The Thinking  /  The Ledger

Everywhere else on this site, I make a case. This is where it is checked.

The four resets, the successes and the failures, the arguments — elsewhere these are claims I ask you to weigh. Here they meet the data. Every reading below is drawn from a primary source — computed where the data can be computed, documented where it is a matter of record — with the workings shown, so you can check it, reproduce it, or carry it into your own work.

The open, computed record of the reset — primary sources, method shown, growing. It grows as the work is done — readings are added as the data is mined, and retired when they go stale. Use anything here — it is built to be cited.

New to the Ledger? Start with these three — the full record, in its own order, follows.
01 Technological × Financial #

AI capital spending has quadrupled since GPT-4 (in Q2 2023).

33.1% vs 32%
AI-era capital intensity has crossed the dot-com peak on filed actuals (aggregate capex 33.14% of revenue in Q1-2026, against ~32% at the 2000 peak)
$117bn → $448bn
2023 → 2025; quarterly aggregate capex reached $148.4bn in Q1-2026 alone
$239.1bn → $411.5bn
calendar-year aggregate on the cash-PP&E measure, 2024 → 2025
The reading — and its limits The defensible claim is the growth, which the companies attribute to AI — not a clean "AI capex" figure (no hyperscaler reports AI capex as a separate line; totals include warehouses, offices, other bets). The dot-com-peak crossing was previously stated as a 2026 guidance expectation; it is now an actual (Q1-2026, 33.14%). The five report on three different fiscal calendars, so every observation is mapped to the calendar quarter it predominantly covers.

Method Capex computed by us as a quarterly series from each firm's own SEC EDGAR XBRL filings (Alphabet, Amazon, Meta, Microsoft, Oracle), discrete quarters derived from year-to-date facts and reconciled to each issuer's annual figure; the growth rate and the dot-com-peak anchor (Morgan Stanley) corroborated.

Computed 1 August 2026
Permanent page →
02 Technological × Financial #

The price of machine intelligence has fallen about 10x a year, while enterprises' AI costs are growing exponentially.

$60 → $6 → $0.06
cost per million tokens, for the same performance (2021 → 2023 → 2024): ~1,000x cheaper in three years
$1.7bn → $37bn
enterprise AI spend, 2023 → 2025 (~22x)
The reading — and its limits The curve is per-token for equivalent performance — the falling price of a fixed capability, not the cost of any one task. Unit price collapses; total spend explodes anyway as usage surges. (Spend estimates vary by scope across analysts; Menlo's ~$37bn generative-AI figure is not the same as broader all-AI estimates and shouldn't be added to them.)

Method Cost curve from a16z "LLMflation" + Epoch AI (reproducible, published code); enterprise spend from Menlo Ventures' 2025 report (named).

Computed 1 August 2026
Permanent page →
03 Technological #

Klarna — a case of AI deployment success and business-outcome failure.

5,527 → 3,422 FTEs
Dec-2022 → Dec-2024 — ~40%, attributed to AI plus a hiring freeze
"the work of 700 agents"
the AI assistant handled ~2.3m conversations (~2/3 of all chats) in its first month
then reversed
by mid-2025 the CEO said quality had fallen and began rehiring humans
The reading — and its limits The deployment worked — the AI did the volume. The business outcome didn't: in the CEO's own words, cost had become "too predominant" a factor and service quality suffered, so the company began rehiring. The savings were modelled; the cost of unwinding was not.

Method Headcount from Klarna's IPO prospectus (Mar-2025); the reversal from the CEO on record (CNBC, Bloomberg, May-2025).

Computed 1 August 2026
Permanent page →
04 Technological × Financial #

AI fear is repricing the Indian IT industry.

