Surface Laptop Ultra: Which Configuration to Buy for Local AI

Microsoft put prices on the Surface Laptop Ultra on October 7, 2026: preorders from $2,599.99, availability from October 16. The headline price is the part everyone repeats. For anyone buying this laptop to run AI locally, the more important number is how much memory comes with it, because that decides whether the machine can run the models Microsoft is advertising.

This guide does the memory math, lines it up against the new MAI-Code-1.1 Flash local model, and says who should buy which tier, who should wait, and who should skip it. It builds on our broader look at the Surface RTX Spark hardware lineup, which said configuration pricing was not yet public. It is now, partly.

TL;DR: the questions people are asking

The configurations and prices

Pricing and specs are reported from launch coverage of the preorder page. We did not see Microsoft's configurator ourselves, and not every combination has been published, so treat the middle of this table as incomplete.

The shape of the pricing is the story. Going from 24GB to 32GB, with a storage upgrade, costs about $1,100. Going to 128GB costs roughly $3,300 more than the base. Memory is where Microsoft is making its margin, and it is also where the capability is.

Other reported hardware features: a 15-inch PixelSense mini-LED touchscreen with up to 2,000 nits of peak HDR brightness (measured over a small window, so sustained brightness is lower), up to three 4K external displays, magnetic USB-C charging, a full-size SD card reader and a removable drive. Fingers work on the touchscreen, but Surface Pen and Slim Pen are not supported.

Why memory decides everything for local AI

A language model has to fit in memory before it can run. The rough size of a quantized model is its parameter count times the bits per weight divided by eight, plus working memory for the context (the key-value cache). These are estimates, not Microsoft figures.

Two practical consequences follow. A 24GB machine shares that memory with Windows, your browser and everything else, so a 30B model at 4-bit is a squeeze and anything bigger does not fit. A 128GB machine fits the 70B and 120B class and Microsoft's 137B coding model with room left over.

Speed is a separate limit. Once a model fits, generation speed is mostly bound by memory bandwidth, not the headline compute figure. Petaflop-class numbers are usually quoted at low precision, so wait for independent tokens-per-second results before you assume interactive speeds. Our explainers on MacBook versus a dedicated GPU for local LLMs and the DGX Spark local LLM setup cover the capacity-versus-bandwidth trade.

Which tier for which job

24GB base ($2,599.99): a fast Windows laptop with light local AI

You get a capable Windows on Arm laptop and enough memory for small models, speech, image tools and a 7B to 14B coding assistant. It will not run the models in Microsoft's keynote. If you are buying it for local AI specifically, the base model is the wrong tier.

32GB and 1TB ($3,699.99): small and mid models, tight

This tier handles models in roughly the 20B to 30B class at 4-bit with modest context. It is a reasonable fit if your workload is a local coding assistant plus normal office work. It is not enough headroom for a 70B model, and at nearly $3,700 you are close to the price where the 128GB machine starts to look like the better long-term buy.

128GB (up to $5,899.99): the one that matches the marketing

This is the machine Microsoft used to show MAI-Code-1.1 Flash. Reference measurements put peak memory at about 75.5GB at the full 256K context, decode at 923.5 tokens per second at 64K context and 769.8 at 128K, per reporting on Microsoft's numbers. Those are vendor figures. If you want to run a 70B or 120B-class model, or the 137B coding model, locally and keep working in other apps, this is the tier.

What about the MacBook Pro M5?

The feed headline claiming the Surface "beats Apple M5 in AI speed" does not match anything we could verify. We found no independent head-to-head benchmark and no Microsoft table showing the comparison with methodology. What we can say from reports:

• Both platforms top out at 128GB of unified memory, so capacity is similar on paper at the high end.
• Apple says its M5 Pro and M5 Max deliver up to 4x faster LLM prompt processing than their M4 predecessors, and cites memory bandwidth as a strength.
• The Surface runs Windows on an Arm chip with an Nvidia Blackwell GPU, so CUDA-style tooling and Windows-only software are the draw; macOS tooling and battery life are the Mac's.
• We did not find a current matched-configuration price for the Mac, so we cannot say which is cheaper.

Until reviewers publish tokens per second on the same model, treat any "beats Apple" line as marketing. Our posts on the Mac Studio M5 Max and Ultra for local AI and the M6 Mac mini show how Apple's side looks.

What the laptop does not solve

• Heat and battery. Sustained local inference is demanding. We have no independent thermal or battery data for this machine under load.
• Software maturity. Windows on Arm has improved, but check that the tools you use run natively. Local inference runtimes, Windows ML and local endpoints are the paths Microsoft named for Copilot.
• Policy. Microsoft has not published a routing rule for when Copilot uses the local model, and business plans need an admin to turn model policies on. A fast laptop does not change that.
• Resale and generation churn. First-generation hardware on a new platform tends to drop in value quickly.

Buy, wait or skip

Buy now if you have a deadline, a funded team, and a specific model that you know fits in the 128GB tier, and you accept that first-week buyers get the least information.

Wait if you are deciding between tiers. Independent reviewers get units from October 16. Tokens per second on 70B and 120B models, fan noise and battery life under load will tell you whether the 128GB machine is worth about $5,900.

Skip if your models are under about 30B parameters. A conventional laptop with a modest GPU, or a cheaper high-memory machine, will serve you for less. Skip also if your work depends on a Pen or on macOS tools.

What we could not confirm

• Full configuration table. We saw three price points reported from launch coverage, not Microsoft's complete configurator.
• Independent performance. No review benchmarks exist before October 16.
• The Apple comparison. No credible head-to-head source found.
• Battery life and thermals under sustained load.
• International pricing and availability. We saw US figures only.

What this means for what you build or pay

1. Decide the model first, then the memory. Look up the model size and add context overhead before you pick a tier.
2. Do not buy the base tier for local AI. Its price is the headline; its memory is the limit.
3. Compare against renting. If you need a 120B-class model a few hours a week, cloud inference may cost less than $5,900 of hardware. If you run it daily on sensitive data, local wins.
4. Keep the cloud path. Copilot routes between local and cloud, so even a 128GB machine will send some work off the device unless you configure otherwise.
5. Wait for the October 16 benchmarks before spending above $3,700.

Related reading on explainx.ai

• MAI-Code-1.1 Flash runs locally on Windows
• Surface RTX Spark Dev Box and Laptop Ultra preorders
• Microsoft Windows hybrid intelligence: local and cloud agents
• NVIDIA DGX Spark: best local LLM setup
• MacBook vs dedicated GPU for local LLMs
• Apple Mac Studio M5 Max and Ultra for local AI
• Closed-source AI versus local open-source alternatives

Prices and specifications are as reported from launch coverage on October 8, 2026 and may change. We have not tested the Surface Laptop Ultra.