NVIDIA Brings the Grace Blackwell Superchip to the Desktop

NVIDIA opened the year at CES by unveiling Project DIGITS, a personal AI supercomputer that puts the Grace Blackwell platform — until now reserved for data centers — on the developer's desk for about $3,000. The machine will be available in May through NVIDIA and its top partners, the company announced.

The move upends a logic that had seemed fixed until now: training, fine-tuning or running large AI models meant paying for cloud compute time or owning server infrastructure. With Project DIGITS, NVIDIA is betting that work can start on the desktop and only move to the cloud once it needs to scale.

What Project DIGITS Actually Is

At the heart of the system is the GB10 Grace Blackwell Superchip, a system-on-a-chip (a single piece of silicon integrating several components once kept separate) built on NVIDIA's Grace Blackwell architecture. The company says it delivers up to one petaflop of AI performance at FP4 precision — a quadrillion operations per second, with calculations compressed into a low-precision numerical format that trades fine-grained accuracy for speed and energy efficiency, a common technique in AI inference.

The GB10 pairs a Blackwell GPU with latest-generation CUDA cores and fifth-generation Tensor Cores, linked via NVLink-C2C interconnect to a Grace CPU built with 20 power-efficient Arm-based cores. MediaTek, a specialist in Arm-based SoC design, collaborated on the chip with an eye toward squeezing out every bit of power efficiency.

That detail matters for home use: NVIDIA points out that Project DIGITS runs off a standard electrical outlet. No special installation, no server-rack cooling or power setup required.

Each unit packs 128 GB of unified, coherent memory and up to 4 TB of NVMe storage. With that configuration, the company says, a developer can run language models with up to 200 billion parameters. And there's room to grow: using NVIDIA's ConnectX networking, two Project DIGITS units can be linked to handle models with up to 405 billion parameters.

The Pitch: Bringing Frontier Compute Within Reach

"AI will be mainstream in every application for every industry. With Project DIGITS, the Grace Blackwell Superchip comes to millions of developers," said Jensen Huang, founder and CEO of NVIDIA. "Placing an AI supercomputer on the desks of every data scientist, AI researcher and student empowers them to engage and shape the age of AI."

Beyond the rhetoric, the move has clear industrial logic. NVIDIA is pitching a seamless workflow: prototype and fine-tune a model locally on Project DIGITS — which runs NVIDIA DGX OS, a Linux-based operating system — then deploy it without friction to DGX Cloud, to accelerated instances, or to a data center, all using the same Grace Blackwell architecture and the NVIDIA AI Enterprise software platform.

That "same environment, start to finish" is the real product. By sharing architecture between desktop and cloud, NVIDIA cuts the friction of porting a model from one place to another — one of the recurring headaches in AI development — while also tying developers into its ecosystem at every stage.

A Software Ecosystem Included in the Box

Project DIGITS isn't sold as hardware alone. Buyers get access to NVIDIA's AI software library for experimentation and prototyping: development kits, orchestration tools, frameworks and models available through the NGC catalog and the developer portal.

Among the specific pieces the company highlights: the NeMo framework for fine-tuning models, RAPIDS libraries for accelerating data science, and compatibility with widely used tools such as PyTorch, Python and Jupyter notebooks. For those building agentic AI — applications where models act autonomously, chaining tasks together — NVIDIA Blueprints and NIM microservices are available through the developer program. Once an application moves from experimentation into production, an NVIDIA AI Enterprise license provides enterprise-grade security, support and software releases.

The message: buy the machine, and you're plugged into a full pipeline — from the first experiment to production deployment, all within NVIDIA's tools.

A Critical Read: Power and Lock-In in the Same Package

The $3,000 starting price puts Project DIGITS in interesting territory. It's not a hobbyist's indulgence, but it's nowhere near the six-figure outlay of server infrastructure either. For a university lab, a startup or an independent researcher, being able to run models with hundreds of billions of parameters locally changes the cost calculus: less dependence on cloud bills that grow with every hour of compute.

That's also where the catch lies. The same architecture that makes the workflow so seamless — desktop to cloud without switching environments — reinforces NVIDIA's dominant position across the entire chain. A developer who prototypes on Grace Blackwell will tend to deploy on Grace Blackwell. The hardware is the entry point; the software is what locks in the relationship.

The numbers are also worth reading carefully. The one-petaflop figure is measured at FP4, the most compressed precision available, which inflates the numbers relative to more demanding formats. And running a 200-billion-parameter model — or 405 billion by linking two units — is a feat of inference and fine-tuning, not necessarily of training a model that size from scratch, a task that still belongs to large data centers.

What to Watch Between Now and May

Project DIGITS goes on sale in May through NVIDIA and its top partners, starting at $3,000, with notification sign-ups already open. Between now and then, the open questions are the usual ones for a CES announcement: how real-world performance compares to lab figures, what configurations partners will offer, and whether that starting price holds once memory, storage and accessories are added in.

What's already clear is the direction NVIDIA wants to set for the year ahead: decentralizing the starting point of AI development, moving it from the data center to the desktop — without ever loosening its grip on the ecosystem that work runs through.