NVIDIA RTX Spark: the ARM chip that wants to end the PC as we know it
For years, if you wanted a machine capable of running a serious AI model without a cloud connection, you bought a Mac. Apple's unified memory, the coherence of its ecosystem, the battery life of its M chips: all boxes that Windows simply didn't check.
RTX Spark changes that. And not at the margins.
What RTX Spark actually is
NVIDIA is not releasing a new graphics card. RTX Spark is a complete system chip, co-developed with MediaTek, Microsoft, and ARM: a Blackwell RTX GPU with 6,144 CUDA cores fused with a 20-core Grace CPU, sharing up to 128 GB of unified LPDDR5X memory.
Announced AI compute performance: 1 petaflop in FP4.
For context: a MacBook Pro M5 Pro runs on up to 64 GB of unified memory. RTX Spark announces twice that, inside a 12.9 mm laptop from ASUS.
This is not an incremental improvement. It is a different architecture.
Unified memory: the real story
Unified memory is what made Apple Silicon powerful since 2020. CPU and GPU draw from the same RAM pool, without the transfer latencies of a discrete architecture. The result: AI models run faster, consume less power, and you are no longer bottlenecked by your GPU's VRAM.
Until now, Windows had no credible equivalent.
With 128 GB of unified memory, RTX Spark can run LLMs with 120 billion parameters locally, with no cloud connection, no subscription. On a laptop. That is the kind of capability only fixed workstations allowed 18 months ago.
For an AI developer prototyping, a creative generating assets, or an agency keeping its data in-house: this is concrete.
The ecosystem taking shape
NVIDIA is not launching RTX Spark alone. At Computex 2026, 10 OEM manufacturers announced machines:
- ASUS ProArt P16 / P14: targeting professional creatives, OLED Lumina Pro displays (1,600 nits, Delta E < 1, 100% DCI-P3), large batteries (99.9 Wh on the P16)
- Dell XPS 16: premium business/creative positioning
- HP OmniBook X 14: mobility focus
- Lenovo Yoga Pro 9n: versatile
- Microsoft Surface Laptop Ultra: deep Windows integration
- MSI Prestige N16 Flip AI+: creative/gaming hybrid
Adobe has optimized Photoshop and Premiere Pro for RTX Spark. The ProArt Mini PC (150x150x51 mm) is a direct answer to the Mac Studio. The industry convergence here is real, not cosmetic.
Microsoft breaks its own lock-in
For two years, Microsoft reserved its AI features in Windows 11 to Copilot+ PCs only — machines with a dedicated NPU delivering at least 40 TOPS. Result: only 0.6% of Windows sales in 2024 involved those machines.
On June 11, 2026, Microsoft reversed course.
The Phi Silica model (3.3 billion parameters, text rewriting and summarization) now runs on any GeForce RTX 30 series card or newer, with a minimum of 6 GB VRAM. If you have an RTX 3060 bought during the 2021 shortage, you now have access to Windows 11's local AI features.
This reversal is not altruistic. The arrival of RTX Spark made it untenable to keep an NPU barrier in place when NVIDIA was about to offer a chip delivering 1 petaflop of AI compute. Microsoft anticipated the contradiction.
Recall and Click to Do remain restricted to integrated NPUs, for now.
What we still don't know
Early RTX Spark performance benchmarks were published online, then pulled shortly after. Broken embargo? Results deemed insufficient? Publishing error?
NVIDIA has provided no official scores on its product page. Terms like "all-day battery life" and "most efficient RTX chip" remain unverified claims as of this writing.
The hardware launch is expected in autumn 2026 in select regions. Pricing has not been announced. What we have is a positioning, an architecture, and a promise.
That is not a reason to ignore it. It is a reason not to buy before independent testing is available.
What changes in practice
If RTX Spark delivers on its promises this autumn, three things shift for tech professionals:
For creatives: a Windows laptop becomes comparable to a MacBook Pro for rendering, 4K editing, and local AI generation, with the added benefit of the Windows/Adobe software ecosystem.
For AI developers: prototype with 70B-120B LLMs without a server GPU, without cloud costs, without network latency. Iteration becomes local and immediate.
For teams: sensitive AI workflows (client data, intellectual property) stay on the machine. No data transiting through an external API.
Should Apple be worried?
Yes and no.
Apple has a five-year head start on CPU/GPU/NPU integration and on the software maturity of its ecosystem. macOS optimizes at a level Windows ARM has not yet reached. Real-world MacBook battery life remains a reference.
But unified memory, long the Mac's trump card, is no longer a monopoly. And Windows retains real advantages: broad software compatibility, hardware flexibility, the professional market.
RTX Spark does not kill the Mac. It makes the choice genuinely difficult for the first time since 2020. That alone is a win for NVIDIA.
References
- Asus ProArt with RTX Spark: a new era for Windows PCs : Computex 2026 coverage, ProArt P16/P14/Mini PC specs, L'Éclaireur Fnac
- From RTX Spark to RTX 3060: Microsoft opens Windows 11 local AI to NVIDIA GPUs : analysis of Microsoft's Copilot+ reversal, Clubic, June 11 2026
- NVIDIA RTX Spark — official product page : official specs, OEM partners, marketing positioning
- NVIDIA RTX Spark: early performance benchmarks have mysteriously disappeared : investigation into pulled benchmarks, Les Numériques