NVIDIA RTX Spark Explained: What Local AI on a Windows PC Could Mean for Users

01 / The story

The short answer

RTX Spark, announced May 31, 2026, is a hardware platform: a Grace CPU and Blackwell RTX GPU combined in Windows-on-Arm systems with up to 128GB of unified memory and a claimed peak of up to one petaflop of FP4 AI compute. It is not software that upgrades your existing RTX desktop, and it is not the same product as NVIDIA's separate DGX Spark. That peak figure is a precision-specific marketing throughput number, not a measured model speed. [1]

Laptops from six manufacturers—ASUS, Dell, HP, Lenovo, Microsoft and MSI—went on sale October 16. Compact desktops and additional models follow later, so do not assume every announced form factor is shipping. [2, 3]

02 / The story

Watch NVIDIA's announcement

Video in this section · scroll down to watch

These are NVIDIA's own videos—the manufacturer's presentation of its platform, including its vision and its claims.

03 / The story

What 'local AI' means, and why memory is the whole story

Running AI locally means the model lives in your machine's memory and runs on its chips, instead of sending your prompts to a data center. The appeal: privacy, offline use, no per-token fees, and full control over the model. The catch: AI models are enormous, and memory is the binding constraint.

This is where configuration matters more than the headline. Entry RTX Spark machines start around $2,599 with 24GB of unified memory. An ASUS ProArt P16 with 64GB starts at $4,499.99, and the 128GB/2TB configuration runs $6,999.99 in US launch reporting. The memory pool is shared between the operating system, your apps, the model, and its working context—so 24GB does not mean a 24GB model. Do not shop by the entry price while planning for the maximum capacity. [3]

04 / The story

What independent testing found

Tom's Hardware tested a Surface Laptop Ultra with 20 CPU cores, 6,144 CUDA cores, and 128GB of memory. Running a 27-billion-parameter coding model (Qwen 3.8 27B, quantized) through llama.cpp, the Spark system completed an agent inference benchmark about 25% faster than a 14-inch M5 Max MacBook Pro with the same memory. Important caveats: the two machines used different speculative-decoding methods and different power envelopes, so this is not a controlled chip-versus-chip comparison. [4]

Gaming told the opposite story. In Gears of War: E-Day at High settings with frame generation enabled, the tester measured 77.7 average FPS—but with a 39.9 FPS 1% low, 83.5ms of PC latency, and a reported hang. The review identifies x86 emulation overhead and 300GB/s shared memory bandwidth as constraints. Frame-generation-enabled numbers are not native rendering performance, and the review's conclusion was blunt: local AI impresses, gaming is not its strength. [4]

05 / The story

Software: vendor claims versus reality

Video in this section · scroll down to watch

NVIDIA says the platform offers a Windows-on-Arm PyTorch-CUDA build, over 2,000 Conda packages, and Arm-compatible versions of more than 97% of the top 15,000 Python packages, and names llama.cpp, vLLM, Ollama, Unsloth and LM Studio as supported for local models. Those are vendor compatibility statements. The independent reviewer specifically hit emulation overhead in games, and '97% of top packages' is not a promise that your particular extension, driver, or x86-only tool works. [5, 4]

NVIDIA's advertised gains—4.6x on coding agents, 2.7x on fine-tuning, 1.8x on compilation—are vendor comparisons against a specific Intel-based system, not reviewer findings. [5]

06 / The story

Who should consider one

Our analysis: RTX Spark makes sense for developers building with local models, AI enthusiasts who want to customize and experiment without per-token costs, and CUDA developers who want one machine for coding, testing, and deployment. The 128GB configurations are the interesting ones—and the expensive ones.

It makes less sense as a gaming PC (emulation overhead and shared bandwidth hold it back), as a budget machine (entry configs have entry memory), or as an automatic cloud-AI replacement—cloud models remain smarter and faster for many tasks, as the independent reviewer noted. Buy it for what it verifiably does well today, not for the platform vision.