Product image

Manufacturer product image. The selected configuration may vary.

Product photo & illustrative 3D · configuration may varySource

Sourced specificationsReviewed 2026-09-07

Desktop AI system

MSI EdgeXpert

GB10 Grace Blackwell Superchip

A compact GB10 workstation offered in several SSD and cable bundles. This listing uses the single-unit 4 TB Gen4 configuration.

128 GB unifiedReference memory

arm64CUDA ecosystem

Models & capacity Share your setup

Compared withNVIDIA DGX Spark

BEFORE IT ARRIVES

Power, cooling & your space.

What this configuration needs on your premises.

Requirements reviewed 2026-09-29

Lab assessment · based on the evidence below

Office candidate

A desktop office candidate with documented airflow provisions for stacking. Check total heat and fan noise when several boxes run continuously.

Power rating

Published detail

Approximately 240 W USB-C external adapter capacity. [1]

Electrical input

Published detail

External adapter feeds the USB-C power input. AC input range is not specified in the reviewed product table. [1]

Outlet & circuit

Not verified

No machine-specific branch-circuit requirement verified. Check the nameplate, regional cord and other loads on the same circuit.

Machine cooling

Published detail

FanDock airflow system, SSD vapor chamber and raised feet for airflow between stacked units. [1]

Operating limits

Published detail

0–35°C operating ambient; 10–90% relative humidity, noncondensing. [1]

Noise in the workspace

Not verified

No comparable noise measurement under an AI workload verified. Ask for dB(A), distance, workload and ambient conditions; audition it before placing it beside people.

ACTUAL CONSUMPTION

Power at the wall

PSU capacity, chip TDP and measured consumption describe different things. Only measurements with a stated configuration and test method appear here.

Idle and AI-load wall power: not verified.

No attributable wall-power test for this configuration was established in this review. Measure at the outlet with the exact model, workload, runtime and power mode recorded.

Share a power measurement

Room cooling

Plan for the computer and its power supply to heat the room. Check ventilation during long jobs and when several machines share a small office.

Lab planning guidance. Office suitability depends on the installed configuration, shared circuits, heat removal and acceptable noise; a workstation label alone does not settle it.

Space & installation

Enclosure: Confirm selected configuration
Weight: Not verified

Keep air inlets and exhausts unobstructed, with room for cables and servicing. An enclosed cupboard needs its own ventilation; the enclosure dimensions are not an airflow clearance.
Keep the factory feet in place and the airflow gap clear when stacking; budget each unit’s adapter and room heat.

Estimate electricity use & room heat

Enter average power at the wall for the hours you are modeling, or an explicit planning assumption. Supply wattage and chip TDP are not consumption measurements.

Wall power per machine (W)MachinesHours per dayDays in periodPrice per kWh (optional)Tariff currency

Add wall power to calculate a scenario.

Scenario only. Hours outside this period are excluded: include idle time for an always-on system. Displays, networking, UPS losses and room air conditioning are extra. Heat assumes the computer and power supply release their heat into this room; it does not size a circuit or an HVAC system. Currency selects your tariff’s unit, with no conversion.

Sources & review scope (1)

Manufacturer documentation and attributed first-hand hardware inspections, scoped to the named model or configuration. “Not verified” identifies a gap in this review, not a claim that the manufacturer has no documentation. Installation assessments are the Lab’s interpretation.

MSI · EdgeXpert MS-C931 power and thermal design

THE MACHINE × THE MODEL × THE WORKLOAD

Models, capacity & economics

Connect the model, serving stack, workload and cost of operating this box.

Contribute a configuration

Model	Serving configuration	Capacity planning	Evidence
openai/gpt-oss-20b	MXFP4SGLang	No measured latency or throughput published for this configuration	Vendor-listed validationNVIDIA · source below
openai/gpt-oss-120b	MXFP4SGLang	No measured latency or throughput published for this configuration	Vendor-listed validationNVIDIA · source below
nvidia/Llama-3.3-70B-Instruct-FP4	NVFP4SGLang	No measured latency or throughput published for this configuration	Vendor-listed validationNVIDIA · source below
nvidia/Qwen3-8B-FP8	FP8SGLang	No measured latency or throughput published for this configuration	Vendor-listed validationNVIDIA · source below
nvidia/Qwen3-14B-FP8	FP8SGLang	No measured latency or throughput published for this configuration	Vendor-listed validationNVIDIA · source below
nvidia/Qwen3-32B-FP4	NVFP4SGLang	No measured latency or throughput published for this configuration	Vendor-listed validationNVIDIA · source below

SELECTED CONFIGURATION

gpt-oss-20b · NVIDIA DGX Spark

Checkpoint / model artifact: openai/gpt-oss-20b
Precision: MXFP4
Runtime / environment: SGLanglmsysorg/sglang:latest-cu130
Context setting: 8,192 tokensShared guide starting value; not a measured limit for this model.

