NVIDIA DGX Spark and DGX Station: Desktop supercomputing for AI models with up to 1 trillion parameters

The world of artificial intelligence has just taken a giant leap: now, the power of a data center literally fits next to your desk.

During CES 2026, NVIDIA revealed new supercomputers DGX Spark e DGX Stationdesigned for developers and researchers to run the world’s most advanced AI models locally and efficiently.

The presentation took place at the Fontainebleau Hotel, in Las Vegas, in the United States. Jensen Huang, CEO of NVIDIA, took the stage in front of more than two thousand people present at the event. The Adrenaline team was also present following the conference.

DGX Spark e o DGX Station: the future of data centers

Imagine that, before, to build a building (train a giant AI), you needed to rent a fleet of trucks and cranes from a distant company (cloud).

O DGX Spark e o Station They’re like having your own kit of high-tech magic tools in your garage: they take up little space but have the power of an entire construction site, allowing you to build and test your ideas whenever you want, without having to wait for transport to arrive.

DGX Spark and DGX Station are desktop supercomputers that eliminate exclusive reliance on the cloud for complex model development. They are equipped with the architecture NVIDIA Grace Blackwellwhich offers level performance petaflop and a massive unified memory.

• NVIDIA DGX Spark: focused on developers and data scientists, it allows you to run models of up to 100 billion parameters.
• NVIDIA DGX Station: a more robust solution for companies and cutting-edge laboratories, capable of processing monumental models of up to 1 trillion parameters.

Local performance: why does it matter?

Until recently, large-scale models required huge data center infrastructures. With DGX Spark and Station, NVIDIA brings this capability to the on-premises environment, enabling faster iterations and greater control over intellectual property.

Compression innovation: NVFP4

The Blackwell architecture introduces the data format NVFP4. Simply put, this technology allows compress AI models by up to 70%increasing performance without compromising the model’s intelligence.

For content creators, this translates into speed: the DGX Spark can even be 8 times faster in video generation workloads than a high-end MacBook Pro with M4 Max chip.

Support for the open-source ecosystem

NVIDIA is collaborating heavily with the open source community. Thanks to optimizations with libraries such as llama.cppthere was an average gain of 35% performance when running high-end models.

The DGX Station, equipped with the superchip GB300 Grace Blackwell Ultra e 775GB of coherent memorysupports the heaviest models on the market, such as:

• DeepSeek-V3.2
• Mistral Large 3
• Meta Llama 4 Maverick
• OpenAI 120b 120b

Real-world applications: from robotics to mobility

The power of these machines is not restricted to code alone. They are being used as the “brain” for several innovations:

• Embodied AI Agents: in partnership with the Hugging Facethe DGX Spark powers the Reachy Mini robot, allowing it to see, hear and respond with expressive movements completely locally.
• Urban mobility: the vehicle TRINITY uses DGX Spark to process vision and language models in real time, functioning as a “brain on wheels”.
• Enterprise security: giants like IBM e JetBrains highlight that local processing ensures that source code and sensitive data remain secure and under the full control of the organization.

When will DGX be available?

If you’re ready to take your AI development to the next level, purchasing options have already started to appear:

• DGX Spark: Now available through partners such as Acer, ASUS, Dell, HP, Lenovo and others.
• DGX Station: availability is expected for the second quarter of 2026through selected partners such as Supermicro and Gigabyte.

You can review the NVIDIA conference at CES 2026 in full at this link.