eu-compute-index

EU-CRI — a daily, fully reproducible reference price for renting AI compute in the EU/EEA. Headline series EU-CRI-H100: one NVIDIA H100 SXM 80GB GPU-hour, on-demand, per-GPU, ex-VAT, from an EU/EEA data centre, in USD with a EUR companion at the ECB reference rate (T-1). Companion series by market segment (-MKT marketplace, -NC neocloud, -HS hyperscaler catalog, -SOV EU-incorporated operators), by generation (H200, B200, B300, A100, H100-PCIe), and EU-CRI-COMPUTE, a chain-linked composite that follows the observed market across hardware generations.

The credibility strategy is reproducibility, not scale: every parameter is a visible config value, the methodology document is generated from that config, raw observations are immutable, every print carries a queryable constituent set, and the whole thing can be rebuilt from public sources by anyone with the repo. Design follows the IOSCO Principles for Financial Benchmarks (2013) as voluntary best practice — see GOVERNANCE.md, METHODOLOGY.md, and SOURCES.md.

EU-CRI is a price-transparency benchmark, not a settlement benchmark. It is a research publication, is not investment advice, and may not be used as a reference price in financial instruments. It has no transaction feed and does not claim one. The conditions that would have to be met before settlement use is credible are published in GOVERNANCE.md; as of v0.3.0, six of the seven are unmet.

Why it was rebuilt in v0.3.0, and what was wrong before: research/composition-vs-price.md.

Quickstart

python -m venv .venv
.venv/Scripts/activate            # Windows; source .venv/bin/activate on Linux
pip install -e .[dev]
python -m eucri.run migrate       # create data/eucri.db
python -m eucri.run daily         # collect today's observations + compute all series
python -m eucri.run constituents --date 2026-07-18
pytest

Commands

Command	Purpose
migrate	apply database migrations
daily [--date D]	run collectors (idempotent per source+day), compute all series, regenerate the site, CSV and charts
constituents --date D [--series S]	full audit table for a print (IOSCO P13/P16)
backfill --from D --to D	recompute prints from stored observations (never re-collects)
weights [--date D]	show the stored weight review for a date (v0.3.0 weights providers by tier; reviews are retained for audit, not used in the calculation path)
validate	source-dropout sensitivity + optional check-series correlation
post	regenerate the paste-ready Substack post
docs	regenerate METHODOLOGY.md + METHODOLOGY.lock

Layout

• config/ — all methodology parameters (factors.yaml), sovereign constituent list, static provider price entries with last_verified dates
• src/eucri/ — collectors (fail-soft, 1 request/source/day, honest User-Agent), normalisation, index calculation, outputs
• data/eucri.db — SQLite, committed; observations and prints are append-only (trigger-enforced)
• site/ — the published site, regenerated from the database on every daily run by src/eucri/outputs/site.py; do not hand-edit the HTML. Six pages: index.html (dashboard), methodology.html, governance.html, research.html and research/*.html. Plus assets/ (the design system: tokens.css, site.css), data/ (CSV history + latest.json), and charts/ (PNGs used by the Substack post, not by the site — the site draws its own inline SVG).
• site/components.html — the design-system component gallery. A reference artefact, not linked from the site.
• DESIGN.md — the design system spec: tokens, chart rules, density, contrast ratios.
• research/*.md — research notes, rendered to site/research/ by the generator.
• config/source_links.yaml — reference weblinks for the Sources panel and per-constituent links. Presentational only; not part of METHODOLOGY.lock.

Deploying the dashboard

site/ is a self-contained static site plus one optional serverless endpoint, so it can be hosted anywhere that serves static files. Two targets are wired up:

Both targets serve byte-identical content. Every link in the generated site is relative, so the site works equally at a subpath (markrusch.github.io/Compute-Index/) and at a domain root (Vercel), and there are no external requests to differ between them.

• GitHub Pages: .github/workflows/pages.yml deploys site/ on every push that touches it. Served at https://markrusch.github.io/Compute-Index/.
• Vercel: import the repo with Root Directory set to site. Served at the project root.

site/api/refresh.js is a Vercel-only serverless endpoint that can trigger daily.yml on demand. The v0.3.0 dashboard does not surface it — deliberately, so the two hosts behave identically rather than one carrying a button the other cannot honour. The endpoint still functions if called directly; to use it, set two Vercel Environment Variables (Project Settings → Environment Variables — never committed to the repo):

• GITHUB_DISPATCH_TOKEN — a token scoped to just this repo's Actions (read/write), e.g. a fine-grained PAT limited to markrusch/Compute-Index
• GITHUB_REPO — markrusch/Compute-Index

The refresh endpoint refuses any date other than today (in UTC): these collectors report live market prices, not history, so a past date can never be honestly re-collected — see GOVERNANCE.md and STYLE.md on why a gap stays a gap.

Changing the methodology

Not casually. Any change to config/factors.yaml, config/sovereign.yaml, src/eucri/index.py, src/eucri/normalise.py, or src/eucri/weights.py fails CI unless the version is bumped, the CHANGELOG has an entry, and the lock is regenerated — and takes effect only after one publication's notice. Scheduled weight reviews execute a fixed published formula and are data updates, not methodology changes. Procedure: GOVERNANCE.md.