Back to index

What Is Actually Broadening

Part 1: Same Demand Signal, Three Different Qualities of Stock

The most honest way to write about April 24, 2026 is to start with what the market did not do. It did not simply reward Intel. It did not simply chase semiconductors. It ranked them. Intel rose 23.6 percent. AMD rose 13.9 percent. NVIDIA rose 4.3 percent. The same earnings print — a Q1 beat of $0.29 EPS against a $0.02 estimate and $13.58 billion in revenue against a $12.42 billion estimate, paired with a Q2 revenue guide of $13.8 to $14.8 billion — produced three quite different reactions. Anyone writing about "the semi rally" as a single object is already one level too high. The interesting work is in why the three names moved in inverse proportion to the quality of the underlying business.

The consensus that has formed over the weekend is coherent and, on its own terms, correct. Intel's beat is real. Revenue guidance that runs materially above the prior street number is not a narrative event; it is an earnings-power event. D.A. Davidson did not upgrade AMD on sentiment. They upgraded because the Intel print is, for the first time in this cycle, direct evidence that data-center CPU demand is broadening rather than concentrating further into accelerators. NVIDIA's new record close — its first since October 29, 2025 — matters precisely because it is not a lonely print. The AI-compute stack is no longer visibly a single-name story. That is the cpu-bottleneck-readthrough-broadens-semis thesis in one sentence, and I do not think it is wrong.

What I think the consensus is missing is the gap between what the print demonstrates and what the stocks priced. Every name that rallied on Friday is pricing something slightly different. Intel is pricing validation of a multi-quarter re-acceleration in the data-center group, the survival of the 18A ramp, and — implicitly — a foundry narrative that the analyst community specifically refused to pay for on the same day the stock rallied 23.6 percent. It is worth sitting with that. The same news flow that drove the price contained explicit caution on 18A yields, Terafab structure, and cost pressure. The stock did not just move faster than the fundamental evidence. It moved faster than the sell-side's willingness to revise. That is a different thing, and historically a less durable thing.

The closest analog is not another semi rally. It is the mirror image of the intel_7nm_delay episode from July 23, 2020, where Intel fell 17 percent in a week on a process-node miss against a narrative of on-track execution. That episode worked at a weekly-magnitude of -0.16. A decade of Intel earnings events tells you that the stock's multi-standard-deviation moves in either direction tend to outrun the subsequent one-quarter revisions. The October 28, 2022 foundry-losses-accelerate print took the stock down 22 percent in the same direction. These are not academic footnotes. They are the population the current 23.6 percent gap-up belongs to, and it is a population with a characteristic half-life.

The technical work from our own round-002 level-interaction pass makes the same point in mechanism language. Intel is described on Friday's tape as "post-breakaway hold rather than a clean new-leg trend," with distance of roughly 0.8 ATR up to the 85.22 intraday breakout high and only 0.29 ATR down to the 80.64 reclaim level. That is not the geometry of a settled new trend. It is the geometry of a tape that has to prove acceptance above 85.22 before it earns the next leg. The round-001 reflection adopted exactly this view: an adversary pass flagged the initial long as R/R parked against the stop, and the reflection moved the commitment to flat pending level resolution. The chart layer and the fundamental layer are saying the same thing at different vocabularies. Intel is the most fragile of the three expressions because its stock has moved faster than either the analyst revisions or the level acceptance that would ratify the move.

AMD is a different animal, and the market, to its credit, priced it that way. Friday's 13.9 percent gain is large but it sits on a base where the company already reported $1.53 EPS against $1.32 and $10.27 billion revenue against $9.67 billion — numbers that were posted before Intel's print. AMD did not need Intel to validate server-CPU demand abstractly. It needs AMD's own May 5, 2026 report to validate it specifically, with pricing commentary and EPYC booking color that either confirms or damages the D.A. Davidson upgrade. That is a ten-day confirmation window, not a sympathy squeeze. The round-002 reads show AMD on the 60-minute tape at +2 bps above the 350.88 resistance shelf after the latest bar traded as low as -111 bps below it — resistance not yet accepted, support at 334.54 held for three straight closes above. The chart layer calls this "unresolved." The fundamental layer calls this "awaiting May 5." Both are honest; neither is the kind of pattern you trade ahead of.

