Semiconductors, 2026-09-03: eight names through the framework #

Framework: semiconductor-framework.md (written today; resources/semiconductor-framework.md links to it). Data: eight stock-data-analyst passes run in parallel on 2026-09-03 (Finnhub, yfinance, SEC 6-K/8-K/10-Q, company releases; consensus from yfinance because Finnhub's estimate endpoints are paywalled). Prices are intraday 2026-09-03. Positions held: INTC 14 sh @ 90.10 (rule-006 cap breached, hold re-authorised 2026-09-03). Open decision: GFS-DEEP45 fired 2026-08-28 at 44.88. VWCE already holds NVDA, TSM, ASML, AMD and MU.

Answer first #

None of the eight is cheap. Three are good businesses at a fair price near a cycle peak (TSM, ASML, NVDA). One is a mediocre business at a fair price (GFS). Four fail the good test, and three of those (AMD, MU, ARM) fail it on the owner's own deal-breaker: cluster selling by the CEO and other officers.

EV / price is the framework's C7: probability-weighted value two fiscal years out at a multiple that fits the model, divided by today's price. 1.25 or more is cheap, 0.90 to 1.25 fair, below 0.90 expensive. "Cheap begins" is EV ÷ 1.25. Every scenario input is listed in section 4 so the owner can argue with it.

1. Where the sector is in its cycle #

The May 2026 supply-chain screen (30 names, zero passes) read "late-cycle euphoria". Three months on the gauges have moved further, not back:

Memory pricing, the sector's most sensitive gauge, is still rising but decelerating: DRAM contract prices +95% in Q1 CY26, +58-63% in Q2, forecast +13-18% in Q3 (TrendForce). Micron's own management says "meaningful moderation in the rate of price increases" is ahead. The framework's rule stands: a name at Peak cannot be cheap on current-year numbers.

2. Good? Scorecard (0 / 1 / 2 per test, see the framework for thresholds) #

Reading the zeros:

• AMD, MU and ARM fail on the owner's rule. All three show the CEO selling alongside two or more other officers within 90 days and no open-market buy by anyone. AMD's Lisa Su sold 335,000 shares (~$130M) in March, May and June under a 10b5-1 plan; the CFO, CTO and two SVPs sold in the same window. Micron's Sanjay Mehrotra sells a fixed 40,000 shares every month (~$136M since May) while the Chief People Officer sold $57M. Arm's Rene Haas and CFO Jason Child both sold, Child four times, the last on 27 August. Planned sales are still sales; the unanimity is the signal, as the May MU note already said.
• ARM does not generate cash for shareholders. Company-reported FCF minus SBC was −$163M (FY24), −$721M (FY25), −$170M (FY26). SBC is 21-33% of revenue and there are no buybacks. The 97% gross margin is real; the equity holder does not receive it.
• INTC is a turnaround, not a business, by this checklist: no self-funded capex, 20% dilution in a year, zero return on capital. Yesterday's thesis said the same; nothing in today's data changes it.
• GFS fails on returns. Net cash, flat share count, no insider cluster, but ROE of 6-7% on a specialty foundry that renegotiated its long-term agreements downward in 2025. A P/B of 2.06x is above the justified P/B of ~0.6x at today's ROE and ~1.5x at the 2028 target (40% GM). The stock is not cheap on book; the book is not earning enough.

3. Cheap? Scorecard #

Two readings matter more than the rest of the table:

• Forward P/E is the sector's favourite lie. MU at 6x FY27 and NVDA at ~15x FY28 look like the cheapest stocks in the market. Both are at Peak on four of five gauges. Micron's FY27 consensus of $155 spans $107 to $221 across 39 analysts; the low end alone would put the P/E at 9x on earnings that then fall. The reverse DCF says the same thing from the cash side: every name needs 20%+ FCF growth for a decade to justify its price, and none has a free-cash-flow yield above 2.3%.
• TSM is the one name where Graham nearly clears. Break-even growth of 11% on normalised EPS against a demonstrated 22% (5-yr) or 16.5% (10-yr). Value at 15% growth is $520, a 20% margin at today's price, five points short of the 25% the framework asks for at a cycle peak. It is the closest to cheap of the eight and still not there.

4. Scenario inputs (C7), stated so they can be argued with #

EPS two fiscal years out (FY27 for Dec-year names, FY28 for NVDA and ARM) × multiple; probabilities bear / base / bull. SBC-adjusted where non-GAAP was the only consensus.

Multiples: memory at peak earnings has historically cleared at 5-8x (MU 2018: 5.4x), hence 7-9x on FY27; foundry 17-24x; equipment 25-32x; fabless 16-30x on the cycle read. The bear probabilities are 0.30 (0.35 for MU) because every gauge in section 1 reads late or peak.

5. What each name is, in a paragraph #

TSM. The best business here and the benchmark for the rest: 67.7% gross margin at 60% operating margin, $80B net cash, flat share count, dividend raised 17% in August, wafer prices going up 5-10% in 2027 with a further 10-15% surcharge on HPC. Insiders filed 151 purchases against two sales. Two customers are 36% of revenue, Nvidia now the largest at 19%, and both have nowhere else to go until Intel 14A or Samsung reach parity. Capex is 45% of revenue and rising, which is why FCF yield is 1.7% on a business earning 30% on equity. Fair at 415; cheap at 316, roughly where it traded in January.

