[Sorry for the delay in posts, I’ve been really busy lately with my wife on travel, though you’d be excused if you thought I was doing a victory dance about about the SpaceX IPO.]
If you’ve read my monthly updates, you might have noticed that I have been writing more about AI as my main hobby. I use it to do all kinds of side projects. Some side projects don’t see the light of day. For example, I still have a book series that I’m working on (but haven’t touched it in two months). I have a life guide for my kids that is basically done. I could go back and adjust it infinitely, though.
I’ve become more interested in image and movie generation. For the most part, I just want to play around with it and see what’s possible outside of the big companies’ tools. The subscriptions for those add up very quickly when you are making video.
Most of all, I’ve become interested in AI agents. They are more than chat. You can give them things to do, and they go do it. OpenClaw and Hermes Agent are two popular examples. For these, I’d want a separate computer that’s more powerful than my laptop. There’s a reason why Nvidia has one of the largest market caps in the world, and it’s because their GPUs are perfect for this kind of AI work.
When I was looking into buying this desktop, I did what I always do before spending real money, which is research the thing to death. That trait annoys my wife to no end, but it has paid dividends time and time again. Of course, I did this research with an AI, which has a fun circular irony.
Here’s the part that made me laugh at myself. I’ve spent the better part of two decades writing about making money while I sleep. So I decided to look into whether this computer could do just that.
What I Actually Wanted It To Do
I didn’t want a gaming PC, but that’s what I ended up buying. I wanted something specific, and the software decided the hardware.
The big one is running image (and maybe video) generation locally. Flux is an open-weight model you can run yourself through ComfyUI. I’m not sure what I’d do on it that I couldn’t do with Google’s Gemini or OpenAI’s ChatGPT, but it’s something I want to play with. My son is getting interested in learning Blender. Wikipedia describes it as “a free and open-source 3D computer graphics software suite used for 3D modeling, animation, and rendering.” A great graphics card makes it work much, much better.
I also wanted to run reasoning models on my own machine instead of paying API fees by the token. I have read about companies spending their annual AI budget on tokens in a couple of months. This makes my token cost just the cost of energy. And since we have solar power, electricity is very cheap.
Open-source models like gpt-oss-20b and Qwen 3 are typically in the 20 to 30 billion parameter range. To be able to run these, you need a graphics card with about 16GB of memory. The cheapest Nvidia one with that is an RTX 5070 Ti, which is typically about a thousand dollars.
I wanted to try OpenClaw, the open-source agent that blew up earlier this year — the one that lives on your machine, remembers things between sessions, and actually does stuff instead of just answering questions. Hermes Agent seems to be very similar, and I’m looking into that. Get Rich Slick shows a lot of what it can do, and it’s pretty amazing.
The Machine
I went with the MSI Aegis R2 AI (model B2NVV9-1411US), which is a pre-built. I figured I was paying a convenience premium for that, and I’d made peace with it. I thought about building a computer myself, but AI convinced me that it was not going to be a fun project to learn. In fact, it would have probably taken me many months to get the best deals on components.
The important parts:
- Intel Core Ultra 9 285K with 24 hybrid cores – This is a much faster processor than I needed. In fact, it seems to be the fastest consumer processor that Intel makes. While a lot of AI work will be done in the GPU, the main brains will definitely be busy. On one popular speed test, my 2-year-old top-of-line laptop’s processor has a multithreaded CPU Mark of 13,870. This processor’s multithreaded CPU Mark is 67,260.
- NVIDIA RTX 5080 with 16GB of VRAM – This is an upgrade over the 5070 Ti that I mentioned above. Like the processor above, I didn’t need the upgrade. However, it wasn’t a system that I could customize and save money, so I’ll take the upgrade.
- 64GB of DDR5-6000 RAM – Most of the systems came with 32GB of memory, which is fine for gaming. For AI work, the recommended minimum is 64GB. Due to RAMageddon, memory prices are through the roof. It was critical to find this included.
- 2TB NVMe Gen4 drive – I only need a 1TB drive because I set up a NAS. However, I wish I could have gotten a Gen5 drive.
- Other stuff – An 850W power supply, a 360mm liquid cooler, Wi-Fi 7, and Windows 11. These are necessary for running all this powerful equipment. I don’t know anything about the liquid cooler, which is exactly why I didn’t want to build my own machine.