−48.9%
foreign holdings in Indian IT nearly halved (₹712,646 Cr → ₹363,843 Cr, Oct-2024 → Jul-2026), while total foreign equity holdings fell just −7.9% and financials, capital goods, telecom and metals grew
−12.7% vs −0.2%
trailing year: Nifty IT vs Nifty 50
The reading — and its limits The −48.9% is the fall in foreign IT holdings through capital exiting and the rest being repriced down (~72% of the fall is price, ~28% net selling). Crucially, the capital that left IT largely stayed in India — financials, capital goods, telecom and metals all grew — so this is rotation within India, a verdict on the IT sector, not flight from the country. The driver is read from what the companies and the market say, not from price alone: Indian IT's own filings name an "AI productivity impact," and the correction is described as a structural re-rating of the labour-arbitrage, headcount-to-revenue model. In addition to AI, cyclical concerns are at work — weak US discretionary spend, tariff drag, BFSI (~30% of the sector) under pressure — so the honest reading is AI-structural fear layered on a real cyclical slowdown, with AI being what makes this a re-rating rather than a dip.

Method Foreign-holding rotation computed by us from NSDL's fortnightly sector-wise FPI reports (equity AUC). Index levels from NSE Indices' own historical file. The rotation is window-free; the index comparison is the trailing year (NSE's public cap).

Computed 1 August 2026
Permanent page →
05 Technological × Financial #

Foreign money is leaving India for the AI trade it can't buy at home — ~$58bn out since 2024, while Taiwan and Korea swell.

~$58bn
net foreign equity outflow from India since the Sept-2024 peak (~₹4.7 lakh cr), computed from NSDL
MSCI EM weight Sept-2024 Jun-30-2026
Taiwan 18.77% 27.34%
South Korea 11.67% 23.72%
India 19.9% 11.09%
China 24.42% 19.03%

0 Indian companies in the MSCI EM top 10 — first time in 26 years (3 at the Sept-2024 peak: Reliance, Infosys, ICICI). TSMC alone (~14.5%) is now more than all of India.

The reading — and its limits India's weight roughly halved while Taiwan and Korea — the AI-hardware markets — swelled, because index weight follows market value and ~$750bn of passive money mechanically follows the index. India isn't in the AI-hardware trade (its largest names are banks, energy, consumer), so as the world rotated to AI there was nothing in the Indian index for that money to land on. The $58bn is the cumulative outflow since the peak; the calendar-year figures (2025 ≈ −$19bn, 2026 YTD ≈ −$24bn) measure a different window and are not additive with it.

Method Outflow computed by us from NSDL depository data (net FPI equity flows, cumulative since the Sept-2024 peak). Index weights from MSCI Emerging Markets Index factsheets read directly — Gold where a factsheet (including archive-recovered), Silver where computed from the iShares EEM fund proxy for the five gap months, tier marked per row.

Computed 1 August 2026
Permanent page →
06 Financial × Technological × Risk #

When you buy an index, you're not buying diversification — you're increasingly buying concentration into one trade (AI/semiconductor).

Index Const Eff-N Top-10 AI-core Top-10% Largest
KOSPI (Korea) 830 6.9 64.9% 53.0% 90.9% Samsung Electronics 29.2%
DJIA (US, price-wtd) 30 20 55.4% 15.9% n/a Goldman Sachs 11.7%
Nifty 50 (India) 50 24 52.9% 0% 36.8% HDFC 10.6%
NASDAQ-100 103 33 46.1% 53.9% 46.1% Apple 8.2%
EURO STOXX 50 55 33 41.1% 14.0% 28.7% ASML 8.7%
S&P 500 505 48 37.2% 35.7% 62.5% Apple 7.7%
MSCI EM 1,190 32 35.5% 31.9% 65.2% TSMC 14.9%
MSCI World 1,273 94 26.1% 26.7% 59.7% Apple 5.5%
TOPIX (Japan) 1,638 107 23.2% 8.0% 79.2% MUFG 3.6%

concentration measures, computed by us from each index's own full constituent file (July basis) Eff-N = how many equal-weight stocks the index really behaves like. AI-core = share of the whole index in the AI complex — chips + hyperscalers, now tracked as a live series per index.

