The AI Workbench banner for Roger's Rabbit Hole: the RRH rabbit in a red apron installs a graphics card in an open PC tower whose power cord runs into a glowing rabbit hole
Welcome To The AI Workbench

The AI Workbench

Agents, local AI, open models, the platforms worth paying for and the projects worth stealing ideas from. Tested on my own bench, mistakes included.

I've been in tech since 1995, and I've watched plenty of "this changes everything" come and go. Lately, most of what's on my bench is AI. I run my businesses with a staff of AI agents. I host my own AI on hardware I put together myself. And I keep a pile of open models downloadable for anybody who wants them.

Most of what you read about AI is a demo, a sales pitch or both. This section is the other version. I tell you what I tried, what it's actually good for, what it costs and what broke. When something doesn't live up to the hype, I'll say so. When something does, I'll say that too, and if I can make a buck off a link, I'll tell you before you click.

What's on the bench

Agents at Work

My businesses run with a staff of AI agents, each with one job, a written persona and a list of things it's never allowed to do without me. Yes, they're named after characters from HBO's Silicon Valley. Between them they cover writing, art, video, research and operations. The one rule: agents propose, I approve. Nothing gets published, emailed, bought or deleted until I say yes myself. This is also where you'll find Field Notes, short real stories from the work, mistakes and all.

Local AI

CyberTron is my homebuilt local AI server. It started as a $150 computer and turned into a $592.33 workstation. It runs open-source tools like Ollama and Open WebUI, plus the Kokoro text-to-speech voices that narrate our YouTube Case Files. I picked that narrator blend by ear, and there's no per-minute voice bill. The whole build story lives in Building CyberTron.

Open Models

Hugging Face is the big hub for open models, and I pull a lot from it. I also help keep them downloadable through PirateFace, an outside project that turns eligible open models into torrents. My agents download and verify the files, and I run every submission myself.

Platforms

Which tool is actually for what, compared from my own use: agent platforms, local versus paid voices, model hosts and the rest. If I haven't used it, it doesn't get a verdict.

Projects

Code worth a look, mine and other people's. My own is the CyberTron Agentic Stack, a local-first agent memory project I forked from Agentic Stack and extended. It's public on GitHub. Worth a Look roundups tell you plainly whether I ran something or just read about it.

Tools and Products

The hot new stuff, put through a Bench Test: what it is, what I tested, what worked, what didn't, who it's for and what it costs. Some of these I may recommend or resell. When that's the case, you'll see it right at the top, and it doesn't change the verdict.

Start here

Nobody paid for this section. I use Grok Bot, Hugging Face and plenty of other tools, but I'm not a spokesman for any of them, and nothing here is official from xAI, X, SpaceX, Cursor, Hugging Face, GitHub or anybody else.

Latest From The Workbench
Building CyberTron — So What Did the $150 AI Server Really Cost?

Building CyberTron — So What Did the $150 AI Server Really Cost?

A full breakdown of the upgrades that turned a $150 computer into a $592.33 local AI workstation.

Author: Roggie D/Saturday, August 22, 2026/Categories: Building CyberTron, Build Journal, Hardware & Upgrades

Rate this article:
No rating
TECH · Article 87

Building CyberTron — So What Did the $150 AI Server Really Cost?

CyberTron started with a $150 computer.

That statement is still completely true.

It is also becoming increasingly misleading.

When I bought the original machine on August 4, 2026, I wasn’t buying a finished AI server. I was buying a starting point.

For $150, I got an older AMD FX-era desktop with an eight-core processor, 16 GB of DDR3 memory, a 1 TB mechanical hard drive, a 500-watt power supply, and an NVIDIA GTX 1060 with 6 GB of VRAM.

It wasn’t new.

It wasn’t fast by modern desktop standards.

And it certainly wasn’t designed with local artificial intelligence in mind.

But it worked.

Ubuntu Linux ran on it.

Ollama ran on it.

Open WebUI ran on it.

Local AI models ran on it.

