Surface Laptop Ultra style thin laptop in use

Surface Laptop Ultra for Local AI: RTX Spark Brings CUDA to a Laptop, With Caveats

Microsoft's Surface Laptop Ultra is the first RTX Spark laptop. What it runs, how fast, what it costs, and the Windows on Arm and stability caveats from early reviews.

This review is built from Microsoft’s and NVIDIA’s published specifications, retail pricing and third-party testing rather than hands-on use. Figures are attributed in the text. RTX Spark is new, and drivers and software support are changing quickly.

The Surface Laptop Ultra is the first laptop built on NVIDIA’s RTX Spark chip, and for local AI it does something no Windows laptop could do before: it runs models far larger than any laptop graphics card can hold, with CUDA, on a machine you can carry.

The honest verdict is that the Surface Laptop Ultra is a capacity machine, not a speed machine, and it is very early. On models that fit in 24GB, a gaming laptop with an RTX 5090 is more than twice as fast. Windows on Arm still trips up some software, Linux support is not there yet, and one major reviewer had review units lose power under sustained load. If you need 64GB or more of GPU memory in a Windows laptop today, it has no rival. Everyone else can afford to wait a few months.

Table of Contents

Surface Laptop Ultra class laptop on a cafe table
Not a Surface in this photo. The RTX Spark pitch is capacity you can carry: 120B-class models in a bag, at about a third of a desktop GPU’s bandwidth.

What the Surface Laptop Ultra Offers

Microsoft and NVIDIA revealed the Surface Laptop Ultra at Computex 2026 and gave full details and pricing at Microsoft’s San Francisco event on October 7. Machines reached shelves in mid-October, with the Surface leading the first wave of RTX Spark laptops.

It is a 15-inch Windows 11 on Arm laptop with a mini-LED display, user-replaceable NVMe storage and up to 128GB of unified LPDDR5X memory shared between CPU and GPU. Microsoft places it in a new category it calls builder PCs, aimed at developers, creators and AI builders rather than general buyers.

The local AI pitch is simple. A conventional gaming laptop’s GPU has its own fixed pool of video memory, at most 24GB on an RTX 5090 laptop. The Surface Laptop Ultra’s GPU can use most of the system’s memory. StorageReview’s unit had 64GB of GPU memory by default, adjustable up to 96GB, and NVIDIA says 128GB systems can hold models up to about 120 billion parameters entirely on the GPU.

RTX Spark N1X Explained

RTX Spark is NVIDIA’s first system-on-chip for Windows PCs. i-Scoop described it as essentially the GB10 chip from NVIDIA’s DGX Spark mini computer, tuned for laptop power levels and running Windows instead of Linux.

Specification Top N1X bin Second N1X bin
CPU 20 Arm cores 18 Arm cores
GPU 6,144-core Blackwell 5,120-core Blackwell
Maximum memory 128GB LPDDR5X 64GB LPDDR5X
Memory bus 256-bit 256-bit

The GPU is roughly comparable to a desktop RTX 5070 in core count, according to NGXP Tech. The more important number is bandwidth. ServeTheHome explained that the chip’s 256-bit LPDDR5X bus delivers up to 300 GB/s with LPDDR5X-9400 memory. That is very high for a laptop’s system memory and modest for a GPU: an RTX 5090 laptop’s dedicated memory is several times faster.

That one fact explains every benchmark that follows. Generation speed on large models is limited mainly by memory bandwidth, so the Surface Laptop Ultra loads huge models and runs them at moderate speed.

Surface Laptop Ultra Pricing and Configurations

TechRadar reported a starting price of $2,599.99 for the 18-core model with 24GB of memory and 512GB of storage. PCWorld’s review unit was priced at $3,699.99. Higher-memory configurations cost more, and the 128GB model is the one that delivers the headline capabilities.

For local AI, the entry configuration is the wrong one. With 24GB shared between Windows, apps and the GPU, it holds roughly what a much cheaper desktop graphics card holds. The 64GB model is the realistic starting point for serious model work, and 128GB is where the Surface Laptop Ultra has no direct competition in the Windows world.

Memory is soldered, so choose carefully. i-Scoop made the point well: the choice you make at purchase lasts for the life of the device, and open models tend to grow rather than shrink.

Surface Laptop Ultra Benchmarks

Independent results arrived in October from StorageReview, The New Stack, PCWorld and others. They line up well with what the bandwidth predicts.