The reading — and its limits Three exceptions prove the point — Japan (TOPIX) is genuinely diversified (behaves like 107 stocks, AI-core just 8.0%); India (Nifty) is concentrated but in the old economy (AI-core 0% — its "tech" is IT services, the AI-disrupted category, not AI hardware); and the Dow (DJIA) sits near the top by top-10 (55.4%, second only to Korea) but this is a price-weighting artefact — its weights track nominal share price, not company size, so two high-priced blue-chips (Goldman, Caterpillar) dominate and the AI names (Nvidia, Microsoft) are muted. The Dow's low AI-core (15.9%) does NOT mean it escapes the AI trade — it holds the same names, down-weighted by share-price accident. Everywhere else, buying the index increasingly means buying the same handful of AI names. (July: the AI-hardware pullback trimmed AI-core in six of nine indices and lifted Apple back above Nvidia as the largest single holding in the S&P, NASDAQ and MSCI World — the concentration stands, but the trade breathes.)

Method Effective-N and concentration measures computed by us from each index's own full constituent file (index-owner or full issuer-holdings files). AI-core on a named-basket definition (semiconductors + hyperscalers) that crosses GICS sectors.

Computed 1 August 2026
Permanent page →
07 Geopolitical × Technological #

India sits on one of the world's largest rare-earth endowments and turns under 1% of it into output.

Rank Country Rare-earth oxide
1 China 44.0 Mt
2 Brazil 21.0 Mt
3 India ~8.52 Mt
4 Australia 6.3 Mt
5 Russia 3.8 Mt
The reading — and its limits The constraint was never geology. Comparator countries are on USGS reserves; India's figure is its own AMD in-situ resource (adjacent measures, so the rank is order-of-magnitude) — and we use India's own primary because the originator outranks the aggregator (USGS itself currently lists India as "NA"). India's decade of rare-earth trade: nil imports, 18 tonnes exported.

Method Endowment from India's AMD/DAE, read from the PIB Parliament record; output and world-share from USGS Mineral Commodity Summaries 2026, read directly.

Computed 1 August 2026
Permanent page →
08 Financial × Risk #

The 60/40 portfolio is no longer valid — it was never universal, and the regime that supported it has turned.

Market pre-2022 2022-on What it means
US −0.34 +0.08 inverted hard — bonds now fall with equities
UK −0.17 +0.14 inverted hard
Germany −0.27 ~0.00 hedge neutralised
Japan −0.09 −0.02 no hedge to lose (rates pinned near zero)
India +0.02 +0.15 never a 60/40-style hedge

63-day rolling stock-bond correlation (negative = bonds cushion equities, the 60/40 promise)

The reading — and its limits The hedge inverted hard where it was deepest (US, UK) when the 2022 hiking cycle hit; it was neutralised in Germany and never existed in Japan or India. A correlation that "always held" turns out to have been a feature of one rate regime in a few markets — not a universal law the world could safely build a portfolio on. (Bond legs are long-government-bond fund proxies, not cash bonds.)

Method 63-day rolling correlation of daily equity-index vs long-government-bond returns, per market, pre-2022 vs 2022-on, computed by us from primary price series (same method as the US base reading).

Computed 1 August 2026
Permanent page →
09 Financial #

The price of money has doubled — and it isn't coming back down.

~2.1x
the US 10-year yield today (4.30%) against its 2012-2021 average (2.04%)
~8.5x
the US 10-year's rise from its 0.52% COVID-era low to 4.46% now
3 years
the US 10-year has held above 4%, even as the Fed cut 175bp
The reading — and its limits "Doubled" is measured against the 2012-2021 decade average; against the longer 2010s it is ~1.85x, so the window is stated. The deeper point is not the level but the persistence: the long end rose through an entire Fed easing cycle, and the Fed's own convergence-to-2% forecasts have missed six years running. Two independent witnesses — the bond market and the forecaster's own record — say the same thing: a re-based regime, not a cycle.

Method Annual-average and current US 10-year Treasury yields, computed by us from the US Treasury's own daily par-yield series (1990-2026). The policy-rate cut is the Fed's path over Sep-2024 → Jun-2026.

Computed 1 August 2026
Permanent page →
10 Geopolitical × Financial #

The dollar isn't being dethroned, but select countries are hedging with gold.

Gold as % of reserves 2018 2025
India 5.5% 16.2%
China 2.4% 8.5%
Poland 4.5% 28.2%
Turkey 21.6% 61.1%

Under the table: the dollar's share of global reserves drifted down from 69.7% to 56.8% (2000-2025).

