The AI Margin Compression Index is the macro thesis Palanor has been writing into — that the software-sector margin structure of the 2010s and early 2020s is unstable in the AI era. The v1 composition was built on proxy signals because we did not have a clean public-data path to the most important input: hyperscaler AI capex. The v2 composition closes that gap.
What changed
The v2 composition adds four manual-entry signals — Microsoft AI capex, Alphabet AI capex, Meta AI capex, and Amazon AWS capex — each sourced from the most recent 10-Q filing. The four together carry 50 percent of the index's total weight. The v1 proxy components (Nasdaq strength, ISM Services, GitHub trending, wage growth, PPI, industrial production) remain in the composition at the other 50 percent. Nothing was removed.
Capex components transform via 12-month percentage change because the level grows over time — the index reads acceleration, not level. Proxy components retain their 24-month z-score transforms.
What this means for the reading
When v1 read 25 and v2 reads 55, that is not noise. The divergence is the point. v1 saw the proxy market still pricing software margins as safe; v2 sees capex accelerating into the supply build that compression depends on. Both reads are correct given what they measure. The gap between them is where the thesis lives.
Expect v2 to be more volatile quarter-to-quarter than v1 because capex is reported quarterly with material variance. The index renormalizes around any missing component, so a single late filing does not break the reading.
Refreshing the data
Capex values are entered manually each quarter from the four hyperscalers' 10-Q filings. The admin entry surface is planned for a follow-up release; for now, the data is seeded through Q3 2025 and refreshes via direct database update by Tim until the entry surface ships.
A v3 of this index is planned that will automate ingestion via SEC filing parsers — at that point the manual-entry components become automated, the index becomes daily-refreshable, and the v3 reading will diverge again. Each version's history stays anchored to the composition it was scored against, so past readings do not silently revise.
What this is not
The index is not a GPU spot price tracker. GPU spot prices are not cleanly public; what is public is the dollar amount the hyperscalers are putting into the AI infrastructure stack, which is one step downstream of GPU prices but cleaner to measure. When direct GPU pricing becomes accessible, it will join the index — versioned again.
The index is not a software-stock signal. Holding the index high or low says nothing about how to position equities; it reads the macro pressure on margins, full stop. Scenarios and Posture are where that read converts into operating decisions.
Carrying it into a meeting
The v2 reading remains a one-sentence deliverable. "AI Margin Compression is at 55 — the capex acceleration thesis is materializing alongside proxy market calm." That sentence tells the room more than v1's 25 ever could, because the v2 number knows about the capex side of the thesis.