RTX Spark: Nvidia puts an AI SoC on the desktop and laptop

Nvidia brought to IFA 2026, in Berlin, the first real devices running the RTX Spark superchip, three months after unveiling the chip and promising laptops built around it. According to WIRED's report, you can now see assembled hardware: a Lenovo laptop and a batch of mini PCs from Acer and Asus, all sold with the same promise, running agentic AI workflows locally, without relying on the cloud.

For the developer who today needs a professional GPU (or a cloud account) to run models at home, the interesting angle isn't that the laptop is thin: it's the architecture underneath.

What RTX Spark actually is

The technical point that changes the game is the SoC. Instead of an x86 CPU + discrete GPU talking over PCIe, RTX Spark packs everything together: a Grace CPU based on Arm with up to 20 cores, a Blackwell RTX GPU with up to 6,144 cores, and memory, in a single system-on-a-chip. This is conceptually closer to Apple silicon than to a traditional Windows laptop, with the added benefit of unified memory shared between CPU and GPU.

Why does this matter for whoever runs LLMs locally? Because the bottleneck of running a large model is usually VRAM. On a consumer discrete GPU, you're stuck with the 12, 16, or 24 GB soldered onto the card. On a unified-memory SoC, the model can use the system's entire RAM pool. Hence Nvidia and its partners' insistence on the up to 128 GB of memory configuration, which is the kind of headroom that lets you load models that don't fit on a consumer card today.

Based on core count alone, WIRED estimates that RTX Spark's integrated GPU lands between a current laptop RTX 5070 Ti and RTX 5080, with the caveat that Nvidia's own language ("up to" 20 and "up to" 6,144 cores) indicates weaker configurations will exist, without it being known yet how wide the performance range is.

The announced devices

The only laptop WIRED saw up close was the Lenovo Yoga 9n 2-in-1, a 16-inch model with a 2880 x 1800 OLED screen at 120 Hz, 0.69 inches thick (practically tied with the 16-inch MacBook Pro). A detail that might frustrate anyone expecting a beefy AI machine: the Yoga 9n only goes up to 64 GB of RAM. It's the Yoga Pro 9n, a 15-inch model, that climbs to 128 GB, while keeping the rest of the performance.

The list of RTX Spark laptops expected for the northern hemisphere's fall includes, besides the Lenovo models:

• Dell XPS 16
• Asus ProArt P16
• Microsoft Surface Laptop Ultra
• HP OmniBook X 14 (so far the only 14-inch model announced)

On the desktop side, the design is even more straightforward for anyone who wants an "AI box" on the desk. The Acer SFF RTX Spark and Asus's ProArt Mini PC are described as being the size of a Mac Mini, with up to 128 GB of memory and a "petaflop of AI performance", always with the same superchip at the center.

Nvidia isn't alone in this

The bet on unified-memory SoCs for local AI isn't exclusive to Nvidia, and this is where Brazilian developers can compare paths. Lenovo announced at the same IFA the AMD-powered ThinkCentre X Ultra, with the Ryzen Max+ Pro 495 chip and also 128 GB of memory available to AI agents.

That same Ryzen already shows up in other machines aimed at local AI, like the Framework Desktop, and AMD now sells a Developer Kit with 128 GB of RAM. In other words: comparable hardware is already on the market while RTX Spark hasn't actually arrived yet. And there's Apple on the other side, which refreshed the M6 Mac mini and M5 Ultra Mac Studio just days earlier, the same small, efficient desktops that were already being used precisely to run models locally.

| Machine | Chip | Max. Memory | Reported Base Price | |---|---|---|---| | Lenovo Yoga 9n 2-in-1 | RTX Spark (Grace + Blackwell) | 64 GB | not disclosed | | Lenovo Yoga Pro 9n | RTX Spark | 128 GB | not disclosed | | Acer SFF / Asus ProArt Mini | RTX Spark | 128 GB | not disclosed | | Lenovo ThinkCentre X Ultra | AMD Ryzen Max+ Pro 495 | 128 GB | US$ 3,699 (base) | | MacBook Pro (reference) | M5 Pro | 24 GB (US$ 2,349) / 128 GB (US$ 6,139) | US$ 2,349+ |

The gap: price, and it won't be small

The most important piece of information still hasn't been given. No RTX Spark system has had its price announced. WIRED uses AMD's ThinkCentre X Ultra as a gauge: US$ 3,699 already in the base configuration, without the full 128 GB. And it notes that RAM has never been as expensive as it is necessary, which makes betting on anything cheap sound naive.

To put it in perspective: a MacBook Pro with 128 GB costs US$ 6,139 today. It's not hard to imagine where the top RTX Spark configurations will land.

What changes for those building software in Brazil

None of this has an official price or launch date in Brazil, and the bill in reais for any of these machines, with import costs and exchange rates, is going to hurt. So the immediate value here is understanding the industry's direction, not rushing out to buy.

The technical takeaway is clear: the architecture that Nvidia is pushing into the mainstream (Arm + GPU + unified memory in a single SoC) is the same one that already makes a Mac with lots of RAM a decent option for running local LLMs. If you're weighing investing in a machine for local inference instead of paying for a cloud API, the decision axis is no longer "how many GB of VRAM does the card have" but has become "how much unified memory does the system offer".

What remains open is exactly what decides the purchase: final price, battery life (Nvidia hasn't disclosed numbers, although an Arm SoC tends to be more efficient than a discrete GPU) and, above all, the actual performance running the models you use, not the core count on a slide. Until any third-party benchmark on the final hardware, everything remains in the realm of promise.