Model and setup Generation speed Source
gpt-oss:20b and Qwen3.6-35B-A3B 58 to 89 tok/s across the two StorageReview
Nemotron 3.5 Bolt 56 to 60 tok/s The New Stack
Gemma 4 26B-A4B, 8-bit About 40 tok/s The New Stack
Qwen 3.8 27B dense, 4-bit 25 to 28 tok/s, about 18 on battery The New Stack
Qwen 3.8 27B dense, 8-bit 8 to 9 tok/s The New Stack
PCWorld local AI benchmark 39.9 tok/s PCWorld via Tech Insider

The New Stack’s reviewer was frank about the dense 8-bit result: not quite usable, and almost certainly down to memory bandwidth. Dropping to 4-bit roughly tripled the speed. That is the trade you will make constantly on this machine.

For the biggest models, there are fewer measurements. vdf.ai estimated about 36 tokens per second for gpt-oss-120b on the laptop, and noted that the 140W DGX Spark with the same chip measured 58.7, well above its own prediction. Laptop power limits will keep the Surface below the desktop box, but probably above the estimate.

Why MoE Models Change the Picture

Every fast result in the table above comes from a mixture-of-experts model. These models have many parameters in total but activate only a few billion for each token, so the chip reads far less memory per step. The New Stack called the technique really what you need to run virtually any model locally at reasonable speeds, and on this hardware that is true.

In practice, it means the Surface Laptop Ultra suits the current generation of open models very well. Qwen’s A3B models, Gemma 4’s A4B variants, gpt-oss and NVIDIA’s own Nemotron models all use this design. Dense 27B to 70B models are where the machine struggles, and fewer new releases use that design each month.

Choose models with that in mind. A 35B MoE model at 4-bit gives you strong quality at 60 to 90 tokens per second. A 70B dense model at 4-bit will crawl.

Surface Laptop Ultra vs MacBook Pro M5 Max

The obvious rival is Apple’s MacBook Pro M5 Max with 128GB. Notebookcheck framed RTX Spark as NVIDIA’s answer to the M5 Pro and M5 Max, and the comparison is fair.

Question Surface Laptop Ultra MacBook Pro M5 Max
Maximum memory 128GB 128GB
Memory bandwidth Up to 300 GB/s Up to 614 GB/s
Software ecosystem CUDA on Windows on Arm MLX, llama.cpp, Ollama
Price From $2,599.99 From $3,899 before upgrades
Maturity Weeks old Third generation of the approach

The Mac has about twice the bandwidth, so on the same large model it should generate text noticeably faster. It also runs on a software stack that has been tuned for this kind of hardware for years. The Surface’s advantages are CUDA, a lower starting price and Windows. If your tools assume CUDA, that settles it. If they do not, the Mac is the faster and safer choice today. Our MacBook Pro M5 Max review covers it in detail.

Against a Gaming Laptop With an RTX 5090

StorageReview tested the Surface against a Razer Blade 16 with an RTX 5090 laptop GPU. The Blade was about 2.3 times faster on anything that fits in 24GB, and could not run the larger models at all without spilling into slow system memory.

TechRadar reached the same conclusion from the other direction: gamers, heavy browser users and anyone whose models fit in 24GB are better served by a laptop with a dedicated GPU. RunAIHome added a desktop comparison: a used RTX 3090, at $1,190 to $1,400, has 936 GB/s, about three times this platform’s bandwidth, and wins comfortably on any model that fits.

So ask one question before anything else. Do the models you run actually need more than 24GB? If not, the Surface Laptop Ultra is an expensive way to run them slowly.

Windows on Arm and Linux

The Surface runs Windows 11 on Arm64. NVIDIA provides a full graphics and CUDA driver stack, and ServeTheHome confirmed that drivers were in place at launch. PCWorld noted that ComfyUI’s latest version explicitly supports the RTX Spark architecture. llama.cpp and Ollama run natively.

The rough edges are elsewhere. Some Python packages, developer tools and older utilities still lack Arm64 Windows builds, and x86 emulation does not help with GPU code. ComputingForGeeks was blunt about Linux: if you need it on day one, do not buy one hoping for it. WSL2 is the nearer option, and NVIDIA’s NemoClaw already has an Ubuntu aarch64 WSL path, though its own issue tracker shows it still breaks.

If your workflow is Ollama, LM Studio or llama.cpp plus a few Python scripts, you will probably be fine. If it depends on a long tail of x86-only tools, test before you commit.

The Stability Question

One finding deserves attention. NeoTeo reported that on October 14, StorageReview recorded power losses on two Surface Laptop Ultra review samples during a sustained OpenBLAS workload. An earlier sample had failed under the same load, and a later retry on the second unit passed.

That may be a firmware issue fixed in an early update, or a problem limited to pre-release units. It may also be a real limit of a thin laptop running a chip designed for a 140W desktop box. We do not know yet. If you plan to run long batch jobs or overnight agent sessions, wait for more owner reports before buying, or buy from a retailer with an easy return policy.

