Microsoft puts 120B models on your desk

If you have spent the last year building AI applications, your cloud bill probably looks like a second rent. Every prompt tweak, agent loop, and evaluation run chips away at your budget, even when you are just running basic tests.

At Build on June 2, Microsoft introduced a dedicated hardware play to stop the bleeding: the Surface RTX Spark Dev Box. Following the announcement of the Surface Laptop Ultra earlier in the week, this compact desktop machine is built entirely around local-first AI development.

A Blackwell superchip on your desk

The beating heart of the machine is the NVIDIA RTX Spark superchip, which pairs an NVIDIA Blackwell RTX GPU with an ultra-efficient NVIDIA Grace CPU.

It packs 128 GB of unified memory and delivers up to 1 petaflop of AI compute using theoretical FP4 with sparsity. Those are server-like numbers in a desktop form factor.

According to Microsoft, that is enough headroom to run 120B+ parameter models with 1 million tokens of context locally at interactive speeds. It can also handle model fine-tuning jobs that previously demanded dedicated cloud GPU instances.

To handle sustained loads like multi-step agent pipelines and long training jobs, Microsoft designed the aluminum chassis to double as a giant heatsink, keeping sustained performance steady.

A stripped-down, CUDA-ready Windows

Instead of making developers spend days fighting drivers, Microsoft pre-configured Windows 11 Pro at the image level.

Out of the box, you get dark mode, a clean taskbar with Widgets removed, Do Not Disturb turned on, Developer Mode active, and PowerShell 7 as the default shell.

Under the hood, WSL 2 is pre-configured with GPU passthrough and native CUDA support. VS Code, Git, Python, Node.js, and GitHub Copilot are pre-installed.

It integrates deeply with Microsoft's developer ecosystem, including Windows Copilot Runtime, Windows ML with TensorRT, the AI Toolkit for VS Code, and Microsoft Foundry for moving local prototypes into production.

Microsoft says the Dev Box will launch later this year in the U.S. exclusively through Microsoft.com, with pricing details still under wraps.

Why it matters

This is not just another Surface; it is a strategic counterattack aimed directly at local AI workflows.

For developers and technical teams, the value proposition is obvious. Prototyping complex autonomous agents over cloud APIs is slow, rate-limited, and expensive. Having 128 GB of unified memory with CUDA capability lets you run and fine-tune large open models locally with zero token fees, saving API calls only for truly frontier-tier queries.

For the broader industry, this is Windows and NVIDIA finally delivering a true answer to Apple's high-end Apple Silicon Macs. For the last couple of years, developers turned to Mac Studio rigs primarily because consumer GPUs lacked unified memory to load massive models. By bundling Grace and Blackwell into a compact desktop with proper CUDA tooling, Microsoft is making a serious bid to reclaim the developer workstation.

We think the execution here looks unusually thoughtful. Stripping bloat from Windows and shipping day-one CUDA support addresses developers' real friction points. The real question mark remains pricing. If it is priced within reach of boutique studios and serious builders, it could quickly become the benchmark hardware for agent engineering.