NVIDIA is where the analysis gets most interesting, because NVIDIA's 4.3 percent rally is the piece of Friday's move I am most sympathetic to and most worried about on different time horizons. The short-horizon reading is straightforward: a broader compute-demand readthrough is, if anything, better for NVIDIA than a narrow accelerator-only story, because it means the capex pool underwriting the NVDA order book is less concentrated and less single-customer than the bear case requires. Friday's 209-handle close against a 209.92 cited resistance reflects exactly this — the market is paying for continuation, not for optionality.

But NVIDIA is also the name where the bilateral-monopoly-squeeze and nvidia-margin-residual work still applies. Gross margin compressed 390 basis points from 75.0 percent in FY2025 to 71.1 percent in FY2026. Q1 FY2027 guidance is 73-74 percent. TSMC and HBM suppliers have under-exercised pricing power. Hyperscaler ASIC programs — TPU v7, Trainium2/3, Maia — are maturing as BATNA. A broader readthrough helps the revenue line and does nothing for the margin line. At 37x trailing and a fresh record close, the stock is paying for margin durability, not for margin acknowledgment of structural squeeze. The contradiction is not whether NVIDIA is the highest-quality expression here. It is whether "highest quality of the three" is an adequate description of a name that is simultaneously the most beat-up on the specific margin mechanism that is already eroding.

This is where the psychology matters. Marks would note — and I am writing in his voice because the temptation is exactly the one he has warned about most often — that the market is doing an interesting thing: it is revising up the durability of AI demand at precisely the point the sell-side is quiet about the cost side. The 2020-2021 semiconductor_shortage_supercycle worked for 96 weeks; the SOX doubled; TSMC, ASML, and NVIDIA all hit ATHs. The 2023-2025 HBM supercycle has worked for 104 weeks and counting. Supercycles are real. They are also the setups that produce the nvidia_crypto_mining_blowoff of November 22, 2021 (-66 percent over 52 weeks, with the underlying demand thesis intact the whole way down) and the amd_crypto_mining_blowoff of October 3, 2018 (-58 percent over 52 weeks, same pattern). The demand was right. The price priced too much of it.

The honest second-level read is not that Friday was a momentum blow-off. It is that Friday was both a legitimate broadening event and a valuation extension event, and these two facts are inseparable. The question is not whether to believe the thesis. It is how many quarters of follow-through are already in the three stocks, and at what price each one becomes a reload rather than a chase.

Here is where the chart layer is orthogonal to — not confirming of — the narrative layer. The narrative says "CPU bottleneck is real, the whole stack is getting revised up." The chart says "QQQ is rejected at 663.42, AMD is marginally cleared at 350.88 but unresolved, NVDA is one bar into a new high with 0.13 ATR of named upside, and the only name where stretch versus structure looks clearly unearned is Intel." These are not the same statement. They are complementary. The fundamental layer tells you what is true; the chart layer tells you what has been accepted as true by the price. On April 25, 2026, those two are about half a move apart.

The conclusion I would hold, if I had to hold one, is probabilistic. If AMD confirms on May 5 with demand color and pricing commentary that matches the D.A. Davidson view, the thesis survives and the ranking stays NVDA > AMD > QQQ > INTC. If AMD's print is merely in-line or shows pricing softness, the sympathy squeeze will have reached past what the earnings can carry, and Intel will be the first name to give back its gap because it has the least fundamental cushion underneath it. In either case, the current setup does not argue for pressing fresh risk at Friday's close. It argues for waiting on either AMD earnings or a support reload in the stronger names — the round-002 trade commitments name this as wait-for-X long on QQQ into 656.92 and on NVDA into 203.83, with Intel flat until 80.64 either fails or 85.22 accepts. That is the discipline the chart supports. It is also the discipline the fundamental calendar supports. The May 5 print is the separator. Between now and then, the cheapest thing a careful investor can do is nothing.

