No Bursting Bubble
The Rotation Inside the AI Trade, in Four Panels, With PPI as the Hinge
The claim, stated plainly
The rotation inside the AI trade is real. There is no bursting bubble — but there has been a rotation, from AI capex to AI opex and AI applications, and it has been running for months in plain sight.
The reason it gets misread as a top is that most people watch one index. Watch four purpose-built ones and the move resolves into something much more ordinary: capital re-deciding where in the stack the margin lives, while the total keeps growing.
Four panels, one message
Our index family tracks the rotation cleanly. The Rubin Build-Out 100 — the capex layer — has gone sideways for a few months and still sits on +109.8% for the year at 2,097.53: priced, not broken. The Agentic Ecosystem index — AI opex, the compute-renters and token-meterers — has trended up consistently and closed at its high, +63.5% YTD at 1,634.51.
Agentic Winners 40 — the application layer — has climbed off its July floor to its best level since the January break, still −12.0% on the year: what an early-cycle recovery looks like rather than a finished trade. And HALO Growth — growth unrelated to AI — has firmed to +8.5%, back near the top of its range.
Sideways at the capex layer, new highs at the opex layer, a recovery high at the application layer, and a bid under non-AI growth. That configuration is not what a bursting bubble looks like. It is what a handoff looks like.
The forecast line that embarrasses the bear case
Bad times for AI permabears: the capex forecasts have now moved significantly higher again. Initial 2026 estimates sat below $300 billion in late 2024. By late 2025 they had risen to roughly $525 billion. Big Tech has now guided to more than $700 billion of 2026 capex — about 150% above where the original estimates started.
And the least speculative dollar in that chain keeps confirming it: TSMC — the foundry whose capacity is pre-sold years out — raised its own budget twice this year and committed a further $100 billion to Arizona. We graded that name today in Part III of the handoff series, and the interconnect layer around it in Part V: the spending is contracted, not hoped for.
Overnight, the same message in single names
Wednesday’s close paid the rotation’s challengers: Nebius +34.1% to 259.20, straight through a famously disclosed short entry near 212, on a print that merely reiterated guidance; CoreWeave +19.3% to 107.73, holding and extending its overnight gap. After the bell, the market charged the incumbents for their beats — Coherent sold about 5% after its eleventh straight EPS beat, Cisco about 4.7% after its twenty-first consecutive double beat and a raised guide.
And Seoul went vertical a third session: KOSPI +3.8% into a bull market, roughly +22% off its July 30 low, Samsung and SK Hynix leading. The memory complex — the piece of the capex layer where pricing is tightest — is being re-bought at size while the broad capex basket digests. Rotation, again, not exit.
The hinge at 14:30
Everything above is positioned on the dovish branch that Tuesday’s in-line CPI opened — and July PPI at 14:30 CET is the first number that can contest it. The setup: consensus carries +0.2% headline on the month after −0.3%, core at +0.3%, and the year-over-year pair still hot — 4.9% headline expected after 5.5%, core 4.2% after 4.7%. Initial claims print alongside, 202K expected, the tight range intact.
The tells to watch into the print: IEF closed 92.96, a second close beneath the broken 93.17 floor — and dollar-yen sits at 159.35, pressing the 159.5 line after refusing the dovish CPI outright. PPI sets the rate path, the rate path sets the yen, and the yen sets the stress.
The number came in cooler. Headline wholesale prices were flat in July against the +0.2% consensus, and the annual rate stepped down to 4.7% against 4.9% expected, from 5.5%. Core ran a tenth under consensus on the month at +0.2%, with annual core easing from 4.7% to 4.2% — in line, but the direction is the story.
The one measure that leaned the other way barely did: ex food, energy and trade printed +0.4% against roughly +0.3% expected — a tenth hot on the narrowest cut, against a tenth cool on the standard core. The labor side: initial claims 209K against a ~202K consensus, continuing claims better than expected at 1.78M — softening at the margin, not cracking. Net: a good number. The slate landed on the dovish side, and the branch Tuesday’s CPI opened gets its ratification from the pipeline.
One detail inside the release belongs to this page’s own thesis: the AI-adjacent line items ran a touch inflationary while the broad index sat flat — construction inflation above all, with manufacturing categories alongside. The build-out is now large enough to leave fingerprints in the producer-price data itself; the $700 billion doesn’t just buy chips, it bids up the inputs around them.
And the tell answered within minutes, the same way it answered Tuesday: dollar-yen sat at 159.4 through the release — a cool print, refused again, the second refusal in twenty-four hours. That keeps the chain live rather than resolved: if US inflation data this soft cannot move dollar-yen off the 159.5 line, the move is a yen story wearing a US-rates costume, and the Bank of Japan question stays at the top of the risk list. Into the close, the checklist is unchanged — whether IEF can reclaim 93.17 on a friendly pipeline print, whether the semis bid and Seoul’s follow-through survive the session — and then the docket takes over tonight.
C · members block
Into the close
Scoring rules for the day, written before the number.
monday.com’s window closes tonight. Cumulative −7.7% through yesterday; the three-session verdict lands at the bell and goes into the print record as written.
Lumentum has until Friday to pick its lump. +13.6% on day one is exactly the in-between its record says never survives — sixteen or better, or minus nine or worse, nothing else in ten prints.
Applied Materials prints after the close as the anti-Coherent. Ten-for-ten on EPS, one paid reaction in ten tries, 1-3-6. If it beats and gets sold again, the week’s sorting rule — pay the challengers, charge the champions — closes its case. If it gets paid, the rule has its exception.
Day-two retention for the pair. CoreWeave and Nebius hold their gains through a second close or they don’t — day two is where paid prints have been failing this season.