style service availability
An AI workstation is designed to run models, develop applications, process private data, and test AI workflows locally. It is not automatically a multi-tenant enterprise platform.
That is why “DGX Spark for teams” should be understood as a shared local AI workstation with a team access layer—not a fully managed cloud service with unlimited concurrency.
Can you share DGX Spark with multiple users?
Yes. But you need a team access layer between the users and the hardware.
A personal setup might look like this:
Individual user
↓
Terminal or local application
↓
Model runtime
↓
DGX Spark
A team deployment needs a different structure:
Team members
↓
Browser access
↓
Team AI Box workspace
↓
User accounts and application access
↓
Local AI applications, team files, messaging, and virtual devices
↓
NVIDIA DGX Spark
Team AI Box turns DGX Spark into a browser-accessible workspace. Employees do not need to install a complete AI environment on every computer, and most users do not need to manage Docker containers, model files, or Linux services directly.
See How to Share One DGX Spark With Your Team for the account, file, messaging, and browser-access workflow.
What kind of AI team can one DGX Spark support?
Good fit: shared and asynchronous workflows
One DGX Spark can be a practical starting point for:
- Internal document processing
- Meeting transcription
- OCR for images and PDFs
- Translation review
- Product documentation
- Technical writing
- AI image generation
- Customer-support material
- Local AI application development
- Virtual device testing
These workloads do not always require every user to receive an immediate response.
For example:
Meeting recording
↓
Speech transcription
↓
Human review
↓
AI-assisted meeting notes
↓
Approved file in the team drive
↓
Team discussion
The work can be processed, reviewed, saved, and reused. For this kind of workflow, one shared DGX Spark may be easier to operate than separate GPU systems for every employee.
Requires testing: several heavy tasks at once
A DGX Spark workload becomes more demanding when users simultaneously run:
- Long-context conversations
- Large local models
- Image generation
- Large OCR batches
- Long audio or video transcription
- Large translation jobs
- Virtual Android devices
- Windows or Linux virtual machines
- Model fine-tuning
- Background agent workflows
DGX Spark uses unified memory. The operating system, model weights, context windows, containers, file-processing tasks, virtual devices, and user sessions all share the same physical memory pool.
A model loading successfully does not mean that the entire team can use it reliably at the same time.
Poor fit: high-concurrency production serving
A single DGX Spark may not be the right architecture when a team needs:
- Large numbers of simultaneous users
- Strict response-time guarantees
- High-volume API serving
- Dedicated GPU capacity for each project
- Elastic scaling
- Automatic failover
- A formal production SLA
- Cloud-style service availability
In those cases, compare multiple DGX Spark systems, a larger NVIDIA workstation, dedicated GPU servers, cloud inference, or a hybrid architecture.
Benchmark starting points—not Team AI Box guarantees
Public DGX Spark measurements can provide useful starting points, but they should not be treated as Team AI Box capacity guarantees.
One public coding-agent benchmark reported approximately 37–47 tokens per second per user with 4–8 concurrent users on one Spark. At 16 concurrent users, the same benchmark reported approximately 22–24 tokens per second per user, with around 350 tokens per second aggregate throughput. The benchmark author also states that 16 users was a test point, not a hardware ceiling. See the benchmark methodology and results.
Single-stream results also vary substantially by model, runtime, quantization, and prompt shape. One benchmark roundup reports 33.5–60.7 tokens per second for gpt-oss-120b on the same DGX Spark hardware across different engines and test setups. Review the published DGX Spark benchmark comparison.
Use these numbers as test starting points:
| Test condition | Starting reference |
|---|---|
| 1–2 interactive users | Measure single-user latency and first-token time |
| 4–8 concurrent users | Test per-user speed, queueing, and task completion |
| 16 concurrent users | Treat as a stress test, not a default target |
| Long-context workloads | Expect lower concurrency |
| OCR, transcription, images, and VMs | Test separately from text generation |
The correct production number depends on the team’s model, runtime, prompt length, output length, and workload mix.
A practical interactive-speed reference
Some published local-LLM capacity analysis uses approximately 15 tokens per second per user as a practical reference point for comfortable interactive text generation. This is not an NVIDIA specification and is not a Team AI Box guarantee. It is a useful test line for a team that wants to compare responsiveness across models and concurrency levels:
| Per-user speed | Practical interpretation |
|---|---|
| Around 15 tok/s or higher | Often comfortable for interactive text work |
| Around 10–15 tok/s | Usable, but slowdown may be noticeable |
| Around 5–10 tok/s | Better suited to light or asynchronous work |
| Below 5 tok/s | Usually poor for interactive-heavy workflows |
These ranges are workload-dependent. Context length, time to first token, output length, batching, quantization, and the user’s task all affect the experience. Treat the 15 tok/s line as a measurement reference, not a promise. Review one published concurrency analysis.