What would change this view is the simplest sentence in the whole memo: if Intel holds its 80.64 reclaim and accepts through 85.22 before AMD reports, with analyst revisions catching up to the stock, the skepticism above turns into acknowledgment. If Intel instead gives back half its gap in five sessions while AMD reports soft, the whole April 24 move becomes a historic one-day event that led a narrative but not an earnings cycle. Both paths are live. Neither is resolved. Holding a view that is conditional on a ten-day catalyst window is not weakness. It is the shape the evidence currently supports.

Part 2: The Arithmetic the Weekend Read Skipped

Part 1 is right about the facts and maybe a half-step too polite about the trade. Three things are missing that, once you see them, change the framing.

Start with Intel, because Intel is where the weekend analysis was most careful and most wrong about the shape of the risk. The stock went from roughly $66 into the print (where it had been consolidating after the Google-deal run) to $81.84 on Friday. That is a 23.6 percent gap on a $351 billion market cap. The beat itself — $0.29 versus $0.02 on EPS, $13.58B versus $12.42B on revenue — is a revenue beat of $1.16 billion and an EPS beat of $0.27. Call it something like $1 billion of incremental quarterly earnings power if you extrapolate Q1's margin. Annualized, that's maybe $4B of additional earnings on a business that lost $591M in Q4 2025 and is still running foundry at a loss. The stock added roughly $67 billion of market cap on Friday.

That is the specific number the consensus is quietly choosing not to do the math on. The market paid approximately 17x for the incremental earnings improvement on one print, from a company whose gross margin is 36.1 percent GAAP and whose foundry segment is still a multi-billion-dollar drag. Intel's forward P/S is 5.67x. If Q2 revenue hits the midpoint of the $13.8-$14.8B guide, Intel is running at a roughly $57B annualized revenue rate. At a market cap of $380B post-gap, that is a P/S of about 6.6x. Intel is now trading at a multiple that the market would not have given the company when it had a profitable foundry roadmap. And it got there in one day.

The closest price precedent is not another semi rally. It is the October 28, 2022 Intel foundry-losses-accelerate print, which took the stock -22 percent in the opposite direction on the opposite kind of news. These moves are symmetric. They have a half-life. The 23.6 percent does not decay to zero, but the history says to expect 30 to 60 percent of the initial gap to retrace within 10-15 sessions if follow-through evidence (analyst revisions, next-quarter guidance update, foundry customer disclosure) does not arrive. Our own intel-partnership-execution-gap work names this exact pattern.

AMD is where the per-share math actually argues for the trade that Part 1 hedged around. The company reported $1.53 EPS against $1.32 on $10.27B revenue. Using the TradingView-sourced market cap of roughly $567B after Friday and the street's rough forward estimates, AMD is around 42x forward earnings. That is expensive. It is also cheaper than the 40x+ multiple NVIDIA traded at through the 2023-2024 run that worked. AMD's Q4 revenue grew 11.08 percent QoQ. The data-center segment inside that print has been the fastest-growing piece. If the May 5 guide raises server-CPU revenue outlook — which is what Intel's print implies is coming — AMD's incremental 2026 EPS could revise up into the $7 range, implying a forward multiple closer to 36x on the post-revision number at Friday's close. That is not cheap, but it is inside the range where you can size a position without getting laughed at.

The arithmetic the weekend coverage skipped: D.A. Davidson's $375 target from Friday implies roughly 8 percent upside from the $348.90 close, which sounds underwhelming until you put it next to Intel's 23.6 percent one-day move. AMD is being priced as if the readthrough is partial. The actual confirmation on May 5 — if it comes — would justify a multiple that Friday's close is still 10-15 percent away from. The asymmetry is:

The expected value on holding AMD into the print is modestly positive from Friday's close if you size small enough that the disappoint case is a 0.5 percent portfolio drawdown rather than a 2 percent one. This is not a screaming buy. It is a defensible hold-into-confirmation, which is different from either the "wait-for-breakout" framing or the "fade the overshoot" framing the TA work considered.

