TL;DR: I built a $1,000 PC to run OpenClaw on my existing Claude/ChatGPT subscriptions. Mid-build, both providers cut off subscription access for agent tools — so a $250 GPU deal turned this into a local-model-hosting project instead.

Why I Built Instead of Bought

In March 2026, there was a lot of buzz around OpenClaw and what it could do. I wanted to see it for myself, and looked at a few ways to run it: a hosted VPS, a Mac mini, a mini PC, or a custom build.

Each had tradeoffs, and any of them would have worked for the experiment. But a VPS is just another subscription, and as a home-lab enthusiast I mostly avoid those — LLM subscriptions are my one exception, since the value is high enough that I already pay for both Claude and ChatGPT. I wasn’t looking to stack a second subscription just to try an agent layer. Mac minis and most mini PCs aren’t upgradable either, which mattered more than usual with memory prices climbing at the time. After weighing cost, current needs, and upgrade room, building my own PC won out.

Picking the Parts

The use case was clear: run OpenClaw now, and leave room to add a GPU for local model hosting later. After digging through Reddit threads, blog posts, and a few rounds of back-and-forth with LLMs, I landed on this build — target budget under $1,000:

Component	Store	Price	Link
CPU — AMD Ryzen 7 9700X	Micro Center (combo deal)	$550 combo	Bundle deal
Motherboard — Gigabyte B650E (the one part I regret)	Micro Center (combo deal)	bundled above	—
Memory — 32GB Corsair DDR5 (2×16GB)	Micro Center (combo deal)	bundled above	—
Storage — 1TB Samsung 990 Pro	Best Buy	$200	Product page
PSU — Corsair RM850x	Best Buy	$125	Product page
Case — Lian Li Lancool 216	Amazon	$100	Product page
Cooler — Thermalright Phantom Spirit	Amazon	$30	Product page

Total: $1,005 + tax — right at the target.

Prices reflect what I paid at the time of purchase (March 2026); deals and listings above may have changed since.

CPU, motherboard, and memory came bundled in a single $550 Micro Center combo deal — here’s the exact bundle I bought.

That completed the original plan: build the PC, install OpenClaw on it in its own space for security, and wire it up to my existing ChatGPT and Claude subscriptions.

Then the Rules Changed

Around the same time, Anthropic and OpenAI both announced that OpenClaw could no longer run on subscription plans — it had to go through the API gateway instead. Even just exploring OpenClaw would now add $10–20/month on top of the LLM subscriptions I already paid for. I’d already made one subscription exception for the LLMs themselves; stacking a second one just to try an agent layer felt like the wrong direction.

I still wanted to test the capability, though. Which raised a new question: could I run everything locally on the hardware I was already building?

The $250 GPU Deal

While the parts were shipping, I found a deal I hadn’t planned around at all — a Sapphire RX 9060 XT 16GB, usually $400+, listed for $250. I grabbed it before I could talk myself out of it. (The universe wanted me to buy it.)

By the time everything arrived, the GPU was in the box too, so I assembled the whole thing at once — about 1.5 hours, since it was my first build, but LLMs and YouTube filled in the gaps.

The GPU also changed the software plan. I’d originally intended to dual-boot Windows and Ubuntu, but after reading up on AMD ROCm, I went with Ubuntu 24.04 only, per the Reddit community and AMD’s own docs.

Software Stack: Ubuntu + ROCm + Ollama

After installing Ubuntu, the first job was getting the GPU working. Fair warning: AMD is not plug-and-play — it took real time and effort, and a fair amount of forum-and-docs archaeology, before I had it sorted. I installed ROCm 7, and the system picked up the GPU correctly.

Next was OpenClaw itself. It wanted to connect to a provider, but I didn’t want to hand it a subscription or API tier — the whole point now was to use the GPU. That meant deciding how to actually host the models locally.

Research pointed me to the easy path: Ollama. Simple, one command, done. I installed it and started looking at which models to run, landing on Llama 3 8B as a starting point. OpenClaw worked — but it was slow. Over the next 2–3 days I tried a handful of different models and sizes, and the conclusion was consistent: only the smallest models were usable on this setup.

More on what “usable” actually looked like — benchmarks, model sizes, and where the RX 9060 XT hit its limits — in Part 2.