# Scrutica Compute Cost Index — Data Dictionary

Column-level reference for `compute-cost-index_v2026.3.csv` — version 2026.3, released 2026-07-17, 508 rows, CC-BY-4.0.

Cloud GPU pricing normalised to $/petaFLOP-day across on-demand, spot, and reserved tiers. Throughput uses dense BF16 specs from Epoch AI ML Hardware (no 2:4 structured sparsity).

Generated from the dataset's JSON Schema (`compute-cost-index_v2026.3.schema.json`); this Markdown is the human-readable view of that machine-readable contract.

## Conventions

- **Nullability** — every column is optional unless its **Required** cell says `Yes`. A null is written as an empty CSV cell.
- **Lists** — multi-value columns are semicolon-joined (`a; b; c`).
- **Currency** — monetary columns are USD unless a `currency_of_record` column on the same row states otherwise.
- **Dates** — ISO 8601 (`YYYY-MM-DD`).
- **Estimates** — columns suffixed `_is_estimated` flag analyst-derived values (`true`) versus primary-source values (`false` / null), per Scrutica's transparency convention: no derived number is presented as if measured.
- **Provenance** — every row carries `data_source` and, where the source is public, `source_url`, so any value traces back to a primary source.

## Columns (15)

| # | Column | Type | Required | Description |
|--:|--------|------|:--------:|-------------|
| 1 | `provider` | string | Yes | Cloud provider slug (aws, azure, gcp, oci, coreweave, lambda) |
| 2 | `instance_type` | string | Yes | Provider instance / SKU identifier (e.g., p5.48xlarge) |
| 3 | `gpu_type` | string | Yes | Accelerator model (e.g., H100_SXM, H200, B100) |
| 4 | `num_gpus` | number |  | Number of accelerators in the instance |
| 5 | `region` | string |  | Provider region code (e.g., us-east-1) |
| 6 | `region_name` | string |  | Human-readable region name |
| 7 | `pricing_model` | string | Yes | Pricing tier (on_demand, spot, reserved_1yr, reserved_3yr) |
| 8 | `price_usd_per_hour` | number |  | Instance price in USD per hour |
| 9 | `cost_per_pflop_hour` | number |  | Derived cost per dense BF16 PFLOP-hour (no 2:4 sparsity) |
| 10 | `cost_per_pflop_day` | number |  | Derived cost per dense BF16 PFLOP-day |
| 11 | `cost_per_pflop_year` | number |  | Derived cost per dense BF16 PFLOP-year |
| 12 | `note` | string |  | Provider-specific caveat (e.g., SKU does not offer 1yr RI) |
| 13 | `data_source` | string | Yes | Primary source attribution (e.g., AWS pricing page, ec2.shop JSON API) |
| 14 | `source_url` | string |  | URL of the primary source |
| 15 | `retrieved_at` | string |  | ISO 8601 retrieval date for the row |

## Integrity & citation

- **SHA-256** of `compute-cost-index_v2026.3.csv`: `9d4ad130620644787a23f75d6b4e8b84ffee19e38782383a8d3ba80dc07180d5`
- **Canonical page** (BibTeX · APA · DOI once minted): https://scrutica.com/datasets/compute-cost-index
- **Methodology**: https://scrutica.com/datasets/compute-cost-index → Methodology, and https://scrutica.com/methodology

_Scrutica · compute-cost-index · v2026.3 · 2026-07-17 · CC-BY-4.0_