Execution backend: CUDA 13 / FlashInfer
Configuration notes: Follow the checkpoint requirements in the upstream guide.
Exact model revision / OS / driver: To record in a reproduction
Usable concurrency / task quality: Load and quality results not attached

Mutable container tag; pin its digest and model revision.

Reference-platform evidence does not establish exact-OEM performance.

Support evidence · NVIDIA Upstream instructions Reviewed 2026-09-10

THE OPERATING RECORD

A model that loads is the beginning.

Each measured result belongs to one box configuration, one model, one serving stack and one workload. Publish latency, per-stream speed, request rate, errors and task quality together.

Opening model configurations…

01 / PRIVACY

Your data has a home.

Keep sensitive processing on your infrastructure. Inspect every connection in the complete stack.

02 / SECURITY

You set the boundaries.

Control access, connectivity and updates. A local machine still needs a secure deployment.

03 / TOKEN ECONOMICS

Own the capacity.

Generate tokens on your own compute. Understand the full cost of the useful work it delivers.

DEPLOYMENT NOTEBOOK / NVIDIA GB10

From a desktop box to a managed service

Arm software compatibility

Check that every container and native dependency supports arm64. Model compatibility alone does not validate the application.

Shared-service operations

NVIDIA now documents fleet lifecycle integration and customized installation. Its example tools need adaptation to your IT environment; they are not preinstalled.

A meaningful capacity test

Record prompt length, output length, cache state and simultaneous requests. Repeat after warmup and report the latency target and failure rate.

Lab synthesis of supplier guidance and deployment questions · reviewed 2026-09-13. Reference-platform details may differ from this machine.

NVIDIA enterprise manageability

SHARED WORK AROUND THIS PLATFORM

Setups you can build on.

Reference-platform guidance and field records. Exact OEM configurations need their own validation.

All setups

Understand TechField record

An AI application stack on two office Sparks

UT’s move from rented cloud GPUs to machines in its own office, and the engineering record behind it.

Office inferenceApplications

GB10Reported · technical detail pending

NVIDIA · indexed by the LabRecipe

A browser chat interface, served by your Spark

NVIDIA’s Open WebUI and Ollama guide, connected to the hardware and operating questions it raises.

OllamaOpen WebUI

GB10Source guidance · not UT reproduced

NVIDIA · indexed by the LabRecipe

An inference endpoint for your local applications

Connect applications to a model server on a GB10 or GB300 reference platform using vLLM.

vLLMModel serving

GB10 / GB300Source guidance · not UT reproduced

THE PLATFORM BEHIND THE MACHINE

GB10 / DGX Spark

8 catalog configurations

Shared silicon does not guarantee identical performance. Compare the memory, cooling and configuration of each machine.

Explore this platform

BEYOND THE SPECIFICATION

Lessons for running this machine.

Published guidance and checklists to help you investigate the full deployment.

Share your experience

Infrastructure

Ollama, vLLM, NIM: nine months of deployment learning.

UT’s reported journey across serving stacks, hardware integration and customer feedback. A starting record for the decisions and evidence still to publish.

UT field note

Startup & recovery

A NIM container that took about 30 minutes to start.

An operator-reported startup observation from UT’s GB10/GB300 work. The next step is to identify the exact configuration and separate the startup phases.

UT field note

Engineering applications

Local AI beside a Wi-Fi module test bench.

An execution agent and confidential engineering test assets show how on-premises AI extends beyond office applications.

