What Is the Same and What Is Different Between the 128GB Surface Laptop Ultra and the DGX Spark? A Look at the RTX Spark Laptop Through Official Information and Third-Party Measurements

October 21, 2026

If you are looking for a small machine with 128GB of memory, the options now extend even to laptops. Microsoft’s “Surface Laptop Ultra", released on October 16, 2026, carries NVIDIA’s new chip, RTX Spark, and its top configuration has 128GB of memory. That is the same 128GB as the DGX Spark I looked into earlier on this blog, but what is the same and what is different between a laptop and a small desktop machine?

This reflects the state of things as of October 17, 2026.

I summarized the RTX Spark announcement in this article, and the DGX Spark price and the 64GB version in this article.

Method: what I looked into and where the numbers come from

I read the official pages from Microsoft and NVIDIA (specifications, price lists, and footnotes) and compared the 128GB Surface Laptop Ultra with the 128GB DGX Spark in terms of configuration, power consumption, usable memory, OS, and where you can buy them. I did not run either machine. I state the source of every number, and I label the manufacturers’ own measurements as “manufacturer’s figures".

Surface Laptop Ultra configurations

It went on sale in Japan at 21:00 on October 16. The main consumer configurations are as follows.

As of October 11, the Microsoft Store configuration page showed the 128GB version as “out of stock". I could not confirm stock after the release. Electronics retailers carry configurations from 24GB (excluding the 1TB one) up to 48GB. I could not confirm whether Amazon carries it.

The memory is mounted on the board, and it is reported that it cannot be added to or replaced later (ITmedia). The SSD is removable. The screen is a 15-inch touch-only display and does not support a pen. It weighs about 2.0kg and is about 18mm thick.

What is the same and what is different between the 128GB version and the DGX Spark?

The fact that the CPU configuration and memory capacity are the same, and that the memory bandwidth is close, shows that the two have a very similar design. On the other hand, neither NVIDIA nor Microsoft says outright that the two chips are the same thing. The stated bandwidth figures differ, and the Surface is described as carrying an NPU (a compute unit dedicated to AI). I could not confirm whether the silicon is the same.

The difference in power consumption is a big one. The Surface chip is rated at 45 to 80W, while the DGX Spark’s GB10 is rated at 140W. For a laptop, the concern is whether it can hold the same speed over a long run, but Microsoft only says the thermal capacity is 2.5 times that of the previous model and does not publish a sustained-performance figure. Among third parties, StorageReview ran a continuous load for about 30 minutes on a 128GB machine and reported that the final score was 1% higher than the starting one. The same review, however, says that when it applied a heavy all-core computation (OpenBLAS), the machine shut off suddenly three times on October 14. It says an earlier test unit showed the same symptom and was replaced. PCWorld, which tested a 32GB machine, wrote that performance dropped 6.3% after 10 minutes of sustained load in Cinebench 2024.

How much of the 128GB can the AI use?

A footnote in Microsoft’s material says, “The maximum amount of memory available to the GPU depends on the configuration and usage, and is less than the total." In a Task Manager screenshot in Microsoft’s blog, GPU memory was shown as 110GB (74.9GB dedicated to the GPU and 35.5GB shared). The memory capacity of the machine in the screenshot is not stated, and I infer it is a 128GB machine. On StorageReview’s test unit (128GB), 64GB was allocated to the GPU, and the review says it can be changed up to a maximum of 96GB.

Not all of the 128GB becomes space for the model. The OS and apps need some too. NVIDIA’s porting guide explains the mechanism in which a GPU-dedicated area and a shared area are separated.

How fast can we expect it to be?

Writing speed (token generation) is determined almost entirely by memory bandwidth. The Surface’s 300GB/s is close to the DGX Spark’s 273GB/s, and also close to the EVO-X2 (256GB/s) on this blog. So I expect it to be of a similar speed, but this is an inference from bandwidth, not a measurement. After the release, third-party measurements appeared. On a 128GB machine, StorageReview measured Llama 3.1 8B at 56.4 tok/s and Phi-4-mini at 97.5 tok/s, and wrote that gpt-oss:20b and Qwen3.6-35B-A3B “can write 58 to 89 tokens per second". On a 32GB machine, PCWorld measured Qwen 3.8 27B (LM Studio) at 39.9 tok/s. I could not find a measurement with gpt-oss 120B.

The models and conditions differ in every row, so you cannot line them up and rank them. The DGX Spark’s 58.7 tok/s is for gpt-oss 120B, while the Surface’s 58 to 89 tok/s is for gpt-oss 20B and Qwen3.6-35B-A3B. Please treat them as a rough guide. tok/s is the number of tokens written per second; in Japanese, roughly 1 to 2 characters make up 1 token.

The “up to 1 petaflop" that Microsoft states is a theoretical value that uses sparsity (a mechanism that skips part of the computation). It is not a number that shows actual AI speed.

What to know before using it

• The OS is the Arm version of Windows. Apps not built for Arm are run through translation, but for special drivers and old utilities, checking beforehand is recommended.
• CUDA (NVIDIA’s mechanism for its GPUs) is supported on the Arm version of Windows. Ollama also added CUDA support for Windows on Arm in its July release. The distributed llama.cpp build (b11702, October 17) includes a CUDA build for Windows on Arm. On GitHub it is treated as a prerelease.
• StorageReview writes that it used a CUDA-enabled Arm64 build of llama.cpp and Ollama on the actual machine. The official Ollama release notes (v0.40.5, October 14) do not yet mention RTX Spark.
• The battery life (up to 15 hours of video) is a figure for a prototype with 64GB of memory, not for the 128GB version.

Who is it for?

The value of the Surface is that 128GB of memory fits in a single machine you can carry. It is a candidate for people who want to run large models on the go, or who want to use it on the same machine as their Windows apps.

For stationary use, the DGX Spark and mini PCs such as the EVO-X2 are also candidates. The EVO-X2 (128GB) on this blog was considerably cheaper than the Surface’s 128GB version as of October 9. The DGX Spark has the same 128GB and a dedicated interconnect (200Gbps) for linking multiple units. I think the choice comes down to whether you put priority on portability or on price and expandability.

What I do not know

• I have not run the Surface on this blog to check it. All numbers are figures published by the manufacturers and third parties
• A measurement of the Surface with gpt-oss 120B (needed to compare it with the DGX Spark on the same model)
• Fan noise, and the battery life of the 128GB version (I have not been able to confirm the original source of third-party figures)
• Whether RTX Spark and GB10 are the same silicon
• The official DGX Spark price in Japan
• Stock of the 128GB version after release

Summary: carry the weight of 128GB, or leave it on a desk

The 128GB Surface Laptop Ultra fits the same capacity and a similar memory bandwidth as the DGX Spark into a laptop of about 2kg. The differences are the OS (Windows versus Linux), power consumption, portability, price, and where you can buy it. As of October 11, the 128GB version was out of stock and could be bought only at the Microsoft Store.

Because third-party speed measurements appeared after the release, I added them to the table above. When a measurement with gpt-oss 120B appears, I will update this article.

Sources

• Surface Laptop Ultra (Microsoft Store Japan)
• Windows Blog Japan (October 9, 2026)
• Microsoft Command Line (article on local models)
• NVIDIA RTX Spark
• NVIDIA RTX Spark Porting Guide
• NVIDIA DGX Spark
• Official llama.cpp DGX Spark benchmark
• ITmedia (Surface Laptop Ultra release, in Japanese)
• StorageReview (review of the Surface Laptop Ultra for Business)
• PCWorld (review of the Surface Laptop Ultra)