Every driver below was read out of a chart image in the brokers' own PDFs, not their text. Plotted against what our engine was scoring on the same days, the lag is visible.
The AI complex. Token spend peaked on 31 May and had fallen 26% by 25 July. Hyperscaler credit began breaking down in mid-June. Our AI conviction went on rising, peaked at +0.62 on 28 June, and has since halved to +0.14 — a 28-day lag behind the first driver to turn.
The energy complex, by contrast, was near-coincident. Brent bottomed on 30 June; our Power Generation conviction troughed on 2 July, 2 days later, and has tracked it up since.
The difference is what the engine reads. Oil is priced in the text of every brief; token expenditure, ETF share creation and SPV bond prices appear only as unlabelled charts. The engine is fast where the number is written down and slow where it has to be seen.
| Turned | Theme | What the chart showed | Source |
|---|---|---|---|
| 31 May | AI | Token spend peaks, −26% since Silicon Data LLM Token Expenditure Index 190 → 139.6 | The Upcoming AI Spend Slowdown |
| 15 Jun | AI | Hyperscaler credit breaks down Meta Hyperion 6.581s of '49: 110 → 94.9, decline accelerates | Goldman 1-Delta desk |
| 30 Jun | Power Generation | Brent bottoms, +41% since Generic 1st Brent 71 → 100.26 in under four weeks | Goldman 1-Delta desk |
| 10 Jul | AI | SOXX: record inflow into a falling tape Shares outstanding at a high while price 350 → 291.5 | The Upcoming AI Spend Slowdown |
| 17 Jul | Power Generation | Sept hike odds ramp to 70.9% Chart labels oil as the driver of the current leg | Hypershrinkers & Hormuz Hell |
| 17 Jul | Minerals | Silver leads the complex, +4.3% vs gold +0.94%, copper +1.21%, platinum −0.15% | Tehran, Tech & Tariffs |
| 20 Jul | AI | Bear flattener: 2Y +18bp, 30Y +10bp 30Y nominal at 5.17%, level with the 2007 high | Hypershrinkers & Hormuz Hell |
| 23 Jul | AI | MegaCap tech gaps −4.5% on the week GS TMT MegaCap basket 615 → 587.65 | Tehran, Tech & Tariffs |
| — | Minerals | OECD commercial oil stocks 48th → 8th percentile; global SPR 22nd → 1st | Energy & AI (Debt) Cost Anxiety |
| — | AI | $31tn of S&P market cap added since Nov 2022 vs a $9tn baseline AI revenue PDV ($28tn even in the combined-positive case) | Goldman: Size Of The Pie |
| — | AI | AI valuation is a rates problem first: PDV $22tn at a 10% discount rate, $5tn at 20% — a wider swing than adoption speed, productivity or capital share | Goldman: Size Of The Pie |
How to read this, and what it cannot tell you.
Conviction is the engine's daily theme score, −1 to +1, averaged across each major's
constituent themes and smoothed over five days; the raw series moves more than half a
point between consecutive days, so a single print is not a trend. Scored output begins
18 June, which is after the token-spend turn on 31 May: our conviction before that date
is unobserved, so the 28-day figure is the gap between the driver turning and the peak of
the window we can actually see. What the data does support is the direction — through the
three weeks after token spend rolled over, our AI conviction was still rising. Turn dates
are read off chart images by eye, not from a data feed, and are accurate to a few days
rather than to the session. Every reading is traceable to one cropped figure in one
document via research/chart-readings/2026-07-26.json.