Microsoft is betting that the Surface Laptop Ultra can combine development, AI, and gaming workloads in a single device, relying on the NVIDIA RTX Spark N1X platform based on a Grace processor, Blackwell graphics, and unified memory. Ars Technica’s review shows that the bet clearly succeeds at the hardware level, but runs into software that has not yet caught up with the platform.
The device starts at $2,599.99, while the price reaches $5,899 depending on the processor and memory. The base model comes with 24 GB of LPDDR5x unified memory and Windows 11 Home, while higher-end models use Windows 11 Pro. The tested model had 64 GB of memory and 1 TB of SSD storage, and cost $4,299.99.
Hardware That Combines Gaming and AI
The device features a 15-inch touch mini-LED display with a resolution of 3270×2180 and a refresh rate of up to 120 Hz, with peak brightness of 2,000 nits. The review praised the color quality, black levels, and viewing angles, while noting that the glossy surface can cause reflections in direct lighting.
More important is the RTX Spark N1X platform, which does not rely on a dedicated NPU for AI, but instead uses unified memory and the Blackwell graphics unit. The device ran the 70-billion-parameter Llama 3.3 model through Ollama at 6.26 tokens per second, even though the model requires about 43 GB of memory. The reviewer considers running a model of this size on a laptop a notable achievement, with the fans reaching nearly their maximum speed during testing.
For image generation, Stable Diffusion v1.4 produced an image in approximately three to four seconds when using the graphical NvTensorRT path. The device also showed strong gaming performance: Gears of War: E-Day ran on the ARM architecture with DLSS and ray tracing support, although the first launch required four or five minutes to build shader files.
What Limits the Experience in Practice?
The tests reveal that the main problem is not the hardware’s capability, but the software’s ability to take advantage of it. Most of the conversational models available in AI Dev Gallery run solely on the CPU, while GPU and NPU paths remain limited to some image-generation models and embedded models. The application also froze and then crashed twice during testing.
It was also impossible to run Phi Silica directly without registering as a developer and creating an app identity, and the CUDA path recommended by Microsoft failed after a problem appeared with the onnxruntime-gpu package compatible with Windows ARM64. The test does not determine whether the cause is a shortcoming in the packages, the software environment, or the settings, but it demonstrates that the local AI experience is not ready with the simplicity suggested by the device’s specifications.
The same pattern appeared in the HandBrake test: video decoding through Media Foundation succeeded, but NVENC and NVIDIA HEVC options were unavailable for encoding, leaving the CPU to handle the task.
Compatibility, Power, and Deferred Features
Microsoft says that 85% of the Game Pass library is playable on ARM, but some applications and games run through the Prism emulator rather than native ARM ports. Problems may arise with multiplayer games that use kernel-level anti-cheat systems, as shown by a warning related to Easy Anti-Cheat during testing of Gears of War: E-Day.
The device will receive Copilot+ certification, but Microsoft expects full software support to arrive in late 2026 or early 2027. Features such as Recall, Click to Do, and the new search bar will not be available at launch, while the Copilot button currently opens the M365 Copilot window for cloud AI services.
In an unscientific battery test, the device retained 66% after three hours of varied use, but running intensive AI workloads on the GPU caused faster drain during the final 30 minutes. Magnetic charging over USB-C adds practical protection against pulling the device, but it is not enough on its own to justify the price.
Why Does This Test Matter?
The Surface Laptop Ultra offers the first convincing indication that ARM can be combined with high graphics and computing performance in a single laptop. But it also shows that the success of this type of device does not depend only on the number of GPU cores or the amount of memory; drivers, model execution paths, application compatibility, and support for professional tools are critical factors.
The review’s editorial conclusion is that the device is particularly suitable for a CUDA developer or a user who wants to test large local models with high portability, provided they accept the high price and current limitations. Those looking for a ready and stable computer for all their work, gaming, and AI needs may be better off waiting until the software matures and the announced Copilot+ support arrives.