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.
The same-hour spread across European H100 sellers was 7.40x on 7 September, and no regression on that panel can say how much of it is quality. One venue publishes a performance score on every offer, and a design built on it is identified in principle; the book it would be estimated on held 7 datacenter-verified H100 offers from 3 hosts on 12 September, and the score moved 30.50% overnight on one o
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 nothing the index had stored described a fabric, a power envelope or a thermal limit.
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.
Design rules from oil, power, freight and repo, and which of them compute can actually satisfy
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
One email every two to three weeks, plus a short print note whenever something breaks or moves. The archive is public; the index is published here whether you subscribe or not.