AI supercomputer shrinks to desktop size

A supercomputer that fits under a desk could change how companies handle artificial intelligence.

At CES 2025 last week, NVIDIA announced Project DIGITS, a personal AI supercomputer featuring the company’s new GB10 Grace Blackwell Superchip. The system, which starts at USD 3,000  and launches in May, can run AI models with up to 200 billion parameters and delivers a petaflop of AI computing performance.

"AI will be mainstream in every application for every industry,” Jensen Huang, founder and CEO of NVIDIA, said in a news release. “With Project DIGITS, the Grace Blackwell Superchip comes to millions of developers," "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."

The desktop device packs 128 GB of unified memory and up to 4 TB of storage, targeting companies needing instant data processing without cloud delays. Two systems can link together to handle models of up to 405 billion parameters.

"Think about industrial applications, where on the factory floor you want compute to be right next to where the manufacturing is happening," said IBM’s Shobhit Varshney, a VP at IBM Consulting, on a recent episode of the Mixture of Experts podcast. "There's a lot of latency between calling a server, or a cloud API, and getting responses back."

Security-sensitive operations are driving demand for local processing capabilities. Varshney outlined multiple scenarios where this is essential. "You want to make sure that the data, especially if it is related to something sensitive, doesn't leave your premises," he said. "The same thing goes for defense applications, where you are doing something more tactical in the field. You want to be able to compute all the images coming in from all the drones at the particular place, because you may be in a territory where you don't even have a server connection."

The Project DIGITS system runs on the Linux-based NVIDIA DGX OS, letting users develop locally and then deploy to cloud or data center infrastructure. Manufacturing facilities can implement "real-time monitoring, predictive maintenance and process optimization powered by edge computing," says Ashok Reddy, CEO of the software company KX. Financial firms gain tools for "real-time trading, risk management and predictive insights at unprecedented speed." Meanwhile, healthcare organizations could tap the technology for "accelerated diagnostics, imaging and drug development through advanced AI analytics," Reddy says. The system processes "massive volumes of structured and unstructured data instantly" for "real-time decision-making and low-latency analytics."

University of Pennsylvania engineering professor Benjamin Lee said he sees NVIDIA’s supercomputer as part of a broader shift toward agentic AI, where systems manage entire jobs rather than individual tasks.

"An intelligent AI will then strategize and determine the set of tasks needed to complete that job," Lee says, adding that this approach "could significantly improve productivity by allowing humans to request large jobs from an agent rather than request individual tasks from a model."

Varshney notes that NVIDIA open-sourced technology from its recent USD 700 million Run AI acquisition. "They're trying to ensure that the entire industry moves closer to this physical AI and agentic AI and autonomous driving era,” Varshney says “They want to be the backbone across each one of them.”

Price cuts rippled across product lines. "Last year, the 40 series of their chips used to be USD 1,600," Varshney says. “They just released an equal compute for USD 550.”

The price reductions suggest technical advances. "GPU design is largely defined and is not being aggressively optimized for performance, energy and cost," Lee says. It’s possible, however, that NVIDIA has reached that point "much earlier than other chip designers, allowing it to compete on both performance and cost," Lee says.

"Models could be specialized for specific tasks, computing the same answer with much fewer calculations and much less energy,” Lee says. “Agents could then learn to select the right specialized model for each task, something a human might struggle to do.”

This push toward specialized computing marks a dramatic shift from current limitations. "We've been constrained by some cutesy little small models running on mobile devices," Varshney says. "They want to make compute as easily accessible and democratized as plugging into electricity, but they want to be the electric superpower of the entire world."