NVIDIA is the position that the valuation extension argument applies to hardest and where Part 1 was mostly right. The per-share numbers: $1.62 EPS against $1.54, $68.13B revenue against $66.13B. The next-quarter consensus is $78.62B revenue and $1.74 EPS. At Friday's 4.3 percent gain to a new record close and a $5.06T market cap, NVIDIA is trading at roughly 37x trailing and about 33-34x on the next-quarter annualized run rate. That is not the valuation you buy; it is the valuation you own because you already did.

The margin math is the part Part 1 named and the one I want to sharpen. FY2026 gross margin of 71.1 percent, guidance at 73-74 percent for Q1 FY2027, structural drift toward 68-70 percent over 2-3 years per the bilateral-squeeze work. If you re-run NVIDIA's earnings at 68 percent gross margin instead of 71 percent while holding the current revenue run rate, net income falls by roughly 7-9 percent. The stock is not priced for that. It is priced for 73-75 percent margin to hold. The readthrough from Intel does nothing to solve this. It actually worsens it at the margin, because a broader compute demand story gives TSMC and the HBM vendors more pricing leverage, not less.

The trade that is arithmetically defensible: long AMD (size small, 0.3-0.5 percent portfolio, exit either direction on May 5). Hold NVIDIA if you already own it, but do not add at a record close where the margin math is deteriorating. Do not touch Intel long here; if you own it from below, the historically honest move is to trim 20-30 percent into the gap and keep the rest as optionality on the 18A narrative. The round-002 AMD short that the TA pipeline filed — 0.5 ATR short against 350.88 with a target at 334.54 — is defensible as a tactical level-interaction trade, but it is contradictory to the fundamental readthrough framing, which is exactly what the adversary rival in the same round flagged. I would not short AMD into a confirmation catalyst unless my horizon is three to five days and I am running a pure mean-reversion book. At a ten-bar horizon, going short against May 5 is fighting the catalyst.

Downside scenario. Assume Intel gives back 10 percent by May 2, AMD reports in-line on May 5 with no pricing upside, and NVIDIA trades sideways on the news. What happens? Intel at $73.65 is still up from the pre-print $66. AMD at $330 is roughly the 334.54 support test that the TA work named. NVIDIA at $200 is the 203.83 reload that the round-002 wait-for-X commitment is literally sized for. The entire April 24 move becomes a one-quarter spike that broke the previous range but did not reprice the sector. The structural thesis — AI demand is real, the stack is broadening — survives. The trade expressions get reset to cleaner entries. This is not a ruinous outcome. It is the mean-reversion the population of Intel multi-standard-deviation gaps historically produces.

TL;DR.

The move was real. Three of the four expressions of it are already most of the way priced.

Part 3: At What Margin, At What Multiple — The Five-Year View

Neither Part 1 nor Part 2 is the view a long-duration investor should hold. Both are about the ten-to-thirty day window around one earnings print. The right question for a five-year horizon is different: at what terminal margin does each of these businesses clear, what multiple does that terminal margin deserve, and which of them can I own without checking the chart?

Start with what changes when you extend the horizon, because most of what Parts 1 and 2 argued about dissolves. The May 5 AMD print is not a tripwire; it is one data point out of twenty earnings reports between now and 2031. The 350.88 level that matters for the trader is noise for the owner. The 23.6 percent Intel gap is not a valuation event; it is an entry-timing event, and entry timing has a six-to-eighteen-month tolerance on a five-year hold. What does persist across the horizon is the economic structure of each business — the pricing power, the cost trajectory, the capital allocation record, and the moat durability against substitutes that haven't been built yet.

