A few weeks ago, Microsoft and Nvidia finally released Surface Laptop Ultra powered by RTX Spark to the market. It was a major moment for the Windows platform and ecosystem, with a collection of laptops with a large GPU and unified memory pool for running local inference. Microsoft took this as an opportunity to build the Surface Laptop Ultra, what they say is a device designed for developers.
While I would not consider myself a developer, I think it would be fair to say I'm pretty close to one. I've spent the better part of this year developing Creative Strategies Atlas platform, redoing our website, building internal AI tooling, as well as my own hobby projects for fun. All of this was made possible because of AI. With this in mind, I wanted to talk a little bit about the Surface Laptop Ultra and my experience using it.
To start, of course, build quality and what you actually see and touch. This is arguably one of the most important parts of a consumer electronics product. Microsoft did a fantastic job with the design. It feels extremely premium with the aluminum body, the cooling is done exceptionally well, pulling in air from the sides and venting out the back, rather than the bottom on your lap like a lot of other laptops. The hinge feels tight and not flimsy. There's this part as well where the bottom is designed with a unique platform so it feels thin and has this sort of floating look on a table or desk, while also giving the device a better area to pull in air for the fans. It is all around one of the most premium designs I've seen in the Windows ecosystem in years. It feels premium, well built, expensive, and all around like a device I'd actually want to use. It's a bit heavy, but I don't really mind that but obviously would prefer lighter. I'm in the camp when weight is added properly and distributed right, it can make a device feel more premium and durable. I think Microsoft hit that really well.
The parts you look at and touch are also really good, Microsoft has what I'd consider the best trackpad on a Windows laptop. It's a soft and great haptic trackpad, it has fantastic multi-touch and haptic feedback. It's really great, so is the keyboard. Keyboards are very subjective, I like this one a lot. The display is the part that I'm not totally in love with. It has a Mini-LED display with great clarity and PPI. Microsoft added Mini-LED specifically because for a developer machine, text is key and on LCD displays text clarity tends to be a bit better. That is true, and you can absolutely see it. The problem is the dimming zones, the Surface does not have enough of them and they do not react fast enough. When you set it up, you may notice ghosting because Microsoft actually keeps the window you are looking at slightly brighter than the background apps, likely for battery life and that's a really smart way to improve it, but this adds a bit of brightness ghosting that's distracting. You can disable this and enabling HDR does seem to help, but it's just not what I was expecting with Mini-LED and a nice OLED may have been better, but I think this a bit of a nitpick vs. actual issue.
Oh and a few small details I think are really good, the magnetic USB-C port is really cool. I've been mostly using it with standard USB C cables because I just have so many, but it's really useful. I do like it, and you aren't forced into using the USB C cable it came with so there is no downside to it. The other one is actually the speakers. The speakers are great, but there are no speaker grills… it's the air intake ports for cooling on the sides! The laptop has side firing speakers with Dolby Atmos. They sound great and were actually tuned for speech. Video calls, movies, and TV shows all sound really good.
It's the first time in probably 5 years we've had truly premium Windows laptops and I am really excited by this. I'll talk more about the other OEM machines over the coming weeks, but wanted to start out with Microsoft!
What's a little more interesting is actually using the laptop and what it allows. I don't need to go over what Windows is like or Windows on Arm, it's the same. Basically all issues are fixed, it just runs like any other computer. I don't really even think it's worth talking about at this point, but I still get many questions about this, so I wanted to at least bring it up.
The big part of this system is local models, my unit here has 128GB of Unified Memory on the RTX Spark with 6144-core GPU. This part is a bit interesting, because there are two parts to it. One is how well do models run on the RTX Spark, which is to say, very good! I've been using MAI-Code Flash 1.1 in GitHub Copilot and it's blazing fast, if you can use a model like this for your work, it's fantastic. Another part of the system is actually the CPU and speed, it's really quick and as a whole, the best SoC for getting real work done. There are other Windows PCs with better CPUs, but those lack the GPU. There are others with a larger GPU, but those are thicker and heavier with poor battery life. When you want the best of everything, the RTX Spark is the only laptop with that.
I actually used this system during the review period around a month to help contribute Windows on Arm optimizations and support to OpenCode and T3 Code. Doing it on this machine, I was actually surprised how well it worked. It was fast, the fans rarely turned on, and the battery life was genuinely great. It's just a really good developer machine.
Now that other part of local models I want to talk about is capabilities of the local models. Today, they are not good. The best local model to run on this system from a US lab is maybe Recursive Flint, which is a 40B parameter MoE model. The other option would likely be Qwen 3.8 Flash Next from Alibaba, which is slightly more intelligent, but requires the 128GB SKU while Flint can run on a 64GB SKU. These are not models that I would consider capable. A task I would trust these with would be, say, "add support for GPT 6.1 Terra to the Atlas agent" which is one I did recently do. This is the sort of task where it's a few lines of code for me to change, but I'd rather just delegate it. It's small, easy for me to do, easy for the agent to do, and works well. I've tried using these for larger agentic work, it does not work. If I were paying API rates for a model, I'd likely just make the change myself.
The reason I wanted to start with the fact it runs well is the trends, we are trending towards better models running locally. I am not as optimistic as most at there being a proper monetary value proposition here long term, but I could be wrong. What does matter is the models do run, they do work, and if you are concerned about token pricing in the future, there are options!
I would say the former is only relevant for enterprises. If you are a consumer or a hobbyist, I would actually say never worry about token pricing. A statistic I saw recently was on an RTX Spark machine you could generate roughly $1200 worth of equal intelligence tokens locally at API pricing. While this is true, a consumer would never pay API pricing. They would go to OpenAI, Anthropic, or Google and get a $20 to $200 per month plan. These plans offer between $800-$16,000 worth of usage. There have been comments about token pricing going up, that's actually wrong. Token pricing has gone down and been trending down for every model and every model provider, except open models, funny enough.
What has gone up is token usage, because the models become more capable. I'm able to spend more because they models are able to do WAY more, or I can spend the same amount and do more, or I can spend less and do more. Every 45-90 days we get a pretty large jump in how much intelligence and capability $1 gets me. There are obviously concerns with privacy and trust, and there is nothing anyone could say to people with those concerns to justify using a cloud model, and that's totally ok. All I am saying is this laptop is not going to be your way to save money on tokens, it's arguably a stepping stone into better cloud models and using the local compute for more intense work, like rendering 3D models and videos or tuning custom models using the local GPU and memory, rather than running LLM inference on it.
One thing I want to make clear is there is genuine fun and joy in running local models and experimenting with custom finetunes, seeing how well it can run, and all of that. It's just fun. I do it because I enjoy it, not because I'm trying to economically justify these machines existence. I think the idea of downloading and risk-free running models is really fun!
I really like this laptop, I can see why people would want it, I would happily recommend it, and obviously, I think Microsoft is leading the Windows laptop ecosystem with thoughtfully designed products that actually focus on meaningful details. They are the only ones that really do. They should be proud of this product.