Repurposing mining racks for AI inference: Don't just look at cost

Title: Repurposing Mining Racks for AI Inference – Don't Just Look at the Price

I spent the weekend tinkering with converting retired mining racks into AI inference nodes and hit quite a few pitfalls. By "AI inference node," I mean servers dedicated to running model predictions. If you're just a small lab trying to save money on protein structure prediction, it's probably not worth it; but if you happen to have stable power, a server room, and people who can fix servers, this approach is worth copying. What's truly valuable here is electricity, cooling, and reliable delivery capability.

Reports say many Bitcoin miners in Texas are pivoting to AI data centers. Hut 8 secured debt financing to expand its AI facilities, and Core Scientific abandoned its mining plans in Denton to switch to AI. It sounds like you could just change the sign and open for business, but after racking up three retired cabinets, I realized that getting them to run is only the first step; keeping them running daily is where the trouble lies.

Our team has been using a GPU cluster for less than a month. Recently, we connected several server racks retired from mining farms to the network, preparing to run AlphaFold3. A GPU cluster is essentially a bunch of GPU servers computing together.

AlphaFold3 is a protein structure prediction model. You feed it amino acid sequences, and it tries to guess how the protein folds in space. The gap between dry experiments (computational) and wet experiments (lab-based) often gets stuck right here. The prediction results look pretty, but they still need experimental verification afterwards.

Before booting up, I thought it would be a bargain. The racks, fiber optics, and power distribution were all ready-made, much faster than building a server room from scratch. But we got stuck on day one. The power policies in the old machines' BIOS were still set to mining mode, and GPU detection was unstable. Running nvidia-smi—the command to check GPU status—would sometimes work and sometimes not. We swapped cables, flashed firmware, and reconfigured network cards, struggling until nightfall. Permission configuration couldn't be skipped either; one service account had write access to the model cache directory, and accidentally deleting half the cache stopped all tasks. Before even seeing how the model performed on biological data, we got a lesson in operations maintenance.

Heat dissipation was the second pitfall. Mining rigs previously focused mainly on blowing air over the machines themselves. When AI inference runs at full load, servers and network devices heat up together. Reports state that AI data centers require more advanced cooling solutions, such as chip-level liquid cooling (using liquid directly against the chips for heat dissipation) plus additional air cooling. Redundancy requirements are also stricter, meaning systems must keep running even if components fail. We didn't have liquid cooling, only airflow channels, so long-running tasks suffered significant clock throttling. The biggest surprise was throughput. For batch inference on short sequences, running several machines concurrently was more stable than using a laptop plus cloud APIs. However, the output format must be stable; otherwise, it's worse than writing code manually. We eventually enforced fixed field formats. I used field validation scripts for three weeks before daring to integrate them into the workflow.

This approach suits those with electricity, server rooms, and people who can fix machines. It's not suitable for individual hobbyists, small research groups, or anyone expecting cheap compute power. The narrative of mining farms transitioning to AI sounds like democratizing compute power, but in reality, it's more like bringing a pile of messy operational headaches home. Next week, I plan to specifically test recovery time after a power outage restart.

---

📌 This article is compiled from Hacker News. Original source: https://americanbuildout.com/goodbye-bitcoin-hello-ai-data-center/

Copyright belongs to the original authors. This is a compilation and independent analysis based on public reports.