Oct 13, 2026

Surface RTX Spark Dev Box for Local AI in 2026: What $5,999 Buys That a $3,099 GX10 Doesn't

By RunAIHome Team · 12 min read

rtx-sparknvidiaai-hardwarelocal-llmmini-pc

TL;DR: Microsoft opened preorders October 7 for the Surface RTX Spark Dev Box — a $5,999 desktop with the same NVIDIA N1X chip and 128GB unified memory as the $5,899 Surface Laptop Ultra, shipping in November. The silicon family is proven at 46–61 tok/s on gpt-oss-120b, but the ASUS GX10 does that today from $3,099.

	Surface RTX Spark Dev Box	ASUS Ascent GX10	Mac Studio M5 Max 128GB
Best for	Windows-first CUDA dev, zero setup	Same GB10-class silicon, lowest price	Fastest measured decode in class
Price / Cost	$5,999 (ships Nov 2026)	$3,099–$4,150	$5,099
The catch	~$2,000 premium over same-family hardware	Linux (DGX OS), no Windows	No CUDA; Metal/MLX stack

Honest take: Unless your workflow genuinely requires Windows 11 Pro with CUDA preinstalled and a receipt your employer will sign, buy the GX10 and keep the ~$2,000 difference — it buys the same model ceiling on the same silicon family.

Microsoft now sells three shapes of the same chip. The Surface Laptop Ultra put NVIDIA’s RTX Spark N1X in a 15-inch laptop; the Dev Box puts it in a 4.4 lb aluminum desktop and charges $100 more for it. For local AI, the question is never the chassis — it’s bandwidth, GPU-addressable memory, and what the software stack actually delivers. Two of those three are already measurable on sibling hardware, and they’re the reason the price is the story here.

What Microsoft announced

The Dev Box went up for preorder on October 7, 2026 at $5,999 MSRP ($5,999.99 on the configurator), sold exclusively through Microsoft.com in the US, shipping in November (Windows Devices Blog). The hardware, per Microsoft’s store listing and launch coverage:

Spec	Surface RTX Spark Dev Box
Chip	NVIDIA RTX Spark N1X — 20 Arm CPU cores + 6,144-core Blackwell GPU, NVLink C2C
Memory	128GB unified LPDDR5X (shared CPU/GPU)
AI compute	”Up to 1 petaflop” — sparse FP4, the same figure quoted for DGX Spark
Storage	Removable 2TB PCIe Gen 5 SSD
Thermals / power	~100W sustained thermal envelope, 165W external PSU, anodized aluminum chassis doubling as heatsink
Ports	2× USB-C, USB-A (5 Gbps), HDMI 2.1b, Ethernet, 3.5mm jack
Weight	4.4 lbs (2.0 kg)
OS	Windows 11 Pro

(Engadget’s hands-on counted three USB-C ports; Microsoft’s configurator lists two — trust the spec sheet until retail units arrive. Engadget)

The software pitch is “Project Zenith,” Microsoft’s name for a ready-to-code Windows image: VS Code, Git, GitHub CLI, GitHub Copilot, WSL2 with GPU passthrough, Python, and Node preinstalled (Neowin, XDA). That’s genuinely the product’s one unique feature: every other 128GB unified-memory box on the market makes you pick and configure your own stack.

One number Microsoft conspicuously does not publish: memory bandwidth. The N1X platform is rated at roughly 300 GB/s in its laptop configurations, and the DGX Spark — built on the same GB10 silicon family — is rated at 273 GB/s. Every decode-speed calculation below uses that ~273–300 GB/s window, and nothing about a desktop chassis changes it.

The $100 desktop: what you trade against the Laptop Ultra

The 128GB Surface Laptop Ultra is $5,899 and ships October 16. The Dev Box is $5,999 and ships in November. Same chip, same memory, $100 apart — so what does the desktop actually buy?

Sustained thermals. The Laptop Ultra has to live inside a 15-inch chassis; the Dev Box gets a ~100W sustained envelope with the whole case acting as a heatsink. We’ve documented what chassis limits do to the same silicon before — the MacBook Pro M5 Max throttles against its 14-inch case while the Mac Studio doesn’t — and long inference runs are exactly the sustained-load workload where that gap shows. For batch prompt processing and fine-tuning jobs, the desktop should hold clocks the laptop can’t. No independent numbers exist yet to size that gap; treat it as directionally real and quantitatively unproven.
A removable Gen 5 SSD. The laptop’s storage is fixed at purchase; the Dev Box’s 2TB drive swaps out. Model libraries grow — a used 4TB Gen 5 drive down the line is cheaper than buying storage at Microsoft’s config prices.
You give up the 2,000-nit mini-LED touchscreen, the battery, and portability.

