Microsoft Surface Laptop Ultra Review: RTX Spark Excels at Local AI, Falls Short for Gaming

The Microsoft Surface Laptop Ultra makes its strongest case when it runs AI models on the laptop itself. Reviews found Nvidia’s RTX Spark platform especially capable with larger local models, but its price starts at $2,599.99, and its results in gaming and some everyday workloads are less compelling. Buyers need a specific reason to use that extra memory and AI performance to make the cost add up.

The laptops went on sale last week. The entry configuration pairs an 18-core chip with 24GB of memory and 512GB of storage. At the other end, the 20-core version with 128GB of memory and 1TB of storage costs $5,899.99; reported lead times for that configuration stretch to late December. The key AI capability, access to up to 96GB of memory for the GPU, is limited to the 128GB version. The 32GB and 48GB configurations can’t run the same large models.

Local AI is the reason to consider it

The 20-core RTX Spark combines a 6,144-core Blackwell GPU with up to 128GB of unified memory. Up to 96GB can be assigned to the GPU, giving it more room for large local models than a conventional laptop GPU. Reviews found it particularly strong with models in the 20B to 35B parameter range: 35B models ran at up to 89 tokens per second, while models as large as 120B could fit entirely on the GPU.

In agentic AI tests, reviewers placed the Surface ahead of the M5 Max MacBook Pro, with results about 25% faster, and roughly twice as fast as the Ryzen AI Max+ 395 in the cited comparisons. Stable Diffusion 1.5 INT8 scored 11,294, beating the tested RTX 5080, RTX 5070 Ti and RTX 5060 systems by a substantial margin. One review also used Hermes Agent to sort a cluttered downloads folder and build a website locally, without sending the work to the cloud.

That independence can matter to people who work offline, want their data to stay on their machine, or regularly experiment with local AI tools. There’s a limit to the advantage, though: when a model fits within 24GB, a Razer Blade 16 with an RTX 5090 generated tokens 2.3 times faster. The Surface’s extra memory is most useful when the workload actually needs it.

CPU strength comes with uneven results

The 20-core N1X has 10 performance cores and 10 efficiency cores, and its best results came in multi-threaded work. In Cinebench 2026, reviews recorded multi-core scores of 6,885 and 6,963. Windows Central reported a 6,885 score, while StorageReview measured 6,963. The Surface also completed an LLVM compilation in 206 seconds, compared with 327 seconds for the next-fastest system in that test.

Single-core performance was less impressive. The Surface scored 558 in Windows Central’s Cinebench 2026 test and 558.5 in StorageReview’s, narrowly ahead of a Dell Pro Precision 5 16S score of 534.8 in the cited comparison. Across a broader set of tests, StorageReview found single-thread results 15% to 30% behind x64 laptops. It also recorded 251ns memory latency, the highest in its comparison group.

There are more specific compatibility and workload gaps. Arm64 SVT-AV1 encoding reached about one-fifth of the output from the x64 systems in one review. Under a demanding all-core OpenBLAS double-precision solve, the laptop repeatedly lost power. Most buyers won’t run that particular task, and reviewers said normal use felt responsive, including with emulated apps. For browser-heavy or other single-threaded work, though, a less expensive x64 laptop may be the better fit.

Gaming isn’t its strong suit

RTX Spark’s GPU can play games, but reviewers described performance as closer to an RTX 4070 or RTX 5060 laptop than to a top gaming machine. Cyberpunk 2077 reached 60fps at 1440p with frame generation. In Gears of War: E-Day, the laptop averaged 77.7fps, with 1% lows of 39.9fps and measured latency of 83.5ms. Control Resonant reached 109.2fps with 3X frame generation.

Shared memory bandwidth reached 300GB/s, below the 384GB/s reported for a mobile RTX 5060. Reviewers also pointed to limited input responsiveness. For games that need fast controls or consistently high frame rates, a gaming laptop with a dedicated RTX 50-series GPU is the more convincing choice. The same trade-off applies to AI: an RTX 5090 laptop can be faster when the model fits into its 24GB of memory.

The configurations and the buying case

Both N1X versions use LPDDR5X memory running at 9,400MT/s, offer up to 300GB/s of bandwidth, and have a laptop power range of 45W to 80W. The 18-core chip uses nine Cortex-X925 and nine Cortex-A725 cores, plus 5,120 CUDA cores and up to 64GB of memory. The 20-core configuration has 10 of each CPU core type, 6,144 CUDA cores and up to 128GB. Both are made on TSMC’s 3nm process.

The steep pricing narrows the audience further. Tom’s Hardware cited high DRAM and NAND prices as a near-term headwind, while noting that the purchase price could instead cover a lot of cloud AI usage. Windows Central also pointed to less expensive laptops that offer sufficient performance for many people. Those alternatives deserve a close look if local models aren’t part of your daily work.

The Surface Laptop Ultra looks best for developers and analysts who routinely run models larger than 24GB on their own machine. Its local AI speed and memory capacity stand out, and its multi-core CPU results can help with heavily threaded tasks. For gaming, browser-led work or models that fit on a dedicated-GPU laptop, the price is much harder to justify. The deciding question is simple: will you use the extra GPU memory often enough to make it worth paying for?

Source reviews: Windows Central, Tom’s Guide, StorageReview and Tom’s Hardware.