Comparable work. Explicit assumptions.

1. A versioned comparison contract

Model and exact revision, precision, quantization, runtime and version, benchmark protocol and version, and quality target must match. Inference also fixes input/output lengths, batch size, concurrency, caching, token basis and latency constraint. Training fixes the dataset/revision, work definition and completion criterion.

GPU model, GPU count, hardware, interconnect and parallelism may vary between results within that contract, but each result must retain its measured configuration. V1 price binding supports a single node only. No hardware performance scaling is inferred.

2. One formula, explicit units

For whole-system hourly cost C and compatible aggregate throughput T in tokens/second:

USD per 1M tokens = C × 1,000,000 / (T × 3,600)

A per-GPU hourly price is multiplied by the measured GPU count once. A whole-instance hourly price is already a system price. Per-GPU throughput is multiplied by the measured GPU count once; aggregate throughput is used directly. The comparison contract identifies input, output or total tokens.

For training: cost per defined run = system hourly cost × measured run hours. GPU-hours/run is divided by the measured GPU count to estimate elapsed hours, explicitly labeled as an accounting estimate. Repeated-run scenarios assume sequential identical runs.

3. Price compatibility and costs

A founder must attest that the public price observation covers the exact measured configuration and listed compute charge. Original USD GPU/hour and instance-hour units are supported. Other currencies and units remain unavailable. A public KCI aggregate fixing cannot stand in for a purchasable system price.

Only the listed compute charge is included. Setup, separately billed CPU/RAM, storage, networking/egress, tax and support are excluded. Base snapshots do not model commitment minimums, billing increments, utilization losses or availability; realized spend can be higher. These are not API token prices or total cost estimates.

4. Freshness, review and revisions

Benchmark observations retain the existing 180-day performance freshness ceiling. Economics v1 applies a 30-day price ceiling to avoid presenting old list observations as current costs. These are eligibility windows, not confidence scores. Source classes are provider-declared, independent or internal.

Import or capture creates a draft; an explicit review approves it. Withdrawal and approved supersession block new calculations. Immutable snapshots preserve original inputs, measurement dates, acquisition dates, benchmark revisions, group versions, formula version and calculation time. Source unpublication can remove public access without deleting the retained internal record.

5. History and decision tools

Economics history is an informational ledger. It is not an index level and is not automatically inserted into KCI fixing history. Corrections reference earlier snapshots. Changes in configuration, contract version, benchmark revision or formula establish separate segments. No historical backfill is generated.

User volume scenarios preserve the selected workload conditions. Time is estimated from sustained measured throughput or repeated identical training runs. Follow the offering link to inspect commercial details or open Workload Economics to plan with additional assumptions. No reservation or procurement action is performed.

Cost and deadline scenarios

The decision workspace compares two explicitly selected snapshots under the same contract, formula, units and cost coverage. Different commercial terms and source classes are disclosed. Expired, withdrawn or corrected inputs cannot support a new comparison.

Productive hours = snapshot hours per metric unit × requested units
Elapsed hours = productive hours / (utilization percent / 100)
Billed hours = max(elapsed hours, user minimum billed hours)
Compute subtotal = billed hours × system hourly USD charge
Modeled total = compute subtotal + explicitly entered additional USD costs

Each scenario uses one measured system and assumes it is billed during idle time. The billing floor changes cost, not completion time. Additional costs left blank remain unknown; no modeled total is shown. A deadline estimate excludes queueing, startup delays and capacity constraints. Complete contractual obligations, unentered charges and provider availability remain unverified. The result is neither full TCO nor a quote.

The downloadable scenario JSON retains exact evidence snapshots, source timestamps and user assumptions. It is a point-in-time report, not a guarantee of current validity. Refresh evidence before a decision. Scenarios are not saved to an account.

6. Read-only data access

GET /api/index/economics?limit=50&format=json GET /api/index/economics?group=GROUP_UUID&format=csv

Limits: 1–100 snapshots; default 50. Unknown, duplicate or malformed parameters are rejected. Responses include original inputs, units, sources, timestamps, revisions, current status and reason codes. Reads enforce public visibility, exclude review actors/notes, and do not expose the internal run ledger. Bounded responses explicitly flag truncation.

Observation and calculation times

The API effective_at field records the original price observation time. calculated_at records when the reviewed inputs were combined. A later benchmark does not establish economics at an earlier price date; this ledger does not backfill historical performance.