Nvidia's RTX Spark is the AI PC Microsoft promised, at $2,600 and up

Nvidia's RTX Spark is here, and it does the thing Copilot+ PCs spent two years promising. The new lineup of laptops and desktops is built to run serious AI models on the machine itself, not in a data center somewhere. According to Yahoo Tech's review of a 64GB, 20-core model, the hardware doesn't just talk a good game — it delivers.

Microsoft kicked off the AI PC era in 2024 by requiring every Copilot+ machine to carry a neural processing unit capable of 40 trillion operations per second. That bar looks silly in hindsight. A mobile RTX 4050 pushes nearly 200 TOPS, roughly 400% past the cutoff, and plenty of AMD and Intel GPUs from the same era cleared it too. Millions of PCs got locked out of flagship Windows AI features over an arbitrary spec, while the genuinely impressive stuff stayed in the cloud — making an AI PC about as special as any laptop with a ChatGPT tab open.

Recall's rocky launch and Windows' built-in image editing tools didn't help. Local models back then were memory hogs that hallucinated freely, so Microsoft pushed the pitch hard and early, and it soured a lot of people. Nvidia's answer is brute force: the top RTX Spark superchip pairs a 20-core CPU with a Blackwell GPU packing 6,144 CUDA cores and up to 128GB of unified memory. A smaller version drops to an 18-core CPU, 5,120 CUDA cores and up to 32GB of memory.

The flagship silicon borrows its design from Nvidia's DGX Spark personal AI computers, and its memory runs about 10% faster, according to the review. In testing, the RTX Spark beat a DGX Spark by roughly 10% on LLM output speed and handled models far bigger than a 16GB 5070 Ti could touch. On the 64GB unit, a 30B parameter model at 4-bit quantization fit in under 30GB with context, with room left for speech-to-text, image description, editing, OCR and a hobbyist plain-language-to-PowerShell translator. The whole stack stayed under 40GB, leaving 24GB for Windows and everything else — no cloud required.

That headroom matters. The 64GB and 128GB variants can run a smart 30B model as an orchestrator for smaller helpers, or load quantized 100B models that sometimes hang with cloud frontier systems. They're also big enough to fine-tune small models, which was basically off the table before. The catch is the app ecosystem, which the review calls immature and clunky — the best tools are enthusiast-grade and fiddly to set up.

Then there's the bill. RTX Spark devices start around $2,600, and plenty cross $4,000. The AI PC is finally real, and it's priced like a workstation.