Field Report ④: An old mining PC becomes an in-house "private ChatGPT" in two days

AI case studies

"We want to use ChatGPT for work — but we can't send customer data or our own code to an external AI service." This is a worry we hear at many companies. This time, we tooka single used PC left over from crypto miningthat had been sitting in storage, and reincarnated it as an "in-house private ChatGPT" that every employee can use. Additional hardware investment was almost zero. The build was led by an AI agent (Claude), taking2 working days.

The raw material: "four outdated, used GPUs."

We used a PC that had finished its duty in the mining boom of a few years ago. The GPUs on board were four AMD Radeon Vega 56/64 — not the latest AI-oriented GPUs, and not even made by NVIDIA. This is material that falls outside the conventional wisdom of what an "AI server" is.

But the four together add up to 32 GB of video memory. Homing in on that, the AI agent designed an architecture that works around the constraint of not being able to use the AI-standard approach (CUDA) by relying instead on the open graphics/compute API "Vulkan."on par with 2024's GPT-4o. We succeeded in running three AI models for different purposes at the same time, including an open model reputed to be on par with 2024's GPT-4o (on the order of 30 billion parameters).

• For everyday chat (7B): 43 tokens per second — a perceived speed on par with ChatGPT
• For advanced code generation and design consultation (30B class): 30 tokens per second — the smarts of GPT-4o (2024) class
• For in-editor code completion (1.5B): 127 tokens per second — input completion like GitHub Copilot

The interface is a web UI that looks just like ChatGPT. Employees simply open a browser and can switch between models as they use it. Engineers also have Copilot-style auto-completion set up in VS Code.All processing is completed in-house, and not a single byte of data leaves the premises.

The AI agent was the lead builder; humans stuck to making decisions.

This build was AI-driven development as we practice it. OS setup, building the inference engine, designing network isolation, investigating failures — it was the AI agent that did the hands-on work. What humans handled was only the key decisions and the physical tasks (such as re-seating cables).

Worth special mention is the failure handling. During the build, one of the old GPUs kept malfunctioning and repeatedly dropping its connection. The AI agent identified the faulty part from the kernel log and guided the swap by indicating "which physical slot's card it was" via fan spin. It even built ina design where the network won't drop even if a GPU fails in the future(fixed MAC addresses), as well asmonitoring that detects anomalies every 5 minutes and recovers automatically. This server is now in a "leave-it-alone" state where simply powering it on auto-starts every service.

Security assumes "you will be breached."

We published it to the internet so it can be used from outside the office too, but the design assumes "even if breached, the damage won't spread." The server is isolated on a dedicated network (DMZ) and configured so it cannot reach any other in-house system at all. Accounts require administrator approval, and all communication is encrypted.A server where, for an attacker, "there's nothing there even if you get in".

What it cost, and what we got.

• Additional hardware investment:almost zero(reusing an existing idle mining PC)
• Build time:2 working days(including failure handling and security design)
• Running cost: electricity only (already made power-efficient via GPU power limiting).API usage fees are a flat zero yen, unlimited
• What we got: an in-house private AI chat that all employees can use, plus a code-completion platform for developers

Even without the latest GPUs, even without NVIDIA, a "production-grade in-house AI platform" can be built depending on how you go about it. If anything, the essence is, faced with hardware full of constraints,quickly finding the right architecture and building in a mechanism that gets back up even when it falls — and that turned out to be exactly the kind of work an AI agent excels at most.

"Our idle machine" might become an AI platform too.

"We've given up on using AI because we can't send data outside." "We can't predict the API costs of cloud AI." "We have spare servers or gaming PCs." If even one of these applies to you, an in-house AI platform is closer than you think.

Series: "Legacy to AI — Field Reports"

• Overview: 22 systems and 246,000 lines on a single PC — a measured record of 3.5 months of AI development
• Field Report ①: Fully modernizing 20-year-old Java in about a week
• Field Report ②: Adding a new entry point to a core system with zero lines of existing code changed
• Field Report ③: A month-scale architecture study, done with a 2-day prototype + real measurements
• Field Report ④: An old mining PC becomes an in-house "private ChatGPT" in two days (this article)

* The figures in this article are measured values as of August 2026. Build outcomes vary depending on the state and requirements of the hardware.