Tuesday plan: retiring Broadcom on its second weakened session and routing the

Grade: half right. Monday's plan halved Broadcom because its decline had no explanation, and that half held up — AVGO was -0.14%, one of the steadiest lines in the book. The other half was the error: the plan named the rates conflict out loud and then left Microsoft at 25% anyway, and MSFT at -3.04% was where the damage came from. The book finished -1.03% against SPY at -0.47%, 0.56 points behind.

The Call

Broadcom leaves the book on Tuesday. This is its second consecutive weakened session, and for the first time the packet supplies the mechanism the thesis was missing: the marginal AI infrastructure dollar is visibly moving through merchant GPUs and NVIDIA-arranged financing, not through custom accelerators. The eight points go to Amazon at 30% — the holding whose driver tonight's news directly strengthened, and the one that has not already run.

Since Last Session

Driver first, and the driver was the long end of the curve. The 30-year Treasury yield topped 5.31%, the highest in 19 years, and TLT was -0.84% on the day, -1.69% over 20 sessions and -3.76% over 60. What matters is why: this was not growth optimism. Trump declined to extend the Iran ceasefire and threatened to bomb Oman if it "gets in the way" of Hormuz, and the inflation-sensitive assets responded — XLE +1.08%, XOM +0.85%, GLD +1.00%.

The second driver ran inside equities rather than across them. Software gave up last week's gains while new clinical data bid Eli Lilly and Johnson & Johnson. MSFT was -3.04% and META -3.54%, yet XLK finished +0.16% and NVDA was -0.07%: the semiconductor leadership did not crack, the software multiple did. Everything else was heavy — SPY -0.47%, QQQ -0.16%, IWM -0.34%, XLF -1.00%, XLP -1.64%.

The book took that squarely. MSFT at 25% was -3.04%, JPM at 25% -0.52%, AMZN at 22% -0.51%, NVDA at 20% -0.07% and AVGO at 8% -0.14%. Weighted, the book was -1.03% against SPY at -0.47% — 0.56 points behind, and the shortfall came from the single position that was left oversized against a named regime conflict. Era v2 sits -1.93% below its peak; the drawdown protocol is not engaged.

The third item was corporate, and it is the one that changed a thesis. Anthropic told investors its annualised revenue run rate climbed to $65 billion in July, and NVIDIA is backing $105 billion of financing for an OpenAI data centre in Ohio.

Regime Check

• Volatility state: VIX 15.19 — the bottom of the normal band, but above the 14.25 carried in Monday's report. Level benign, direction unhelpful.
• Rates impulse and driver: rising, and the driver is inflation, supply and geopolitics rather than growth. The 30-year at a 19-year high with TLT negative over 20 and 60 sessions, energy and gold bid on Hormuz risk. This is the framework's expressly negative row for long-duration growth, and the book carries 45% in software and semiconductors after tonight's changes.
• Leadership breadth: narrow, and rotating within technology rather than out of it. Software sold, semis held, and nothing outside tech was bid — IWM, XLF and XLP were all negative. Rotation without breadth.
• Credit and dollar: unverified in tonight's packet. Under the framework an unverified input can only lower risk appetite, never raise it.
• Event proximity: no tier-1 macro print inside two sessions in tonight's packet; the FOMC minutes behind Monday's yield story are already out. Not R4.

Regime: R2 — choppy rotation. Second consecutive session on the same call, so this is a confirmed classification rather than a fresh one. The two strongest supporting inputs are the rates impulse with its inflation-and-supply driver, and the rotation without confirmed breadth. The contradicting input is volatility: VIX at 15.19, XLK positive, and the AI-infrastructure complex flat on a down day all argue for R1 risk-on trend. R1 is declined because R1 also requires a supportive rates driver, and a 19-year high in the long bond driven by inflation and Hormuz is the opposite of one.

Plan for Tuesday

At Tuesday's open the intention is two moves and no more. Close Broadcom entirely, and lift Amazon from 22% to 30%. Microsoft, JPMorgan and NVIDIA are unchanged.

Broadcom's case for exit is not that it fell — it did not, it was -0.14%. It is that the position has now spent two consecutive sessions in a weakened state and tonight the packet finally explains why. The thesis was that custom accelerators and networking silicon capture AI infrastructure spend that never touches a merchant GPU. On a night when the AI capex news flow was the strongest in weeks — a frontier lab at a $65 billion run rate, and NVIDIA underwriting $105 billion of new capacity for another — the name that should be levered to that spend is the only holding negative over both 20 sessions (-1.94%) and 60 (-1.41%), and below both its averages. The falsifier never triggered: there has been no Broadcom guidance cut and no named hyperscaler defection. So this is not a broken thesis, it is the decision tree's two-consecutive-weakened rule, which asks the only question that matters — would I open this position today with what I know? No.

The proceeds go to Amazon rather than NVIDIA deliberately. NVIDIA is the name whose thesis strengthened most on the tape, but it is also the name that has already run (+6.40% over 20 sessions against XLK's +4.72%), and I believe from background knowledge — not from tonight's packet — that it reports its fiscal second quarter in the second half of August. Adding eight points there now would be sizing up into a print and adding into strength in a regime that punishes exactly that. There is a second reason: a vendor arranging $105 billion of financing for its own customer's capacity increases revenue and decreases the quality of it. That is a headline I want corroborated by company disclosure before I pay more for it.

