NVIDIA RTX Spark: the ARM superchip that wants to replace your desktop PC

Jensen Huang dropped the bomb at GTC Taipei, end of May 2026. RTX Spark , that's the name. And no, it's not just another GPU in the catalog. It's a complete platform — ARM CPU + Blackwell GPU + unified memory — all in one chip. The kind of thing Apple has been doing since the M1, except this time it's NVIDIA getting in on it. Under Windows.

Concretely, NVIDIA wants us to stop thinking "graphics card" when we hear their name. The RTX Spark is a complete SoC intended for thin laptops and mini desktop PCs. And the pitch is quite clear: it's supposed to be the Mac Silicon killer on Windows. Just that.

So, is it really the revolution announced or just well-crafted Computex marketing? We will dissect all of this.

The heart of the RTX Spark is a CPU-GPU duo connected by NVLink-C2C (the same chip-to-chip interconnect as on NVIDIA servers). On the CPU side, we have a Thanks to 20 ARM cores — based on the Neoverse V2 architecture, co-developed with MediaTek. On the GPU side, it's Blackwell with 6,144 CUDA cores and 5th generation Tensor Cores in FP4.

Memory, that's where it gets interesting. Up to 128 GB of unified LPDDR5X — shared between CPU and GPU, as on Apple chips. Announced bandwidth: around 300 GB/s. It's half as much as Apple's M5 Max (614 GB/s), but well, we're not exactly comparing the same price ranges.

NVIDIA announces 1 petaflop of AI performance in FP32 . It's a marketing figure — no one runs their models in pure FP4 — but it gives an idea of the raw power of the onboard Tensor Cores. For local LLM, NVIDIA talks about models of 120 billion parameters with 1 million context tokens. Directly on the laptop.

No official TDP communicated at the moment. NVIDIA insists on energy efficiency ("all-day battery life" on laptops), but without specific figures.

It is NVIDIA's boldest promise. From AAA game in 1440p at over 100 FPS , with ray tracing, DLSS, and Reflex. On an ARM laptop. If it's true, it's a game changer. Because so far, Windows on ARM and gaming have been somewhat synonymous with a hassle.

The thing is, NVIDIA includes its entire graphics stack. Native CUDA, full RTX, DLSS 4 (and soon 4.5 Ray Reconstruction in August), Reflex, G-SYNC. It's not a castrated mobile GPU — it's Blackwell, the same silicon as in desktop cards, adapted into SoC format.

Microsoft, for its part, is working on Prism - their x86 emulation layer - to improve compatibility with existing games. But let's be honest: x86 emulation on ARM is still the weak point. Anti-cheats breaking, games refusing to launch... We'll see in the fall if NVIDIA and Microsoft have really solved the problem or if it's wishful thinking.

More than 100 software publishers are announced as partners: Adobe, Blackmagic Design, Blender, CapCut, KRAFTON, Riot Games, XBOX... Adobe will even rewrite Photoshop and Premiere from scratch for RTX Spark, with a 2x gain in AI and graphics performance announced.

This is NVIDIA's true positioning with the RTX Spark — and frankly, it may be more important than gaming. Jensen Huang made it clear: "The PC goes from being a tool to a teammate." The idea is that your PC is no longer something you launch apps on, but an assistant that works for you.

128 GB of unified memory + 1 PFLOP of AI compute, it allows to run heavy LLMs directly on the machine. No need to send everything to the cloud anymore. NVIDIA talks about models with 120B parameters locally, with TensorRT for optimized inference.

On the security side, the partnership with Microsoft introduces new Windows primitives — identity, containment, policy — so that the agents run in a sandbox on your machine. NVIDIA adds OpenShell , a runtime that controls what agents can and cannot do, and that can mask your personal data when an agent sends requests to the cloud.

Projects like OpenClaw and Hermes Agent are already developing native Windows apps for RTX Spark. The goal: agents that can perform tasks in your applications, reason about cross-app workflows, generate images and videos, code plugins... all of this locally.

