NVIDIA DGX Spark: local AI in your company without depending on the cloud

Until now, “using AI” in a company has almost always meant the same thing: subscribing to a paid API, sending your data to a server in the United States and paying for every query. It works, but it leaves three questions unanswered: where does your data end up?, how much will it cost you when you really use it?, and what happens if they raise the price tomorrow?

There is another path that makes more and more sense for many SMEs: running the models on your own infrastructure, and you no longer need a server room. NVIDIA has launched the DGX Spark, a device that packs an AI supercomputer into something the size of a book. We’ll tell you what it is, what it can and can’t do, and when it’s worth it for a business.

What the NVIDIA DGX Spark is

The DGX Spark is what NVIDIA unveiled in early 2025 under the codename Project DIGITS: a desktop mini-computer for working with AI locally without depending on the cloud. Inside it carries the GB10 Grace Blackwell chip, which combines in a single piece a 20-core ARM CPU and a GPU with Blackwell architecture (the same family used by AI data centers). These are the figures that matter, without the marketing:

• Up to 1 petaFLOP of AI compute (in FP4 precision), the unit by which the power of these systems is measured.
• 128 GB of unified memory shared by the CPU and GPU. This is the key figure, and you’ll see why in a moment.
• 273 GB/s of memory bandwidth and NVIDIA ConnectX networking to join two units for larger tasks.
• An indicative price of between 3,000 and 4,000 dollars depending on version and storage.

Why do we keep stressing the 128 GB? Because a model needs to “fit” in memory to run smoothly. A good gaming graphics card has 16 or 24 GB and only runs small models; with 128 GB, the DGX Spark runs models of up to 200 billion parameters, and joining two units it reaches 405 billion. High-end open source models, not toys.

What it’s for (and what it isn’t)

Let’s be honest: the DGX Spark is not meant to serve millions of users at once or to train a model from scratch; the big data centers are still there for that. Its turf is another one, and it’s exactly where an SME fits:

• Running open source models locally for your team and your internal processes.
• Prototyping and fine-tuning a model with your own data.
• Building AI agents that work over your documents, your CRM or your ERP.
• Testing before scaling: validate a use case on your desk and, if it works, take it to more capacity.

It’s the ideal piece to start having your own AI with a bounded investment, instead of an API bill that grows every month.

The angle that really matters to an SME

Beyond the specs, the question is: what does your business gain by having AI in-house?

Data privacy and GDPR

When you use a cloud API, your queries (contracts, customer records, invoices) leave your company toward a third-party server, often outside the EU. With a local machine, your data doesn’t leave your office. For sectors that handle sensitive information (law firms, consultancies, clinics, HR), this greatly simplifies GDPR: there are no international transfers to justify nor doubts about who is training on your data.

Fixed cost versus pay-per-use

An API is paid by consumption: the more you use it, the more you pay. Fine to start with, but if you turn AI into part of your everyday work, the bill skyrockets just when the project succeeds. A local machine is an initial investment with a fixed, predictable cost: above a certain volume it pays off, and you control the cost instead of depending on a provider’s rates.

Latency and availability

A local model responds without going through the internet: less latency and independence from provider outages or limits. If your process needs to respond fast and reliably, having the compute right next to you helps.

So, local or cloud?

There’s no single answer: the two worlds coexist and the right question is what you put in each place.

A local server makes sense when you handle sensitive or regulated data, you have a high and constant usage volume where the per-API cost adds up, and a “good enough” open source model is sufficient for your real tasks.

The cloud still pays off when your usage is sporadic or exploratory, you need the most powerful model on the market for a specific very hard task, or your demand peaks are highly variable and you’d rather not manage hardware.

In practice, many SMEs end up with a mixed setup: a local machine for the recurring and the sensitive, and cloud APIs for the occasional or the very demanding. We dig deeper into how this split will evolve in the future of local AI on rented servers.

How we set it up at DominaInternet

A machine like the DGX Spark is just the engine. What turns it into something useful is what you put on top of it, and that’s where we come in: we don’t sell you a box, we leave the AI running over your real processes.

• We choose the right open source model. Families like Llama, Mistral, Qwen or DeepSeek have come a long way and, for most internal tasks, a well-tuned open model is more than enough. You don’t need the most expensive one on the planet, but the one that does your job well.
• We connect it to your data. We build RAG systems and agents that answer using your documentation, your history and your CRM: assistants that read documents, write, sort emails or extract data from invoices.
• We integrate it with your tools. We link the AI to your management software and your automations so it isn’t an isolated chat, but one more piece of your operations, with maintenance and without black boxes.

You can see more about our approach on the AI infrastructure page.

Conclusion

The DGX Spark is a clear sign of where the sector is heading: powerful AI is leaving the data center and arriving on the desks of companies. For an SME this opens up an option unthinkable two years ago without a huge budget: your own AI, with your data kept safe and a cost you control. It’s not the solution for everything, but for more and more businesses setting up AI locally is a sensible decision, not a tech whim.

Does it fit you? At DominaInternet we analyze your case honestly, we tell you whether local, cloud or a mix pays off for you, and we prepare a quote with no commitment. Tell us what you have in mind from contact and we’ll look at it together.