MSI EdgeXpert MS-C931

Published on August 2, 2026

The MSI EdgeXpert MS-C931 is MSI’s build of NVIDIA’s DGX Spark platform, which means the silicon inside is the same GB10 Grace Blackwell Superchip that NVIDIA and its other partners ship: a 20-core Arm CPU (10 Cortex-X925 plus 10 Cortex-A725) tied to a Blackwell GPU by NVLink-C2C, with 128GB of LPDDR5x unified memory that both halves address directly. NVIDIA rates the package at 1000 AI TOPS in FP4, and the memory runs on a 256-bit interface for 273 GB/s of bandwidth. It runs NVIDIA DGX OS, an Ubuntu derivative, on Arm64; there is no Windows option and no x86 fallback. The whole thing occupies 1.19 liters, weighs 1.2 kg, and takes its power from a single USB-C adapter. This is a developer appliance for running and fine-tuning large models locally, not a mini PC in the sense that most of this catalog uses the word.

Pros and Cons of the MSI EdgeXpert MS-C931

Related Videos

MSI EdgeXpert MS-C931 Comparison Chart

	MSI EdgeXpert MS-C931	MSI EdgeXpert MS-C931
Price	List Price: $2,999.00 Amazon Prices: Check Price on Amazon	List Price: $5,999.00 Amazon Prices: Loading prices... Best Buy: Loading price...
Version	128GB/1TB/NVIDIA GB10	128GB/4TB/NVIDIA GB10
Performance Rating	0.6	0.6
Operating System	Linux	Linux
Processor	Twenty-core NVIDIA GB10 Grace Blackwell Superchip	Twenty-core NVIDIA GB10 Grace Blackwell Superchip
GPU	Integrated NVIDIA GB10 Blackwell GPU	Integrated NVIDIA GB10 Blackwell GPU
RAM	128 GB LPDDR5X (128GB unified, shared by CPU and GPU)	128 GB LPDDR5X (128GB unified, shared by CPU and GPU)
Internal Storage	1 TB NVMe SSD	2 TB NVMe Gen5 SSD
Dimensions width x length x thickness	5.94 x 5.94 x 2.05 inches (150.88 x 150.88 x 52.07 mm)	5.94 x 5.94 x 2.05 inches (150.88 x 150.88 x 52.07 mm)
Weight	2.65 lbs (1.2 kg)	2.65 lbs (1.2 kg)
WiFi	Wi-Fi 7 (802.11be)	Wi-Fi 7 (802.11be)
Bluetooth	Bluetooth 5.4	Bluetooth 5.4
Ethernet	1 Ethernet port at 10 Gbps	1 Ethernet port at 10 Gbps
HDMI	1 Full-Size HDMI Port	1 Full-Size HDMI Port
DisplayPort	3 DisplayPorts (DP 1.4a via 3 USB-C ports)	3 DisplayPorts (DP 1.4a via 3 USB-C ports)
VGA	No VGA Ports	No VGA Ports
USB Ports	4 USB-C 4x USB 3.2 Type-C (1 power in, 3 with DP alt mode)	4 USB-C 4x USB 3.2 Type-C (1 power in, 3 with DP alt mode)
Thunderbolt Ports	No	No
OCuLink	No	No
Internal SATA Ports	No SATA ports	No SATA ports
Card Reader	No Card Reader	No Card Reader
Headphone Jack	--	--
Fanless	No	No
VESA Mount	No	No
In the Box	EdgeXpert unit, USB-C power adapter, quick start guide.	EdgeXpert unit, USB-C power adapter, quick start guide.
Expandability	Memory is unified and soldered. One M.2 M-key NVMe from the factory (1TB or 4TB). Two ConnectX-7 QSFP ports pair two units.	Memory is unified and soldered. One M.2 M-key NVMe from the factory (1TB or 4TB). Two ConnectX-7 QSFP ports pair two units.

