Microsoft's new Surface AI PCs pack Nvidia's RTX Spark

Microsoft wants developers back, and it's willing to pay a four-figure bounty to pry Silicon Valley away from their MacBooks.

At San Francisco Tech Week, Microsoft and Nvidia officially launched a new generation of local AI PCs powered by Nvidia's RTX Spark chip. The headline hardware includes the Surface Laptop Ultra, a powerhouse starting at $2,600, and a dedicated 24/7 workstation dubbed the Surface RTX Spark Dev Box that starts at $6,000.

To make their intentions unmistakably clear, Microsoft is dangling up to $1,000 in trade-in credits for anyone handing over a MacBook Pro.

Desktop-class compute inside a laptop

The beating heart behind these machines is Nvidia's RTX Spark superchip. It pairs an up to 20-core Grace CPU with a Blackwell RTX GPU sporting up to 6,144 cores, tied together across a massive 600 GB/s unified memory interconnect.

The hardware delivers up to one petaflop of FP4 compute and supports up to 128GB of unified memory. That is enough headroom to run massive open weights—like the 125B Qwen 3.8 Flash Next model—entirely on-device without cloud API fees or latency.

Prices ramp up quickly: the Laptop Ultra climbs to $5,900 for maxed-out specs, and Microsoft says that top tier is already out of stock. OEM partners like Dell are jumping in too, listing its RTX Spark XPS 16 Creator Edition for $3,800. For non-stop workflows, Microsoft also rolled out the Dev Box, a compact desktop built for continuous tasks, preloaded with VS Code, GitHub Copilot CLI, and WSL.

Sandboxing agents at the OS level

Raw power is only half the pitch. As Microsoft CEO Satya Nadella put it during a fireside chat with Jensen Huang, having a raw model doesn't mean much anymore. To make AI useful, developers need an orchestration layer that manages memory, context, and external action spaces.

To handle that, Microsoft rolled out Microsoft Execution Containers (MXC) to Windows 11. It acts as an operating-system-level sandbox built specifically for autonomous agents, letting them run persistently in the background while keeping them strictly isolated from sensitive OS processes.

Huang didn't hold back his praise, claiming MXC will redefine agent deployment just as DirectX once revolutionized PC gaming.

Why it matters

We think this partnership finally gives Windows a legitimate weapon against Apple Silicon's dominance among AI engineers.

For nearly four years, Silicon Valley developers defaulted to top-spec MacBooks because Apple's unified memory was the only way to run beefy local models on the go. Now, by marrying 128GB unified memory with the undisputed king of AI software—Nvidia CUDA—Microsoft and Nvidia are presenting a compelling alternative. For enterprise devs and creators, the ability to iterate locally on the exact same stack used in GPU clusters is a genuine workflow unlock.

Still, at these stratospheric price tags, everyday users have zero reason to buy in right now. It's strictly an expensive sandbox for professionals.

Strategically, MXC is Microsoft's smartest move here. Instead of trying to own every agent, Windows is positioning itself as the trusted runtime police. By making the OS the execution sandbox, Microsoft guarantees that no matter whose agent wins, Windows remains the essential foundation.