Surface Laptop Ultra Battery and Daily Use

A laptop for local AI still has to be a good laptop. Reviewers were broadly positive about the everyday side. The Neuron’s reviewer, who used one for daily work, called it the most exciting laptop he had owned, and PCWorld’s long review praised its productivity performance while warning that Windows on Arm gaming remains limited.

Battery life under AI load is the weak point, as on every machine in this class. Sustained generation keeps the GPU and memory busy, and The New Stack’s results show a clear speed drop when unplugged. Plan to run heavy jobs on mains power and treat battery inference as a convenience for short sessions, such as a quick question on a train.

The upside of a unified-memory design is that you are not carrying two machines. There is no home server to tunnel into and no remote desktop session to keep alive. For consultants handling client documents that cannot leave the device, that alone can justify the price.

What It Means for Developers

The most interesting part of RTX Spark is not the laptop. It is CUDA on a portable, high-memory machine. Code written for an RTX workstation or a DGX Spark can run on the same software stack, which makes the laptop a reasonable place to prototype agents, retrieval pipelines and evaluation scripts before moving them to bigger hardware.

That said, this is an inference and development machine, not a training rig. Fine-tuning large models still belongs on a desktop GPU or a rented cloud instance. Use the laptop to iterate quickly, then hand heavy jobs to something with more bandwidth and power.

Other RTX Spark Laptops

The Surface is the first, not the only, RTX Spark laptop. Dell has confirmed an XPS 16 Creator Edition with the chip. ASUS has a ProArt P16 version, which Notebookcheck measured drawing up to 85W in performance mode, and The Verge tested machines from several vendors including Lenovo and HP.

Different cooling designs will matter. The same chip at a higher sustained power limit will hold its speed better through long generation runs. If thermal stability is your concern, wait for full reviews of the thicker creator laptops before deciding.

Who Should Buy What

If you are Start with Why Link
Needing 64GB+ of GPU memory with CUDA in a laptop Surface Laptop Ultra, 64GB or 128GB Only Windows option at this capacity Check price
Wanting the fastest large-model laptop MacBook Pro M5 Max Twice the bandwidth Check price
Running models under 24GB RTX 5090 gaming laptop More than twice as fast Check price
Not needing portability DGX Spark Same chip, higher power limit, Linux Check price

Frequently Asked Questions About the Surface Laptop Ultra

Is the Surface Laptop Ultra good for local AI?
Yes, if you need more than 24GB of GPU memory in a Windows laptop. It runs large mixture-of-experts models at 40 to 90 tokens per second in published tests.

How much does it cost?
From $2,599.99 for an 18-core, 24GB configuration. Higher-memory models cost considerably more; PCWorld’s unit was $3,699.99.

Is it faster than a MacBook Pro M5 Max?
Unlikely on large models. The Mac has about twice the memory bandwidth. The Surface’s advantage is CUDA on Windows.

Can it run Linux?
Not natively at launch. WSL2 is the practical route for Linux tools for now.

Which memory option should I choose?
At least 64GB for local AI. The 24GB model holds no more than a cheap desktop GPU.

Are there reliability concerns?
StorageReview recorded power losses on review units under sustained load. It is not yet clear whether retail units are affected.

Setting Up the Surface Laptop Ultra

Update Windows, firmware and NVIDIA drivers before installing anything else. Early RTX Spark systems are getting frequent fixes, and some of them will matter for stability.

If your configuration allows it, raise the GPU memory allocation for AI work, as StorageReview did from 64GB to 96GB. Install Ollama or LM Studio, then start with a mixture-of-experts model such as Qwen3.6-35B-A3B or gpt-oss-20b to check that everything runs on the GPU.

Plug in for heavy work. The New Stack measured a 27B model dropping from 25 to 28 tokens per second to about 18 on battery. And keep an eye on temperatures during your first long runs.

Verdict

Buy the Surface Laptop Ultra with 64GB or 128GB if you need CUDA, Windows and more GPU memory than any laptop graphics card offers. It runs the current generation of mixture-of-experts models at genuinely useful speeds in a portable machine, and nothing else on Windows does that.

The honest caveat is that it is slower than it looks and very new. Bandwidth limits dense models, an RTX 5090 laptop is much faster on anything under 24GB, Linux is missing, and early review units showed power losses under sustained load.

Next step: list the models you need and check whether any exceed 24GB. If they do and your tools need CUDA, wait for the first firmware updates and owner reports, then buy the 64GB or 128GB Surface Laptop Ultra from a retailer with a generous return window.

Sources and Further Reading

On this site: compare the desktop version in our DGX Spark review, see the Apple alternative in the MacBook Pro M5 Max review, size memory with how much VRAM for AI, and pick models from best local LLMs.