By Gavin Stevens · Updated November 2026

Surface Laptop Ultra vs MacBook Pro M5 Max for local AI

Two laptops hold 128GB of unified memory, and both have been measured now. The Surface's bandwidth is 283GB/s against the Mac's 614, the Surface has CUDA and the Mac doesn't, and Apple's redesigned MacBook Pro ships with the same M5 Max. Which to buy in November 2026.

Illustration of two open laptops side by side, one space gray and one platinum silver, with plain blue gradient screens

As an Amazon Associate I earn from qualifying purchases. Product links on this page go to Amazon or the Microsoft Store; both are affiliate links, and the price you pay is the same.

The picks

No ceiling
Illustration of an open thin silver laptop with a plain light-blue screen on a dark backdrop No ceiling

Microsoft Surface Laptop Ultra (RTX Spark, up to 128GB unified memory)

Order the 64GB or 128GB configuration; the 24GB and 32GB ones are the base Mac mini story in a bigger chassis. Measured at 58 tokens a second on gpt-oss-20B and 89 on Qwen3.6-35B-A3B. $5,899 at 128GB.

CUDA in a laptop that holds a 70B model. Bandwidth is unpublished (about 300GB/s derived) and the Windows-on-Arm tools are new, so it's a pick for people who want CUDA on Arm and can wait for the software to settle.

View on Microsoft Store

Illustration of an open 16-inch aluminum laptop No ceiling

MacBook Pro 16-inch (M5 Max, 40-core GPU, 128GB unified memory)

The known quantity: 128GB at 614GB/s, about double the Surface on the same models. Pick the 40-core GPU. The redesigned OLED model shipped November 6 with the same chip. If you don't need CUDA, this is still the laptop to buy.

The fastest laptop for running a 70B model in memory. The 128GB Strix Halo and RTX Spark laptops fit the same models at well under half the bandwidth. It will be loud and warm doing it, and it will do it on a plane.

View on Amazon

Illustration of a small, plain dark graphite desktop computer box No ceiling

Microsoft Surface RTX Spark Dev Box (128GB)

The same platform without the screen, $5,999 with 128GB. The DGX Spark 128GB is $6,950 now, so the Windows box is the cheaper of the two. Ships in November.

The same silicon as the laptop at 140W instead of 45 to 80W, so it should hold its speed under sustained load. If you don't need the screen, this or the DGX Spark is the better buy.

View on Microsoft Store

What changed

Until October 2026 one laptop held a 70B model in memory: the MacBook Pro with M5 Max and 128GB. On October 7 Microsoft opened pre-orders for the Surface Laptop Ultra, built on NVIDIA’s RTX Spark: up to a 20-core NVIDIA Grace CPU and a 6,144-core Blackwell GPU sharing up to 128GB of unified memory, running Windows with the full CUDA stack. Units shipped October 16 and the reviews are in. Acer, ASUS, Dell, HP, Lenovo, MSI and Gigabyte sell RTX Spark machines too; the RTX Spark laptops guide covers the whole line. This page is the Surface against the Mac.

Apple, meanwhile, redesigned the MacBook Pro: a tandem OLED touchscreen, M5 Pro and M5 Max inside, up to 128GB, shipping November 6. The chips are the ones this page already compared, so nothing on the Mac side of the numbers moved. The new 14-inch M6 MacBook Pro is a different machine: 32GB ceiling, 170GB/s, not a local-AI laptop.

What’s confirmed

  • Memory. Up to 128GB unified, shared between CPU and GPU. The 20-core chip comes in 32GB, 48GB, 64GB and 128GB; the 18-core version starts at 24GB. You can dedicate up to 96GB to the GPU in the NVIDIA app; in practice Ollama uses shared memory beyond the split, and a 28GB model ran at 100% GPU with 16GB dedicated.
  • Bandwidth. 283GB/s measured on the GPU (LPDDR5X-9400, about 300GB/s theoretical), per XDA’s test on the sibling HP OmniBook. That’s faster than the DGX Spark’s 240GB/s measured, and under half the MacBook Pro M5 Max’s 614GB/s.
  • Token rates. StorageReview, 20-core 128GB unit, native Arm64 Ollama and llama.cpp on CUDA: gpt-oss-20B at 58 tokens a second with a 64K context, Qwen3.6-35B-A3B at 89, Llama 3.1 8B at 56. A dense 32B runs at about 12 on this platform; a dense 70B will be under 7. Prompt processing is quick: about 7,000 tokens a second on Phi-4-mini.
  • Software. Full CUDA on Windows on Arm. Ollama and llama.cpp run on the GPU in native Arm64 builds. PyTorch works with preview caveats, vLLM has no Windows build, LM Studio and ComfyUI need an Arm64 check before you order, and Smart App Control blocks unsigned builds until you turn it off.
  • Price. $2,599 (18-core, 24GB, 512GB) to $5,899 (20-core, 128GB, 1TB). The 64GB and 128GB configurations that matter here are No ceiling.
  • The rest. 15-inch 120Hz touch display, up to 2TB of user-replaceable storage, three USB-C, USB-A, HDMI, SD slot, under 4.5 pounds and under 18mm.

Two cautions from the reviews

StorageReview saw two separate Surface Laptop Ultra units cut power under an all-core CPU stress test. The logs show a hard cut, not a thermal trip, which points at power delivery or firmware protection; it didn’t happen during GPU inference. Unexplained as of November 6. And every RTX Spark review notes loud fans and heavy battery drain under sustained AI loads; on battery, this is a two-hour machine.

Mac or Surface

MacBook Pro M5 Max if: you want the speed and your tools run on Metal (Ollama, LM Studio and MLX all do, with nothing to configure). 614GB/s generates text about twice as fast as the Surface on the same model, the software has been mature for two years, and the November redesign adds a better screen without changing any of that. Pick the 40-core GPU; the 32-core drops to 460GB/s.

Surface Laptop Ultra if: your work needs CUDA. Fine-tuning, anything built on PyTorch with CUDA kernels, Windows-only pipelines, or a team that standardizes on NVIDIA tooling. You give up half the bandwidth for the software stack, and you get the faster prompt processing. Order 64GB at minimum; 128GB if 70B or 120B models are the point.

Neither if: your models fit in 32GB. A desktop with a 5090 or a used 24GB card generates several times faster on anything that fits. These laptops are for the models that don’t.

The desktop version

The Surface RTX Spark Dev Box is the same platform in a small desktop: 128GB, 2TB, Ethernet, Windows with VS Code, Git, WSL, Python and GitHub Copilot preinstalled, $5,999, shipping in November 2026. When this page was written in October it cost more than the NVIDIA DGX Spark. The DGX Spark 128GB went to $6,950 on October 23, so the Dev Box is now the cheaper of the two, and the only one on Windows. The DGX Spark’s case is NVIDIA’s Linux-based DGX OS with the datacenter tooling preinstalled, and a ConnectX-7 port for clustering two of them. The three-box guide has the rest of that comparison.

Sources