- Intel Core Ultra 9 285K: ~$530 (it launched at $589 and I’ve seen it listed as high as $754)
- RTX 5080: ~$1,246
- Z890 motherboard: ~$200
- 64GB DDR5-6000: about ~$599
- 2TB Gen4 NVMe: ~$275** for a budget drive
- 850W 80+ Gold power supply: about ~$79
- 360mm liquid cooler: about ~$90
- Windows 11 Home license: $139
Their main compromise is the 16GB of VRAM with the card. The big 20-to-30 billion parameter models don’t fit in 16GB. They spill over into system RAM, which is why the 64GB matters. This makes it slower. The RTX 5090 has 32GB and could handle those models. The street price of the RTX 5080 card that I got alone is between $ 1,300 and $ 1,400. The RTX 5090 is around $4,500. The price of the final computer would probably be an extra $3,000 – no thank you.
Can I Make Money with this Machine?
As I was waiting for the machine to arrive, I had a thought. This machine is powerful. Powerful computing is in demand. It’s not a big moat, but not everyone has the equipment or the desire to do AI stuff.
I went to do some research with ChatGPT. I thought about being able to automate Fiverr. It turns out that most people on Fiverr want human work – booooo! AI suggested that I use it to make this blog better. That’s reasonable, except that AI rubbed out the last ember of money in blogging.
Finally, I saw what seemed to be ideal. When the machine is ideal, I could rent the GPU out on a distributed compute network like Salad or GamerHash and collect passive income. It’s a Lazy Man’s dream!
However, I’ve written about enough passive income claims to not get my hopes up too much.
Salad currently pays somewhere around $1.12 a day for an RTX 5080. The card pulls about 360 watts under load, which is roughly 8.6 kilowatt-hours over a full day. At 15 cents a kilowatt-hour, you’re spending about $1.30 in electricity to earn $1.12. While solar power almost makes our electricity free, if we use more than we produce, we have to pay the rates in RI. Last I checked, they were around $0.26 kWh, so even worse.
It said that Vast.ai and RunPod pay more, but the numbers assume your machine actually gets rented. New hosts also start at the bottom of the reputation ranking. It doesn’t seem like it would pay off, but I might give it a try.
No matter what I looked at, it seemed that starting a company was the way to go. I wanted to avoid that, because there’s always a lot of other work with starting a company. Often, there’s no payoff at all. (Kid Wealth, I’m look at you!)
The Cost per Component
I was curious how much the computer would cost if I did buy each component separately. Here’s what AI came up with for the lowest prices:
Add up the absolute cheapest version of every line, and you land around $3,158. That’s before a case.
Originally I bought it for $2750 on Amazon. That’s an affiliate link, and they are currently out of stock. If you go there, you’ll see the catch: it is “Renewed.” That may be a deal-breaker, but I’ve bought a lot of renewed stuff on Amazon and never had a problem. It was sold by Woot, but the warranty checked out. Woot is an Amazon company, so there was significant overlap.
You’ll notice I said “originally” above. My final price was $2,375. Two days after I bought it, but before it shipped, Woot put the EXACT same machine on sale on its website. I was able to cancel the Amazon order and wait a week for the Woot one. When it arrived, the box wasn’t in perfect condition, but that’s the only flaw I can see.
Remember when I gave you the RTX 5080 and 5090 pricing above. Just a few days ago, Digital Trends reported that Nvidia is going to push the price for the 5080 to around $1,845. So the lowest price will be about $600 more than you can get it now. Hopefully, the street price won’t be too much more.
Nvidia realized that they can make more money due to the demand. Also, they make more money selling cards to AI data centers than to consumers. So these cards have quietly become the low-priority product coming out of the same constrained TSMC capacity. We’re not the customer anymore. We’re the leftovers.
At that $1845 price, the graphics card (ALONE!) would cost about 78% of what I spent on the whole computer.
I usually preach patience on big purchases. The rewards go to those who wait for the sale and let the new model depreciate. This is the bizarro world of AI that we are now in.
Get a Similar Deal Yourself
With the prices of components rising, I can’t promise that another deal like this will come around. However, there was one a couple of weeks before that still met my requirements, but had the minimum GPU and CPU that I was looking for. It was about the same $2375 price, but brand new. I didn’t find that deal until it expired, or I would have bought it. It was one of those days that I wasn’t around my computer.
How did I find the deal? I went to Slickdeals.net and added an email alert for “Nvidia.” I got about 3-4 emails a day that weren’t useful. Sometimes they were for a monitor or a laptop. I could have been more specific, such as “Nvidia Desktop,” but I didn’t want to risk the deal being posted without the word desktop in it. I went with the widest net, and it worked.
