PRINTER DEEP RESEARCH REPORT ON MEMORY STOCKS $MU $SNDK
TL;DR
Micron (MU): BUY the dip — high conviction. The July 27 CXMT-driven selloff to ~$871–900 is a sentiment/positioning event (Burry short, profit-taking after a 223% run, a headline IPO), not a fundamental break. CXMT cannot touch Micron’s profit engine — HBM3E/HBM4 is sold out through calendar 2027, backed by 16 take-or-pay contracts worth ~$100B minimum revenue, and MU trades at only ~8–9x consensus FY27 EPS. That is cheap for what it is.
SanDisk (SNDK): BUY but lower conviction / more tactical. NAND is more commoditized and more directly exposed to Chinese competition (YMTC is scaling wafers hard) than DRAM/HBM is to CXMT, and SNDK’s ~505%+ YTD run leaves more froth. But New Business Model contracts with price floors, an enterprise-SSD demand inflection, a zero-debt balance sheet, and a ~8x forward P/E make the selloff a buying opportunity, not a top — size it smaller and respect the higher beta.
The CXMT threat is real long-term but mispriced short-term. It pressures commodity DDR5/LPDDR5, not the leading-edge HBM/enterprise-SSD tier where MU and SNDK now make their money. The memory cycle is not at its top: contract prices are still rising (just decelerating), inventories sit at 2–4 weeks (lowest in modern history), and hyperscaler capex is guided up ~77% to ~$725B in 2026.
Key Findings
1. CXMT is a genuine threat to commodity DRAM — but years away from the HBM/leading-edge segment that drives MU’s and SNDK’s margins. Per Reuters, CXMT shares “soared 466% on their Shanghai trading debut” July 27, lifting its market cap to 3.3 trillion yuan (~$487.73B), with intraday prints as high as +531% (~3.66–3.68 trillion yuan per TechNode); it “raised 57.92 billion yuan ($8.6 billion) in the IPO, the biggest mainland Chinese semiconductor offering on record.” But per CXMT’s own IPO prospectus it holds just a “7.67% share of the global DRAM market by late 2025,” its cost-per-bit is >30% above the incumbents (sustainable only at supercycle prices, per Tech Times), it is DUV-only (no EUV) with a ~2–3 year DRAM technology gap and a ~3–4 year HBM gap, and it carries a U.S. DoD “Chinese military company” (Section 1260H) designation plus an NDAA federal-procurement ban. China-made competitive-volume HBM is a 2028+ story.
2. Even with CXMT’s aggressive expansion, DRAM stays undersupplied. SemiAnalysis models CXMT reaching ~350k wafer starts/month by end-2026 (vs Micron’s ~375–385k) and ~17% global DRAM share by 2028 — yet still projects DRAM undersupplied by a high-single-digit % in 2026, widening to low-to-mid teens in 2027. The Chinese wafer adds are already in the models, and the market is still short.
3. Micron’s fundamentals are the strongest in company history and structurally de-risked. Fiscal Q3 2026 (reported June 24): revenue “$41.46 billion versus $9.30 billion for the same period last year” (+346% YoY), with non-GAAP net income “$28.86 billion, or $25.11 per diluted share.” Q4 guidance (from the earnings call): record $50.0B revenue (±$1B), ~86% gross margin, and $31.00 non-GAAP EPS (±$1). CEO Sanjay Mehrotra: “Micron’s record fiscal Q3 financial results and even stronger outlook for Q4 reflect the strategic value of memory in the AI era.” HBM3E/HBM4 is fully booked through calendar 2027; HBM4 entered volume production a quarter early. Sixteen take-or-pay Strategic Customer Agreements “cover roughly 20% of DRAM volume and one-third of NAND volume through calendar 2030,” with “$100 billion” minimum contractual revenue and “$22 billion in customer financial commitments.” Net cash ~$24.4B.
4. Micron is not priced for perfection — it’s priced for a cycle turn its contracts have structurally delayed. At ~$871–900, market cap ~$1.02–1.04T (EV ~$990B), MU trades at ~12x FY26 EPS (~$73) and ~8–9x the $98.52 consensus FY27 non-GAAP EPS. That is a below-market multiple for a company Wall Street models 50%+ upside on (average target ~$1,492 across 45 analysts, “Strong Buy”). The only stretched metric is price-to-book (~10.3x vs ~2.2x at the 2018 peak and ~2.0–2.5x in 2021–22) — a lagging measure as book value catches up (BVPS jumped from ~$52 in Nov 2025 to ~$89 now).
