Every AI experiment I've wanted to run in the last couple of years has started with the same question. How much is this going to cost me?
I bought an NVIDIA DGX Spark to stop asking it.
Link to Why I bought itWhy I bought it
Most of my work is with AI models. Testing them, fine-tuning them on my own data, building things on top of them.
That usually means renting a GPU by the hour from a cloud provider. It works. But the meter is always running, and I've skipped experiments I was curious about because I didn't want to pay.
I want two things out of this machine.
The first is a place to run open-weight models that belongs to me. Not everything moves here. I'll keep using cloud providers for plenty of work. But when I want to pull down an open model and take it apart, I'd rather do that on hardware I own, with my data staying where it already is.
The second is that I learn this stuff by doing it. Fine-tuning, hosting a model myself, keeping it running. I want to know what that takes for one person, and what the same problem looks like for a lab doing it at a scale I'll never touch. Reading about it only gets me so far.
Link to What it doesWhat it does
It runs the model on my desk instead of in somebody else's data center.
Memory is what makes that possible. A graphics card in a gaming PC has its own memory, usually somewhere between 8 and 32 GB, and if a model doesn't fit in there you can't run it.
The Spark has 128 GB, and the whole machine shares it. Models a gaming PC can't even open fit here with room left over.
It cost about $5,000 delivered. That's expensive. I've also already spent over $1,000 on Modal this year renting GPUs for training runs, and that number was only going up. This one is paid for.
Link to What I'm running on it firstWhat I'm running on it first
Qwen3.8, a 27 billion parameter model. Something that size normally sits behind an API where you pay per token.
It's slower than ChatGPT. You can watch it type.
What I get back is that everything stays on the box. My prompts, my logs, all of it. Running it more costs me electricity and nothing else.
I also put it on Tailscale, which is a private network between my own devices. So I can open it from my phone in any country and it's still the machine in my house answering. More on how that works in the next post.
Link to What's nextWhat's next
The real reason I bought it is my Kalenjin language work. Kalenjin is my first language, and it wasn't one of the languages most speech models were trained on. Getting a model to understand it means running the same training job again and again, changing one thing each time. Every one of those runs costs money when you're renting.
Give it a few months and I should have a real answer on whether buying this beats renting. I'll write it up as I go.