Detailed Insights into the MSI EdgeXpert MS-C931

The EdgeXpert measures 5.94 x 5.94 x 2.05 inches (151 x 151 x 52 mm) and weighs 2.65 pounds, so it is a square metal slab a little larger than a CD case and about twice as thick. The front is a brushed bar over a honeycomb intake mesh, with a Kensington lock slot on the left side and vent slits down both flanks. Every port is on the back: a power button, four USB 3.2 Type-C ports, a full-size HDMI 2.1a output, a 10GbE RJ-45 jack, and two QSFP cages driven by an NVIDIA ConnectX-7 NIC rated at 200 Gbps. One of the four USB-C ports is the power input and three carry DisplayPort 1.4a alt mode, which is how MSI reaches its claim of four independent displays. There is no USB-A, no 3.5mm jack (audio goes out over HDMI), and no card reader, which tells you plainly who this box is for.

The GB10 Superchip is the part worth understanding before buying. Its CPU half is a 20-core Arm cluster, and its GPU half is Blackwell with fifth-generation Tensor cores, but the interesting number is neither of those: it is the 128GB of coherent unified memory that the GPU can use as if it were VRAM. That capacity is the entire argument for the platform, because an RTX 5090 tops out at 32GB. The tradeoff is bandwidth. LPDDR5x on a 256-bit bus delivers 273 GB/s, a fraction of what GDDR7 on an RTX 5090’s 512-bit bus moves, so the EdgeXpert wins on model size and loses on raw token throughput. No PassMark CPU Mark exists for the GB10, so there is no directly comparable score against x86 chips.

Networking is where MSI’s build earns its industrial branding. Alongside WiFi 7 and Bluetooth 5.4, the 10GbE port handles ordinary LAN duty while the two ConnectX-7 QSFP ports exist for one specific job: connecting a second EdgeXpert directly, with no switch in between, so the pair shares 256GB of memory. MSI sells that pairing as a bundle SKU (EdgeXpert-02SKUS, two units plus a QSFP cable) rather than as a storage option, and Tom’s Hardware reported at the Computex 2025 announcement that a GB10 system handles models up to 200 billion parameters alone and up to 405 billion when two are linked this way. Storage is a single M.2 M-key slot holding a self-encrypting NVMe drive, filled at the factory with either 1TB or 4TB.

Running Local Models on the EdgeXpert

The case for this hardware is the same case we walked through in our look at the DGX Spark and local AI hardware in 2026: a machine that holds a very large model in memory is doing something a fast consumer GPU cannot do at any price, and a machine that streams tokens quickly through a small model is doing something the EdgeXpert cannot match. Which of those matters depends entirely on your workload, and it is worth reading our comparison of Ollama Cloud versus local AI hardware before spending four figures on a box whose main job you might be renting for dollars a month.

Because the GB10 is a platform rather than an MSI exclusive, the chip is not what separates the EdgeXpert from its rivals. ASUS sells the Ascent GX10 on the same superchip, and NVIDIA sells its own DGX Spark; the differences that remain are thermals, storage configuration, chassis, and price. MSI’s list prices are $2,999 for the 1TB EdgeXpert-99SUS and $5,999 for the 4TB EdgeXpert-13SUS, so the 4TB step costs $3,000, far more than a 4TB Gen5 NVMe drive is worth on its own. If a second SSD is not the reason you are buying, the 1TB configuration holds the identical compute. Readers weighing this class against x86 alternatives should also see our roundup of mini PCs for running local LLMs and our coverage of dual-unit Strix Halo clustering, which chases the same large-model goal from the AMD side.

Reviewer Insights on the MSI EdgeXpert MS-C931

Independent testing of MSI’s specific build is thin. Most published GB10 coverage evaluates the DGX Spark platform generally, using NVIDIA’s own unit or a different partner’s box, and those results describe the chip rather than MSI’s chassis or cooling. One trusted-channel hands-on covers this exact machine.

