Compute Cost Index

Cloud GPU pricing normalised to $/petaFLOP-day across on-demand, spot, and reserved tiers, using dense BF16 throughput from Epoch AI ML Hardware (no 2:4 structured sparsity). Coverage: AWS, Azure, GCP, Oracle OCI, CoreWeave, Lambda.

Version 2026.2108 rowsReleased 2026-05-19License CC-BY-4.0SHA-256 cfe9766a4e9d…

Download bundle

CSVFlat data, UTF-8 JSON SchemaColumn types MethodologyMarkdown Croissant JSON-LDMLCommons 1.0 SHA-256Integrity

Hosted mirrors

ZenodoDeposition pending — David-action: set ZENODO_TOKEN in .env.local + Vercel, then run scripts/deposit-dataset-to-zenodo.ts --all

Pending

Hugging FaceMirror pending — David-action: install @huggingface/hub + set HF_TOKEN, then run scripts/deposit-dataset-to-hf.ts --all

Pending

Citation

Gringras, D., & Scrutica. (2026). Scrutica Compute Cost Index (Version 2026.2) [Data set]. Zenodo. (DOI pending deposition)

BibTeX

@dataset{scrutica_compute_cost_index_2026_2,
  author    = {Gringras, David and {Scrutica}},
  title     = {Scrutica Compute Cost Index},
  year      = {2026},
  version   = {2026.2},
  publisher = {Zenodo},
  note      = {DOI pending Zenodo deposition; cite by version + SHA-256 hash from /datasets/compute-cost-index until assigned},
  url       = {https://scrutica.com/datasets/compute-cost-index}
}

Methodology

# Scrutica Compute Cost Index — v2026.2

**Release date:** 2026-05-19
**License:** CC-BY-4.0
**Suggested citation:** see `CITATION` block at the bottom.

## Coverage

108 (instance × pricing-tier) rows across 5 providers and 5 accelerator types. Pricing tiers covered: on-demand, spot, reserved 1-year, reserved 3-year.

## Normalisation

Each row's hourly instance price is divided by the instance's dense BF16 throughput to give a $/petaFLOP-hour figure, multiplied by 24 for the day figure. Throughput uses Epoch AI's ML Hardware corpus (e.g., 989.5 dense BF16 TFLOP/s on H100 SXM); **no 2:4 structured sparsity** factor is applied — the index is intended to be comparable with dense training cost. Reserved-tier rows are amortised at the provider's published commitment-term pricing; we do not impute hidden discount tiers.

## Sources and authority tiers

| Source | Authority tier | What it contributes |
| --- | --- | --- |
| AWS pricing page | 1 | On-demand and reserved hourly rates |
| Azure pricing page | 1 | On-demand and reserved hourly rates |
| GCP pricing page | 1 | On-demand and reserved hourly rates |
| Oracle OCI pricing page | 1 | On-demand hourly rates |
| Lambda pricing page | 2 | On-demand hourly rates |
| CoreWeave pricing page | 2 | On-demand hourly rates |
| ec2.shop JSON API | 2 | Cross-verification of AWS rates and spot pricing |
| Epoch AI ML Hardware | 1 | Per-accelerator dense BF16 throughput denominator |

## Refresh path

Pricing is refreshed via the scripts that produce `data/processed/cloud_pricing.json`; this dataset is regenerated by `scripts/package-datasets.ts`.

## Known limitations

- Spot pricing is a snapshot; real-world spot prices fluctuate intraday.
- On-prem TCO (build-vs-buy) is rendered alongside on `https://scrutica.com/cost-index` but is *not* in this CSV — it is a derived view, not a primary measurement.
- Commitments that require enterprise procurement (private negotiated rates, GPU reservation deals) are out of scope; this index is the public-rack-rate floor.

