Cheap eBay Hardware Runs Powerful AI Locally

A tech hobbyist has successfully run a powerful AI model on cheap, second-hand hardware originally used for cryptocurrency mining. He bought FPGA chips on eBay for around $280 each and used them to run a 9-billion-parameter AI model called Qwen3.5. The setup achieved speeds of about 2 to 3 tokens per second, which is slow but functional. Tokens are pieces of words—so 2 tokens per second is roughly one to two words per second.

The project matters because it shows that running advanced AI locally is becoming more affordable. Instead of paying for cloud services that charge by usage, you could own the hardware and run AI privately. The hobbyist also estimated that a larger 27-billion-parameter model could run on a more powerful board, potentially reaching 8 to 25 tokens per second with optimizations. That's fast enough for real-time conversation.

The main catch is that this is still a DIY project. It requires technical skills to set up, and the hardware is not plug-and-play. The chips are old mining cards that need custom programming. Also, the speeds are much slower than what you'd get from a commercial AI service. But for someone willing to tinker, it could be a cost-effective way to run AI without relying on big tech companies.

Overall, this experiment hints at a future where powerful AI runs on inexpensive, repurposed hardware in your home, giving you more control and privacy. It's a small step, but it points to a bigger trend of AI becoming more accessible and decentralized.