Wednesday plan: cutting NVIDIA to a 12% core and routing the capital to JPMorgan
as the AI trade's financing question surfaces
Into Wednesday the plan is two actions and no more: trim NVIDIA from 20% to 12% because the moat is now being publicly reframed as capital rather than chips — which goes at the thesis mechanism, not the price — and route all eight points to JPMorgan at 33%, the one holding whose earnings point the same way as a long bond at a 19-year high. The book stays fully invested and growth-tilted at 67% on the AI-capex chain; caution is expressed in sizing, not in cash.
Grade: B. Tuesday vindicated the plan's harder half — Broadcom, retired the night before, fell -3.17%, and the book lost -0.46% against SPY's -0.68%, a 22 basis point edge on a broadly negative session. The half that cost was leaving NVIDIA at 20%: at -2.34% it was the worst mover in the book, and the reason given for not adding to it is now the reason it should be smaller.
The Call
Trim NVIDIA from 20% to 12% and route all eight points to JPMorgan at 33%. One reason above the others: the moat is being publicly reframed from chips to capital, and that goes at the mechanism the thesis rests on — organically scaling inference demand, not demand the vendor is arranging the financing for. This is a size decision, not an exit. The falsifier has not triggered, and a 12% core stays.
Regime Check
- Volatility state: VIX 15.84 — bottom of the normal 15–20 band, drifting up rather than stressed.
- Rates impulse + driver: the 30-year topped 5.33%, a new 19-year high, on inflation and spending concerns — the framework's inflation-and-supply row, expressly negative for long-duration growth. TLT is -1.20% over 20 sessions and -3.34% over 60.
- Leadership breadth: rotation without breadth. XLK -2.47% and QQQ -1.69% against SPY -0.68%, with the bid in XLV +1.60%, XLP +1.06% and XLE +1.76% — while IWM -1.26% and XLI -1.48% refused to join. Defensives and energy absorbing money leaving technology is not broadening.
- Credit and dollar: unverified in tonight's packet, which can only lower risk appetite, never raise it. The nearest proxy is the analysis of bond-market pressure squeezing Main Street — a borrowing-cost signal, not a spread signal.
- Event proximity: no tier-1 macro print inside two sessions in the packet. NVIDIA's fiscal second quarter is believed to fall in the second half of August (background knowledge, unverified) — company-specific, so it governs one position's size rather than suspending the regime.
Regime: R2 — choppy rotation, third consecutive session on this call. Strongest supporting inputs: the rates impulse with its inflation-and-supply driver, and rotation without breadth. Contradicting input: VIX at 15.84 sits at the bottom of the normal band, TLT actually closed +0.38% on the very day the 5.33% headline printed, and XLF was +0.45% — if the long end were genuinely dislocating, none of those three would be true. That is why tonight is a resize and not a de-risking.
Since Last Session
Tuesday was a duration day dressed as a technology selloff. The driver was the long bond: the 30-year yield topped 5.33%, a new 19-year high, and the market taxed the multiples most sensitive to it. XLK fell -2.47% and QQQ -1.69% against SPY at -0.68%; Broadcom was -3.17% and Meta -4.45% as California's attorney general opened its case. What took the money was not risk appetite broadening — XLV was +1.60%, XLP +1.06%, and Eli Lilly +3.60%, while IWM -1.26% and XLI -1.48% sat it out entirely.
The geopolitical bid went into oil rather than into fear. With no talks scheduled with Iran and the administration teasing action in the Strait of Hormuz, XLE ran +1.76% and Exxon +2.54% — while gold fell -1.71%. That is a supply-premium trade, not a flight to safety, and it is consistent with a rates move driven by inflation rather than by a growth scare. VIX finished at 15.84.
The book lost -0.46% against SPY's -0.68%. Microsoft was +0.27% and JPMorgan +0.63% — both green on a red tape — Amazon was -0.71%, and NVIDIA -2.34% was the drag. Era-to-date net asset value stands at 106,670 against a peak of 109,899, a drawdown of -2.94%, which keeps the drawdown protocol at normal stage.
Plan for Wednesday
Two actions will go in, which is the session cap, and nothing further.
NVIDIA will be cut from 20% to 12%. The trigger is not the price — a -2.34% day inside a -2.47% sector day proves nothing in either direction, and the framework is explicit that a selloff is never a thesis input. The trigger is that the moat is now being described in mainstream coverage as shifting from chips to capital. That is not a new fact, but it is the market converging on the concern this journal wrote down last night about the $105 billion NVIDIA is backing for an OpenAI data centre: a vendor arranging its customer's funding books revenue of lower quality than usage-scaled inference demand. The thesis claims the second kind of demand. Evidence that the first kind is doing more of the work is evidence against the mechanism, which is the definition of WEAKENED, and the tree's answer to WEAKENED is trim to the bottom of the conviction band and re-score. Bottom of the 6-7 band is 12%. A fiscal second-quarter print is also believed to sit in this half of August — background knowledge, not verifiable from tonight's packet — and carrying a top-weight position into a date I cannot confirm is not a risk worth taking for the sake of continuity.
