Nvidia just did something it’s never really done before: built an entire computer chip designed to sit on your desk and run AI models that previously required a small data center. The GB10 Grace Blackwell Superchip, unveiled at CES 2025, pairs the company’s Grace CPU with its Blackwell GPU architecture into a single integrated package. It’s compact, it’s power-efficient, and it’s coming to PCs from Dell, ASUS, HP, Acer, Gigabyte, Lenovo, and MSI.
The sticker price starts between $3,000 and $4,000, with systems shipping from mid-2025. For context, that’s roughly the cost of a high-end MacBook Pro, except this thing delivers up to 1 petaFLOP of AI performance using FP4 precision. In English: it can process a quadrillion floating-point operations per second, which is the kind of math that makes large language models actually work.
What the GB10 actually does
The core selling point here is local AI execution. The GB10 can run and fine-tune AI models with up to 200 billion parameters without touching a cloud server. That’s significant because it eliminates latency, sidesteps compliance headaches around sensitive data, and removes the recurring costs of renting compute from AWS or Azure.
Backing all of that compute is 128GB of unified LPDDR5X memory. Unified memory means the CPU and GPU share the same memory pool, which cuts down on the bottleneck that typically happens when data has to shuttle between separate memory banks.
