Surface Laptop Ultra: Is it worth it for local AI in SMEs?

Microsoft has unveiled the Surface Laptop Ultra, a laptop that competes directly with Apple's MacBook Pro and is specifically aimed at AI developers. With an Nvidia RTX Spark GPU based on the Blackwell architecture, up to 128 GB of unified memory, and a price that, according to the company, is lower than Apple's machines, this device promises to run AI models locally without relying on the cloud. But is it really a sensible option for an SME or an IT team with limited resources? We break down what The New Stack tells us after a week of testing.

Local AI performance: highs and lows

The model tested by The New Stack is the high-end configuration, with 64 GB of RAM and a 1 TB SSD, priced at $4,300. With tools like Unsloth Studio, Ollama, or Hermes Agent, the laptop performs capably. In fact, the journalist notes that the Blackwell GPU was, overall, somewhat faster than his Mac Studio with an M4 Max chip in local inference tasks.

Results vary depending on the model. With the recent Nemotron 3.5 Bolt, a quantized model with 9 billion active parameters, it reached between 56 and 60 tokens per second, an excellent speed for use in code agents. However, with dense models like Qwen 3.8 27B in 8-bit quantization, performance dropped to 8-9 tokens per second, which is barely usable. Dropping to 4 bits raised the speed to 25-28 tokens per second, which is acceptable. The conclusion is clear: to get reasonable speeds locally, it's best to opt for models with a mixture-of-experts (MoE) architecture or aggressive quantization.

The price of power: battery, heat, and noise

One of the most striking aspects is the drop in performance when the charger is disconnected. Microsoft claims the device retains 99.6% of its performance on battery, but in the AI tests the loss was notable: around 20-25% less speed. For example, the Qwen 3.8 model in 4 bits went from 25-28 to 18 tokens per second. In addition, battery life drops drastically if AI workloads are run continuously; although Microsoft promises 12 hours of web browsing, in these tasks it wouldn't last more than a few hours.

Heat and noise are also part of the package. The two fans kick in and generate audible, though not extreme, noise when the GPU is pushed for prolonged periods. For normal office use, the device is quiet, but if you're going to work intensively with local models, it's worth keeping this in mind.

Memory: is 64 GB enough?

The 64 GB RAM configuration offers headroom for the KV cache and other applications, but most current small models are around 27-30 billion parameters, so the full potential isn't used. Jumping to 128 GB raises the price to $5,900, a considerable investment for an SME. In our opinion, for most business use cases, 64 GB is more than enough, and the extra cost of 128 GB is only justified if you plan to run very large models on a regular basis.

What does this mean for an SME?

The ability to run AI models locally without relying on the cloud is attractive for reasons of privacy, latency, and recurring costs. However, the Surface Laptop Ultra is an expensive device with thermal and battery limitations that take it away from intensive use on the go. For an IT team that wants to explore local AI, it can be an interesting development machine, but it doesn't seem like the most cost-efficient option for deploying across the entire workforce.

If the goal is to have a testing lab or a machine for a specific developer, the Surface Laptop Ultra offers notable performance and good integration with tools like Ollama. But for production workloads, it's probably wiser to opt for dedicated servers or cloud GPU instances, or even the Surface RTX Spark Dev Box, which shares the platform but has better cooling.

ForgeNEX recommendations

• Evaluate the use case: not all SMEs need to run AI models locally. If privacy isn't critical, the cloud remains cheaper and more scalable.
• Prioritize memory and model type: for local AI, choosing optimized models (MoE, 4-bit quantization) matters more than having a large amount of RAM.
• Consider the form factor: if you need mobility, the Surface Laptop Ultra is powerful but heavy and with limited battery life under load. For a desktop machine, a server or a workstation with better cooling may be more cost-effective.
• Don't forget security: running models locally eliminates dependence on third parties, but it also shifts the responsibility for updates and patching onto you. As we discussed in our analysis of the CISO facing AI, risk management must adapt to this new paradigm.

In short, the Surface Laptop Ultra is one more step toward the democratization of local AI, but its price and compromises place it as a tool for very specific profiles. At ForgeNEX we recommend carefully analyzing your needs before investing in dedicated hardware, and considering hybrid alternatives that combine the flexibility of the cloud with the sovereignty of local processing when it's essential.

Source: The New Stack. Analysis and adaptation: ForgeNEX.