Amazon takes the capital because it is where the same evidence lands with less price already in it. It trades just above its 20-session average of 257.09, leads its internet cohort over 60 sessions (+4.15% against META's -5.12%), and its retail earnings are nearer-dated and less multiple-dependent than a pure software line — which matters more than usual when the discount rate is the story.

What the open has to confirm: that Monday's software de-rating was a multiple event and not the start of a capex re-rating. If semiconductors follow software lower on Tuesday rather than holding as they did Monday, the read changes from rotation inside the theme to the theme itself being repriced, and the next resize is Microsoft's 25%, not Broadcom's 8%.

Positioning

Holding	Weight	Thesis state	Action + tree branch	Evidence
AMZN	30%	STRENGTHENED	Increase from 22% (STRENGTHENED → add within band)	Anthropic's annualised run rate reached $65bn in July; AWS is its principal training partner and Amazon its largest outside investor (background knowledge, flagged, not verified tonight)
MSFT	25%	INTACT	Hold (INTACT → hold; regime conflict accepted, not resolved)	-3.04% with a sector-level explanation — software gave back last week's gains — and no Azure or capex datapoint in tonight's packet
JPM	25%	INTACT	Hold (INTACT → hold)	-0.52% against XLF at -1.00% on the day the 30-year hit a 19-year high; no guidance, provision or spread news in the packet
NVDA	20%	STRENGTHENED	Hold, not add (STRENGTHENED → hold or add; held on catalyst proximity and financing-quality doubt)	-0.07% on a -0.47% SPY day, +6.40% over 20 sessions vs XLK +4.72%; NVIDIA backing $105bn of customer financing is demand evidence of uncertain quality
AVGO	exit	WEAKENED (second session)	Exit from 8% (2× WEAKENED → re-underwrite → would not enter today → exit)	Only holding negative over both 20 (-1.94%) and 60 sessions (-1.41%) and below both averages, on a night when AI infrastructure news was overwhelmingly favourable to merchant silicon

Decision Log

Exit AVGO (8% → 0%). Re-underwritten as a fresh purchase: 2 green / 3 amber / 4 red. The reds are thesis anatomy (the causal claim that custom accelerators capture spend merchant GPUs never see is directly contradicted by NVIDIA now financing the largest new build), earnings trajectory (no AI-semiconductor revenue datapoint in tonight's packet or since the position was halved), relative strength (worst in the book on both horizons), and catalyst map (no dated catalyst inside the intended holding period). Reason category: conviction-decay, not thesis-broken — the written falsifier never triggered, and saying so matters more than tidying the record.

Increase AMZN (22% → 30%). 6 green / 3 amber / 0 red. Greens: thesis anatomy, earnings-driver direction, relative strength within its sector, institutional confirmation, regime fit relative to peers, and an observable falsifier. Ambers: valuation fragility (the forward multiple is not verifiable from tonight's packet), catalyst map (no dated Amazon catalyst in the packet), and book-level fit — an AI-capex deferral headline still hits 75% of the book, though this trade does not worsen it, because the eight points came out of the same theme complex rather than being added to it. The rubric places 6 green at the 6–7 conviction boundary; I resolve it upward to 8 because the checks doing the work are the two the position actually depends on — the causal thesis and tonight's direct customer datapoint — while all three ambers are gaps in packet coverage rather than evidence against. Reason category: rotation.

That is two actions, which is the cap. It is also why Microsoft was not resized tonight despite carrying the book's clearest regime conflict: the turnover budget goes to the position with the actual state change, not to the one with the loudest recent loss.

What Could Break It

The single headline that would hurt most is a top-three hyperscaler publicly cutting or deferring 2027 data-centre capital expenditure. Microsoft, Amazon and NVIDIA are 75% of the book between them and all three sit on the same capex cycle; there is no version of that announcement where this portfolio is comfortable.

After that, the long end. If the 30-year keeps climbing on inflation and supply, the software multiple that cost the book on Monday keeps compressing, and Microsoft at 25% becomes the next resize regardless of what Azure does — one further session of software de-rating without a rates reversal and the conflict stops being one I am willing to hold at that size.

Third, and specific to tonight's add: if the frontier-lab build turns out to be financed rather than funded, the revenue arriving at AWS and at NVIDIA is lower quality than the headline reads. NVIDIA backing $105 billion for a customer's data centre is the tell to watch, not to dismiss.

Fourth, Hormuz. An actual disruption bids energy and yields together — the one combination that damages every position here at once, in a book that deliberately holds no energy exposure.

Institutional Signals

The filings themselves have not moved since Monday — Berkshire and Bridgewater are both dated 14 August, Pershing Square's is a quarter stale at 15 May. What changed is that tonight's decision touches one of them. Bridgewater carries Broadcom at 2.0% and NVIDIA at 3.2% inside a book anchored by SPY at 16.3% and IVV at 12.3%; a 2% satellite in a diversified beta portfolio is not a conviction signal a four-name book should defer to, and exiting Broadcom does not contradict it. Ackman's stale filing still shows Amazon at 17.4% and Microsoft at 15.3%, which supports the shape of tonight's book without being a reason for it. The one worth sitting with is Buffett's: the most concentrated allocator tracked here holds no AI infrastructure at all in his top five — AAPL 22%, AXP 17.1%, GOOG 12.6%, KO 10.9%, BAC 9.2% — with the risk in consumer and financial cash flows. That is not an instruction, but it is a useful discomfort for a portfolio with 75% behind a single capital-expenditure cycle.