And that was enough to prove the basic idea behind this entire project:

You don’t necessarily need an expensive new computer to start experimenting with local AI.

But then I started doing what I said I was going to do from the beginning.

Run the machine.

Find the bottleneck.

Understand the bottleneck.

Fix it.

And CyberTron had bottlenecks.

A lot of them.

The $150 Starting Point

The original CyberTron cost me exactly:

$150.00

That bought the entire functioning computer.

Case.

Motherboard.

AMD FX processor.

16 GB of DDR3 RAM.

1 TB mechanical hard drive.

500W power supply.

GTX 1060 6 GB graphics card.

And all of the miscellaneous pieces that make an old computer an actual computer instead of a pile of parts sitting on a workbench.

That matters when looking at the economics of the project.

If I had decided to build something from individual components, I would have needed a motherboard, processor, case, cooling, storage, RAM, power supply, GPU, cabling, and all the little odds and ends that tend to magically add another hundred dollars to a “cheap” computer build.

The $150 machine already had all of that.

More importantly, it booted.

That gave CyberTron a huge head start.

Unfortunately, it also had 16 GB of RAM.

And that became bottleneck number one.

Upgrade Number One — 32 GB of RAM

On August 6, 2026, I ordered:

Patriot Signature 16 GB — 2x8 GB DDR3-1600 PC3-12800 Desktop RAM Model: PSD38G16002

Grand total:

$25.00

That brought CyberTron from 16 GB to 32 GB of system memory.

Running local AI isn’t only about GPU memory.

The operating system needs memory.

Docker needs memory.

Open WebUI needs memory.

Databases need memory.

Containers need memory.

Models being loaded, manipulated, or partially offloaded can consume memory.

And then there are all the other things I inevitably decide to install because apparently simply having a functioning computer isn’t complicated enough.

For twenty-five bucks, doubling the system memory was an easy decision.

CyberTron was now sitting at:

$175.00 total investment

And honestly, at that point the project was still ridiculously cheap.

Then I started looking harder at the GTX 1060.

The Upgrade That Changed CyberTron’s Brain

The GTX 1060 wasn’t a bad graphics card.

In fact, it turned out to be much more useful than I originally expected.

It had 6 GB of VRAM, CUDA support, and it was perfectly capable of proving that GPU-accelerated local AI would work on CyberTron.

But six gigabytes of VRAM becomes restrictive very quickly once you start playing with larger models.

I could run AI.

The problem was that I wanted to run more AI.

So on August 11, 2026, I ordered the biggest single upgrade of the entire project:

Zotac GeForce RTX 3060 Twin Edge OC 12 GB GDDR6 Model: 9288-5N630-301Z8

Grand total:

$285.04

That one purchase changed CyberTron more than anything else I had done.

The RTX 3060 doubled available GPU memory from 6 GB to 12 GB.

For gaming in 2026, an RTX 3060 isn’t exactly the hottest GPU sitting on the shelf.

For what I was trying to accomplish, however, those 12 GB of VRAM were far more important than having the newest GPU architecture.

CyberTron wasn’t being built to chase frame rates.

It was being built to run local AI models.

And VRAM is workspace for the AI.

The RTX 3060 gave CyberTron a much bigger desk.

After the upgrade, I was able to run Qwen3 8B entirely on the GPU and experiment with larger context windows.

I also ran Qwen3 14B entirely on the GPU, using roughly 9.6 GB of VRAM.

That was the point where CyberTron stopped feeling like an old computer that happened to run AI.

It was becoming an actual local AI workstation.

But there is another part of the RTX 3060 purchase that doesn’t show up if I simply list the CyberTron parts.

The GTX 1060 didn’t go in the trash.

One GPU Purchase — Three Computers Upgraded

This is where the economics of upgrading old hardware get interesting.

When the RTX 3060 went into CyberTron, the GTX 1060 6 GB that came out of CyberTron was still a perfectly functional GPU.

So I moved it into another workstation.

That machine previously had an NVIDIA graphics card with 4 GB of VRAM.

The GTX 1060 was an upgrade for that machine.