As a pure trade, $100 for better sustained thermals and serviceable storage is fine. The problem is that neither machine is competing against the other — both are competing against boxes that cost thousands less.

The bandwidth math: what 128GB at ~300 GB/s actually runs

Decode speed on local LLMs is memory-bandwidth-bound. The ceiling is bandwidth divided by bytes read per token, and real hardware lands at 70–95% of it:

Dense 70B: fits, crawls. Llama 3.3 70B Q4_K_M is 42.5GB of weights; 300 ÷ 42.5 ≈ 7 tok/s ceiling. The capacity to hold it is not the speed to enjoy it.
MoE 120B-class: the actual use case. gpt-oss-120b reads ~3GB of active experts per token: 300 ÷ 3 ≈ 100 tok/s ceiling. This model class is why 128GB unified boxes exist.
The Windows cut. NVIDIA’s own porting guide documents that Windows splits the unified pool — a GPU carveout, a shared region sized by formula (post-carveout capacity minus 16GB, clamped to 50–80%), and a CPU-only remainder (NVIDIA RTX Spark Porting Guide). On a 128GB machine that works out to roughly 105GB GPU-touchable, not 128GB. gpt-oss-120b (~61GB) plus generous context fits comfortably. Qwen3-235B-A22B at Q4 (~141GB) does not fit — for that class you need the Framework Desktop 192GB’s ~160GB Linux pool or a Mac Studio tier. Run your own model-plus-context arithmetic in our VRAM calculator before paying for capacity you can’t address.

What the same silicon measures today

No independent Dev Box benchmarks exist — preorder hardware, November shipping. But the GB10 family has been in buyers’ hands for months as the DGX Spark, and the llama.cpp community benchmark thread is the best public data on what this platform does (llama.cpp discussion #16578). On the stock DGX OS kernel, build 6816:

$ llama-bench -m gpt-oss-120b-mxfp4.gguf -fa 1 -ub 2048
| model                   |      size | test   |             t/s |
| ----------------------- | --------- | ------ | --------------- |
| gpt-oss 120B MXFP4 MoE  | 59.02 GiB | pp2048 | 1737.17 ± 81.66 |
| gpt-oss 120B MXFP4 MoE  | 59.02 GiB | tg32   |    45.87 ± 0.74 |

46 tok/s decode on a 120B-class MoE — about half the theoretical ceiling. The same thread shows the headroom is software: one user swapped the stock kernel for mainline 6.17 on Fedora 43 and the identical benchmark jumped to 60.57 tok/s decode and 1,956 tok/s prefill — a 32% speedup from a kernel change alone. The Dev Box runs Windows, a path with zero published llama.cpp numbers on this silicon. Expect launch performance below the Linux figures, improving with driver updates — and expect WSL2, which Microsoft preconfigures, to be the serious path on day one.

Two failure modes from the DGX Spark community are worth knowing before reviews land, because Dev Box reviewers will hit them too:

The FP8 default trap. NVIDIA’s NIM container serves Llama 3.3 70B at FP8 by default — ~70GB of weight reads per token through a ~273–300 GB/s pipe. Users reported exactly the ~3 tok/s the math predicts (NVIDIA Developer Forums). The fix is the same on every GB10 machine: run Q4_K_M GGUFs, or better, run MoE models that only read active experts. A launch review showing one dense-70B-FP8 number and calling the box slow has measured the configuration, not the hardware.
The driver question. The only leaked Surface-platform benchmark so far — a September Laptop Ultra engineering sample — ran Qwen3.5 9B at 22.45 tok/s on Vulkan because the CUDA path failed outright. NVIDIA has had months since, and the DGX Spark proves the silicon; the thing to verify in first retail reviews is llama.cpp or Ollama running on the CUDA backend at a sane fraction of the ~100 tok/s MoE ceiling.

