$50,300.32

The measured shock absorber of this book, not a label: beta of -0.189 against NVDA and the only position that rose on all three stress tests (+2.07% on the ten worst TLT days, +1.13% on the ten worst SPY days). Cash conversion 1.77x, net debt 0.22x EBITDA, 3.2% of shares retired a year. Trailing valuation reads 82nd percentile, but energy is cyclical and the governing measure is the multiple on the average margin: 50th percentile. The three-year margin trend of -9.7pp is the base year sitting at the oil peak, not deterioration. Held at full weight with one honest caveat: the negative beta is a property of the current war-premium regime and reverses if Hormuz reopens.

Breaks if: A quarter shows all-in FCF breakeven above $55 WTI, or net debt/EBITDA above 1.0x, or the 5% oil growth guidance is dropped while Brent is above $75.

Cash conversion 1.35x, operating margin up 5.0pp over three years to 46.8%, net debt 0.10x, dilution -0.1%, and 18th percentile trailing, 23rd on the average margin. One question does not come back clean, and August's claim that this position cleared all six without a single breach was wrong: return on invested capital has fallen from 32.0% to 26.6% over the same three years. That is the capex programme arriving in the denominator rather than a business deteriorating - invested capital is growing faster than operating profit while the data centres are built - but it is a real decline and it belongs in the record. It also carries the lowest AI-factor loading of the AI names at 0.217 against NVDA, which is why the cut to the AI chain came from elsewhere.

Breaks if: Azure growth below 30% y/y in any quarter, or FY free cash flow below $60bn, or quarterly capex above $40bn without concurrent backlog growth, or either cluster rule.

Demand is confirmed and the balance sheet is the new question. FQ2 revenue $96.2bn (+106%) with guidance of $108bn against $104.2bn consensus, and the stock sits at the 1st percentile of its own valuation history. But cash conversion is 0.78x against our own reject threshold of 0.80: receivables went $38.5bn to $63.1bn and inventory $21.4bn to $31.6bn in six months. And in June 2026 the company issued $25bn of senior unsecured notes for general corporate purposes, taking total notes from $8.5bn to $33.5bn, alongside a $25bn commercial paper programme; commitments total $366bn with $120bn falling inside twelve months. Our own macro view names the financing channel as the main risk, and our largest position is now part of it. Conviction cut from high to medium and the weight from 11% to 7% for exactly that reason.

Breaks if: Cash conversion below 0.8x in two consecutive quarters, or commitments falling due within twelve months above $150bn, or a new debt issue above $10bn; or two consecutive quarters of declining data centre revenue q/q; or Rubin shipments not started before the end of October 2026; or either cluster rule.

Measured on its own filings this is the strongest business in the book: operating margin from 42.0% to 60.3% in three years, gross margin 67.7%, net cash of NT$2,486bn, and monthly revenue running +44.7% y/y in July against a cumulative +37.0% for the year. Capex guidance was raised twice, to $60-64bn, not cut. The statements are reported in TWD and carry no SEC XBRL, so our pipeline suppresses the multiples; the honest input is the company's own diluted EPS per ADR in USD, $13.86 TTM, which puts the ADR near 29.8x. Trimmed anyway, from 9% to 6%, purely because it belongs to the factor this round is reducing.

Breaks if: Three-month rolling revenue growth below 25% y/y (a deceleration trigger; the old 15% level was never going to fire in time), or capex guidance cut below $60bn, or gross margin below 60% for two quarters, or either cluster rule.

August's thesis quoted a backlog up 103%. That figure was Electrical Global including acquired Boyd backlog; organically it is +54%, and this round corrects it in public. What the company actually steers on held up: segment margin 23.1%, down 80bp but still above the top of guidance, with most of the decline explained by intangible amortisation rising from $129m to $255m after Boyd and Ultra PCS. Organic revenue +14%, book-to-bill 1.3, guidance raised on 31 July. Leverage is the real change: the Boyd acquisition closed 12 March for $9.55bn net, debt went from $9.9bn to $20.6bn, quarterly interest expense from roughly $70m to $201m, and buybacks stopped. Two leverage figures matter here and they are not the same measure. On trailing twelve months, net debt is 3.30x EBITDA; on annualised second-quarter earnings it is 2.4-2.8x. The break condition below is written on the annualised quarterly basis, because the trailing figure sets three pre-Boyd quarters of earnings against a fully post-Boyd balance sheet and therefore overstates the leverage the company actually runs at. Interest cover is 12.0x trailing and roughly 6.9x annualised. Trimmed from 9% to 6% on leverage and a 92nd-percentile multiple, not on the price drop: the -11.2% in August came on 18 August with no company filing at all, a rate-driven de-rating of the whole AI-power complex.

Breaks if: Net debt divided by annualised quarterly GAAP EBITDA above 3.0x in two consecutive quarters (no date clause: the old trigger deferred to Q2 2027 while the level was already passed), or segment margin below 22.0%, or organic growth below 5% in any quarter.

Down 9.6% in month one and now at the 1st percentile of its own valuation history, 69th on the average margin. The proof of AI revenue is still the strongest in the group and free cash flow is the open question, exactly as it was in August. Kept at half position rather than added to, because the instruction this month is to reduce the factor rather than to buy the cheapest expression of it.