NVIDIA at a five-year horizon is the most interesting and the hardest case. The short answer is that the current 37x multiple on $5.06T of market cap is a bet on two things simultaneously: that revenue continues to compound at 30-percent-plus rates, and that gross margin holds in the low-70s. The first is defensible. The AI capex pool is real and broadening, as Friday's print corroborates. The second is where the bilateral-monopoly-squeeze work has force. Gross margin went from 75.0 percent to 71.1 percent in a single fiscal year while the stock went sideways-to-up. The market has already absorbed one margin step-down without re-rating. It will probably absorb a second — from 71 percent to mid-68s — without a crash. What it will not quietly absorb is the structural acknowledgment that NVIDIA's economic rents are being split three ways (TSMC packaging, HBM vendors, hyperscaler ASICs) rather than captured end-to-end. That acknowledgment comes when hyperscaler capex mix shifts by more than five points toward custom silicon, or when TSMC raises CoWoS pricing by more than 20 percent on a single contract renewal. Both are live. Neither has happened decisively.

On a five-year hold, NVIDIA probably compounds, but at a lower rate than the past five years and with a wider distribution of outcomes. A reasonable base case is 12-15 percent annualized total return to 2031, assuming 25 percent revenue growth, margin drift to 68 percent, and multiple compression to ~25x — which is still premium to the semi sector. The bull case is 20-percent-plus if CUDA switching costs prove stickier than the ROCm and PyTorch-backend maturation curves imply. The bear case is a nvidia_crypto_mining_blowoff-style drawdown of 40-60 percent on a customer-concentration shock or an ASIC-share revelation, with fundamentals staying fine the whole way down and the stock recovering over 2-3 years. The 2021-2022 episode is the relevant precedent: the demand thesis was right, the valuation multiple collapsed, and the patient owner who held through made everything back and then some. You can own NVIDIA at this multiple, but only if you genuinely have the temperament for a 50 percent drawdown without selling. Most people who say they do, don't.

AMD at a five-year horizon is the cleanest quality-at-a-price story of the three. The structural case is that AMD is now the best-run fabless semi company by execution record. Lisa Su's capital allocation — the Xilinx integration, the server-share march, the measured GPU strategy that avoids bidding against NVIDIA on training and concentrates on inference where the economics are better — has been rational in a way Intel's has not and in a way NVIDIA's cannot be because NVIDIA has never had to optimize against a binding budget constraint. AMD's terminal margin is structurally lower than NVIDIA's — call it 55-58 percent gross against NVIDIA's eventual 65-68 percent — but its revenue base can compound faster from a smaller number, and its customer concentration is healthier. EPYC is now genuinely competitive with Xeon in both performance and pricing, and server CPU share is a business where share gains compound over multiple refresh cycles. The MI-series GPU strategy is a credible number-two in a market big enough to support two winners.

The question for the long-term AMD holder is not the May 5 print. It is whether AMD can sustain 15-20 percent annualized revenue growth through 2031 without the margin structure that compounding at NVIDIA scale requires. At a current 42x forward and assuming that multiple compresses to 25-28x over five years as the growth rate normalizes, you need about 18 percent annualized EPS growth to earn an 11-13 percent total return. That is not a layup, but it is inside the band of what AMD's execution history supports. The downside scenario is not an amd_crypto_mining_blowoff repeat; AMD's current revenue is not dependent on a speculative end-market the way it was in 2018. The downside is a slower grind: the stock does nothing for three years while earnings grow into the multiple, which is an acceptable outcome for a long-duration holder and a painful one for anyone who paid Friday's close expecting momentum.

Intel at a five-year horizon is the pure optionality trade, and it needs to be sized as one. The business that reported on April 24 is still fundamentally unprofitable at the segment level once foundry is included. The 18A ramp is real but has not produced the yields that the Street has been willing to pay for. The Google agreement is the first major foundry customer win and it is two years from being revenue-material. The Terafab structure is an accounting construct that may or may not survive the shareholder-accountability pressure it will face in 2027-2028. None of this is resolved by a single-quarter revenue beat.

What the long-term Intel investor is buying is a call option on US domestic leading-edge foundry capacity. That option has real value — it is arguably the highest-value single piece of semiconductor optionality in the public market — but the strike price is not $81. The strike price is the probability-weighted distribution of 18A foundry-customer revenue, foundry segment margin, and the CHIPS Act policy environment in 2027-2029. Honest probability weights on that distribution put the base-case five-year IRR between 5 and 9 percent, with a fat left tail if 18A disappoints and a fat right tail if Intel Foundry captures a second major hyperscaler customer. You own Intel as a 1-3 percent portfolio position against the specific scenario where TSMC's Taiwan exposure becomes a first-order geopolitical concern, and you size it as insurance — not as a core semiconductor holding.