## Citation

```bibtex
@dataset{scrutica_compute_cost_index_2026_2,
  author       = {Gringras, David and {Scrutica}},
  title        = {Scrutica Compute Cost Index},
  year         = {2026},
  version      = {2026.2},
  publisher    = {Zenodo},
  note         = {DOI pending Zenodo deposition; cite by version + SHA-256 hash from /datasets/compute-cost-index until assigned},
  url          = {https://scrutica.com/datasets/compute-cost-index}
}
```

Data Provenance

Primary100%

108 records

Sources

Source	Tier
AWS	T1
Azure	T1
GCP	T1
Oracle OCI	T1
Epoch AI ML Hardware	T1

Processing

Bundle regenerated by scripts/package-datasets.ts from data/processed/. SHA-256 cfe9766a4e9d49ee… signs the CSV.

Reading this data

~valueEstimated — derived, modeled, or inferred. Hover for method and source.

valueDerived from another measured field (e.g. power from GPU count).

Source conflict — multiple sources disagree. Click for competing values.

hoverHover any annotated value to see authority tier, method, source, and vintage.

Full methodology

Built for the AI governance research community

@dataset{scrutica_compute_cost_index_2026_2, author = {Gringras, David and {Scrutica}}, title = {Scrutica Compute Cost Index}, year = {2026}, version = {2026.2}, publisher = {Zenodo}, note = {DOI pending Zenodo deposition; cite by version + SHA-256 hash from /datasets/compute-cost-index until assigned}, url = {https://scrutica.com/datasets/compute-cost-index} }

# Scrutica Compute Cost Index — v2026.2 **Release date:** 2026-05-19 **License:** CC-BY-4.0 **Suggested citation:** see `CITATION` block at the bottom. ## Coverage 108 (instance × pricing-tier) rows across 5 providers and 5 accelerator types. Pricing tiers covered: on-demand, spot, reserved 1-year, reserved 3-year. ## Normalisation Each row's hourly instance price is divided by the instance's dense BF16 throughput to give a $/petaFLOP-hour figure, multiplied by 24 for the day figure. Throughput uses Epoch AI's ML Hardware corpus (e.g., 989.5 dense BF16 TFLOP/s on H100 SXM); **no 2:4 structured sparsity** factor is applied — the index is intended to be comparable with dense training cost. Reserved-tier rows are amortised at the provider's published commitment-term pricing; we do not impute hidden discount tiers. ## Sources and authority tiers | Source | Authority tier | What it contributes | | --- | --- | --- | | AWS pricing page | 1 | On-demand and reserved hourly rates | | Azure pricing page | 1 | On-demand and reserved hourly rates | | GCP pricing page | 1 | On-demand and reserved hourly rates | | Oracle OCI pricing page | 1 | On-demand hourly rates | | Lambda pricing page | 2 | On-demand hourly rates | | CoreWeave pricing page | 2 | On-demand hourly rates | | ec2.shop JSON API | 2 | Cross-verification of AWS rates and spot pricing | | Epoch AI ML Hardware | 1 | Per-accelerator dense BF16 throughput denominator | ## Refresh path Pricing is refreshed via the scripts that produce `data/processed/cloud_pricing.json`; this dataset is regenerated by `scripts/package-datasets.ts`. ## Known limitations - Spot pricing is a snapshot; real-world spot prices fluctuate intraday. - On-prem TCO (build-vs-buy) is rendered alongside on `https://scrutica.com/cost-index` but is *not* in this CSV — it is a derived view, not a primary measurement. - Commitments that require enterprise procurement (private negotiated rates, GPU reservation deals) are out of scope; this index is the public-rack-rate floor. ## Citation ```bibtex @dataset{scrutica_compute_cost_index_2026_2, author = {Gringras, David and {Scrutica}}, title = {Scrutica Compute Cost Index}, year = {2026}, version = {2026.2}, publisher = {Zenodo}, note = {DOI pending Zenodo deposition; cite by version + SHA-256 hash from /datasets/compute-cost-index until assigned}, url = {https://scrutica.com/datasets/compute-cost-index} } ```