ASML. The only monopoly in the set, with a backlog that funds two more years: FY26 guide raised three times to €43-45B, 2027 orders "substantially secured", capacity up 30% for 2027. China fell from 42% to 14% of system sales in four quarters, so the export-control hole is mostly behind it, except that the US is now pushing the Netherlands toward a total ban including servicing of installed tools. Net cash €5.6B, share count down 4.6%, SBC immaterial. It stopped publishing quarterly bookings in 2026, the one metric an equipment framework leans on. Insider transactions were not checked (AFM register, not SEC), which is the one gap to close before any buy. Fair at €1,423; cheap at €1,076.

NVDA. By the checklist a good business: FCF after SBC positive even in the FY23 trough, $68B net cash, share count down 4.5% in five years, a $80B buyback and a 25x dividend raise in May. By the cycle gauges the most extended name here: inventory up 128% in five quarters against revenue up 85%, purchase commitments doubled to $279B, gross margin guided down to 71-72% on HBM cost, three customers at 54% of revenue and every one of them designing its own silicon. Zero open-market insider buys; a director sold $445M. Management guided FY28 revenue +70%. On forward earnings it is the cheapest of the eight; on the cycle it cannot be called cheap. Fair at 230; cheap at 182.

GFS. Fell 52% from a May high that coincided with Mubadala's 22M-share block; then a soft Q3 guide, flat adjusted FCF in Q2 and a sector sell-off. Net cash, buyback, first dividend, no insider cluster, utilisation in the high 80s. But mobile revenue has fallen four quarters running, automotive flipped to −10% in Q2 after +24% and +30%, the LTA backlog was renegotiated lower, and ROE is 6-7%. At 44 the price is fair for what it earns; it is not the "2x trough earnings, genuine Graham margin" the June alert imagined because the E was the problem, not the P. The fired DEEP45 rung is a PASS. Cheap begins near 32.

INTC. Held. Today's data confirms yesterday's thesis line for line: 18A is in the numbers (revenue +18%, gross margin 40%, DCAI +59%), foundry still loses $2.1B a quarter, no external 14A customer, the company sold $20B of stock at 95 in August. CEO and CFO buying, no cluster. Six of sixteen on the good test, expensive at 0.86. The plan stands: trim at 115, kill list in INTC_plan.json, cheap at 65.

MU. The textbook case for the framework. Trailing P/E 21x, FY27 P/E 6x, and every cycle gauge but inventory at an all-time extreme: 84.6% gross margin against a ten-year range of −9% to 59%, revenue tripled on price (DRAM bits +low single digits, ASP +60% in one quarter), P/B 10.6x where no prior peak exceeded 2.5x, capex guided from $27B to over $45B, CXMT at 10% of DRAM and shipping HBM3E. The CEO sells 40,000 shares a month. Through-cycle, nine years of FCF after SBC total $12B against a $1.07T market cap. The May note said the entry is the trough, below 2x tangible book; that is still the entry.

AMD. Revenue +47% and data centre +107%, but gross margin stuck at 54-57% because AI accelerators earn below the company average, FCF after SBC yields 0.9%, share count up 35% in five years plus a 160M-share warrant handed to OpenAI at a cent. The insider cluster is the widest of the eight. Expensive at 0.70 before the warrant, 0.64 after.

ARM. The best franchise in the set (97% gross margin, 350bn chips shipped) attached to the worst equity: FCF after SBC negative three of four years, SBC 22% of revenue, no buybacks, GAAP P/E 251x, FY28 non-GAAP P/E 80x, break-even growth of 35% forever. It cut FY27 royalty growth to "high teens" in July, stopped disclosing its two key metrics, is moving into selling its own chips against its licensees, and the CEO and CFO both sold. SoftBank holds 87% with $18.5B of margin loans against the shares.

6. Actions taken today #

• REVIEW alerts at the "cheap begins" price for the three good businesses. Each note says it is a deliberate overweight of a VWCE holding, requires the rule-007 same-day check and settled cash, and names this document: TSM-REVIEW316 (below 316), ASML.AS-REVIEW1076 (below €1,076), NVDA-REVIEW182 (below 182).
• No alerts for AMD, MU, ARM: they fail the good test on insiders, so no price is a buy until that changes. Lead generation is done; the tooling stays.
• GFS: the framework's verdict on the fired GFS-DEEP45 is PASS (not cheap, returns too low). The triggered row and GFS-DEEP40 are left for the owner; the framework's cheap price is 32.
• INTC: no change. Ladder, kill list and INTC-REVIEW65 stand.

7. What would change a verdict #

• Cycle turn. Two quarters of falling DRAM contract prices, TSM or ASML guiding below the prior year, hyperscaler capex guides flat or down. That is when the good names (TSM, ASML, NVDA) approach their cheap prices and the framework re-runs with a Trough read.
• Insider reversal. An open-market buy by the CEO or CFO of AMD, MU or ARM lifts G6 from 0 and reopens the analysis. Until then no price is cheap enough.
• GFS returns. ROE above 12% for two quarters (the 2028 target) would lift G3 and make 32 a real entry.
• INTC. The INTC_plan.json kill conditions and rungs; a named 14A customer is the bull trigger.

Data gaps (rule-002) #

ASML insider transactions (AFM register) unchecked. GFS China revenue % not found. NVDA FY28 consensus unreliable (management's +70% revenue guide used as the bull). MU FY27 capex ">$45B" is a secondary source. Finnhub's insider feed missed the INTC CEO's August Form 4 entirely and lags a quarter on financials; yfinance P/B and yield fields are wrong for every ADR. All balance-sheet ratios above were recomputed from filings.