Global Semiconductor Cycle Sector IC Memo

Executive Summary & Action Plan

Verdict

NEUTRAL — do not call the bottom yet; add sector risk only after price and breadth confirm. The earliest tradable bounce window is 2026-07-30 to 2026-08-03. Our base case for a durable turn is late August to mid-September, but the level matters more than the date: SOXX must close above $534.30 with broader participation, then reclaim the $561.77–568.86 moving-average band.

The selloff has damaged trend, not yet earnings. Seven of seven Tier-1 names still have non-negative 30-day EPS revisions, while just one is above its 20-day average and none is above its 50-day average.

Loading chart

Abstract

Global semiconductors are in a valuation, financing and China-risk correction rather than a verified demand break. NVIDIA, Broadcom, TSMC, ASML and Micron all retained extraordinary revenue or AI guidance, while Samsung and SK hynix remain at record memory profitability and TrendForce still expects tight server DRAM. That fundamental resilience is necessary but not sufficient: SOXX is 21.2% below its June peak, below both its 20- and 50-day averages, and breadth is only 1/7 above the 20-day line. The first tactical turn can be tested after the July 29–30 event cluster, but a durable turn requires a weekly reclaim of $568.86, a successful retest of $561.77, and at least five of seven leaders above their 50-day averages. Memory should be treated separately: scarcity supports near-term earnings, yet record margins make Micron, SK hynix and Samsung the most exposed to normalization. We therefore wait for confirmation, carry no unconditional Top Pick, and favor quality enablers over peak-margin beta.

Forecast Path

The path uses SOXX's $516.23 adjusted close and an outside-view cohort of four completed drawdowns; the sample is too small to calibrate probabilities. The 80% bands are empirical 2016–2026 rolling-return ranges, not guaranteed confidence intervals.

Loading chart

The probability-weighted path is roughly flat at 1M and +18.8% at 12M. That is a conditional recovery prior, not permission to catch a falling tape.

Action Plan

Open no full-sized position at $516.23. Take a 25% tactical position only after a close above $534.30, Tier-1 breadth reaches at least 3/7 above the 20-day line, and up-day relative volume is at least 1.2×; add after a weekly close above $568.86, a $561.77 retest holds, and breadth reaches 5/7 above the 50-day line.

The ladder separates a trading stop from thesis falsification. At the $534.30 trigger, reward to the $610 base path versus a $499 close stop is about 2.1:1; a weekly break of $398.36 is the structural failure line.

Loading chart

Earliest bounce: July 30–August 3, after FOMC and the hyperscaler/memory prints settle.
Durable-turn base case: late August–mid-September; accelerate only if all price/breadth rules trigger earlier.
Memory confirmation: September–October contract negotiations; rising earnings without price confirmation is not a buy signal.
Prob-weighted 12M return: +18.8% | P(thesis wrong): 45% | Confidence: Medium-Low.

State of the Sector

The 2026-07-27/28 break was attributed to NVIDIA customer-financing concerns, reported Chinese DUV progress, CXMT competition and crowded valuations. It is not yet corroborated by issuer guidance: NVIDIA guided Q2 revenue to $91B ±2%, TSMC guided Q3 to $44.6–45.8B, and ASML raised its full-year range to €43–45B.

Layer	Frozen evidence	Investment read-through
Compute — NVDA	Q1 revenue $81.6B; Data Center $75.2B; Q2 guide $91B ±2%	Platform demand intact; financing/circularity raises the evidence bar
Custom/network — AVGO	Q2 AI semiconductor revenue $10.8B, +143% YoY; Q3 AI guide $16B	Durable alternative path, but expectations and debt are high
Foundry — TSM	Q2 revenue $40.2B; Q3 guide $44.6–45.8B	Cleanest cross-customer read; consumer/mature-node demand is weaker
Equipment — ASML	Q2 sales €9.33B; Q3 guide €11–12B	Bottleneck quality; orders and export policy remain lumpy
Memory — MU/Samsung/SKH	MU Q3 GM 84.6%; Samsung Q2 OP guide ₩89.4T; SKH Q1 OP margin 72%	Scarcity is real, but peak-margin extrapolation is the core risk

SOXX is only a proxy for this global map. Samsung and SK hynix are absent from the ETF, so Korean memory can diverge materially from the scoreable target.

Forecast Drivers & Cycle Map

Price and breadth lead the call: $534.30 is the bounce trigger; $561.77–568.86 plus 5/7 breadth is the trend gate.
Revisions have not broken: compute revision breadth is 0.982 and memory breadth 0.941; magnitude is strongest in MU and ASML, which also carry the highest expectation risk.
Hyperscaler capex decides compute: two of Meta, Microsoft and Amazon maintaining or raising capex keeps Base/Bull alive; two cuts of at least 10% move the view to Bear.
Contract pricing decides memory: TrendForce still sees a 3Q server-DRAM surge and 2027 demand above supply, while softer spot pricing is the earliest caution flag.
China evidence must move from headline to yield: verified commercial DUV throughput and CXMT customer qualification matter; an undisclosed tool specification does not yet reset the cycle.

