Rail fees, not a GPU landlord
Cautus Intelligence is a rail — gate, meter, Litheum settle. We do not rent you GPUs. Platform take is flat across model class (2.5–6%). Class only moves the creator royalty on matched jobs. Access policy is a control, not a surcharge. One take per call; chain fees are absorbed.
Buyer brings their own GPUs to a listed model. Platform take is the lowest rail fee.
Same host brought the model and the GPUs. Remainder settles 100% to that host.
Model listed; compute from the pool or a third party. Remainder splits by model class.
Listing GPUs is free. No second take. They earn the compute share of matched jobs.
Remainder after the rail take
Same host brought both, or the buyer already paid their GPUs: 100% of remainder to the model owner. Matched jobs (model_only + pool / third-party compute) split the remainder by class:
| Model class | Model owner | Compute provider |
|---|---|---|
| Stock | 20% of remainder | 80% of remainder |
| Fine-tune | 30% of remainder | 70% of remainder |
| Specialist | 40% of remainder | 60% of remainder |
Stock = base / open weights. Fine-tune = continued train on a named base. Specialist = proprietary / special datasets. Shown on every passport.
Worked $1.00
- Platform
- $0.05
- Host
- $0.95
- Compute
- —
- Platform
- $0.06
- Model
- $0.282
- Compute
- $0.658
- Platform
- $0.025
- Model owner
- $0.975
- Compute
- none
Tiny calculator
Approved rates only. No wallets. No second percentage.
How a call is priced
- Buyer selects a model with $/1M rates and an access policy they pass. Gating is not a fee.
- Gateway checks compute source (self / pool / buyer) and online status.
- Cost = tokens_in/1e6 × input_rate + tokens_out/1e6 × output_rate.
- Fee engine returns one platform take from contribution mode, then model / compute bps.
- Usage + settlement append to Litheum. Production targets PoP. Chain fees absorbed.
Prefer a scripted walkthrough?
Guided demo