UT field note

THE CONFIGURATION

Inside the machine

Configuration: EdgeXpert-11SUS · 128 GB / 4 TB Gen4 · one unit
CPU / accelerator: GB10 Grace Blackwell Superchip
Memory: 128 GB unified
Memory bandwidth: 273 GB/s
Architecture: arm64
Operating system: NVIDIA DGX OS
Storage: 4 TB PCIe Gen4 NVMe

ON YOUR PREMISES

Space, power & connectivity

Form factor: Desktop AI system
Dimensions: Confirm selected configuration
Weight: Not verified
Power rating: Approximately 240 W USB-C external adapter capacity.
Connectivity: Confirm selected configuration
Availability: Listed for order · recheck with supplier

Power supply ratings and processor TDP are not measured wall consumption. Supplier stock and regional delivery can change.

Electrical, cooling & installation details

BEFORE YOU CHOOSE

The US store offer is the EdgeXpert-11SUS single unit with 128 GB unified memory and a 4 TB Gen4 drive. MSI also markets paired-unit configurations; confirm networking accessories for a two-node setup.

The EdgeXpert shares NVIDIA's GB10 platform. Platform guidance applies; no Lab measurement exists for this exact chassis.

Memory capacity alone does not establish speed, runtime compatibility or production capacity.

Manufacturer specifications

Help improve this record. Suggest a correction

IN THE FIELD

Deployment experiences

Operating records from customers, contributors and the Lab.

All deployments

ADD TO THE SHARED RECORD

Where does your AI run?

In an office, a factory, an engineering lab or beside your test equipment. Share a deployment and keep the people behind it credited.

Contribute a deployment

THE SHARED TECHNICAL LIBRARY

Knowledge around this machine.

Supplier recipes, community experience and UT’s deployment lessons, connected to the hardware.

Explore 18 resources

OperationsSupplier

Manage a Spark fleet and prepare offline installation

NVIDIA documents enterprise lifecycle integration and custom installation patterns.

GB10

For your deployment

Example tools are not preinstalled. Adapt to your management system and check the documented support boundaries.

NVIDIASupplier reference · not reproduced by UTIndexed 2026-09-13 · versions live at source

Open originalDiscuss

RecipesSupplier

NIM inference on DGX Spark

NVIDIA’s model-serving playbook.

GB10NIM

For your deployment

Check the selected NIM profile, image digest, memory use and readiness time.

NVIDIASupplier reference · not reproduced by UTIndexed 2026-09-12 · versions live at source

Open originalDiscuss

RecipesSupplier

Serve models with vLLM

NVIDIA’s vLLM setup reference.

GB10vLLM

For your deployment

Pin the container and test your model, prompt lengths and simultaneous requests.

NVIDIASupplier reference · not reproduced by UTIndexed 2026-09-12 · versions live at source

Open originalDiscuss

Related platform guidance does not establish compatibility with this exact OEM configuration.

THE LOCAL AI WATCH

Updates for this machine

All updates

Loading updates…

MEMBER RATINGS · SELF-DECLARED, NOT MEASURED

How members rate the EdgeXpert.

One rating per member, tied to a Lab profile. Stars count immediately; written verdicts are published after a quick review by the Lab team. A rating is an opinion about fit for a job, not a benchmark.

Loading ratings…

THE PEOPLE BEHIND THE MACHINES

Experience with EdgeXpert?

0 contributions

Questions, ideas and experience from putting AI machines to work. You do not need a finished deployment to contribute.

Loading the discussion…

MSI EdgeXpert

Power, cooling & your space.

Office candidate

Power rating

Electrical input

Outlet & circuit

Machine cooling

Operating limits

Noise in the workspace

Power at the wall

Room cooling

Space & installation

Models, capacity & economics

gpt-oss-20b · NVIDIA DGX Spark

A model that loads is the beginning.

Your data has a home.

You set the boundaries.

Own the capacity.

From a desktop box to a managed service

Arm software compatibility

Shared-service operations

A meaningful capacity test

Setups you can build on.

An AI application stack on two office Sparks

A browser chat interface, served by your Spark

An inference endpoint for your local applications

GB10 / DGX Spark

Lessons for running this machine.

Ollama, vLLM, NIM: nine months of deployment learning.

A NIM container that took about 30 minutes to start.

Local AI beside a Wi-Fi module test bench.

Inside the machine

Space, power & connectivity

Deployment experiences

Where does your AI run?

Knowledge around this machine.

Manage a Spark fleet and prepare offline installation

NIM inference on DGX Spark

Serve models with vLLM

Updates for this machine

How members rate the EdgeXpert.

Experience with EdgeXpert?

Add to the conversation

Models, capacity & economics

gpt-oss-20b · NVIDIA DGX Spark

A model that loads is the beginning.