Winners, Losers & Top Pick Decision

There is no unconditional Top Pick at the cutoff. TSMC is the cleanest quality candidate and NVIDIA is closest to an inherited house mark, but neither clears both valuation and trend gates; ASML lacks a fresh house fair-value workup and is excluded from the chart.

The comparison carries forward cycle-normalized marks from the fresh 2026-07-21 compute and memory work, with current prices re-frozen after the >5% shock. These are valuation gaps, not 12M price forecasts.

Loading chart

Relative winners if the turn confirms: TSMC, ASML, then Broadcom — diversified demand, bottlenecks and less direct commodity-price exposure.
Platform winner, higher proof burden: NVIDIA — superior economics, but new customer-financing questions make August 26 decisive.
Fundamental winners, valuation losers: Micron, SK hynix and Samsung — current earnings revisions are strongest, while normalized value is most exposed to falling ASPs and record margins.

What to Watch Now

Date	Gate	Bull settlement	Bear settlement
Jul 29	SK hynix; FOMC; Meta/Microsoft	capex and memory margin outlook maintained	financing/ROI constraint or weaker conventional DRAM
Jul 30	Samsung; Amazon	no capex cut; memory pricing breadth intact	two demand nodes weaken together
Jul 30–Aug 3	Tape response	close >$534.30, breadth ≥3/7, RVOL ≥1.2×	close <$499 or breadth stays ≤1/7
Aug 26	NVIDIA Q2 FY27	revenue at/above guide and next guide sustains growth	revenue below $89.2B, GM <74.4%, or material financing support

What to Watch Next (Technology/Structure)

HBM4/Rubin and advanced packaging: capacity conversion must create volume, not merely higher ASPs.
N2, CoWoS and High-NA: TSMC/ASML remain the physical bottlenecks, but capex and overseas-ramp dilution cap valuation.
Custom ASIC plus Ethernet: Broadcom can gain profit pool even if merchant GPU share falls.
China DUV/CXMT: require specifications, yields, repeat orders and two-fab adoption before treating substitution as proven.

Fair Value Notes

No company was fully re-underwritten from scratch in this timing memo. The peer chart preserves the latest specialized house marks: NVDA/AVGO/TSM from the 2026-07-21 compute map, and Samsung/SK hynix/Micron from the 2026-07-21 memory-China work; the latter supersedes the general map for memory. The current shock was cause-checked and did not reveal a post-cutoff earnings or DRAM-pricing break. Any July 29–30 guide change forces immediate re-underwriting.

Catalysts, Leading Indicators & Risks

The operating data and the tape are presently in conflict. Use the following indicators as explicit view-change rules, not as descriptive dashboards.

Indicator	Current	Bull threshold	Bear threshold
SOXX price/breadth	$516.23; 1/7 above 20D	>$568.86 weekly; retest holds; 5/7 above 50D	<$499 tactical; <$398.36 weekly structural
30D EPS revisions	7/7 non-negative; median +7.9%	≥5/7 positive and median >0%	≤2/7 positive and median <-5% for two reads
Server DRAM contract	rising in 3Q	4Q remains sequentially positive	4Q flat/down with customer inventory >10 weeks
Hyperscaler capex	strong but under ROI scrutiny	≥2 of 3 maintain/raise	≥2 of 3 cut ≥10%
China substitution	reported; performance undisclosed	no verified advanced-node deployment	commercial DUV adoption at two fabs plus falling DRAM ASPs

Principal tail risks are correlated: AI customer financing, China substitution, hyperscaler ROI, memory-margin normalization, rates and crowded ETF positioning. Do not add their probabilities together.

Falsification & Review Rules

Tactical setup fails: after entry, SOXX closes below $499; exit the tactical position but allow re-entry on a new trigger.
Durable-turn timing fails: no weekly close above $568.86 with 5/7 breadth by 2026-09-15; move the base-case turn window to Q4.
Demand thesis fails: at least two major hyperscalers cut AI/data-center capex by ≥10%, or revision breadth is ≤2/7 with a median below -5% for two monthly reviews.
Memory thesis fails: 4Q server DRAM contract pricing is flat/down and customer inventory exceeds 10 weeks; exit pure-play memory exposure.
Structural recovery thesis fails: SOXX closes a week below its $398.36 200-day average and both SOXX and SPY remain below their 200-day averages for five sessions.

References

Method: prices are frozen at the 2026-07-27 US close and 2026-07-28 13:35 KST for Korean listings. Forecast probabilities are judgmental and uncalibrated; empirical intervals use overlapping observations and the conditional drawdown cohort has only four completed episodes.