Nvidia puts a personal AI supercomputer on your desk

Nvidia just brought its massive data center architecture straight to your desk.

At CES 2025 in Las Vegas, CEO Jensen Huang revealed Project Digits, a compact "personal AI supercomputer" designed to run frontier-sized models right from a standard wall outlet.

Starting at $3,000 and shipping in May, the device looks remarkably similar to a Mac Mini, but it is built to handle workloads that usually require a cloud cluster.

Supercomputing in a lunchbox

Under the hood sits Nvidia's new GB10 Grace Blackwell Superchip, co-developed with MediaTek.

The chip combines a Blackwell GPU—complete with latest-gen CUDA cores and fifth-gen Tensor Cores—with a 20-core Arm-based Grace CPU. They are tied together via Nvidia's high-speed NVLink-C2C interconnect.

Spec-wise, each unit packs 128GB of unified coherent memory, up to 4TB of NVMe flash storage, and cranks out 1 petaflop of FP4 compute performance.

That is enough local horsepower to run models with up to 200 billion parameters on a single box. If you need more, you can link two units together to tackle massive 405-billion-parameter models like Meta's Llama 3.1.

The system runs Nvidia's Linux-based DGX OS, offering native support for PyTorch, Python, Jupyter notebooks, and the NeMo framework. You can use it as a standalone workstation or tether it to your existing Mac or Windows PC.

Why it matters

For independent researchers, data science students, and small startups, Digits fills a massive gap between underpowered local PCs and expensive, data-leaking cloud rentals.

A $3,000 price tag isn't pocket change, but compared to the compounding costs of cloud GPU instances or the headache of building a multi-GPU desktop rig, it is an accessible dedicated workstation. If your workflow involves constant prototyping and local fine-tuning, the math looks pretty compelling.

Strategically, this is classic Nvidia. By teaming up with MediaTek to optimize Arm power efficiency, Nvidia is extending its Grace Blackwell architecture from massive enterprise server rooms directly onto developer desks. Everything built locally on Digits deploys seamlessly to Nvidia-powered clouds, deepening the CUDA lock-in before competitors can catch up.

The main question mark is FP4 precision. Squeezing a 200B model into 128GB of memory relies heavily on aggressive quantization, which can degrade reasoning quality in complex tasks. We will have to see in May whether the real-world accuracy holds up.