What your GPU could earn.
Every row is a model deployment. Every column is a GPU cluster configuration. The number inside is the upper bound of what one cluster of that shape can earn per day serving that model on UOMI.
Pick a model. Pick a GPU.
How to read. Each cell shows the maximum daily revenue one cluster of that GPU configuration can earn serving the model on UOMI. Background color encodes the value on a log scale: pale gold = low, deep red = high, green = exceptional. The ★ marks each row's best configuration.
| Model | A 2×4090 | B 4×4090 | C 2×5090 | D 4×5090 | E 2×L40s | F 4×L40s | G 1×Pro6K | H 2×Pro6K |
|---|---|---|---|---|---|---|---|---|
| Qwen3.6 27B | up to$49.7 | up to$74.5 | up to$74.5 | up to$149★ | up to$49.7 | up to$74.5 | up to$74.5 | up to$149★ |
| Gemma 4 31B | up to$15.2 | up to$22.8 | up to$25.8 | up to$36.9★ | up to$13.6 | up to$22.8 | up to$25.8 | up to$45.5 |
| GLM 5 | — | — | — | — | — | — | — | up to$39.7★ |
| MiniMax M2.5 | — | — | — | — | — | up to$21.4 | — | up to$37.5★ |
| Gemma 4 31B | up to$17.6 | up to$29.3 | up to$29.3 | up to$44★ | up to$17.6 | up to$29.3 | up to$29.3 | up to$44 |
| Gemma 4 26B A4B | up to$34★ | up to$34 | up to$34 | up to$34 | up to$34 | up to$34 | up to$34 | up to$34 |
| DeepSeek V4 Flash | — | — | — | — | — | up to$10.7 | — | up to$21.3★ |
| Qwen3.5-9B | up to$10.5 | up to$21★ | up to$21 | up to$21 | up to$10.5 | up to$21 | up to$21 | up to$21 |
| Qwen3.6 35B A3B | up to$90.3 | up to$136 | up to$136 | up to$136 | up to$67.8 | up to$136 | up to$136 | up to$271★ |
| Qwen3.5-35B-A3B | up to$79 | up to$119 | up to$119 | up to$237★ | up to$79 | up to$119 | up to$119 | up to$237 |
| MiniMax M2.5 | — | — | — | — | — | up to$21.3 | — | up to$35.5★ |
| Llama 3.3 70B | up to$2.36 | up to$13.7 | up to$3.97 | up to$21.6 | up to$8.39 | up to$13.7 | up to$15.1 | up to$25.2★ |
| DeepSeek V4 Flash | — | — | — | — | — | up to$8.86 | — | up to$17.7★ |
| Gemma 4 26B A4B | up to$48 | up to$48 | up to$96★ | up to$96 | up to$48 | up to$48 | up to$96 | up to$96 |
| gpt-oss-120b | up to$4.75 | up to$9.5 | up to$6.33 | up to$9.5 | up to$9.5 | up to$19★ | up to$9.5 | up to$19 |
Why 2× RTX Pro 6000 dominates. With 96 GB on a single GPU, a 1T-parameter MoE that needed 28 clusters per instance on 2× 4090 only needs 1, and the interconnect penalty stops compounding. The bigger the model, the more bigger GPUs pay off.
Caveats. Throughput estimates are batched-aggregate at FP8 with sublinear scaling for multi-cluster sharding (interconnect_factor = 1/√(1 + 3.0·(N−1))) calibrated for public-internet latency (~50 ms RTT, ~500 Mbps effective). A provider running entirely within one cloud region with private peering would see closer to K=0.5; the K=3.0 used here is for cross-region operator distribution. Single-cluster numbers (where N=1) are unaffected and most reliable.
Plug in. Get paid.
80% of every dollar earned lands directly in the wallet of the GPU that served the request. The other 20% buys back $UOMI on the open market.