The uncomfortable but structurally correct observation, one I think Parts 1 and 2 both danced around, is that for a patient owner the best semiconductor exposures probably are not in this basket at all. The bilateral-monopoly-squeeze and cowos-packaging-monopoly-repricing and hbm-supply-constrained-supercycle-2028 work this desk has already completed points at TSMC, Micron, and SK Hynix — the suppliers who control the bottlenecks NVIDIA must pay through. Those names trade at 20-30x earnings with pricing power that is still under-exercised and customer concentration that works in their favor. They will probably capture more of the next decade's AI compute rents than any of the three names in this memo. The ai-infrastructure-valuation-problem-2026 memo made this argument on a sector basis; Friday's move is additional evidence rather than a refutation of it.

The long-term investor's ranking is different from Part 1 and Part 2. On a five-year hold: Micron and TSMC over AMD over NVIDIA over Intel, with the first two being considered-core and the last being sized as optionality. AMD is the quality-at-a-full-price name of the three discussed here. NVIDIA is a compounder whose compounding rate is structurally declining and whose multiple has not yet adjusted. Intel is a call option. None of this changes if AMD prints well on May 5; none of it changes if Intel gives back half its gap in ten sessions. The May 5 print matters for the trader, not for the owner.

Three practical implications for allocation. First, if you hold NVIDIA, the five-year question is not whether to trim — it is whether to reallocate some of the NVIDIA weight to the suppliers who will eventually capture the margin NVIDIA loses. Doing so before the transfer is visible is the definition of long-duration investing. Second, if you do not yet hold AMD, the right entry is not "into the May 5 print" — it is a pullback to the 334.54 TA-named level or lower, which would correspond to a forward multiple under 38x. That is patience, not market timing. Third, Intel is the only name here where the long-term case improves if the stock gives back its Friday gap. Lower entry, same optionality, better asymmetry. The short-horizon bad news is the long-horizon good news.

The durable mistake in weekend coverage of any semi rally is confusing "the demand thesis is right" with "these three stocks are good long-term investments at today's price." The first is a statement about the world. The second is a statement about valuation, multiple compression, competitive positioning, margin trajectory, and capital allocation over a time frame that no amount of earnings-beat arithmetic resolves. On April 25, 2026, the thesis is broadening. The investments, for the patient holder, are mostly upstream.

Part 4: Reflection — Where This Memo Is Weak

Applying the six-test adversary discipline from references/adversary.md to the memo itself. A reflection round is not a grade; it is an attempt to file a rival memo strong enough that the current one becomes untenable if the rival survives pressure. The tests are repurposed from their symbol-level use, with one addition (a seventh test) the user's prompt forced.

Test 1: Direction-vs-view contradiction

Parts 2 and 3 disagree operationally without naming the handoff. Part 2 says trim Intel into the gap. Part 3 says Intel's long-term case improves if it gives back the gap. The memo holds both views simultaneously but never names the re-entry price or the trigger evidence. A committed memo would say: add Intel below $72 on a 60m close that reclaims; otherwise flat. The current memo hedges across horizons instead of picking at each horizon.

Test 2: R/R parked against the stop

Part 2's AMD long is 0.3-0.5 percent of portfolio for a trade whose plausible range is ±15 percent. That is 5-8 basis points of expected contribution in either direction. The position is too small to matter to a portfolio and just large enough to let the author claim the upside if it hits. Either this deserves 2-4 percent sizing with defined exits, or it deserves no trade. "Small long" is narrative cover, not a decision.

Test 3: Symmetry

The memo's citations support the opposite stance readily. Intel's post-gap ~6.6x P/S is still below TSMC's foundry comparable multiple; AMD's 42x forward compresses cleanly to 25-28x on two years of server-CPU share gains; NVIDIA's 390bp margin compression came alongside revenue growth, so gross profit dollars rose; QQQ's 663.42 rejection is one bar against three bars of 656.92 acceptance. The same evidence supports "Friday was under-priced for the cycle, not over-priced for the names." The memo picked caution without acknowledging that it picked.

