TCI‑CRI‑H100$3.25TCI‑CRI‑H100‑7D— GAPTCI‑CRI‑H100‑MKT— GAPTCI‑CRI‑COMPUTE100.0TCI‑CRI‑H100$3.25TCI‑CRI‑H100‑7D— GAPTCI‑CRI‑H100‑MKT— GAPTCI‑CRI‑COMPUTE100.0TCI‑CRI‑H100$3.25TCI‑CRI‑H100‑7D— GAPTCI‑CRI‑H100‑MKT— GAPTCI‑CRI‑COMPUTE100.0
AS OF 2026-09-08 06:23:22 UTC
The Compute Indices
Governance

Methodology notices

Every change that would alter a published print is announced here before the first print computed under it, as required by GOVERNANCE.md §1. Prints keep the methodology version they were computed under, so a transition is auditable after the fact rather than only announced before it.

Norway restored to the EU/EEA constituent population, and the collection-time node-size floor aligned to the published floor

Notice 2026-N1 · announced 8 Sep 2026 · methodology v0.4.0

Takes effect 15 Sep 2026

Two defects found in the source audit of 7 September, both of which have been silently narrowing the constituent population against the published methodology. They are announced together because they take effect in the same version bump.

Norway. config/factors.yaml lists the EU/EEA constituent countries, and it has never contained Norway. The token NO is read by YAML 1.1 as boolean false, so eu_eea_countries holds False where it should hold the string NO. Every Norwegian observation collected since Azure's norwayeast region began reporting on 2026-08-16 has been dropped by the country filter: 126 observations as of the date of this notice. Norway is in the EEA, the published unit definition says EU/EEA, and the index has therefore been narrower than its own stated scope.

Node-size floor. filters.min_gpu_count has been 2 since v0.3.0, which is the floor the methodology publishes. The gpuhunt_ collector has continued to discard offers below 8 GPUs at collection time, before storage. Because collectors record live prices that cannot honestly be re-collected for a past date, those offers are not recoverable; the correction applies from the effective date forward.

Both are corrections toward the published methodology rather than changes to it. Neither alters the unit definition, the estimator, the trim or the weighting. They are handled as a minor version bump because each changes the constituent population, which is a parameter change under GOVERNANCE.md §1.

Prints computed before the effective date remain valid under the version they record in daily_index.methodology_version and are not restated. A gap stays a gap.

Expected effect on the level

Expected to lower the headline slightly or leave it unchanged. Both defects currently exclude supply, so correcting them can only add offers to the panel, and added offers at the marketplace end of the distribution pull a weighted median down rather than up. The size cannot be stated in advance: it depends on how many Norwegian and sub-8-GPU offers qualify on the day, which is not known until the first print under v0.4.0. Both legs will be quantified against the prior version in the CHANGELOG entry for that print.

What a notice does not cover

Scheduled weight reviews run a fixed published formula on a fixed schedule with no discretion. They are data updates, not methodology changes, and are not announced here. Corrections to an individual print are handled under the correction policy and published as a new revision, never as an edit.