For the past several months, I’ve been leaning heavily into AI to help process the overwhelming amount of information that passes through my computer. Between FOIA responses, old image libraries, government meeting packets, audio recordings, videos, spreadsheets, various platform exports, websites, and historical documents, the volume of material to review has grown to the point where automation is no longer just helpful. It’s becoming essential.

As I’ve expanded those workflows, I’ve also watched another number steadily increase: token costs.

I’ve spent a lot of time optimizing how I use OpenAI. I’ve changed models depending on the task, redesigned prompts, broken work into smaller pieces, and become much more selective about what information is actually worth sending to a language model. My processing pipeline already has numerous safeguards that monitor costs, skip overly expensive jobs, and stop processing once predefined thresholds are reached. Rather than trying to finish enormous jobs in a single pass, I chip away at them over multiple runs so smaller work can continue moving forward while larger jobs gradually make progress.

Even with all of those improvements, the cost of processing thousands of documents starts to add up.

A local business owner has mentioned OpenClaw to me a couple of times over the past few months. That recommendation finally pushed me to spend some time researching what running AI locally looks like today. The more I looked into it, the more it seemed like the next logical step. I started looking into the specifications of what I would need, and it seemed like I would need to purchase a gaming machine, but then add a ton of RAM. Ironically, I already own a machine that’s almost perfect for the job.

Several years ago, the SSD in my gaming computer failed, forcing me back onto a traditional hard drive. Since then, I’ve done nearly all of my work on my MacBook because the old desktop became painfully slow. Additionally, the motherboard doesn’t include the security hardware required for Windows 11, so the computer has mostly been collecting dust.

It turns out that an aging gaming computer with a dedicated graphics card is still a very capable AI workstation.

Instead of replacing the entire computer, I decided to invest in a new SSD. The cost of the drive is small compared to what I’ve been spending on API tokens, so it feels like a worthwhile investment. Once it arrives, the plan is to install Ubuntu, Ollama, and OpenClaw, allowing many of my existing AI workloads to run entirely on my own hardware.

I’ve experimented with Ollama before on a Raspberry Pi, but inexpensive hardware simply isn’t powerful enough for anything beyond very small models. This desktop should be in a completely different class.

The nice thing is that my workloads don’t require instant responses. Most of my processing happens overnight while I’m asleep. Documents are already broken into small tasks, queued, and prioritized so the system can continuously make progress. Speed isn’t nearly as important as cost, consistency, and the ability to process large volumes of information without worrying that every request will consume additional tokens.

If everything works as expected, I should be able to point many of my existing AI API calls at my own local machine instead of cloud services. For routine extraction, classification, entity recognition, summaries, and structured analysis, that could dramatically reduce operating costs. Then, for the handful of tasks that genuinely benefit from the newest cloud models or deeper reasoning, I can still send those final jobs to OpenAI.

In other words, AI won’t disappear from my workflow. It will simply become another tool running in my own home office.

The SSD is scheduled to arrive on Sunday. After that comes installing Ubuntu, configuring Ollama, learning OpenClaw, and seeing how well everything fits into the processing pipeline I’ve been building over the past year.

If it works the way I hope, this could become one of the biggest improvements I’ve made to my research workflow. Not because it’s faster, but because it allows me to process more information, more often, while keeping costs under control.

I’ll be sure to post updates as the experiment progresses.