ZDNET Key Takeaways
- PCs built on the Nvidia RTX Spark architecture are excellent for AI tasks.
- There are no per-token fees or usage limits for running a local open-weighted model.
- The biggest benefit of running a local AI model is privacy.
At the dawn of the PC era, legendary PC Magazine editor Bill Machrone coined Machrone’s Law: “The machine you want will always cost about $5,000.” » This figure may be a little higher today due to the current AMR crisis, but adjusted for inflation, Machronus’s law remains true. The proof? I currently have a $7,000 laptop on my desk, and I have to admit that I really, Really I want it.
Also: I put Lenovo’s RTX Spark laptop through a crazy AI coding test – and it passed
The HP OmniBook Ultra is an absolute beast. It’s powered by Nvidia’s RTX Spark, with 128GB of unified memory, 2TB of fast NVMe Gen5 storage, and a stunning 16-inch OLED display. It weighs almost five pounds, but it’s not really a laptop but more of a personal data center, delivering 1 petaflop of FP4 AI performance.
I’m not the target market for this PC. But after a few weeks of use, I’m seriously tempted. For AI-related work, its GPU has several times the power of a backyard Copilot+ PC, and it’s able to take full advantage of unified memory.
How local AI established itself in my project
One of the reasons my project finished so quickly was because I was able to run it without interruption. I initially used Copilot to help me set up the working environment, using the open source inference runtime llama.cpp, the large open weights language model Qwen3.8-27B, and a coding harness called Cline, working as an extension within Visual Studio Code. With a little extra help from Copilot, I set up these tools, wrote a detailed spec for the app I wanted, took it to Cline, and then mostly stood back and looked.
I had a good general idea of what all of these components did, but it would have taken weeks of reading, studying, and manual trial-and-error installation to get everything set up and working. Thanks to the PowerShell scripts written by Copilot, I launched the project in half an hour.
As the project unfolded, I was able to observe the “reason” of AI in real time. Every now and then he would ask me to approve something, but most of the time he would write code, run tests, observe the results, modify the code, run more tests, etc. The whole process was a classic example of “human agent-in-the-loop coding”, and all I had to do was check in from time to time to see how things were going.
I knew it was busy, because when the LLM was running, I could hear the fan spinning from the other side of the room. Surprisingly, the machine never got too hot; Still, this is the kind of work that’s best done with the laptop on a desktop, not on your lap.
Also: I can access my home computers remotely for free – how you can do it too (no VPN needed)
I gave the LLM a spreadsheet filled with links to my articles, which I had painstakingly constructed over the past two years. Not only did he find and retrieve almost every article on that list, but he also created a separate routine to retrieve pages from the Internet Archive’s Wayback Machine for about 90 links returning 404 errors, then presented me with a list of about 90 more pages that he couldn’t retrieve at all. It took maybe 30 minutes to review and resolve this last set of issues.
After all, I had two folders, each containing 1,951 files – one in raw HTML, the second in Markdown format. And finally, I asked my local AI to create a simple script that I can run to add new articles to the archive as they are published.
I can’t begin to tell you how happy this result made me.
How much did it cost?
The entire project used 479 LLM inference calls and processed nearly 50 million logic tokens, for a total net cost of exactly $0.00.
One of the most frustrating aspects of using these modern AI coding harnesses is that there is no way to estimate how many tokens they will use.
I could have used a commercial AI subscription to do this work, but I would have been flying blind. None of the companies that sell these plans can tell you how much a given project will cost. Your actual consumption – and therefore the final bill – depends on factors that you cannot know in advance. How much code should the agent keep in context? How many iterations and debug cycles will this require? What models will you use?
Also: Is Claude Code really faster with discussion threads? I did the calculation on my logs
ChatGPT Plus (with Codex), Claude Code Pro, or GitHub Copilot all cost between $10 and $20 per month. This is more than enough for a simple task, but for this project I would have quickly exceeded the usage limits with these entry-level plans. To continue working on a $20 per month plan, I would have had to take frequent breaks, purchase additional credits, or upgrade to a $100 per month plan with higher usage limits, like ChatGPT Pro, Claude Code Max, or GitHub Copilot Max.
After looking at these numbers, I got an idea of where Anthropic, OpenAI, and Microsoft’s business models are headed. Even at $100 per month, if I had decided to take on even more ambitious projects, I could have easily found myself rate-limited, with work having to wait hours or days until the next cycle begins.
For serious coders, these bills can get very expensive. In a recent episode of Nilay Patel’s Decoder podcast, analyst Joey Brookhart offered a surprising statistic: “The top 5% of Claude Code users spend an average of $18,000 per year. Non-Claude Code users spend well under $75 per month.”
Now I know why.
The biggest advantage of all
Using a free, open weight model with an open source coding bundle gives you the freedom to experiment without worrying about the bill. But that’s only part of the story. The much bigger benefit is that you’re not sending potentially sensitive or confidential data to someone else’s cloud; With a completely local AI stack, your prompts, documents, and inference workloads don’t need to leave the PC at all.
This wasn’t a big problem for my project, which used content already published on the web. But I have a few other apps in mind that I’d rather not share with anyone else until they’re ready. The first involves a large stack of financial documents, credit card statements and bills. I’m happy to feed this data into an LLM on my own PC, but I’m not comfortable uploading these documents to a cloud-hosted coding service.
Also: This Free Windows Tool Is My Secret Weapon for Setting Up New PCs
Will I do enough of this kind of work to justify spending $7,000 on a PC? I would love to say yes, but my CPA would stare at me for a full minute and then patiently explain to me how “return on investment” works. “No, Ed, you can’t justify that expense.”
The good news is that Machrone’s Law works in concert with Moore’s Law, which describes a reliable technology trend: computing power doubles approximately every 18 months. That means in three years I’ll be able to drool over a PC with four times the computing power for $7,000. But by then, a configuration like this HP OmniBook Ultra will likely be a business-class consumer machine, available for a fraction of its current price.
In the meantime, I will wait.
See full bio
Gn tech