ETA PRIME

ETA PRIME spent a first-look video with the EdgeXpert, confirming the hardware on camera by opening the RPCS3 GPU configuration and reading back the NVIDIA GB10. He framed the appeal in physical terms up front: “MSI and NVIDIA have created a supercomputer that you could throw in your backpack.” Walking the rear panel, he confirmed the four USB 3.2 Type-C ports with the end one reserved for power input, the full-size HDMI, the 10GbE jack, and the ConnectX-7 200 gig cages, and he noted that pairing two units needs only a direct cable rather than a switch.

On what the box is actually for, he was direct: “I do think that this is a great platform for AI developers, and it’s definitely not a consumer product.” His reasoning was the memory, not the compute: “The main things this has going for it versus, let’s say, the RTX 5090 and an x86 setup is 128 GB of RAM that the GPU can also access.” He ran GPT-OSS-20B through Open WebUI and generated images in ComfyUI without trouble, and he credited NVIDIA’s onboarding, saying “This is made for AI development and NVIDIA has actually made it really easy to get started with this.”

He also hit the rough edges. The bundled DGX Dashboard showed memory and GPU utilization but nothing else, and he said “I have not found a way to add anything else like CPU utilization.” PS3 emulation through RPCS3 held 60 FPS at 1080p but was not perfectly clean: “Every once in a while, I do notice a little bit of a stutter.” Getting a Steam client running on the Arm build defeated him during the video, and he reported “I’ve been trying to get Steam installed on this. Having a few issues.”

Customer Reviews of the MSI EdgeXpert MS-C931

The 4TB EdgeXpert averages 4.6 stars across 16 Amazon ratings, with 77% at five stars, 15% at four, and 8% at two. The written reviews are unusually detailed for a sample that small, and several of them are careful to separate MSI’s build from NVIDIA’s platform.

The praise clusters on thermals and memory capacity. B. Jester, a verified buyer writing in April 2026, said MSI’s version of the DGX Spark is “definitely one of the best” and that under concurrent inference load “it stays under 80C and performs well,” though he added a 140mm fan underneath for extra airflow. Another verified buyer, santa, listed “better thermal design” as the headline pro and reported the box “seems to work fine during heavy load without thermal throttling.” Haruyuki Shini described running NVIDIA’s Nemotron 3 Super alongside Stable Diffusion, writing “You can easily get a 400k context window while also running Stable Diffusion and generating images that are 1080p or bigger.”

The complaints are almost all about software maturity rather than the hardware. santa’s con was blunt: “Official distribution of Pytorch does not support Blackwell GB10 yet,” with NVIDIA’s own PyTorch container as the workaround. Rob M., who raised his own rating from one star to four over a month of use, still concluded that “It is immature and the software landscape is in such a state of flux that every document you find was obsolete a week before you read it,” and described renting B200 or dual-H100 time to convert a 70B model into the NVFP4 format vLLM wanted. B. Jester made the same bandwidth tradeoff explicit: “what you’re really getting is the 128GB of unified RAM at the cost of slower bandwidth,” and recommended a Blackwell card instead for anyone who wants raw speed. Patrick Bolen, five stars, still wished “there was faster bandwidth or cheaper price but it is what it is.” Read more owner reviews on Amazon.

Conclusion

The EdgeXpert makes sense for one reader: a developer or researcher who needs a large model resident in memory on hardware they own, who is comfortable on Arm64 Linux, and who values NVIDIA’s CUDA and container tooling enough to accept a narrower software catalog than x86. For that person, 128GB of coherent unified memory in a 1.19-liter box that a single USB-C adapter powers is a genuinely unusual offer, and the owners quoted above credit MSI’s cooling for holding up where they had heard other GB10 boxes struggle.

Everyone else should be skeptical. Token throughput on small and mid-size models will disappoint anyone coming from a discrete Blackwell card, the software stack was still moving under buyers’ feet through mid-2026, and nothing inside is upgradeable. If your models fit in 32GB, a GPU tower is faster for less money. If you want a general-purpose desktop, this is the wrong machine at any price. And if you do want the platform, the 1TB configuration carries identical compute for $3,000 less, so pay for the larger SSD only if you actually need it.

For those comparing options, see our Mini PC Comparison Chart to find the best match for your needs.

Frequently Asked Questions