Surface RTX Spark Dev Box $5,999: price local AI first

Microsoft put a price on local AI yesterday. The Surface RTX Spark Dev Box is $5,999, and the Surface Laptop Ultra starts at $2,599. Both run on Nvidia's RTX Spark chip, and Microsoft says both run models over 120 billion parameters on the device. The pitch is AI with no per-token bill. The real question for a small business is simpler: how many months of API spend does $6,000 buy?

What actually happened

At its October 7 event in San Francisco, Microsoft opened preorders. Per the Windows devices blog:

• Surface Laptop Ultra: from $2,599, up to 128GB of unified memory, available October 16. Business configurations are on preorder now. A trade-in offer gives up to $1,000 back on an eligible MacBook Pro through November 23 (US and Canada, Microsoft Store online).
• Surface RTX Spark Dev Box: $5,999 with 128GB of unified memory, sold only on Microsoft.com in the US, shipping in November.
• Performance claim: up to 1 petaflop of AI compute. Microsoft's own footnote says that is theoretical FP4 with sparsity, not a sustained number.

TechCrunch reports the top Laptop Ultra configuration runs $5,900, and that Microsoft says the highest-end model is already out of stock. The Dev Box ships with VS Code, GitHub Copilot CLI, WSL, and PowerShell 7.

On the software side, Microsoft's Windows Experience blog says Windows ML now supports llama.cpp, and that local-model routing through GitHub HydraFusion reaches an experimental preview later in October.

Why it matters for your business

Do the break-even math. If your team spends $300 a month on API tokens, a $5,999 box takes 20 months to pay back. That assumes the local model does the job as well, and it ignores power and your time. At $50 a month, the math does not work. Cheap cloud models like Claude Haiku 5.5 move the bar up again.

Privacy is the better reason. If you handle client files, medical notes, or contracts you can't send to a third-party API, a local box can pay off even when the token math doesn't.

128GB is shared memory. Microsoft's footnote says the GPU can address less than the total. A 120B-parameter model at low precision fits. A 120B model at full precision does not. Ask what quantization you will run before you buy.

Hardware is the easy part. You still need patching, backups, and someone who owns the box. A local model server is a server.

Sources: Microsoft Windows Blog, TechCrunch.