What Copilot+ actually certifies
Microsoft's Windows 11 specifications page says Copilot+ features require a neural processing unit (NPU) capable of 40+ trillion operations per second (TOPS) to function as intended. The same page lists a compatible processor or System on a Chip (SoC) with an NPU capable of performing 40+ TOPS as part of the minimum hardware. CNET describes the full definition as an NPU above 40 TOPS, 16GB or more of RAM and at least 256GB of storage. CNET also notes that the memory and storage requirements are fairly basic, which makes the 40-TOPS NPU the real dividing line.
The certification is therefore a statement about one block of silicon. It says nothing about how much memory a model can use, and it doesn't check whether the PC has a GPU capable of hosting a large model. Early reporting also said Microsoft wanted the 40 TOPS met on an NPU rather than a GPU, for efficiency and power savings on laptops. That was a reasonable engineering trade-off for battery life, though it left capable GPUs outside the badge.
Lewis goes further. He says Copilot+ was oversold and that the NPU rule locked out millions of GPU-equipped PCs. Those are his opinions. He also says Windows AI features couldn't easily be pointed at a discrete GPU. Microsoft's developer documentation confirms that NPU-targeted features need software written for the NPU. It does not independently verify his claim about reconfiguring specific Windows features, so treat that one as his experience.
What RTX Spark is
Nvidia announced RTX Spark on May 31, 2026, as a Windows PC platform for local agents, creative work and gaming. Its figures are manufacturer claims:
- Top configuration: a Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores with FP4 precision, linked by NVLink-C2C to a 20-core Grace CPU. Nvidia lists up to 128GB of unified memory and up to 1 petaflop of AI compute.
- Cut-down configuration: Nvidia's specifications page lists an 18-core CPU with a 5,120-core GPU, with the memory ceiling for that part at 32GB, matching Lewis's article. Nvidia's product page lists 64GB for the same part, so the company's own pages disagree. Check the exact model before you buy.
- Design partner: MediaTek collaborated with Nvidia on the custom CPU design.
- Software stack: CUDA, TensorRT, DLSS and related Nvidia technologies. Nvidia says the platform is built to run 120-billion-parameter models with up to a 1-million-token context locally. Treat that as a vendor capability claim, not a guarantee for every model, quantization or context length.
On October 16, 2026, Nvidia said RTX Spark laptops were available globally from ASUS, Dell, HP, Lenovo, Microsoft and MSI. Pricing starts at $2,599, and configurations vary by region. The named models are the ASUS ProArt P14 and P16, Dell XPS 16, HP OmniBook Ultra and OmniBook X 14, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra, and the MSI Prestige N16 Flip AI+ and Prestige N16. Compact desktops are also part of the platform. The launch announcement covers laptops only.
The reviewer's hands-on results
Lewis tested a 64GB, 20-core RTX Spark laptop. These are his observations, not independent benchmarks:
- A 30B-parameter model at 4-bit quantization fit in under 30GB including context.
- Separate models for speech-to-text, image description, editing and OCR ran at the same time.
- An experimental plain-language-to-PowerShell translator built on Qwen Coder also ran in the background.
- The whole stack used under 40GB, leaving roughly 24GB for Windows and other apps, with no cloud involved.
- He reports about 10% more tokens per second than a DGX Spark he tested against. His text gives too little methodology to generalize that figure.
The practical point is that unified memory lets several models run side by side. A 16GB GPU, like the 5070 Ti he mentions, can't do that. He says the best tools are still enthusiast-grade and clunky to set up. Nvidia says guided installers for agents such as GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy remove the need for a command line. Both statements can be true at once: setup is easier than it was, but the ecosystem is young.
Why the TOPS comparison needs care
Lewis puts a mobile RTX 4050 at "just shy of 200 TOPS", which he calls nearly 400% above the Copilot+ minimum. That comparison is shaky. TOPS figures depend on the processor type and numeric precision, and GPU and NPU numbers are not interchangeable as a universal measure of speed. Microsoft's own documentation also shows that NPU, GPU and CPU inference are all possible through ONNX Runtime. The better case for RTX Spark is memory capacity, shared CPU and GPU memory, and access to the CUDA ecosystem. It does not rest on a bigger TOPS number.
It also doesn't follow that ordinary Copilot+ PCs can't run useful local AI. They are designed for efficient, always-on tasks such as search, captioning and effects, which suit a low-power NPU. RTX Spark is aimed at heavier workloads: large models, agents, 3D and video. One platform optimizes for battery life and the other for capacity.
Security and agents
Nvidia says its work with Microsoft includes Windows primitives for identity, containment and policy. It also covers NVIDIA OpenShell, which can route queries to local models according to privacy policy and disguise personal information sent to cloud models. The October announcement adds Microsoft Execution Containers and OpenShell as tools for controlling agent permissions. Microsoft's Yusuf Mehdi described the goal as hybrid intelligence, where AI runs locally when it makes sense and reaches the cloud when needed. That is an announced capability. Admins should still review each agent's permissions and not assume safe defaults.
Cost and who should care
Nvidia confirms the $2,599 starting price. Lewis says many devices exceed $4,000, but Nvidia's announcement does not corroborate that. Check specific models against retailer listings. The expensive configurations are the ones with the memory that makes the platform interesting.
- Likely fit: Developers, data analysts and creators who want private local models and CUDA on a Windows laptop, and who can use 64GB or more of memory.
- Probably not needed: People who want cloud-backed Copilot, Windows Studio Effects or Click to Do on a modest machine. A Copilot+ PC covers that for much less.
- For IT teams: Test your agent and security tooling on Windows on Arm first. Nvidia says it has readied PyTorch-CUDA for Windows on Arm and native Python and Conda builds, but app compatibility varies.
Bottom line
RTX Spark is a credible answer to the question Copilot+ left open: what hardware do you need to run serious models on a Windows laptop? Its strength is memory capacity and the CUDA stack, backed so far by vendor claims and one reviewer's testing. Whether anyone can afford it, as Lewis asks, depends on a price tag that starts at $2,599.
Correction (November 5, 2026): An earlier version of this article said Nvidia's specifications page lists the cut-down RTX Spark configuration at 64GB of memory. The specifications page lists it at up to 32GB, matching the reviewer's figure; the 64GB figure appears on Nvidia's product page. The text has been corrected.