Palanor: AI Margin Compression

## What this measures

When the cost of intelligence falls faster than the price of software, margins compress. The AI Margin Compression Index reads 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. v2 (this version) adds quarterly hyperscaler AI capex to the composition; v1 relied entirely on proxy signals.

## Why it matters

AI capex by the top four hyperscalers is the closest public-data proxy for the cost of building intelligence at scale. Rising capex sustains the supply build that ultimately compresses unit cost of inference — and therefore compresses software pricing power. The v1 proxies remain because they capture the demand side: if Nasdaq is strong and ISM Services is strong, the compression pressure is being absorbed; if both weaken alongside accelerating capex, the thesis is materializing.

## How to read it

A reading above 65 with rising trend is the macro environment Palanor has been warning about — capex accelerating into a demand stack that cannot price the new supply. A reading below 35 with stable trend is the opposite — the thesis is being absorbed by services strength and the proxy market reading is holding.

## What it does not say

Manual-entry data drifts when admins do not refresh. Future versions will automate ingestion via SEC filing parsers. Capex categorization is also imperfect — hyperscalers do not always cleanly separate AI capex from broader infrastructure capex; we use the most generous bucket each company reports and document the source quarter-by-quarter in the admin panel.

## What changed in v2

v1 had no direct read on hyperscaler AI capex — the most important input to the compression thesis. v2 adds four manual-entry components — Microsoft, Alphabet, Meta, and Amazon AWS AI capex per quarter, sourced from public 10-Q filings. These now carry 50 percent of the index's total weight; the v1 proxy components (Nasdaq, ISM Services, GitHub trending, wage growth, PPI, industrial production) retain the other 50 percent. Expect v2 readings to diverge from v1 — that divergence is itself informative.

## Methodology

Capex components transform via 12-month percentage change — the level grows over time, so we read acceleration rather than level. v1 proxies transform via 24-month z-score (level-based, mean-reverting). Weighted sum, sign-flips on negative-direction components, logistic scaling to 0-100. Thesis bands: 0-35 weak, 35-65 baseline, 65-100 accelerating. Manual-entry values refresh quarterly — see the manual entry surface in the admin panel (planned).

## Discourse evidence layer (added 2026-06-04)

The AI Margin Compression Index now incorporates two Discourse-derived components — the average per-call AI mention count across the scored S&P 100+ universe (which rises as the corpus talks more about AI capex pressure on margins) and the inverse of public sentiment (which falls as confidence in the AI-investment thesis broadens). Together these add a public-corpus evidence layer to the existing macro signal stack. When the macro signals point to compression but the discourse layer disagrees, we surface that disagreement explicitly in the index commentary.

## Revision — 24 September 2026

The four hyperscaler capex components were hand-entered estimates of each company's AI spending, last updated for the quarter ending September 2025, so for a year the index scored a frozen snapshot. They now read each company's filed capital expenditure from its 10-Q and 10-K (SEC EDGAR), updated weekly: Microsoft, Alphabet, Meta and Amazon, each scored on its twelve-month change, with the same weights as before. Filed figures are total capex rather than an estimated AI share; at these four companies the two now move together. History was recomputed.

## Source

Computed daily by Palanor from 12 component signals in the lattice. Each component's weight, transform and current contribution is shown on the index's methodology page.