NVIDIA does not make PCs themselves (well, except for the DGX Spark for professionals). For the general public, it is the usual manufacturers who take care of it. ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI are confirmed for launch this fall, with Acer and GIGABYTE to follow. Over 30 laptop models are already in development.

Two formats announced:

• Laptops 14 to 16 inches — fins, with long-lasting battery and premium screens. The direct competitor to the MacBook Pro.
• Compact mini desktops — ultra-efficient, designed to run AI agents 24/7 on your desk. The Mac Mini format, NVIDIA version.

No official price at the moment. Given the positioning (128 GB, Blackwell, premium...), don't expect €999. We would rather bet on €2,000-€3,000 for entry-level laptops, and potentially more for maxed-out configurations. But all of this is still speculation until the manufacturers' announcements.

Let's not lie to ourselves. The biggest risk of the RTX Spark is not the hardware — it's the software. Windows on ARM, in 2026, is better than before. But it's still far from perfect.

Qualcomm paved the way with the Snapdragon X Elite. Result: machines that perform well for office work and multimedia, but struggle as soon as you step off the beaten path. x86 apps go through Prism (Microsoft's emulator), and it works... most of the time. Some professional software still refuses to run. Online game anti-cheats are often incompatible. Third-party drivers remain a black spot.

NVIDIA has a significant advantage: CUDA runs natively . This means that the entire AI, creative, and 3D ecosystem that depends on CUDA (and it's huge) will work without emulation. For gaming, NVIDIA drivers are also native. This is a real advantage compared to Qualcomm, which has a decent Adreno GPU but not in the same galaxy.

But the CPU remains ARM. And as long as the catalog of native ARM Windows apps does not expand massively, there will always be an emulation layer between you and some software. It's not prohibitive, but you need to know.

Be careful not to confuse the two. The DGX Spark This is the pro/workstation model sold directly by NVIDIA — aimed at AI developers and data scientists. RTX Spark it is the consumer platform declined by manufacturers into laptops and mini-PCs.

Silicon is identical (Grace + Blackwell). The difference is the packaging: the DGX Spark is a standalone NVIDIA mini-PC with fixed specs (128 GB, pro connectors), while the RTX Spark will be available in several configurations by ASUS, Dell, and company — probably with options for 32, 64, or 128 GB of RAM.

If you plan to buy a new laptop or mini-PC by the end of 2026, the RTX Spark clearly deserves to be on your shortlist. The efficient ARM CPU + complete Blackwell GPU + unified memory up to 128 GB combination is unprecedented under Windows.

For creatives and AI developers, this is potentially huge. Native CUDA on a thin laptop with all-day battery life? If the promises hold up, it changes the game.

For gamers, it's more nuanced. The hardware looks solid, but Windows on ARM compatibility remains a big question mark. We recommend waiting for the first independent benchmarks before pulling out the credit card.

And the price? Without an official price, it's difficult to judge. But one thing is certain: if NVIDIA really wants to compete with Apple Silicon, the RTX Spark machines will have to be competitive in price. A laptop at €4,000 that "almost" performs as well as a MacBook Pro at €3,000 won't cut it.

See you in the fall of 2026 for the final verdict.

It is a superchip (SoC) that combines an ARM CPU with 20 cores and a Blackwell GPU with 6,144 CUDA cores in a single chip. It is intended for thin Windows laptops and desktop mini-PCs, with up to 128 GB of unified LPDDR5X memory.

NVIDIA has announced availability in the fall of 2026. Laptops and mini-PCs will be manufactured by ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, with other manufacturers to follow.

NVIDIA promises AAA gaming at 1440p at over 100 FPS with ray tracing and DLSS. The onboard Blackwell GPU is powerful, but the compatibility of Windows on ARM with existing games remains to be proven in practice.

No official price has been communicated. Given the premium positioning of the platform, laptops starting around 2,000-3,000€ are expected. The 128GB configs will probably be much more expensive.

This is Windows on ARM, so native ARM apps run perfectly. Classic x86 apps go through Microsoft's Prism emulator, which works well in most cases but can cause issues with certain professional software and game anti-cheat systems.