NVIDIA Project DIGITS: a 1 PFLOP personal AI supercomputer for $3,000
In one sentence Announced at CES 2026, NVIDIA Project DIGITS packs a GB10 Superchip, 128 GB unified memory, and 1 PFLOP FP4 into a desktop device priced at $3,000, enabling local inference of frontier models like Llama 4 405B without the cloud.
Picture having a data-center-class AI machine sitting on your desk at home. That is exactly what NVIDIA announced at CES 2026 with Project DIGITS — a compact device roughly the size of a Mac mini that costs $3,000 and can run the world's most powerful open AI models entirely locally, with zero data leaving your machine.
The magic is inside the new GB10 Superchip, co-designed with MediaTek, which fuses a Grace CPU and a Blackwell GPU into a single package. They share 128 gigabytes of unified memory, eliminating the usual bottleneck of shuttling data between processor and graphics card when dealing with enormous models.
The raw compute tops out at 1 petaFLOP in FP4 precision — one thousand trillion operations per second — enough to run Llama 4 at 405 billion parameters without cutting corners. Need more? Connect two units together and you get 2 PFLOP and 256 GB of memory in a dual-node setup: enough to handle virtually any open-weights model that exists today.
Until now, models of this size needed server racks costing tens of thousands of dollars. Project DIGITS changes that equation entirely, putting frontier-model capability in the hands of researchers, developers, and technically minded users at a price point that is genuinely accessible. For anyone working with AI, this is a paradigm shift.
Companies
NVIDIA, MediaTek
Tools
Project DIGITS, GB10
Tags
Sources