The competition: five boxes, one job

Every machine below holds gpt-oss-120b-class MoE models. Prices are October 2026 street prices we’ve verified for earlier coverage:

Hardware	Price	Bandwidth	gpt-oss-120b decode	CUDA	OS out of box
Surface RTX Spark Dev Box	$5,999	~300 GB/s (platform)	unproven (sibling: 46–61 tok/s)	Yes	Windows 11 Pro, dev stack preinstalled
ASUS Ascent GX10	$3,099–$4,150	273 GB/s	46–61 tok/s measured	Yes	DGX OS (Linux)
NVIDIA DGX Spark	$6,950 (Oct 2026)	273 GB/s	46–61 tok/s measured	Yes	DGX OS (Linux)
GMKtec EVO-X2 128GB	$3,499–$3,649	256 GB/s	~31 tok/s measured	No (ROCm/Vulkan)	Windows 11
Mac Studio M5 Max 128GB	$5,099	614 GB/s	65–88 tok/s measured	No (Metal/MLX)	macOS

Read the table cold and the Dev Box has the same problem as its laptop sibling: it’s the most expensive machine in the class on a silicon family the GX10 sells from $3,099. If raw decode per dollar is the metric, the Mac Studio M5 Max is $900 cheaper with double the bandwidth and the best measured numbers in class — the trade is losing CUDA, which matters for fine-tuning and the broader tooling ecosystem but not for llama.cpp inference. And if you just want the capacity at the lowest price, the EVO-X2 is $2,350–$2,500 less and runs the same models at two-thirds the speed.

What the Dev Box uniquely offers is Windows 11 Pro + CUDA + a preconfigured dev environment in one SKU — a machine you can hand a Windows-first team where ollama run and VS Code + Copilot work out of the box, with a local 120B-class model as a BYOK backend for your coding tools. That’s a real product. It’s just a ~$2,000 convenience fee for most home-lab buyers, and this niche is already crowded with cheaper, measured alternatives — our self-hosting guides cover the same stack setup on any of them in an afternoon.

What to actually buy

Prices as of October 2026, all taken from the comparison above:

Your situation	The machine	Price	Where
Want CUDA + 128GB at the lowest price, Linux is fine	ASUS Ascent GX10	$3,099–$4,150	Check price
Want 128GB today at the absolute floor, no CUDA	GMKtec EVO-X2 128GB	$3,499–$3,649	Check price
Fastest measured 120B-class decode, macOS is fine	Mac Studio M5 Max 128GB	$5,099	Check price
Windows + CUDA + zero setup, company card	Surface RTX Spark Dev Box	$5,999	microsoft.com (US exclusive)
Everything you run fits in 24GB — speed over capacity	Used RTX 3090 desktop	$1,399–$1,450 (card)	Check price
Undecided — test the workload on rented hardware first	Rented 3090, from $0.07/hr	pay per hour	Vast.ai

FAQ

Is the Dev Box faster than the Surface Laptop Ultra?

Same chip, same memory, same ~300 GB/s ceiling — peak decode speed should be identical. The Dev Box’s ~100W sustained envelope should win on long batch jobs where the laptop throttles, but no independent measurements exist yet for either machine.

Can it run Qwen3-235B or DeepSeek-class models?

No. After Windows’ memory split, roughly 105GB of the 128GB pool is GPU-addressable, and Qwen3-235B-A22B at Q4 needs ~141GB. The cheapest boxes that hold that class are the Framework Desktop 192GB ($6,799, ~160GB Linux pool) and Mac Studio tiers above 128GB.

Why is there no Amazon listing?

Microsoft is selling the Dev Box exclusively through Microsoft.com in the US for the preorder window. If that changes after the November ship date, street pricing below MSRP would change this comparison — check back.

Should I preorder or wait for reviews?

Wait. The GB10 family’s Linux numbers are proven, but the Dev Box’s Windows/WSL2 inference path has zero published benchmarks, and the one leaked test on Surface hardware (September, Laptop Ultra prototype) had CUDA failing outright. Shipping starts in November; reviews will exist within days. A few weeks of patience against $5,999 is cheap insurance.

Is 1 petaflop real?

It’s sparse FP4 math — the same marketing number NVIDIA quotes for the DGX Spark. It says nothing about LLM decode speed, which is bandwidth-bound: ~100 tok/s ceiling on 3GB-per-token MoE models, ~7 tok/s on dense 70B. TOPS and petaflops predict prompt-processing speed, not generation.

Recommended Gear

ASUS Ascent GX10 — same GB10-class CUDA silicon from $3,099
GMKtec EVO-X2 128GB — cheapest 128GB unified-memory box
Mac Studio M5 Max — fastest measured decode in the 128GB class
Used RTX 3090 — 936 GB/s for everything that fits in 24GB

Sources

Last updated October 13, 2026. Prices and specs change; verify current rates before purchasing.

Was this article helpful?

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.