Track Record
LiveThe book's return since the June 26 inception, with holdings marked at each nightly close and the benchmark you choose below quoted live.
Cumulative return since inception: Book vs. benchmark
Allocation & Performance
Exhibit IHow the book's weight sits by sub-sector, and which names have actually driven the return.
Allocation
Contribution to return
Run vs. room: return against upside
Earned vs. remaining upside
Construction & Concentration
RiskHow the book is built, and how much of it rides on one theme.
Because the book is really one bet, it carries a break at the book level on top of each position's own rule. Either trigger de-risks the whole book, not a single name:
- Fundamental. If the four largest hyperscalers guide AI-infrastructure capex flat-to-down for two consecutive quarters, the factor thesis under ~94% of the book is broken and exposure comes down across the board.
- Mechanical. If the book draws down 25% from its peak, net exposure is cut regardless of individual theses, a hard floor beneath the per-name 20% breaks.
The Book
Exhibit IIEvery name in the book: its weight, current mark, upside left to target, and thesis status.
Position Theses & What Would Change My Mind
Exhibit IIIThe bull case for each name in a sentence, paired with the evidence that would break it.
The one mega-cap platform worth conviction weight. Azure+OpenAI is the only hyperscaler where AI monetization is already in the P&L (Copilot ARPU, Azure consumption) rather than just capex, with no search-disruption tail-risk, and it has lagged YTD. The cheapest way to own the hyperscaler beta the index over-owns elsewhere.
The single irreplaceable AI-compute standard: CUDA lock-in plus the Rubin ramp. Held as our only GPU name and sized as a deliberate overweight (14% vs the index's ~9%), a conviction bet on the one name we won't fade. Whether that overweight adds or costs alpha is what the record will show; the book's differentiation comes from the off-index names it owns instead (TSM, CRDO, ORCL, PANW), not from underweighting the index leader.
The only at-scale custom-AI-silicon franchise outside Nvidia: AI revenue inflected to $10.8B in Q2 FY26 (+143% YoY, ~half of total) across a six-customer cohort (Google, Meta, Anthropic, OpenAI + two), with management guiding FY27 AI to >$100B, layered on a high-margin VMware software annuity. The post-Q2 ~26% drawdown de-risked the entry, though the bear case ~$344 now sits ~6% below spot.
The cleanest non-mega AI-cloud exposure the index under-weights. OCI's RPO backlog inflection (~$638B, +363% YoY) is re-rating the multiple from legacy-database toward hyperscaler, a re-rating the cap-weighted index does not price. Sized at conviction weight but not co-top, since the thesis depends on that backlog converting cleanly.
The cornerstone and the purest embodiment of the mandate: genuinely NOT in QQQ (NYSE ADR), the indispensable ~fairly-valued toll every accelerator vendor pays, ~66% gross margins with 5-10% leading-edge price hikes, and the lowest-correlation quality name available. It captures the entire AI build-out without picking a winner.
The book's sole NON-CAPEX leg, the vector that keeps this from being a pure AI-capex proxy. Platformization (one vendor consolidating network + cloud + SOC) plus AI-SOC drives NGS ARR ~60% with RPO +36%, a recession-resilient, index-under-represented grower, chosen over Hold-rated CRWD's premium multiple. Now trading around our re-rated $365 target (up from $310 on the Aug-14 dashboard refresh, after a run to ~$384), so it stays a 6% Medium-conviction diversifier hold: kept for the non-capex ballast it provides, not for target upside.
The second platform, held ONLY for its distinct driver: AWS reaccelerated to ~37% (fastest in 18 quarters) plus Trainium custom silicon, and a retail-margin inflection adds a non-AI earnings vector the rest of the book lacks. A real second theme, not a third hyperscaler clone.
The purest off-benchmark small-cap the Nasdaq-100 entirely misses: active electrical cables (AECs), the under-owned connectivity bottleneck of scale-up AI clusters. Revenue tripled past $1.3B with >80% FY27 growth guided and optical scaling past $600M, on 5 hyperscaler customers. Sized small for the customer-concentration risk.