NVIDIA’s own performance material also shows that benchmark context matters. For example, NVIDIA reports dual DGX Spark systems running a 235-billion-parameter Qwen3 model at 11.73 tokens per second in a specific reference workload. That result is useful evidence of model and system capability, but it is not a single-Spark user-capacity promise. Read NVIDIA’s performance reference.
Why DGX Spark concurrency falls: 273 GB/s of memory bandwidth
DGX Spark provides up to 128 GB of unified memory and 273 GB/s of memory bandwidth. The operating system, model weights, KV cache, containers, files, virtual devices, and user sessions share the same memory architecture. See NVIDIA’s DGX Spark hardware guide.
During generation, the inference engine repeatedly reads model weights and related state from memory. Multiple active sequences can share some work, but they still compete for a finite memory path.
This creates a practical pattern:
- Aggregate throughput can increase when more requests are batched.
- Per-user speed usually falls as more users share the same model.
- Queueing and time to first token become more important at higher concurrency.
- Long-context requests consume more KV cache and reduce practical capacity.
- A system may remain technically healthy while becoming unpleasant to use.
This is why “the model fits in 128 GB” is not the same as “the team can use it concurrently.” The 273 GB/s bandwidth figure is not a user limit, but it helps explain why adding users eventually produces diminishing returns.
One published analysis associates the bandwidth constraint with roughly 15 tok/s in a particular 30B-class setup. That is a model- and runtime-specific inference, not a universal DGX Spark limit. Teams should reproduce the measurement with their own model, context length, and serving stack.
This interpretation is an inference from the unified-memory architecture and published benchmark behavior. Teams should validate it with their own models and workflows.
What does Team AI Box add to DGX Spark?
DGX Spark provides the local compute layer. Team AI Box provides the workspace layer that helps a team use that compute.
The Team AI Box capabilities page describes features including:
- Local AI assistants
- AI image generation
- Text recognition from images and PDFs
- Speech transcription
- Translation
- Team files
- Direct and group messaging
- Wireless casting
- Virtual Android phones
- Virtual Windows and Linux computers
- User accounts and task-status visibility
This makes Team AI Box more than a model endpoint. It brings AI applications, working files, team conversations, and virtual devices into one browser-based environment.
The workspace should not be described as automatic cross-user memory or a complete enterprise knowledge base. It gives people and supported applications a shared place to work with approved files, messages, and task results; whether an application can use a particular file or conversation still depends on its workflow and the configured account controls.
Teams still need to manage file organization, source review, external model-service rules, access boundaries, backups, recovery, operating-system licenses, virtual machine media, model updates, and heavy-task scheduling.
Team AI Box should not be presented as automatically providing a complete enterprise knowledge base, per-user GPU isolation, fixed concurrency, or a full enterprise audit system.
How to evaluate DGX Spark team capacity
Do not begin by asking, “How many users can one DGX Spark support?” Start by defining the actual team workload.
Step 1: List the team’s tasks
Write down the workflows the team expects to use:
- Document questions
- Meeting transcription
- Image OCR
- File translation
- Image generation
- Technical writing
- Virtual Android testing
- Windows or Linux application testing
Step 2: Record resource requirements
Measure:
- First-start time
- Task completion time
- Memory behavior
- Impact on other users
- Long-context performance
- Concurrent response behavior
- Remaining capacity after starting virtual devices
- Recovery time after a failed task
Step 3: Test with real users
A useful test might look like this:
User A: text conversation
User B: PDF OCR
User C: speech transcription
User D: document translation
User E: virtual device testing
The goal is not to run one model continuously for a few minutes. The goal is to observe whether the full workspace remains responsive when different applications run at the same time.
Step 4: Set team operating rules
If the team uses one DGX Spark, define which tasks receive priority, which tasks should run during quieter periods, which models ordinary users can access, whether large file uploads are allowed, whether external model services are permitted, and who handles system recovery and updates.
DGX Spark for teams: one system or multiple systems?
| Team requirement | Recommended starting point |
|---|---|
| Shared documents and light AI tasks | One DGX Spark with Team AI Box |
| Several users with mostly asynchronous tasks | One DGX Spark with workload testing |
| Several users running large models simultaneously | Multiple DGX Spark systems or a larger GPU workstation |
| Dedicated GPU capacity for every user | Multiple devices or a dedicated GPU server |
| Elastic scaling | Cloud or hybrid infrastructure |
| AI inference alongside virtual devices | One Spark may work, but test memory headroom |
| Formal production SLA | Data-center or managed GPU infrastructure |
NVIDIA’s announced 2-to-8-node expansion path
For workloads that exceed one DGX Spark, NVIDIA documents a guided path for connecting two to four DGX Spark systems using ConnectX-7 networking and QSFP links. Two or three systems can be connected directly; four systems use a compatible network switch. See NVIDIA’s official multi-Spark connection documentation.
The practical expansion path is:
One Spark
↓
Two to four Sparks through the currently documented guided path
↓
Up to eight Sparks through NVIDIA’s announced advanced path
↓
Custom cluster or data-center infrastructure
A four-Spark configuration can increase available memory and support larger distributed workloads, but it does not automatically provide per-user GPU isolation, enterprise identity management, fixed response-time guarantees, automatic failover, or a complete multi-tenant service layer.
