NVIDIA DGX Spark Gold 2025

The NVIDIA DGX Spark is a compact AI workstation built on the GB10 Grace Blackwell Superchip with 128GB of unified memory and up to 1 petaFLOP of FP4 performance. It can run and fine-tune AI models up to 200 billion parameters locally, making it a compelling alternative to cloud GPU rental for developers. It's expensive and specialized, but for AI work, there's nothing else like it in this form factor.

Let's talk numbers. The 128GB of unified memory puts this in the 99th percentile for RAM among mini PCs, which is frankly ridiculous. The 4TB SSD is 98th percentile. The 20-core ARM CPU lands at 97th percentile. These are top-of-the-charts specs for this form factor. The GPU sits at 78th percentile, which sounds less impressive until you remember this is a Blackwell architecture chip with 128GB of VRAM effectively shared with system memory. That's not a gaming GPU. It's a compute monster.

In practice, what does 1 petaFLOP of FP4 performance get you? NVIDIA says you can prototype, fine-tune, and run inference on models up to 200 billion parameters locally. For context, that covers a huge range of open-source LLMs. You're not running GPT-4 class models on this thing, but you can absolutely work with Llama 3 70B, fine-tune smaller models, and do serious development work without touching a cloud instance. The 10GbE networking also means you can offload larger jobs to a cluster when you need to. The port selection is middle of the pack at 57th percentile, but you get Thunderbolt, DisplayPort, four USB-C ports, and HDMI 2.1, which covers most setups.

The DGX Spark doesn't really have direct competitors in the mini PC space. The HP OMEN GT22-3080 and Lenovo Legion 90Y6003JUS are gaming desktops with discrete GPUs. They'll crush the DGX Spark in gaming benchmarks, but they can't run a 200B parameter model locally. The ASUS Republic of Gamers GM700TZ-BS978 is in the same boat. These are different tools for different jobs. The Apple Mac Studio M4 Max is the closest philosophical competitor. It's a compact, powerful workstation aimed at creators and developers. The Mac Studio has better software compatibility for general use and a more mature ecosystem, but its unified memory tops out at 128GB and it doesn't have the same AI-specific hardware acceleration. The MSI MEG Vision X AI is interesting because it's also AI-focused, but it's a full desktop tower, not a mini PC. If you want something that fits on a desk and does serious AI work, the DGX Spark is basically alone in its class right now.

Here's the thing about value with the DGX Spark. If you're comparing it to other mini PCs, it looks absurdly expensive. But you're not really buying a mini PC. You're buying a local AI development workstation that replaces cloud GPU rental costs. A single A100 instance on AWS can run you $3-4 per hour. If you're doing serious AI work, this thing pays for itself in a few months. The price spread across vendors is wild though. We're seeing $4,680 on the low end and $10,500 on the high end. Best Buy currently has the best deal we've found, so if you're buying, start there. Compared to something like the Apple Mac Studio M4 Max, which is also aimed at creators and developers, the DGX Spark is more specialized. The Mac Studio is a better all-around machine, but it can't touch the DGX Spark for local LLM work.

The NVIDIA DGX Spark is not your typical mini PC. This is a machine built for one thing: running AI models locally. If you've been searching for a compact AI workstation that can handle large language models without renting cloud GPUs, this is the kind of hardware that shows up on your radar. The headline spec is the GB10 Grace Blackwell Superchip, which pairs a 20-core ARM CPU with NVIDIA's latest Blackwell GPU architecture and a whopping 128GB of unified LPDDR5x memory. That's the kind of memory capacity you'd normally find in a server, not something that weighs 1.2kg and sits on your desk.

NVIDIA claims up to 1 petaFLOP of FP4 AI performance, which is a number that would have been absurd for a desktop machine just a few years ago. The 4TB NVMe SSD gives you plenty of room for model weights, datasets, and whatever else you're cooking up. And the ConnectX-7 Smart NIC with 10GbE means you can move data around fast, whether you're pulling from a NAS or pushing models to a cluster. The whole thing runs DGX OS, NVIDIA's Linux-based operating system tuned for AI workloads.

Pricing is all over the place right now. We're seeing this listed anywhere from $4,680 to $10,500 depending on the vendor, which is a massive spread. That's the kind of price variance you see with new, supply-constrained hardware. If you're serious about buying one, shop around. A lot.

Skip the DGX Spark if you're not doing serious AI or machine learning work. Gamers will be disappointed by the ARM CPU and AI-focused GPU. General content creators should look at the Apple Mac Studio M4 Max, which offers better software compatibility and a more mature ecosystem for creative apps. Business users who just need a reliable office machine should buy something far cheaper. And if you're on a tight budget, the $4,680 to $10,500 price range is hard to justify unless AI development is literally your job or your primary hobby. For everyone else, this is a specialized tool that will collect dust.

Should you buy the NVIDIA DGX Spark? If you're an AI developer, ML engineer, or researcher who wants to run large models locally without dealing with cloud costs and latency, yes. This is the most capable compact AI workstation we've seen. The 128GB of unified memory alone puts it in a class by itself. You can prototype, fine-tune, and deploy models up to 200B parameters from your desk, and the 10GbE networking makes it easy to scale to a cluster when you need more horsepower.

If you're a gamer, a general content creator, or someone who just wants a fast desktop, look elsewhere. The ARM CPU means some software won't run, the GPU isn't optimized for gaming, and you'd be paying a premium for AI capabilities you won't use. The Mac Studio M4 Max is a better fit for most creators, and any of the gaming desktops we mentioned will serve gamers better. But for the AI crowd, this thing is a glimpse of the future, and it's available now.