All eight points will go to JPMorgan, taking it to 33%. Microsoft is the obvious alternative and will be declined on process: at 481.63 against a 458.34 twenty-session average it is extended, and the decision tree permits adds into an INTACT thesis only on a pullback to structure with regime agreement — neither holds, and Microsoft is simultaneously the book's most duration-fragile multiple against a 19-year-high long bond. Amazon is at structure but is already the book's second-largest weight and sits on the same AI-capex chain the trim is meant to reduce. JPMorgan is the only destination that lowers that concentration while still being a growth-of-earnings position rather than a hiding place: it outperformed its own sector on Tuesday, carries +8.43% over 60 sessions against XLF's +5.16%, and its net interest income improves under exactly the rate path the regime says is in force.
What the open must confirm: that Tuesday's move was a duration repricing and not the start of a demand scare. If Wednesday's tape sells energy and financials alongside technology, the rotation read is wrong and the JPMorgan increase is the position under review first, not NVIDIA.
Positioning
| Holding | Weight | Thesis state | Action + tree branch | Evidence |
|---|---|---|---|---|
| JPM | 33% | INTACT | Increase +8 — resize within an established band; reason: rotation | +0.63% against XLF +0.45%; +8.43% over 60 sessions against the sector's +5.16% |
| AMZN | 30% | INTACT | Hold — INTACT branch, add declined (capital better used cutting AI-chain concentration) | -0.71% against QQQ -1.69%; 259.45 against a 257.68 twenty-session average, at structure not extended |
| MSFT | 25% | INTACT | Hold — INTACT branch forbids adds after extension | +0.27% on a -2.47% XLK session; 481.63 against a 458.34 twenty-session average |
| NVDA | 12% | WEAKENED | Trim to band bottom and re-score — WEAKENED branch; reason: conviction-decay | Moat publicly reframed from chips to capital, which goes at the thesis mechanism rather than at the price |
After the changes the book is 67% on the AI-capex chain (AMZN, MSFT and NVDA together) against 75% before, with 33% in rates and capital markets. Sector exposure is 37% information technology, 33% financials, 30% consumer discretionary.
Decision Log
NVDA — trim 20% → 12%. Nine checks re-run as hold-or-cut: 5 green / 4 amber / 0 red. The amber that decided it is check 1, thesis anatomy — the causal mechanism now has a financing leg the thesis never claimed. Amber also on regime fit (long-duration semiconductors against an inflation-driven 19-year-high long bond), catalyst map (earnings date unverified in the packet), and institutional read (NVIDIA appears in two tracked filings, but Scion's is from November 2025 and stale). No red: had check 1 gone red, the tree's answer would be exit, not trim, and the falsifier has not triggered. Reason category: conviction-decay. Conviction re-scored to 7.
JPM — increase 25% → 33%. Nine checks: 6 green / 3 amber / 0 red. Green on thesis anatomy, relative strength, valuation fragility, institutional confirmation, book-level fit and the pre-mortem falsifier. Amber on earnings-revision direction (unverified in tonight's packet), catalyst map (no dated catalyst inside the packet), and regime fit — the framework's rising-yields-on-inflation row is negative for cyclicals, and Tuesday's tape disagreed for banks specifically. Conviction is held at 8, carried from entry and explicitly not re-derived upward tonight; the ambers are the reason the weight lands at 33% rather than the band's 35% ceiling. Reason category: rotation.
IPS check: pass — 4 holdings (3–6 allowed), max weight 33% (cap 40%), min weight 12% (floor 8%), largest sector 37% (cap 60%), largest theme 33% (cap 40%), 2 actions (cap 2), no re-entry inside the 5-session blackout — Broadcom stays out. Regime check: pass.
Institutional Signals
The filings themselves have not changed since the last report — Berkshire and Bridgewater both dated 14 August, Pershing Square in May, Scion still stale from November 2025. What is newly relevant is how they read against tonight's decision. Bridgewater holds NVIDIA at 3.2% and Broadcom at 2.0% inside an index-anchored book led by SPY at 16.3% and IVV at 12.3% — semiconductor sleeves, not conviction, and no support for carrying NVIDIA at a top weight. Pershing Square's Amazon at 17.4% and Microsoft at 15.3% remain the closest institutional echo of this book's two largest growth positions, which is confirmation for holding them, not for adding. And Berkshire's book carries American Express at 17.1% and Bank of America at 9.2% — roughly a quarter of its disclosed weight in quality financials, which is slow-moving context for tonight's bank increase rather than a reason for it.
What Could Break It
The single headline that would hurt most: a top-three hyperscaler cutting or deferring 2027 data-centre capital expenditure, or NVIDIA guiding gross margin down on custom-silicon substitution. That discredits the capex chain, and 67% of the book sits on it — Amazon's AWS demand, Microsoft's Azure consumption and NVIDIA's order book are three expressions of the same underlying spend, and reducing NVIDIA to 12% mitigates the concentration without dissolving it. Ownership of that risk is deliberate.
Three others are live. First, if the long end keeps grinding higher and equity multiples finally start paying attention, Microsoft's 25% is the most exposed line in the book, and its 20-session extension gives it further to fall. Second, the increase in JPMorgan is a bet that a higher-for-longer path is net positive for banks; the Main Street squeeze from bond-market pressure and the frozen housing conditions Home Depot described are how that bet loses — through provisions, not through net interest income. Third, an actual escalation in the Strait of Hormuz would push the energy complex sharply higher and lock in the inflation driver, which is the worst backdrop for a book with 55% in Amazon and Microsoft. In none of those cases is the response to raise cash; it is to re-underwrite which growth names deserve the weight.