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 2026-09-07 19:44:27 UTC
TCI Research

Research

The newsletter behind the index: notes on what it measures and what it cannot. Each one is reproducible from the published history in site/data/index_history.csv and the stored constituent sets; where a note makes a numeric claim, the query that produced it is printed with it.

4 notes Methodology v0.3.0-dev As of 2026-09-07
7 Sep 2026

A GPU-hour is a part number, not a unit

The index records an interconnect for all 7,798 observations and has observed one for none of them, 805 rows assert an SXM part and a PCIe bus at the same time, and no source in the panel has ever disclosed a fabric, a power envelope or a thermal limit.

PublishedTCI Research · methodology v0.3.0-dev
6 Sep 2026

Compute prices barely move. Where you buy moves everything.

Five of nine European sellers never changed their H100 price once in fifty days, the same-hour spread across sellers was 8.56x, and the hyperscaler rate cards are region-flat to four decimal places, which is the argument for regional benchmarks and not the argument most people make for them.

PublishedTCI Research · methodology v0.3.0-dev
16 Aug 2026

When an index measures its own sampling

How a 10% move in a compute benchmark turned out to be entirely composition noise, and what the unit definition had to do with it

PublishedTCI Research · methodology v0.3.0-dev