Test 4: Path-specific invalidation (Part 3 has none)

Part 2 names level-break invalidations for each trader-scale recommendation. Part 3 names none. What specific path would kill the "suppliers over NVDA" thesis? The memo does not say. Candidates it should have named: TSMC CoWoS pricing flat for two consecutive contract cycles; hyperscaler ASIC mix rising past 30 percent of AI inference capex; NVDA gross margin holding above 72 percent for four consecutive quarters. Without a named path, the five-year view is an essay, not a position a portfolio can hold.

Test 5: Cross-sectional beta-as-alpha

Part 3 pivots to "suppliers capture more rent" without establishing that the supplier singles outperform the semi-cap ETFs. If SMH compounds at 14 percent annually over five years and a hand-picked MU + TSM book compounds at 15 percent, the 100bp of supposed alpha does not pay for the concentration risk. The memo never weighs its singles recommendation against the benchmark implementation. That is the exact mistake Part 3 accuses Parts 1 and 2 of.

Test 6: Sizing-vs-conviction mismatch

Part 3 calls MU/TSM "considered-core" without a size band. "Core" spans 3 percent to 15 percent across reader mandates. The memo's sizing discipline drops sharply between Part 2 (explicit 0.3-0.5 percent) and Part 3 (narrative). A committed version would name a portfolio: 15 percent combined MU+TSM, 5 percent NVDA (half of a prior holder's weight), 2 percent AMD on a reload below 334.54, 1 percent INTC on a reload below $72, zero in SMH because the ETF underweights the suppliers. That is a position. The current Part 3 is a ranking, which is a different thing.

Test 7: The missing-humans test

The memo does not mention a single person running any of these businesses. At leading-edge process engineering — 18A, N2, 14A — yield is a team outcome, not a tool outcome. The variables the memo omits are concrete and material:

This is not a decoration oversight. It is the reason the memo's five-year view feels spreadsheet-shaped. Businesses that compound at differentiated rates over five-plus years do so because the people running them make differentiated decisions, and a memo that ranks them by terminal margin and multiple without ranking them by management quality has skipped the dominant variable.

Closest Alternative Memo

A rival memo using the same evidence and honoring the missing-humans test would hold:

Friday was under-priced for the cycle, not over-priced for the names. We are 3 years into a 4-5 year semi supercycle with HBM and CoWoS constraints intact and a broadening-compute print to ratify the back half. The correct five-year allocation is overweight semiconductor capital equipment and memory, equal-weight NVIDIA, underweight AMD until confirmation, and sized-up Intel below $72 — not as geopolitical optionality but as a specific bet that Lip-Bu Tan's capital discipline plus ex-TSMC process talent shortens the 18A maturation curve by 2-4 quarters versus consensus. The supplier thesis is correct; the naming mistake is MU and TSM rather than ASML, Applied Materials, and KLA, who benefit from both supercycle extension and capex broadening to Intel foundry.

The current memo does not address this rival. It is the most compact honest statement of where the analysis is weak.

Which Tests Caught What

The memo survives as a ten-to-thirty-day trader-scale analysis. It does not survive as a five-year allocation memo. The next revision should adopt the rival above, refute it with new evidence (most likely on the capital-equipment names versus suppliers choice), or synthesize — not dismiss it.

Related Research

Appendix: Chart-State Tripwires

From the round-002 recursive-TA artifacts, the following level-interaction conditions are load-bearing for the memo's stance. Blind review is not yet run on this round, so confidence in the chart reads is tempered. Each tripwire is a condition that, if breached, should reopen the relevant commitment:

The round-002 adversary filed one live rival in this run: the AMD short faces a direction-vs-view rival arguing the tape is still in breakout-acceptance rather than settled rejection. The memo's AMD framing explicitly sides with the adversary's view, not the trader-pass view, which is why the memo's AMD position is long-into-May-5 rather than short-against-350.88.