The reading — and its limits A hedge, not a replacement — the trigger was the 2022 freezing of ~$300bn of Russian reserves (dollar reserves held abroad can be seized; domestic gold cannot). The share rise is partly buying and partly gold's price rise; for these four, the tonnage rise confirms real accumulation. China's reported holding is a likely-understated floor.

Method Computed by us from IMF primary (SDMX API) — COFER for the dollar share, IRFCL gold-value/total-reserves per country. India cross-validated against the RBI half-yearly report (880t).

Computed 1 August 2026
Permanent page →
11 Financial × Risk #

India spends a third of its revenue just servicing its debt — the most among major economies.

Country Interest-to-revenue Debt-to-GDP
India ~37% ~84%
United States ~19% ~124%
United Kingdom ~13% ~102%
Japan ~13.5% ~207%
Germany ~3-4% ~63%

Interest-to-revenue = the share of what the government collects that goes straight to interest.

The reading — and its limits The two columns don't line up — and that's the point. Japan owes far more relative to its economy (~207%) yet pays a smaller share of revenue in interest than India does on ~84% debt. The reason is the rate at which the debt was raised: India borrows at much higher interest rates, so a given amount of debt costs far more to service each year. The burden is set by the rate, not just the size of the debt.

Method Interest-to-revenue from each country's own treasury/budget primary (India = Union Budget interest payments / revenue receipts). Debt-to-GDP from the IMF (general government gross debt, 2025).

Computed 1 August 2026
Permanent page →
12 Geopolitical × Financial #

Supply chains didn't shorten — they re-routed.

Share of US goods imports 2017 2024 Change
China 21.9% 13.8% −8.1pp
Mexico 13.1% 15.2% +2.1pp
Vietnam 2.0% 4.2% +2.2pp
India 2.1% 2.7% +0.6pp

Mexico overtook China as the #1 source of US imports in 2023.

The reading — and its limits China's true share fell by less than the US figures show — tariff-avoidance routes Chinese goods through Vietnam and Mexico, so part of their "rise" is Chinese value re-routed, not displaced. The data shows where goods ship from, not their ultimate origin. This is re-routing, not retreat.

Method Computed by us from UN Comtrade primary bilateral data (US reporter, full continuous 2017-2024), cross-validated against US-Census-based reads.

Computed 1 August 2026
Permanent page →
13 Risk × Financial #

The risk reset is being priced — insurance premiums are rising far faster than their own history, across lines and across continents.

Line Where Recent pace vs its own history
Motor US +8.8% vs +2.6% = 3.4x
Motor EU +6.0% vs +1.4% = 4.1x
Health EU +4.1% vs +2.3% = 1.8x
Health India +16%/yr (~3.5x inflation); the rise is price, not volume
Home EU +4.2% vs +2.4% = 1.7x

India's standalone health insurers pay back just 69 paise per rupee of premium, vs 100+ for public insurers — a gap held six straight years.

The reading — and its limits The same regime break shows up across two statistical authorities and two continents — motor sharpest (3.4-4.1x), health and home both clearly accelerating. The US homeowners line is absent because US CPI folds it into shelter (which is why State Farm's home repricing doesn't show in US inflation — the EU dwelling series fills that gap). The US health-insurance CPI is deliberately not used — its methodology produces wild swings (−27% in one year) and isn't reliable. India's public-vs-private payout gap is a persistent level, not a decline; India FY25-26 figures are provisional pending the next IRDAI report.

Method US from BLS, EU from Eurostat (motor / health / home), India from IRDAI Annual Reports + MoSPI — all read directly. "Pace" = recent (2022-25) annual rate vs the 2006-19 average.

Computed 1 August 2026
Permanent page →
14 Risk #

The safe-haven metals have turned volatile — silver is having its most volatile year on record.