Breaks if: Cloud growth below 35% y/y, or backlog declining sequentially, or FY free cash flow below $40bn, or either cluster rule.

The calmest expression of electrification and the cheapest, at the 19th percentile trailing and 53rd on the average margin, with roughly 60% utility revenue and net debt at 1.42x EBITDA. Held unchanged: it neither drove the month nor broke anything, and it is the one grid name whose valuation does not depend on the AI story staying intact. One caveat belongs with these figures: Hubbell's most recent quarter in the SEC filing feed ends 31 March 2026, a quarter behind every other position in the book, so its leverage and returns are measured on older data than the rest.

Breaks if: Utility segment organically negative, or net debt/EBITDA above 3.5x, or adjusted Utility margin below 23% in a quarter.

This is a distillate position, not a refining one. The diesel crack is $105 against $31 at the start of January, distillate inventory sits 14% below its five-year average at the lowest August level since 1951, and refineries are running at 98%. The war spike of 1 September pushed the gasoline crack down while distillate kept climbing, which is the distinction that matters. It earns its place on measured behaviour: beta -0.177 against NVDA, and it rose on the worst SPY and TLT days. Two things must be said plainly. Our pipeline returns no revenue in any of the eight quarters and no debt in six, so the operating margin, ROIC, leverage and normalised percentile here are hand calculations on a secondary source and are not reproducible from our own data. And on that hand calculation the normalised multiple is at the 100th percentile: the low headline P/E is the peak-earnings signal, which is why this is exploratory at 4% and not a medium position at 7%.

Breaks if: Adjusted refining operating income per barrel of throughput below $8.00 in two consecutive quarters (Q2 2026: $16.56). The figure appears in every quarterly earnings table.

The only position in the book that breaches both valuation measures: 86th percentile trailing and 84th on the average margin, where for a cyclical the second one governs. The business itself is unchanged, with net leverage at 3.29x EBITDA against the partnership's own 3.0x target and our 3.5x break, interest cover 5.2x, and a distribution funded internally; it contributed +0.28pp in month one. But this was always an explicit rates position, and the 10-year moved from 4.63% to 4.75% while the 30-year touched 5.31%. Cut from 7% to 4%: the valuation says trim and the rate path says the short leg is running against us.

Breaks if: Distribution coverage below 1.4x or net leverage above 3.5x in a quarterly report, or the 10-year closes a quarter above 5.25%.

Held as uncorrelated volatility, and explicitly not as a hedge. We measured it: beta of 0.232 against NVDA, higher than MSFT's 0.217, and it fell on both the ten worst NVDA days and the ten worst SPY days. It only held up on the worst TLT days. Gold rose 10-13% in August depending on the basis and gold miners had their best August since at least 1994, but the metal still sits about 22% below its January high, so this is a volatile asset that happens not to move with the book, not an airbag. Physical rather than miners: no company risk, no multiple that can de-rate, 0.10% expense ratio, and a ten-year record ahead of the senior miners.

Breaks if: Central bank buying below 150 tonnes in two consecutive quarterly WGC reports (Q2 2026: 288.9t), or the 10-year real yield above 2.75% together with a lower quarterly close in gold, or tracking worse than -25bp a year against the LBMA PM fix.

The purest physical expression of the data centre build-out, with the best cash conversion in the book at 1.94x and virtually no debt. It is also the highest factor loading per point of weight we own, at 0.833 against NVDA, which is precisely why it is cut from 5% to 3%: reducing the AI factor means cutting the names that carry the most of it per unit, not the names that are easiest to sell. Note the valuation asymmetry: 6th percentile trailing but 82nd on the average margin, so the cheapness depends on the current peak margin holding.

Breaks if: Book-to-bill below 1.0 for two consecutive quarters, or backlog declining sequentially for two quarters, or adjusted operating margin below 18%, or either cluster rule.

The one position that gains from a cost the book already pays elsewhere: NVIDIA named memory pricing as the reason its gross margin peaked and guided down to 71-72%, and that cost is Micron's revenue. DRAM contract prices rose 93-98% in Q1 and 58-63% in Q2, and the company reports net cash of $20.3bn with ROIC around 63% and revenue growth of 70.6%. Three objections stand on the record. It breaks question five on the average margin at the 93rd percentile. At 3% it carries more AI-factor beta (+0.028) than MSFT does at 10% (+0.022). And our own macro conclusion this month says not to add to the AI chain before the 16 September FOMC, which this position contradicts. It is held at an exploratory 3% for those reasons, not despite them.

Breaks if: TrendForce DRAM contract prices negative q/q (a deceleration trigger: the series has already gone from +93-98% to +58-63% to +13-18% without any level trigger moving), or two consecutive quarters guiding gross margin below 70%, or a revenue midpoint below the preceding quarter.

Zero to three month T-bills at roughly 3.9%. With cash at 18% this is the second half of a defensive sleeve that now stands at 30% of the book, the widest it has been, because the macro verdict moved from neutral to low. Cash in this model earns nothing; this does.

Breaks if: A Fed cutting cycle underway (two cuts) with the 2-year below 3.25%: reconsider toward duration or equities.