Every host for Qwen3.8 27B
| Host | Blended | Input | Output | Cached | Precision | Uptime |
|---|---|---|---|---|---|---|
| Tenstorrent | $0.14 | $0.02 | $0.50 | $0.009 | fp4 | 97.6% |
| Chutes | $0.342 | $0.039 | $1.25 | $0.0039 | fp8 | 99.9% |
| Near AI | $0.368 | $0.04 | $1.35 | $0.018 | fp8 | 99.1% |
| Ionstream | $0.518 | $0.04 | $1.95 | $0.02 | fp8 | 98.6% |
| Darkbloom | $0.576 | $0.035 | $2.20 | $0.018 | fp4 | 98.7% |
| Reka | $0.58 | $0.04 | $2.20 | $0.02 | fp8 | 100.0% |
| DeepInfra | $0.581 | $0.15 | $1.88 | $0.038 | bf16 | 99.4% |
| Modal | $0.605 | $0.04 | $2.30 | $0.02 | — | 99.9% |
| Wafer | $0.62 | $0.06 | $2.30 | $0.059 | fp8 | 99.0% |
| Dekallm | $0.655 | $0.04 | $2.50 | $0.02 | — | 99.5% |
| AkashML | $0.664 | $0.225 | $1.98 | $0.05 | fp8 | 100.0% |
| Parasail | $0.73 | $0.24 | $2.20 | $0.05 | fp8 | 99.7% |
| Mancer 2 | $0.775 | $0.20 | $2.50 | — | fp8 | 98.2% |
| Modelrun | $0.775 | $0.20 | $2.50 | $0.09 | fp4 | 98.5% |
| Phala | $0.775 | $0.20 | $2.50 | $0.05 | — | 97.8% |
| Alibaba Cloud | $0.956 | $0.425 | $2.55 | $0.085 | — | 100.0% |
| CoreWeave | $1.05 | $0.40 | $3.00 | $0.15 | fp8 | 100.0% |
| Weights & Biases | $1.05 | $0.40 | $3.00 | $0.15 | — | — |
| Novita | $1.07 | $0.42 | $3.00 | $0.085 | — | 99.8% |
| Nebius | $1.09 | $0.45 | $3.00 | — | — | — |
| Cerebras | $1.12 | $0.99 | $1.49 | $0.99 | fp16 | 98.5% |
| Cloudflare | $1.14 | $0.45 | $3.20 | $0.05 | — | 99.2% |
| Venice | $1.14 | $0.45 | $3.20 | — | fp8 | 100.0% |
| Groq | $1.60 | $0.80 | $4.00 | — | — | — |
Cheapest price over time
- Input
- Output
What your workload costs
- Tenstorrent$56.00
- Chutes$136.70
- Near AI$147.00
- Ionstream$207.00
- Darkbloom$230.50
- Reka$232.00
- DeepInfra$232.50
- Modal$242.00
- Wafer$248.00
- Dekallm$262.00
- AkashML$265.50
- Parasail$292.00
12 more hosts cost more.
Rent GPUs or pay per token?
- Renting the GPUs
- $16,878/mo
- Break-even volume
- 121B tokens/mo
- Needed around the clock
- 45.9K tok/s
8 × H100 cost $16,878 a month. The cheapest API (Tenstorrent) charges $0.14 per 1M blended tokens, so the GPUs only pay off above 121B tokens a month: the box must serve 45.9K tokens a second, every second.
GPU price: cheapest on-demand offer per GPU-hour × 730 hours. No throughput is assumed: it depends on the model size, precision, batch size and serving software, so measure yours. Power, storage, egress and engineering time are not included.
LMArena rating over time
Rated about the same, by price
Quality data: LMArena (CC BY 4.0), Epoch AI (CC BY 4.0). Prices: OpenRouter, LiteLLM (MIT), provider pages. List prices in US dollars before discounts and taxes; other currencies converted at the European Central Bank reference rate. The figures on these pages are free to reuse under CC BY 4.0 with a link to this page.