Microsoft unveiled the Surface RTX Spark Dev Box at Microsoft Build 2026, a miniature desktop PC built around Nvidia's Arm-based RTX Spark chips and 128GB of unified memory, enough to run 120 billion parameter models locally. The machine is aimed at developers building and testing AI workloads on-device rather than calling out to a cloud API. It ships later this year in the US through Microsoft's online store, with pricing still undisclosed.
The Dev Box runs inside a 100-watt thermal envelope, noticeably higher than the 45-watt-to-80-watt envelopes Nvidia gives the RTX Spark laptops announced alongside the Surface Laptop Ultra this past weekend. The extra headroom is the point: Microsoft is targeting sustained local AI workloads, not bursty consumer use. The aluminum chassis doubles as a heatsink, and the top surface visually echoes the textured grille of an Xbox Series X.
The hardware slot it occupies has been empty for two years. Qualcomm's Snapdragon Dev Kit, originally meant to ship in 2024 as the reference machine for porting apps to Windows on Arm, was canceled after hardware-quality complications. Microsoft is now stepping into that vacuum directly with its own Surface-branded developer box, this time built on Nvidia silicon rather than Qualcomm's.
Key facts
- 01The Surface RTX Spark Dev Box ships with 128GB of unified memory, enough to run 120 billion parameter models locally.
- 02It runs on Nvidia's Arm-based RTX Spark chips inside a 100-watt thermal envelope, up from the 45–80 watts in RTX Spark laptops.
- 03The aluminum chassis doubles as a heatsink and visually echoes the top of an Xbox Series X.
- 04It directly replaces Qualcomm's canceled Snapdragon Dev Kit, which was meant to ship two years ago.
- 05The Dev Box launches later this year in the US through Microsoft's online store; pricing has not been disclosed.
Out of the box, the system is preconfigured with Windows 11 Pro and tuned for developer use, with Visual Studio Code, GitHub Copilot, and other tooling installed at the image level. Andrew Hill, corporate vice president of Surface, said the goal is to drop developers straight into a working environment without setup overhead.
The local-AI angle is the strategic core. A 120 billion parameter model running on a desk machine without an API bill is the kind of capability that, until recently, required a multi-GPU server. Pairing that with GitHub Copilot and Visual Studio Code locally lets developers iterate on agentic workflows, fine-tunes, and inference code without latency or per-token costs.
Microsoft is not alone in chasing this form factor. Other OEMs are building miniature PCs around the same RTX Spark platform, and the Dev Box will compete with those in the same buyer pool. Microsoft's advantage is the integrated software image and the Surface brand inside enterprise IT, where standardized developer hardware is a recurring procurement line.
The Dev Box also fits into a broader pattern at Build 2026. Microsoft has used the conference to announce its first in-house reasoning model, MAI-Thinking-1, and to unify the agentic AI stack with Nvidia across Windows, Azure, and on-premise — both of which we covered earlier this week. The Spark Dev Box is the on-device endpoint of that stack: the place a developer writes, tests, and runs the agent before it ever touches a cloud GPU.
For Nvidia, the Surface deal is another win for the RTX Spark platform inside Windows on Arm, an architecture Qualcomm dominated until recently. Microsoft's willingness to ship its flagship developer reference box on Nvidia silicon — rather than wait for Qualcomm to retry the Snapdragon Dev Kit — signals where the company sees the Arm-on-Windows AI story heading.
Open questions remain. Microsoft has not disclosed full specifications, storage configurations, I/O, or pricing, and "later this year" is a wide window. Software compatibility on Windows on Arm has improved sharply but still trails x86 in niches that matter to some developer teams, and the 120B-parameter local-inference claim will need real-world benchmarks against comparable machines once units ship.
The bigger read for the AI market is that local inference on developer hardware is becoming a product category, not a hobbyist curiosity. If a 128GB unified-memory desktop can host a frontier-class model under a developer's desk for a fixed capex, the economics of agent development shift away from per-token cloud spend during the build phase. Microsoft owning that endpoint — and pairing it with Copilot, Visual Studio Code, and its own MAI model line — is how the Windows platform stays load-bearing in the agent era.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
