Microsoft's new Surface can run 120B AI models locally

Microsoft just pulled the curtain back on the Surface Laptop Ultra, and it might be the most aggressive piece of PC hardware the company has ever engineered.

Built in deep collaboration with NVIDIA from the silicon up, the new machine is squarely aimed at developers, creators, and AI builders who have outgrown standard laptops.

Under the hood, the Laptop Ultra packs an NVIDIA Blackwell RTX GPU with full CUDA support and pairs it with up to 128GB of unified memory. That pool of RAM dynamically shifts between the CPU and GPU depending on what is being thrown at it.

According to Microsoft, the system delivers 1 petaflop of AI compute (based on theoretical FP4 TOPS using sparsity). That raw horsepower translates into the ability to run up to 120 billion parameter models locally, right on your desk without touching the cloud.

Beyond AI muscle, the hardware specs lean heavily into professional utility. It features a 15-inch mini-LED PixelSense Ultra touchscreen hitting 2,000 nits of peak HDR brightness, 262 ppi, and high color precision—making it the brightest panel Microsoft has ever shipped.

The chassis comes in Platinum and Nightfall finishes, sporting the largest haptic trackpad yet on a Surface. Microsoft also resisted the urge to strip away ports: it includes HDMI, USB-C, USB-A, an SD card reader, and a 3.5mm headphone jack. Microsoft claims all-day battery life and quiet thermals, with availability slated for later this year.

Why it matters

For the past few years, developers wanting to run massive local LLMs were basically pushed into buying top-tier Apple Silicon MacBooks purely for the unified memory. The catch? You had to live without NVIDIA's indispensable CUDA ecosystem.

Microsoft and NVIDIA are basically attacking that specific pain point head-on. By pairing 128GB of unified memory with a Blackwell RTX GPU and native CUDA, the Surface Laptop Ultra gives AI engineers a legitimate, no-compromise mobile workstation.

If you are training smaller architectures, testing 120B local weights, or running offline agentic workflows, this machine looks like a dream setup on paper.

That said, we remain curious about how reality holds up against the slides. Pushing 1 petaflop and handling 120B models generates immense heat, so whether it truly stays quiet and maintains all-day battery under genuine engineering workloads will make or break this ambitious bet.