Computers that can run serious AI without an internet connection are finally becoming a real product, and Nvidia’s RTX Spark superchip is the engine behind the first wave. The first machines arrive this October, and most people should not buy one on day one. Here’s the honest breakdown.
What is the Nvidia RTX Spark?
The RTX Spark is a “system-on-a-chip,” which is a fancy way of saying Nvidia crammed the brain of a computer into a single package. That package combines an Arm-based CPU called Grace (up to 20 cores), a Blackwell GPU (up to 6,144 cores), and the memory all in one place. Nvidia’s own launch announcement pitches it as the beginning of local AI for everyone, no cloud required.
Why does that matter? Because it works more like Apple silicon than a traditional Windows PC. No separate processor, no separate graphics card, no separate RAM slots. Everything talks to everything else at lightning speed, which is exactly what AI models need.
Nvidia built it with MediaTek, the chip designer behind lots of the processors in phones and routers. Nvidia just invested $3.5 billion in that partnership, so this is a long-term bet, not a one-off product.
The most powerful configurations promise up to 1 petaflop of AI performance, which sounds absurd, and honestly it kind of is. That’s “supercomputer territory” for things like generating images, running assistants, and crunching local models. Translation for regular people: these machines can run AI models that used to require a beefy desktop with a huge graphics card, or a paid cloud service.
Why “AI PC” is suddenly everywhere
Here’s the thing: when you use ChatGPT, Claude, or Gemini today, the actual “thinking” happens on servers in a data center somewhere. Your question travels over the internet, gets processed, and the answer comes back. That works fine, but it has three built-in costs.
First, you usually pay a subscription or hit usage limits. Second, you need a connection, and the AI is useless on a plane or in a dead zone. Third, and this is the big one, your data leaves your machine. Emails, documents, financial spreadsheets, all of it gets sent to someone else’s computer.
An AI PC flips that model. The model runs on your laptop, locally, with no subscription and no upload. That’s why Microsoft and Nvidia teamed up on “personal AI” for Windows, and why every major laptop maker showed up at the IFA trade show in Berlin this week with an RTX Spark machine in hand.
There’s also a timing element. RAM prices have gone insane, which makes the whole “how much computer do I actually need” question more painful than usual. If you want the full picture on buying hardware for local AI right now, our local LLM hardware guide breaks it down.
What you can actually do with an RTX Spark PC
Run AI models without the cloud
The headline use case is running local LLMs, the same kind of models that power ChatGPT, right on your own machine. No per-token fees, no rate limits, no waiting for a server that’s overloaded. Nvidia’s own demo at IFA used these chips to run agent workflows entirely on-device.
Privacy without the trade-off
For anyone who handles sensitive stuff, this is the feature that matters. Bank statements, client records, private messages, those never leave the laptop. If the idea of your chat history sitting on a company server makes you uncomfortable, local AI removes that worry completely.
Agents that work while you watch (or don’t)
The newer AI agents, the ones that plan and execute multi-step tasks on their own, work much better when they don’t need to phone home every few seconds. Persistent AI coworkers are becoming a real workflow, and running them locally is faster and cheaper. We wrote a beginner’s guide to persistent AI coworkers if you want the concept explained first.
The first RTX Spark computers
The launch lineup, announced at IFA 2026, spans laptops and mini desktops:
| Device | Type | What stands out |
|---|---|---|
| Lenovo Yoga 9n | 16″ 2-in-1 laptop | OLED 120Hz display, up to 64 GB memory |
| Lenovo Yoga Pro 9n | 15″ laptop | Up to 128 GB, stylus-ready touchpad |
| Dell XPS 16 | 16″ laptop | Flagship design, creator-focused |
| Asus ProArt P16 | 16″ creator laptop | Built for designers and video editors |
| HP OmniBook X 14 | 14″ laptop | The only compact model announced |
| Asus ProArt mini PC | Desktop | Mac Mini-sized, 1 petaflop, up to 128 GB |
| Acer SFF RTX Spark | Mini desktop | Small chassis, agentic AI focus |
Microsoft’s Surface Laptop Ultra and a few others were also teased. Expect more models between now and October, when Nvidia says the first RTX Spark PCs actually ship.
How much will an RTX Spark PC cost?
Here’s the uncomfortable part: official prices don’t exist yet. Wired, which went hands-on with the Lenovo machines, called price “the most crucial missing element” of the whole launch. That’s rarely a good sign.
What we know: pre-orders reportedly sold out at around $3,200 in some markets before launch, and enthusiast outlets estimate starting prices close to $1,800. Neither number is official, so treat both as rumors.
For context, a MacBook Pro with 24 GB of memory runs $2,349, and the 128 GB version goes for about $6,139. RTX Spark machines with 128 GB of unified memory are going to land somewhere in that neighborhood. This is premium hardware with a premium price tag.
If you want local AI today on a tighter budget, the AMD route is already alive. Lenovo’s ThinkCentre X Ultra, powered by AMD’s Ryzen AI Max+, starts at $3,699 with 128 GB. There are also interesting cloud hybrids, like Perplexity’s Hybrid Compute setup, that run parts of the workload locally on the hardware you already own.
Should you buy one? (the honest answer)
Buy an RTX Spark PC this year if you check at least one of these boxes:
- You work with sensitive data and want AI to process it without leaving your machine.
- You’re genuinely into running local models or AI agents and have the patience for early-adopter software.
- You were already planning to spend $2,500+ on a premium laptop this year.
Hold off if you just want a nice computer for email, browsing, and a bit of ChatGPT. A regular laptop does all of that today, for less money, and the cloud AI ecosystem is still way ahead of anything you can run locally. Software support is also an unknown. History says new chip architectures always launch with rough edges.
The takeaway
The first real AI PCs are almost here, and RTX Spark is legitimately cool hardware. But cool hardware and useful hardware aren’t always the same thing in year one. Let the early adopters beta test it, let prices settle, and check back in twelve months. Your wallet will thank you, and by then the software will actually be ready for the rest of us.
