Microsoft just gave AI developers a reason to stop renting GPU time from the cloud. At the Microsoft Build conference in San Francisco on June 2, CEO Satya Nadella took the stage to announce the Surface RTX Spark Dev Box, a desktop machine purpose-built for local AI model development that packs specs most data centers would have envied just a few years ago.
The headline numbers: 20 CPU cores, 128GB of unified memory, and up to 1 petaflop of AI compute. In English, that means a single box sitting on your desk can run and fine-tune AI models with over 120 billion parameters at interactive speeds, handling context windows of up to 1 million tokens.
What’s under the hood
The Dev Box is powered by NVIDIA’s Arm-based RTX Spark superchip, which uses NVIDIA’s Grace CPU architecture. This is notable because it represents a significant departure from the x86 chips that have dominated Windows PCs for decades.
The 128GB of unified memory is particularly interesting for AI workloads. Unlike traditional setups where CPU and GPU memory are separate pools, unified memory lets the processor and graphics hardware share the same pool of RAM. That matters enormously when you’re loading massive language models that can easily consume 50GB or more just to sit in memory before doing any actual computation.