NVIDIA has also announced official support for eight-node DGX Spark clusters, with up to 1 TB of unified memory across eight 128 GB systems. NVIDIA’s technical article states that support for eight-node clusters in NVIDIA Sync is being rolled out, while the current Cluster Assistant guide still documents guided configuration for a maximum of four devices. Treat deployments beyond four systems as an advanced or manual configuration unless the relevant NVIDIA Sync release provides the guided workflow for your environment. Read NVIDIA’s eight-node technical article.
The eight-node announcement expands the hardware and distributed-workload path; it does not automatically create a multi-tenant team service. Team accounts, workspace permissions, file organization, scheduling, monitoring, recovery, and application-level access remain separate design concerns.
How to prepare DGX Spark for Team AI
Team AI Box is intended to run on a supported NVIDIA DGX Spark. The current Team AI Box deployment guide lists requirements including:
- DGX OS, an Ubuntu 24.04-based ARM64 environment
- Docker Engine
- Docker Compose v2
- NVIDIA drivers and container support
- KVM
- Git
- Python 3.10 or later
- Several hundred gigabytes of temporary free space
The installation runs directly on the DGX Spark. Initial preparation requires downloads for dependencies, models, and build inputs.
If the target environment must operate offline, prepare the complete offline package on a compatible connected machine first. Cloning the source repository alone does not complete the installation.
See How to Set Up DGX Spark Local AI before beginning deployment.
Do virtual devices affect team capacity?
Yes.
Team AI Box virtual Android phones, Windows computers, and Linux computers also use shared system resources, including unified memory, CPU capacity, storage, network resources, display resources, and background containers.
Windows and Linux virtual computers require compatible ARM64 installation media and any required operating-system licenses. They do not replace the DGX Spark host operating system.
If the team plans to run AI inference, OCR, virtual Android devices, and Windows virtual machines at the same time, test the combined workload before opening the system to everyone.
See the DGX Spark Windows and Linux virtual machine guide for guest-environment requirements.
Frequently asked questions
How many users can one DGX Spark support?
There is no universal number. Public tests provide useful starting references—for example, some 30B-class tests report roughly 4–8 concurrent users at interactive speeds—but these are not Team AI Box capacity guarantees. Use the model, context length, file sizes, task types, virtual devices, simultaneous requests, runtime, time to first token, and available memory headroom to test your own limit. A reference line such as approximately 15 tok/s per user can help compare results, but it is not an official threshold.
Can multiple users share one DGX Spark?
Yes. The safer team pattern is to provide browser access through Team AI Box with individual accounts, shared files, and application access instead of giving everyone the same administrator account or direct SSH access.
How is this article different from the DGX Spark team-sharing guide?
The team-sharing guide explains how people access one DGX Spark through Team AI Box, including browser access, accounts, files, messaging, and workspace use. This article answers a different question: how to evaluate concurrent workload capacity, measure response speed, account for virtual-device overhead, and decide when one Spark is no longer enough.
Is DGX Spark a server or an AI workstation?
It is primarily a desktop AI workstation, but it can operate as a shared local AI appliance when a workspace layer is added. It is not automatically a managed multi-tenant cloud server.
Can DGX Spark run large AI models?
Yes, within the supported hardware and software conditions. However, “the model can run” and “the team can use it concurrently” are different questions. Teams also need to account for context windows, KV cache, containers, file processing, and other applications.
When should a team buy a second DGX Spark?
Consider adding another system when users regularly affect one another’s response times, heavy tasks cannot run together, virtual devices compete with model inference, development and daily use need separation, the team needs higher availability, or one appliance has become a single point of failure.
Conclusion: Can one DGX Spark support an AI team?
One DGX Spark can be a practical starting point for a small team’s local AI workflows, especially when many tasks are asynchronous or lightly concurrent. It should not be treated as an infinitely concurrent AI server.
It is best suited for shared internal documents, local document and media processing, meeting transcription, translation, OCR, AI-assisted content work, virtual device testing, and local AI application development.
Team AI Box lets employees access these capabilities through a browser and brings AI applications, team files, messaging, and virtual devices into one workspace.
The correct evaluation question is not simply how many employees the company has. It is how many people will run high-load tasks at the same time.
If the team mostly uses asynchronous workflows, one DGX Spark may be enough as a shared AI workstation. If the team needs strict concurrency, dedicated GPU resources, elastic scaling, or formal service guarantees, plan for multiple systems—including the announced advanced eight-node path where appropriate—a dedicated GPU server, or a hybrid cloud architecture.
Begin with one real workflow. Test it with actual users. Measure memory, task completion time, response behavior, queueing, and recovery before adding more models or users.
For the broader architecture, read the Team AI Box local AI infrastructure guide.
For related implementation details, see the Team AI Box deployment guide and the DGX Spark Windows and Linux virtual machine guide.