AI Companies Refuse Gun Research, So He Built His Own $32K AI
When big AI wouldn't help his gun business, he spent $32,000 on his own machines.
A business owner who runs USA Arms — a site organizing tens of thousands of firearms-related businesses like gun dealers, ammo dealers and ranges — had been using hosted AI to research those businesses and keep the site up to date. About two to three weeks ago, the hosted models started refusing to do normal research: one refused to identify a gun dealer because the license holder might be an individual, and another wouldn't look at an ammo dealer's website at all.
His harness started breaking in random ways, so he built three boxes for about $10,700 each, 128 GB of VRAM apiece and 384 GB total. He tested Luna 6, DeepSeek, GLM 5.3, Qwen 3.8 27B, Gemma, Inkling, Kimi and GLM Flash on his workload.
GLM gave the best results; Qwen gave the best cost-to-performance, so that's what the boxes run. The models check source facts, sort out business identities and pull out things like services and hours, and every answer is verified against what the agent actually saw.
- Major AI services refused to help with gun-related research, breaking a business owner's system.
- He built three custom AI computers for $10,700 each, totaling $32,000—cheaper than a year of cloud AI.
- The machines run open-source AI that never refuses and can also handle software builds.
Why It Matters
Shows how businesses can bypass AI restrictions and save money with local, private AI.