The Surface RTX Spark Dev Box is Microsoft’s first NVIDIA-powered AI mini PC — a compact, Windows-on-Arm developer machine built around the new NVIDIA RTX Spark superchip, with 128 GB of unified memory and up to 1 petaflop of AI compute. Microsoft revealed it at Build 2026 alongside the Surface Laptop Ultra, and it is squarely aimed at developers who want to prototype, fine-tune and run large models locally instead of renting cloud GPUs. This is a preview, not a review: the device is real and officially announced, but Microsoft has been deliberately sparse with the details that matter most for buyers — price, ship date and port speeds.
What is the Surface RTX Spark Dev Box?
It is a passively cooled small-form-factor desktop in an aluminum chassis “engineered to double as a heatsink,” with a distinctive perforated top surface (early coverage counts roughly 1,000 ventilation holes) that reads like a flattened Xbox Series X. Conceptually it is Microsoft’s answer to the wave of Linux-based AI dev boxes built on NVIDIA’s GB10 Grace Blackwell platform — the NVIDIA DGX Spark and the ASUS Ascent GX10 — except this one runs Windows.
Worth clearing up early: when the partnership was first teased, there was genuine confusion over whether Microsoft had actually committed to a desktop, since the headline reveal was the Surface Laptop Ultra and the “Dev Box” name initially surfaced in NVIDIA’s Q&A and a Windows-blog footer. That ambiguity is now resolved — Microsoft published a dedicated Surface Devices blog post and a Surface product page for the Dev Box. It is a confirmed, named product. What remains genuinely unconfirmed are the specifics below.
What’s confirmed: the RTX Spark superchip
RTX Spark is the Windows-on-Arm sibling of NVIDIA’s GB10 Grace Blackwell silicon — the same die family that powers the on-sale DGX Spark, not a next-generation chip. Per the joint NVIDIA/Microsoft materials, the confirmed pieces are:
- 20-core NVIDIA Grace Arm CPU, co-designed with MediaTek, linked to the GPU over NVLink-C2C.
- NVIDIA Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores with FP4 precision.
- 1 petaflop of AI compute.
- 128 GB of unified memory shared between CPU and GPU.
- Headroom to run 120B-parameter models with up to a 1-million-token context window locally “at interactive speeds,” or fine-tune models that previously needed cloud GPU instances.
- A 100 W thermal envelope — Microsoft’s first public hint at the RTX Spark chip’s power draw.
On the software side, it ships with a developer-tuned Windows 11 Pro, preconfigured with VS Code, GitHub Copilot, Git, Python and Node.js, plus WSL 2 with GPU passthrough and CUDA support and PowerShell 7. The pitch is “local-first AI development” with the CUDA stack available natively on Windows — something the Linux GB10 boxes can’t claim.
What’s still unknown
This is where buyers should keep expectations in check. As of mid-June 2026, Microsoft has not disclosed:
- Price. None announced. Given RTX Spark systems are positioned as premium developer hardware — not a successor to the budget Windows Dev Kit 2023 — expect it to land near the DGX Spark’s ~$3,000–$4,000 bracket rather than below it.
- A firm ship date. Microsoft says only “later this year,” consistent with the Fall 2026 platform window for RTX Spark.
- Memory bandwidth. No figure published. If it inherits the GB10 family’s roughly 273 GB/s LPDDR5x ceiling — as documented in our DGX Spark review — then bandwidth, not compute, will be the real limiter on token-generation throughput. Treat any specific number as unconfirmed until Microsoft says otherwise.
- Port speeds and exact I/O. ServeTheHome notes the rear layout (USB-C, USB-A, HDMI, Ethernet, audio) looks “quite similar to NVIDIA’s own DGX Spark box, minus a couple of USB-C ports and the Spark’s defining ConnectX-7 QSFP Ethernet ports” — meaning the high-speed 200GbE clustering that defines DGX Spark is apparently not here. Link speeds are unstated.
- Storage capacity and SSD generation.
What it means for mini PCs
If the platform delivers, the Dev Box matters for one specific reason: it brings a 128 GB unified-memory, CUDA-native AI box to Windows without dropping to Linux. Today, a developer who wants this much coherent memory on the desktop chooses between DGX OS (Arm Ubuntu) on a GB10 box or an AMD Ryzen AI Max system like the Framework Desktop or GMKtec EVO-X2. Microsoft is betting that “stays in Windows, keeps CUDA, keeps Copilot” is worth a premium to a meaningful slice of AI developers.
The honest caveat: the absence of ConnectX-7 clustering and the likely ~273 GB/s bandwidth ceiling mean this is a single-box prototyping and fine-tuning tool, not a throughput monster. For interactive local inference on 70B–120B models it should be excellent; for high-volume serving it will be bandwidth-bound like every other GB10 device.
Expected timeline and price outlook
- Availability: later this year (Fall 2026 window), U.S. only, exclusively on Microsoft.com at launch.
- Price: none announced. Plan for premium developer-hardware pricing, not budget.
- Status: officially announced “pre-release product,” subject to regulatory approval; specs may shift before launch.
Sources & caveats
Everything here is drawn from NVIDIA’s and Microsoft’s own announcements plus first-look coverage from Tom’s Hardware, ServeTheHome and The Register. The chip-level specs (20-core Grace CPU, 6,144-CUDA Blackwell GPU, 128 GB unified memory, 1 petaflop, 120B-param support) are confirmed by NVIDIA and Microsoft. The device-level unknowns — price, exact ship date, memory bandwidth, port speeds and storage — are not, and we’ve flagged each as such above. RTX Spark is the GB10 Grace Blackwell family on Windows, not a new generation of silicon.
We’ll publish a full review when the Surface RTX Spark Dev Box actually ships and we can test local LLM throughput against its real memory bandwidth. Until then, the closest on-sale equivalents we’ve already reviewed are the NVIDIA DGX Spark and the ASUS Ascent GX10 — same GB10 platform, Linux instead of Windows, and available to buy today.