And yes, that computer actually gets used.

I play GTA V on it.

So while CyberTron gained a 12 GB RTX 3060 and significantly increased its AI capability, my everyday desktop also got a graphics upgrade.

Then the 4 GB NVIDIA card that came out of that workstation was moved into yet another functional workstation.

Think about what happened there.

I purchased one graphics card.

CyberTron: GTX 1060 6 GB → RTX 3060 12 GB

Workstation number two: NVIDIA 4 GB card → GTX 1060 6 GB

Workstation number three: Received the displaced NVIDIA 4 GB card

One $285.04 purchase resulted in hardware improvements across three separate computers.

That is one of the benefits of doing this project with older hardware that tends to disappear when you look only at the final CyberTron price.

The RTX 3060 wasn’t just a CyberTron expense.

It started a hardware cascade.

Nothing got thrown away.

Nothing immediately became e-waste.

Each card simply moved down the chain to another machine where it could still be useful.

And CyberTron got the biggest brain increase of the bunch.

With the GPU included, CyberTron’s project total had now reached:

$460.04

Our little $150 AI server was definitely starting to stretch the definition of “$150 AI server.”

And I still wasn’t finished.

More GPU Means More Power

Dropping an RTX 3060 into an old AMD FX-era desktop without looking at the power supply would have been a particularly creative form of stupidity.

The original machine had a 500W power supply.

Could I have tried running everything that way?

Probably.

Was I going to trust an aging power supply of unknown history with an RTX 3060, an old FX processor, multiple drives, and a system I intended to leave running as an AI server?

No.

On August 11, 2026, I also ordered:

be quiet! Pure Power 12 650W PSU 80 Plus Gold ATX 3.1 PCIe 5.1 GPU Support

Grand total:

$62.30

This wasn’t the glamorous upgrade.

Nobody installs a power supply and suddenly gets twice the AI performance.

But infrastructure people understand this one.

Capacity matters.

Headroom matters.

Reliability matters.

And just because something appears to work doesn’t mean it is engineered correctly.

The 650W be quiet! unit gave me additional power capacity, better efficiency, modern GPU support, and significantly more confidence in the electrical foundation of the machine.

That brought the CyberTron investment to:

$522.34

Then there was one final piece of old technology still slowing down the entire machine.

The hard drive.

Goodbye, Spinning Rust

CyberTron originally booted from a 1 TB mechanical hard drive.

And technically it worked.

Linux booted.

Applications loaded.

Docker ran.

Models ran.

Everything functioned.

Slowly.

Very, very slowly.

This was especially noticeable because the rest of the machine had gotten faster.

The additional memory reduced pressure on the system.

The RTX 3060 dramatically improved AI inference.

And sitting underneath all of it was still a mechanical disk moving a physical read/write head across spinning platters.

It had to go.

On August 10, 2026, I purchased:

Patriot P210 512 GB Internal SSD SATA III 2.5-inch Solid State Drive Model: P210S512G25

Grand total:

$69.99

The Patriot SSD became CyberTron’s primary system drive.

The old 1 TB mechanical drive didn’t get discarded either.

It was reformatted and repurposed as secondary data storage.

So once again, an upgrade didn’t necessarily mean throwing the old hardware away.

The SSD handled the operating system and primary workload.

The mechanical drive became bulk storage.

Not elegant.

Not cutting edge.

But functional.

And if you saw the installation from the last article, you already know that “functional” is sometimes the design standard around here.

The SSD brought the final CyberTron project cost to:

$592.33

And there it was.

The answer.

My “$150 AI server” actually cost me just under six hundred dollars.

The CyberTron Cost Breakdown

Here is exactly where the money went:

Original CyberTron PC — August 4, 2026 $150.00

Patriot Signature 16 GB DDR3 RAM Upgrade — August 6, 2026 $25.00

Patriot P210 512 GB SSD — August 10, 2026 $69.99

Zotac GeForce RTX 3060 Twin Edge OC 12 GB — August 11, 2026 $285.04

be quiet! Pure Power 12 650W PSU — August 11, 2026 $62.30

TOTAL CYBERTRON INVESTMENT: $592.33

Those are not advertised prices.

