The Setup: A Parabolic Trade Unwinds
The Philadelphia Semiconductor Index fell roughly 10% over the past week, its worst weekly showing in more than a year, and now sits about 20% below the all-time high it printed in late June. Measured over the trailing four weeks, the SOXX ETF that tracks the index is down more than 13% — the sharpest four-week move since April 2025. Yahoo Finance’s semiconductor basket has shed on the order of $1.5 trillion in market value since June 25.
Context is everything here. From the February low into the late-June peak, the group more than doubled. The index remains sharply positive on the year even after the drawdown. This is not a crash from a standing start; it is a fast, concentrated de-risking of an overextended position. The memory complex — Micron, SK Hynix, Samsung, and the Roundhill Memory ETF — has already fallen more than 20% from recent closing highs, which is the technical definition of a bear market for that sub-sector, even as the underlying supply picture stays tight.
What Actually Broke
Three forces converged, and it helps to keep them separate because they carry different half-lives.
AI capex durability. The first crack was in networking silicon, not memory. Broadcom’s fiscal Q2 guide put next-quarter AI chip revenue around $16 billion against a roughly $17.2 billion expectation, and the company reiterated rather than raised its full-year AI forecast. The read-through was immediate: if the most consistent compounder in the AI supply chain is guiding in-line rather than above, the slope of the upside may be flatter than the tape had priced. Hyperscaler capex has been the single largest source of chip demand for two years, and investors have started questioning its durability. Nothing in the data shows capex actually being cut — the drawdown reflects doubt about pace, not confirmation of a decline.
Margin contagion from memory pricing. With memory prices having roughly doubled over six months, the fear shifted to whether surging chip costs would compress margins across the broader tech complex. That fear was validated when Apple raised prices on MacBooks and iPads by $100–$300, citing memory shortages. What is a tailwind for Micron and SanDisk is a cost line for everyone downstream.
Supply-side scares and geopolitics. A Seoul brokerage note cut SK Hynix’s Q2 estimate roughly 8% below consensus on a slower HBM4 shipment ramp, and China’s CXMT lined up an $8.6 billion memory IPO that revived DRAM oversupply worries. Layer in Strait-of-Hormuz-related risk pulling oil higher and risk appetite lower, plus hawkish Fed commentary, and you have a sentiment reset on top of a valuation reset.
The net result: for the first time since April, semiconductors no longer carry a valuation premium to the broader Nasdaq. A sector that spent months priced for an uninterrupted buildout is now priced closer to the average large-cap tech name.
Compute and Logic Silicon
This is the highest-quality corner of the wreckage and the one where balance sheets are cleanest.
Nvidia (NVDA). Trading near $202 with a market cap around $4.77 trillion, NVDA carries the largest capitalization in the group but, notably, the lowest forward multiple. Earnings scale has finally caught up to the share price. That inversion — biggest cap, cheapest multiple — is the single most important tell in the complex: it means the selling is about sentiment and positioning, not about Nvidia’s earnings power being in question. This is the name you own first on weakness because the downside is cushioned by a valuation that is no longer stretched.
Broadcom (AVGO). Near $375 after the in-line AI guide that started the unwind. The bear case is that the custom-silicon and networking franchise is maturing into a steadier grower rather than an accelerating one. The bull case is that in-line at Broadcom’s scale is still an enormous, recurring, high-margin business, and the stock now discounts a deceleration that has not actually shown up in the order book. Broadcom is the industry standard in Ethernet switching and the leader in custom AI accelerators; that moat did not narrow because one guide came in flat.
AMD. Around $457, swept down with the group. AMD is the clearest second-source play on accelerator demand, and its exposure to any capex normalization is real but secondary to Nvidia’s. It is a higher-beta way to express the same thesis, appropriate for a second tranche rather than a first.
Marvell (MRVL). Near $247 and up over 260% year to date even after falling roughly 15% from its June 3 record. The stock carries the S&P 500 inclusion completed June 22, a $2 billion Nvidia strategic stake, and record fiscal Q1 revenue of $2.42 billion. The overhang is insider selling — the outgoing CFO filed to sell about $65 million in shares — and a valuation that leaves no room for a miss. Marvell is the highest-conviction custom-silicon and optical-interconnect franchise outside Broadcom, but it is also the one most exposed to a cohort derating because the multiple is richest. Own it, but size it knowing it moves more than the index in both directions.
Intel (INTC). Cratered roughly 21% in seven trading days. Intel is a different animal — a turnaround story rather than an AI-demand story — and its inclusion in this selloff is more about beta and sentiment than about capex durability. It is not the way to express a dip-buy on the AI thesis; it is a separate bet on execution.
Memory
Memory is where the drawdown is deepest and where the risk-reward is most polarizing. The supply story remains intact; the market’s patience does not.
Micron (MU). Down roughly 17% from its $1,213 peak, trading in a volatile $860–$1,000 band, and still up more than 240% year to date. The fundamentals are not the problem: fiscal Q3 revenue of $41.46 billion rose 346% year over year at an 85% non-GAAP gross margin, with Q4 guidance around $50 billion. The forward multiple has compressed to single digits against a trailing multiple in the low twenties — the classic peak-earnings optics that make cyclicals look cheapest right before the debate about the cycle top. The overhang is a Michael Burry put position established near $1,051, plus the CXMT oversupply narrative. Micron is the cleanest large-cap expression of the DRAM and HBM shortage; it is also a 2.14-beta name that will whip on every macro headline. This is a first-tranche name for anyone who believes the HBM bottleneck persists into 2027, sized for further downside.