The highest-beta direct lever on AI-infrastructure demand: $104B contracted backlog ($129B with early-Q3 commitments, including a $21B Meta deal), a Platinum-tier neocloud, first to validate Vera Rubin NVL72. Now a Nasdaq-100 member (owned for beta/optionality, not as off-benchmark) and sized down for its ~7x leverage.
The highest-beta way to own the HBM / AI-memory super-cycle: Q3 FY26 printed a record $41.5B revenue at ~85% gross margins with ~$100B of contracted backlog (RPO). Deeply cyclical and owned through the peak: sized small, ride it, don't marry it. A through-cycle DCF (~$575) sits far below the multiples-anchored target, the honest tension a peak-cycle Buy carries.
Decision Log
Audit trailA dated, public record of every move in the book, so the process can be checked, not just the result.
How the view has changed
What I got wrongPublishing theses with dates means being caught out in public. Here is every correction, misjudgment, and change of mind since the June-26 inception, in plain terms: what I thought, what changed, and where I was wrong. Three of these are errors I made and caught myself.
- Error, self-caughtAVGO: I inflated the target with a share-count error. I divided the sum-of-the-parts by 4.72B shares instead of the 4.94B diluted count. A review caught it; I cut the target ~5% to $461, and the bear case now sits below spot rather than on top of it.
- Error, self-caughtMU: I published a model that couldn't be true. The DCF's FY27 EBITDA came out below the net income my own EPS implied, which is impossible, and the forward numbers were stale against consensus. I rebuilt it from Micron's real ~80% peak margin and re-rated to $1,400. I should have caught the inconsistency before it went out.
- Error, self-caughtINTC: I doubted a quote that was right. I flagged Intel's live price as a bad feed. It wasn't: Intel had genuinely re-rated in 2026 (up ~144% year-to-date, and near a 4x peak when I doubted the quote), so the price was real and only my snapshot was stale. The data was correct, and I was wrong to distrust it.
- MisjudgedPANW: I set the target too low. It ran +117% year-to-date, straight through my $310 target. I under-called the upside I was actually playing for, so the edge is gone. Re-rated Buy → Hold and trimmed it 11% → 6%. Right on direction, wrong on magnitude.
- MisjudgedORCL: I under-weighted the credit risk. I built the OCI re-rating thesis without pricing the balance sheet hard enough; the July S&P BBB- downgrade forced a target cut from $235 to $190. The thesis holds, my risk-weighting didn't.
- Changed my mindv1 re-cut: I dropped four names nine days in. The original book held GOOGL, CRWD, ASML and NOW. On a second look none carried enough conviction upside to keep a slot, so I re-cut around higher-conviction, more off-benchmark names.
What isn't here yet: a thesis the market has proven wrong. The book is about two months old, no position has hit its thesis-break, and the sharpest move against me, MU down ~15% from its inception mark, strengthened the case rather than breaking it. Every name carries a written break level and the book a −25% drawdown trigger; when one fires, the re-cut and exactly what I misjudged land here in the same plain terms.
Methodology & Disclosure
Not advice. A written, publicly scored expression of conviction: not a portfolio or a fund, no capital, not investment advice.
Construction. Names are chosen off-benchmark on a thesis I can defend, sized by conviction, each with a written break condition, logged and not rationalized when it triggers.
The record. Only a couple of months old, far too short to separate skill from luck: read it as a live process log, not proof of a repeatable edge.
Returns & prices. Price-only from a June 26, 2026 inception (the book starts at 0.0%; the earlier nine-day v1 is logged in the Decision Log and excluded here). Each target is the canonical 12-month figure from its dashboard. Holdings are marked at the same nightly close Open Coverage shows (from coverage-data.js), so the two pages always agree; only the benchmark is quoted live, for the alpha.
Sources. Valuation context from public filings as of late June 2026; the research and this self-updating book were built with an AI workflow.
Part of the Deep Tech Equity Research Library. Browse the 39 company dashboards →