Proven AI builds: used hardware with published LLM benchmarks
Each build is a hardware configuration that someone has published language-model benchmarks for, priced against today's used market on eBay. The figures are the sources' own, labeled by who reported them. rackbites links the source and the commit, and hasn't measured any of them itself yet.
| Build | Published results (decode speed, batch 1) | Low-end ask |
|---|---|---|
| 8× V100 32GB SXM2 server Eight Tesla V100 32GB modules on an SXM2 NVLink board: 256GB of VRAM | DeepSeek-V4-Flash: 65.1 tok/s decode GLM-5.3-Flash: 53.0 tok/s decode Qwen3.8-Flash-Next: 80.7 tok/s decode on 4 of the 8 GPUs Qwen3.8-Flash-Next + MTP4: 79.0–126.8 tok/s decode on 4 of the 8 GPUs | $7,127 |
| 4× V100 16GB SXM2 server Four Tesla V100 16GB modules on an SXM2 NVLink board: 64GB of VRAM | Qwen3.8-27B: 65.05 tok/s decode Qwen3.8-27B NVFP4 + DFlash2: 100–142 tok/s decode at a 200K-token context | $6,500 |
| Tesla P100 16GB One Tesla P100 16GB PCIe card in any server with a free slot and airflow | Qwen3.6-35B-A3B: 54–60 tok/s generation (prose); 66–72 (code) | $129 |
A price needs 3 live listings. The low-end ask is the one a shopper sorting by price reaches first: one in ten listings ask that or less (under ten listings, it's the cheapest). Each build's page has the setup and workload behind every figure, what to check before buying, and the cheapest live listings. Card prices by model: GPU prices · CPU prices.