SanDisk (SNDK). Off roughly 30% from its 52-week high near $1,505, yet still up more than 600% year to date — the single most explosive move in the complex this year. SanDisk is the cleanest pure-play NAND expression, with none of the DRAM/HBM mix that complicates Micron. That purity cuts both ways: it is the most direct beneficiary of a NAND upcycle and the most exposed if NAND pricing rolls over first. The magnitude of the year-to-date gain is itself the risk — a name up 600% has the furthest to fall in a positioning unwind.
SK Hynix (SKHY). The U.S.-listed ADR trades around $164–$184 after the Korean line fell about 15% in a single session — its largest one-day drop ever — on the KIS note flagging a slower HBM4 ramp. SK Hynix is Micron’s most direct competitor in DRAM and high-bandwidth memory, so the HBM4 shipment scare cuts at the core bull thesis for the entire group. The freshly listed U.S. ADR now trades as a real-time sentiment gauge for the whole memory trade rather than a victory lap.
Western Digital (WDC) and Seagate (STX). WDC near $540 and up more than 230% year to date; STX up over 200% with a 2.07 beta. These are the storage-adjacent ways to play the same demand, with WDC’s July 29 earnings the next real fundamental checkpoint for the group. A clean WDC print would do more to stabilize the memory tape than any amount of price action this week.
The CXMT overhang. China’s CXMT $8.6 billion IPO is the structural bear case in one headline: new domestic DRAM capacity funded at scale, threatening the pricing discipline the entire supercycle rests on. It is a 2027-and-beyond concern, not a this-quarter one, but it is the reason memory cannot simply V-bottom on valuation alone.
Bandwidth and Connectivity
As GPU clusters scale to hundreds of thousands of accelerators, copper runs out of headroom and optics takes over. This is the highest-beta corner of the AI-infrastructure trade and the one that has both run hardest and reset fastest.
Lumentum (LITE). Near $757 and up roughly 98% year to date, Lumentum has the strongest fundamentals of the optical trio: record quarterly revenue around $808 million, up 90% year over year, an optical circuit switch backlog north of $400 million, and an expanded Nvidia collaboration tied to silicon photonics for gigawatt-scale AI factories. If you want the connectivity theme with the most order-book visibility and the least speculative froth, this is the name.
Coherent (COHR). Around $296 and up about 80% year to date, with structural support from a $2 billion Nvidia investment and an indium phosphide scale-up in Texas. Fiscal Q3 revenue of roughly $1.806 billion rose 21% year over year. Coherent is the more diversified, lower-beta way to own the transceiver and materials side of the optics buildout — less torque than the pure-plays, more ballast.
Applied Optoelectronics (AAOI). Near $110 and still up more than 230% year to date after falling as much as 17% in a single session on no company-specific news. With a beta around 3.69, AAOI amplifies every risk-off wave in AI hardware — it fell three-to-four times as hard as the index on down days. The bull case is accelerating hyperscaler demand for 800G and 1.6T transceivers plus new U.S. capacity, with full-year revenue guided above $1 billion versus roughly $456 million in 2025. The bear case is the mirror image: extreme volatility, customer concentration, and 800G ramp execution risk. This is the highest-risk, highest-torque name in the connectivity bucket — a small-position instrument, not a core holding.
POET Technologies (POET). The micro-cap at roughly $14, with fiscal Q1 revenue up 201.9% year over year off a tiny base (about $503,000) and an EPS loss of $0.08. POET is a story stock — optical engine and interposer technology aimed squarely at the interconnect bottleneck — and behaves accordingly, with triple-digit swings in both directions. It belongs only in a speculative sleeve sized to be written to zero without consequence.
How to Buy This
The distinction that matters is between a valuation reset and a demand collapse, and the current evidence points clearly to the former. Hyperscaler capex has not been cut; it has been questioned. The memory shortage has not been resolved; the HBM4 ramp has been pushed, not cancelled. The optics thesis has not been falsified; a crowded, parabolic trade got unwound. Those are different things, and the market is pricing them as if they might be the same.
The framework for adding, ranked by where the risk-reward is cleanest:
- First tranche — cushioned quality: NVDA and MU, the two names where the multiple has already compressed to a point that limits downside relative to the fundamentals underneath them.
- Second tranche — franchise moats at a discount: AVGO, MRVL, and LITE, high-quality franchises whose selloffs are about cohort derating rather than any crack in the business.
- Speculative sleeve — torque, sized small: AAOI, SNDK, and POET, the highest-beta expressions, appropriate only in size you can hold through another 20% down leg.
The single most important discipline here is to size for further downside rather than to size to catch the exact bottom. A 10% weekly drop after a doubling typically needs a base to form before it reverses; single-week bounces in names like this are frequently retraced. Scaling in across levels beats a single all-in entry.
The Position
This reads as a market repricing risk after an overextended run, not a broken thesis. The inversion in Nvidia — largest market cap, lowest multiple — is the clearest signal that the selling is positioning-driven rather than earnings-driven. Memory is in a technical bear market with an 85% gross margin and a persistent HBM bottleneck underneath it. The optics names gave back gains on no company-specific news while their backlogs kept growing.
The next genuine catalyst is not this week’s tape. It is the next round of hyperscaler capex guidance, with Western Digital’s July 29 print as the near-term memory checkpoint. A first tranche into the cushioned-quality names makes sense now; the rest is a question of how much further the unwind runs before the base forms. The demand story has been questioned, not answered — and the market’s patience has run out well ahead of the fundamentals.
Leave a Reply