Large daily moves Silver >5% Gold >3%
2008 (GFC) 27 35
2011 31 9
2014-2019 (the calm) 0-1/yr 0-1/yr
2020 (COVID) 14 8
2022-2025 (calm again) 2-3/yr 0-6/yr
2026 (YTD → annualised) 28 → ~62 11 → ~24

count of large daily moves per year (silver >5%, gold >3%; LBMA/ICE benchmark)

The reading — and its limits Silver's 2026 pace (~62 on an annualised basis) is the most on record (past 2011's 31 and 2008's 27); gold (~24) is on pace for second only to 2008. The calm was real and prolonged — 2014-2019 and 2022-2025 were nearly all low single digits — which is what makes 2026's jump a genuine break, not noise. This is about the frequency of large moves, not their scale (2008 and 2020 had bigger single-day peaks); gold is "second only to 2008," not above it.

Method LBMA/ICE official benchmark fixings (primary), fix-to-fix daily moves, full continuous 2005-2026, computed by us; corroborated by Yahoo futures. 2026 is YTD through 15-Jun; the annualised figures project the current pace across a full year.

Computed 1 August 2026
Permanent page →
15 Risk × Financial #

When risk outruns the models, insurers stop underwriting it — and the rules have to be rewritten to bring them back.

~1 million policies
State Farm's planned California withdrawal (3.1m → ~2m by 2028); stopped writing new home policies in 2023, non-renewed ~72,000 in 2024
Forward-looking pricing, now permitted
the regulator changed the rules to allow forward-looking catastrophe models (and reinsurance costs) in pricing, to stop insurers exiting the market
+17% / $400m
the regulator approved an emergency 17% premium increase on existing policies to keep State Farm's California unit solvent, conditioned on a $400m capital infusion from the parent
The reading — and its limits The chain is unpriceable risk → financial stress → exit. A non-stationary peril (wildfire shifting faster than the historical record) met a pricing framework that required rates be set on past data and barred forward-looking models — so the distribution moved but the rules forbade pricing it, and the insurer's most honest response was to exit. The proof comes from both directions: the insurer that walked away, and the regulator that rewrote the rules — conceding the old framework could no longer price the risk.

Method State Farm withdrawal figures and stated reasons from its own filings and California Department of Insurance filings. The pricing-framework constraint and its 2024-25 reform from the California Code of Regulations directly (10 CCR 2644.25.1 reinsurance; 2644.4.5 / 2644.5 catastrophe models) — not secondary characterisation.

Computed 1 August 2026
Permanent page →
16 Risk #

Elevated volatility has become India's standing condition, not an acute shock.

India VIX: % of days above 15 value
2008-2016 (the norm) 55-100%/yr
2023 / 2024 / 2025 (the calm) 10% / 30% / 25%
2026 (to 15-Jun) 61%

Peak in 2026: ~28 (moderate amplitude, vs 80-plus in the crisis years). share of trading days with India VIX above 15

The reading — and its limits 2012 is the closest single-year analogue (86.5% of days above 15, peak ~29) — but 2012 sat among a run of high-volatility years (2008-2016 were almost all elevated). 2026 is different: it's a snap-back to elevated after the genuine calm of 2023-2025. The reading is about character — persistent and moderate-amplitude — not the absence of a trigger; the point is the moderate intensity (peak ~28, not 80-plus) and the sustained persistence.

Method 2026 figure from NSE primary (the India VIX historical CSV, read directly); historical by-year series from Yahoo ^INDIAVIX. Cross-checked where they overlap (Yahoo 61.1% vs NSE 60.9%, identical peak).

Computed 1 August 2026
Permanent page →
The record grows on demand

Looking for a reading that isn’t here yet?

Tell me what you need read — a market, a country, a claim you want checked against primary data. If the data can carry it, I will compute it, add it to the Ledger, and send the finished reading to you directly.

Received — the request is with me.

If the primary data can carry it, I’ll compute the reading, publish it here, and send it to you the day it goes live.

An invitation

If a reading here raises a question inside your institution, I would welcome the conversation.

No intake process, no gatekeeper. A note reaches me directly, and I reply myself — usually within a day or two.

This comes straight to my inbox. I read every note personally and reply myself — not a team, not an auto-responder.

Message received

Thank you — your note has reached me.

It comes straight to my inbox, and I reply personally — usually within a day or two.

For editors & journalists

Every reading here is available as source material — the number, the method, and the workings. Use it; cite it; write to me if you’d like a reading computed.