Inference
USD per 1M explicitly defined input or output tokens.
- Contracts
- 0
- Usable snapshots
- 0
- Configurations
- 0
No reviewed contracts registered. The next step is acquiring a reproducible benchmark with its full protocol.
Kovara Compute Index · Evidence & Economics
Which workloads can we price reliably?
Each workload needs a versioned protocol, measured hardware configuration and a compatible price. Inference, training and fine-tuning are kept separate.
Training
USD per complete, explicitly defined training run.
- Contracts
- 0
- Usable snapshots
- 0
- Configurations
- 0
No reviewed contracts registered. The next step is acquiring a reproducible benchmark with its full protocol.
Fine-tuning
USD per complete, explicitly defined training run.
- Contracts
- 0
- Usable snapshots
- 0
- Configurations
- 0
No reviewed contracts registered. The next step is acquiring a reproducible benchmark with its full protocol.
Required evidence before a comparison
| Dimension | Evidence required |
|---|---|
| Workload | Exact model and revision, dataset or token basis, precision, context and output lengths, batch/concurrency and quality target. |
| Protocol | Harness version, software stack, measurement conditions and reproducible result reference. |
| Configuration | Exact GPU, GPU count, nodes, interconnect and parallelism. No scaling from unmeasured configurations. |
| Price binding | Matching hardware, service scope, contract and timestamp. Separately billed costs remain explicit. |
Counts cover the loaded public catalog. Checked 2026-09-30T16:06:07.494Z.