Adel el Hallak has a DGX Sparksitting in his house. It runs AI jobs overnight, and if he forgets to give it something to do before bed, he feels bad about it.
“At home I feel guilty if I don’t kick off my job for it to work on all night,” el Hallak, Nvidia’s VP of Product, tells Tom’s Guide. “It doesn’t make noise. And I feel guilty if I don’t do that for something on a personal front — because it’s plugged in anyway.”
It’s a small, endearing detail from a wide-ranging interview — but it also points at where AI is headed, particularly local AI. Because when the person in charge of Nvidia’s AI product strategy is already living with a personal AI computer in his home, treating it more like an appliance than a server, it could mean a lot of us will be doing the same thing in just a few years.
The case for ditching the cloud
Right now, most people interact with AI through a browser or app. That means, when you prompt ChatGPT or Perplexity or any other AI assistant, your data travels to a data center, and a response comes back. It works, but there’s a cost — both financially and in terms of privacy. Every question you ask, every document you upload and every personal detail you share lives on someone else’s infrastructure.
El Hallak thinks local AI changes that equation. “I could run a good enough model on my system at home,” he said. Essentially, if the model runs on hardware you own, your data never leaves your house, and there’s no monthly subscription.
The math is worth thinking about. Nvidia’s DGX Sparkcurrently retails for $4,699 — not cheap by any measure. But aChatGPT Plus subscription runs $20 per month, or $240 per year. Over five years, that’s $1,200 — and you still don’t own anything. The Spark, meanwhile, runs unlimited queries with no per-token billing, no usage caps, and no internet required.
That’s not an apples-to-apples comparison, of course. GPT-6 and Claude are significantly more capable than the open-r of tasks — summarizing documents, drafting emails, searching your own files, running a coding assistant — a local model with 128GB of memory and access to Nvidia’s full CUDA stack is more than enough
It’s not just for developers
The DGX Spark launched last October as what Nvidia called “the world’s smallest AI supercomputer.” At the time, it was squarely aimed at developers, researchers, and data scientists. It runs Linux. It weighs 2.6 pounds. It’s the size of a Mac Mini.
But the ecosystem building around it is starting to look more consumer-friendly. As I wrote last month, Perplexity just launched Portable Computer, which packages a local AI model, agent tools, app connectors, and a sandboxed runtime into a single system that runs on DGX Spark hardware. Every task starts on your machine. If the local model hits something it can’t handle, it asks your permission before sending that one step to the cloud. Your files stay put.
El Hallak sees this as a trend, not an outlier. He pointed to harness providers — the companies building the software layer between a model and a user — making setup dramatically easier. “You’re not doing installst easier — like the old days of using Windows and setting up an application. They’re bringing it back to a GUI.”
If that sounds familiar, it should. It’s the same arc that turned network-attached storage from a hobbyist pursuit into something Synology sells at Best Buy. The hardware existed for years before the software made it approachable.
The RTX Spark laptops change the picture
The DGX Spark desktop is the proof of concept. But the real consumer play may be what’s coming next: RTX Spark laptops, which pack the same GB10 siliconinto notebook form and are expected to ship this fall from Dell, HP, Lenovo, Asus, MSI, Acer, and others. Microsoft is building a Surface Laptop Ultra around the chip.
These aren’t AI-only machines — they run Windows 11, play games, and handle conventional laptop workloads. But their 128GB of unified memory means they can also run large language models that would choke any current consumer GPU. The laptop version of what el Hallak has sitting on his desk at home, but without the Linux-only limitation.
“The fact that you will have the model running locally with a harness and a runtime that are made easier to install — I think it’s going to really usher in local AI,” el Hallak said.
The real argument isn’t performance — it’s ownership
El Hallak told a story from the enterprise world that translates perfectly to consumers: “There was a point where people were proud of the leaderboards and how many tokens they’re using. Then somebody in their IT department got their bills, and all of a sudden I became a cost optimizer.”
Anyone who’s watched their Claude usage creep up — or hit the limits of a free tier — knows the feeling. And as AI agents get more capable and start touching more personal data (your calendar, your email, your finances), the question of where that processing happens isn’t theoretical anymore.
“Local AI is just going to bring more tokens, and tokens equate to intelligence,” el Hallak said. “When you pair intelligence with a harness, you’re able to do things, and then the runtime secures it.”
Or, as he put it more casually: “Data centers in the house.”
- AI is turning ordinary photos of kids into sexual abuse imagery — here’s how to protect your family
- OpenAI paid contractors to read ChatGPT conversations — here’s how to protect yourself
- I stopped reading the fine print myself after discovering this overlooked ChatGPT trick
