Nvidia Spark: the new promise of the local AI PC (again)

The important word: local

Nvidia obviously emphasizes the numbers. In hardware, spec sheets are part of the show.

Today, the best AI models mostly run in the cloud. It is practical, powerful, and simple to use. This dependency also raises several questions: where does the data go? How much do subscriptions cost? What happens if access is limited? Can you really personalize a model or an agent for your own workflows? Can you run long tasks without relying on an external platform?

Spark shows the direction Nvidia wants to push the market toward: machines capable of running heavier models locally, with enough memory not to hit a wall as soon as the use case becomes serious.

Unified memory then becomes central. On a traditional PC, we often think in terms of CPU, GPU, RAM, and VRAM. For local AI, the problem is simple: large models take up a huge amount of space. A fast chip is not enough if the memory does not follow.

With up to 128 GB of unified memory in its maximum configuration, Spark is precisely trying to avoid that ceiling. The goal is both to host the models and to run them properly.

The trap: confusing promise with proof

At this stage, RTX Spark is still largely a promise. The machines will arrive later, the exact configurations will depend on manufacturers, prices still need to be watched, and solid independent benchmarks on RTX Spark PCs themselves are still missing.

Nvidia is announcing a new generation of PCs. But a manufacturer announcement does not replace an independent test. ComputerBase notably points out that Nvidia has not yet provided precise performance figures, especially for CPU benchmarks, and that RTX Spark still needs to face real-world testing.

For now, RTX Spark mainly represents a strategic direction. Nvidia wants to establish the idea that the future of the PC will involve machines capable of running AI agents locally. The idea seems credible, but everyday use cases will still need to validate it.

What the first DGX Spark tests tell us anyway

Even though RTX Spark has not yet been widely tested, independent benchmarks already exist for DGX Spark and the GB10 platform. They provide some useful clues.

DGX Spark is more aimed at AI developers. The product does not have exactly the same positioning as future RTX Spark PCs, but it relies on a similar logic: a lot of unified memory, a recent Nvidia architecture, and strong software integration around local AI.

The first tests show that Spark’s appeal lies mainly in its ability to run large models locally thanks to its massive unified memory. Comparing it with a large traditional graphics card is not enough to understand the product. [Sources: Tom’s Hardware / ServeTheHome]

ServeTheHome, for example, measured around 49 tokens per second on gpt-oss 20B with Ollama, 14.48 tokens per second on gpt-oss 120B, and around 9 to 10 tokens per second on Qwen3 32B out of the box. These figures show that this type of machine can run models locally that many traditional PCs cannot handle comfortably.

Tom’s Hardware reaches a fairly similar reading: DGX Spark is an interesting machine for local AI thanks to GB10, 128 GB of RAM, and the CUDA ecosystem. The outlet also notes that the platform becomes hard to justify if its capabilities are not actually used.