October 16, 2026 — NVIDIA did not launch another graphics card this week. It started selling a different kind of Windows PC. RTX Spark laptops are now available from ASUS, Dell, HP, Lenovo, Microsoft and MSI: nine models, six brands, one launch day, starting at US$2,599.
That kind of coordinated rollout is rare in the PC industry. It usually means the silicon is ready, the OEMs have committed, and the software stack is good enough to ship. So what exactly is RTX Spark, and why does it matter beyond another thin-and-light laptop?
1. What RTX Spark actually is
NVIDIA calls it a "superchip." It is not a CPU with a GPU bolted on, and it is not a discrete GPU with a bigger cooler. It is a single package that combines:
- A Blackwell RTX GPU with DLSS 5, ray tracing, OptiX, Reflex and G-SYNC.
- An NVIDIA Grace CPU, co-designed with MediaTek for power efficiency and connectivity.
- Up to 128GB of unified memory, shared by the CPU, GPU and AI accelerators.
- Up to 1 petaflop of FP4 AI compute, enough for local inference on large models.
- Windows 11 on Arm, with native CUDA, a PyTorch-CUDA build, native Python and Conda.
The important part is the memory. On a conventional laptop, the CPU and GPU have separate pools. A large language model has to fit into VRAM, and data shuttles back and forth over PCIe. RTX Spark gives the GPU direct access to a single 128GB pool. That is the same idea behind game consoles and Apple's M-series chips, and it is the reason a 70-billion-parameter model can run on a laptop at all.
2. The launch lineup
| Brand | Models | Positioning |
|---|---|---|
| ASUS | ProArt P14, ProArt P16 | Creator |
| Dell | XPS 16 (Creator Edition) | Premium creator |
| HP | OmniBook Ultra, OmniBook X 14 | Premium thin-and-light |
| Lenovo | Yoga Pro 9n | Flagship thin-and-light |
| Microsoft | Surface Laptop Ultra | Flagship thin-and-light |
| MSI | Prestige N16 Flip AI+, Prestige N16 | Business creator |
US$2,599 is the starting price, not the price of the 128GB configuration. Expect the top model to land well above US$4,000. For comparison, a 128GB mobile workstation with a 24GB GPU used to cost US$3,500–5,000 and still kept CPU and GPU memory separate. RTX Spark is not cheap, but it moves a server-class idea into a portable chassis.
3. Four reasons this is a big deal
- 128GB of unified memory in a Windows laptop. This is not just "more RAM." The GPU can address the whole pool, which removes the VRAM wall that stops most laptops from running large models.
- CUDA runs natively on Windows on Arm. NVIDIA and Microsoft shipped a PyTorch-CUDA build, native Python, a native Conda installer and support for most of the top 15,000 Python packages. Whether every tool behaves is another question, but the foundation is there.
- Local agents get a reference platform. Hermes, OpenClaw, GitHub Copilot, Perplexity, Pi and WorkBuddy all have guided installers. The pitch is that a laptop can run an agent without a cloud subscription.
- Six OEMs at launch. This is not a limited developer kit. The major Windows laptop brands are shipping at the same time, which tells you the supply chain and the drivers are ready.
4. Who should buy — and who should wait
| Buyer | Verdict | Why |
|---|---|---|
| Developers running local LLMs or agents | Strong fit | 128GB unified memory, native CUDA, data stays on the device |
| Video, 3D and design creators | Good fit | 8K 4:2:2 editing, 90GB+ 3D scenes, Adobe TensorRT acceleration |
| Pure gamers | Not the target | A desktop at the same price will deliver more frames; DLSS 5 is a bonus, not a reason |
| Users of x86-only industrial software | Verify first | Arm compatibility with old drivers, dongles and plugins is the biggest unknown |
| General office and students | Overkill | Too much compute and too high a price for the workload |
5. What it means for buyers in Guangzhou
For the companies ACCPC works with in Guangzhou and the Pearl River Delta, RTX Spark will not replace desktop workstations overnight. But it changes three things:
- The bar for a "mobile workstation" goes up. A 32GB laptop with an 8GB GPU used to qualify. Buyers will now ask whether it can run a 70B model.
- There is a third option between buying GPUs and renting cloud capacity. Instead of a shared server, each engineer can have a laptop that runs models locally, with the data staying inside the company.
- Memory prices amplify the top-end SKUs. 128GB of unified memory is scarce in the current DRAM cycle. The flagship configuration may hold its price better than the entry model.
6. The fine print
- 128GB is the maximum, not the base. Entry SKUs will ship with less. Unified memory is soldered, so choose carefully at purchase.
- 1 petaflop is FP4. FP16 and BF16 numbers will be lower. Do not compare it directly with a data-center training card.
- The Arm ecosystem is still catching up. Drivers, antivirus, input methods, printers and industry plugins decide the day-to-day experience. Keep one x86 machine as a fallback during the first generation.
- Mainland China pricing is not announced. As of October 21, NVIDIA's China site has no date or price. Parallel imports are possible, but check warranty coverage and software compatibility first.
Sources
- NVIDIA Newsroom: RTX Spark PCs Available Today (16 Oct 2026)
- NVIDIA Taiwan blog: RTX Spark PC launch (16 Oct 2026)