Those are not “before shipping” numbers.

Those are not numbers that conveniently forget sales tax.

Those are my actual out-the-door costs.

The original computer accounted for only $150 of the final $592.33.

I spent another:

$442.33

upgrading it.

Which brings us back to the uncomfortable question.

Did I actually save any money?

So Did I Actually Save Money?

Yes.

With an asterisk.

A very large asterisk.

If I had walked into this project knowing from day one that I wanted an RTX 3060 12 GB, 32 GB of memory, solid-state storage, and a reliable 650W power supply, the technically superior decision probably would have been to start with a newer platform.

I would have gotten a faster CPU.

DDR4 or DDR5.

NVMe storage.

Newer PCI Express connectivity.

Lower power consumption.

Better long-term upgrade options.

And significantly less archaeological computing.

But that’s not how CyberTron happened.

I spent $150 because I wanted to see what the machine could do.

Then I found the memory bottleneck.

Twenty-five dollars fixed it.

Then I found the GPU bottleneck.

The RTX 3060 fixed it.

That required me to reconsider the power supply.

So I fixed that too.

Then the mechanical hard drive became painfully obvious.

So the SSD went in.

Each upgrade happened because the previous step exposed the next limitation.

That was the experiment.

CyberTron wasn’t designed on PCPartPicker and assembled in one afternoon.

It evolved.

And that evolution is the entire point of the project.

The Other Thing the Price Doesn’t Show

There is also something missing from the $592.33 total.

Experience.

By building CyberTron this way, I had to actually understand what was happening inside the system.

I watched memory limitations.

I dealt with GPU VRAM constraints.

I tested model sizes.

I experimented with context windows.

I migrated Linux from a mechanical disk to an SSD.

I repurposed the old drive.

I replaced the power supply.

I moved GPUs between three different machines.

I installed and configured Ollama.

I deployed Open WebUI.

I worked through Docker services.

I tested local models.

I learned exactly what this old machine could — and could not — do.

If I had spent more money on a modern prebuilt computer and installed Ollama, I would have gotten a faster machine.

But I would have skipped half the experiment.

And this project was never only about owning an AI server.

It was about building one.

From $150 Computer to $592.33 AI Workstation

So here is where CyberTron stands.

What started as a $150 used desktop is now running:

AMD FX eight-core processor

32 GB DDR3-1600 system memory

NVIDIA GeForce RTX 3060 12 GB

Patriot P210 512 GB SATA SSD

1 TB secondary mechanical hard drive

be quiet! Pure Power 12 650W 80 Plus Gold power supply

Ubuntu Linux

Docker

Ollama

Open WebUI

And an expanding collection of local AI services, models, experiments, and questionable ideas.

Total cash invested:

$592.33

Is CyberTron the fastest AI workstation you can build in 2026?

Not even remotely.

Is it the most power-efficient?

Have you met an AMD FX processor?

Is it the machine I would recommend somebody duplicate component-for-component?

Probably not.

But can a machine that cost less than $600 run meaningful local AI workloads entirely on its own hardware?

Absolutely.

I’ve already proven that.

And that’s where this project becomes more interesting than the price of the parts.

Because now that the hardware foundation is mostly sorted out, I can stop asking whether CyberTron can run local AI.

It can.

The next question is:

What can I actually make it do?

Because having an AI server sitting in the corner running models is cool.

Having one that actually performs useful work is something else entirely.

And that is where we’re going next.

CyberTron has the hardware.

CyberTron has the models.

Now it’s time to start giving CyberTron a job.

Print

Number of views (130)/Comments (0)

Roggie D
Roggie D

Roggie D

Other posts by Roggie D
Please login or register to post comments.

Contact author

x

 

Follow Us on FacebookFollow Us on XFollow Us on InstagramSubscribe to our Youtube ChannelSubscribe to our Rumble Channel