5. SanDisk is cheap on earnings but more exposed and frothier. Fiscal Q3 2026 (press release, April 30): “Third quarter revenue was $5.95 billion, up 97% sequentially” (+251% YoY), with datacenter revenue “$1.47 billion, up 233% quarter over quarter and 645% year over year,” and non-GAAP EPS of $23.41. Q4 guide: $7.75–8.25B revenue, $30–33 EPS. Five NBM agreements lock in “$42 billion in minimum contractual revenue backed by over $11 billion in financial guarantees.” Forward P/E ~8x, zero debt, buyback authorized. But NAND faces YMTC scaling toward 500k wspm and ~15% share, and SNDK’s +505–707% YTD run is the hottest in the S&P 500.
6. History says selling the “China enters memory” headline has usually been wrong — but selling the cycle top has been right. The 2018–19 downturn crushed MU (from ~$64 in May 2018 to ~$28 by year-end, -57%) — but it was caused by cloud inventory unwind, not by China (JHICC was shut down by U.S. sanctions and never scaled). The lesson: the cycle, not the China headline, is what breaks memory stocks. This cycle’s structure (take-or-pay LTAs, HBM scarcity, 2–4 week inventories) is materially different.
Details
Is the CXMT threat priced correctly? — Overpriced in the near term, appropriately feared long term
The selloff treated CXMT as an imminent margin-killer. It isn’t. CXMT’s output is overwhelmingly conventional DDR5/LPDDR5 for phones, PCs and mainstream servers — the low-margin commodity tier the Big Three have been fleeing toward HBM. Independent testing (Hardware Unboxed, Feb 2026) shows CXMT’s consumer DDR5 has essentially closed the gaming-performance gap, and its 64GB server DDR5 module pricing has even overtaken Samsung’s on supply reliability. But:
HBM is a different sport. HBM requires TSV die-stacking, where a single defect ruins the whole stack. CXMT held ~1% of HBM wafer supply in 2025; even bullish projections put it at ~12% by 2028. Samsung, SK Hynix, and Micron hold >99% today. CXMT targets HBM3 samples by end-2026 and HBM3E volume in 2027, but competitive-yield volume is realistically 2028+.
Equipment ceiling. CXMT stockpiled ASML immersion DUV tools and has semiconductor manufacturing equipment for HBM “through 2026 or 2027,” after which export controls bite. The pending MATCH Act would extend DUV restrictions specifically to CXMT and bar ASML from servicing its installed base.
Cost structure. CXMT’s >30% cost-per-bit disadvantage is “sustainable at supercycle prices; not sustainable in a normalized DRAM market” (Tech Times). If prices normalize, CXMT bleeds before Micron does.
On the NAND side, YMTC is the more credible near-term disruptor: ~11.8% share, Phase 3 Wuhan fab starting H2 2026, and a path toward 500k wspm across five planned fabs with >50% domestic tooling — targeting ~15% global NAND share. This is precisely why I hold SNDK to a lower conviction than MU: YMTC in NAND is a far more direct analog than CXMT in HBM.
Where are we in the cycle? — Mid-cycle, not the top
TrendForce data through July 2026: DRAM contract prices rose ~90–95% QoQ in Q1, ~58–63% in Q2, and a still-positive but moderating 13–18% guided for Q3; NAND rose ~55–60%, then ~70–75%, then 10–15%. Prices are decelerating from a blistering pace, not falling. Supporting the “not a top” call:
Inventories at 2–4 weeks globally — the lowest in modern cycles; lead times 40+ weeks.
HBM consumes ~3x the wafer area per bit of standard DRAM and ~70% of new DRAM output, structurally starving commodity supply.
New fab capacity (SK Hynix M15X, Micron Idaho/New York, Singapore NAND) doesn’t arrive in volume until late 2027–2028+.
Hyperscaler 2026 capex guided to ~$725B, up ~77% from ~$410B in 2025 (Amazon ~$200B, Google ~$185B, Meta $115–135B, Microsoft ~$120B), with 2027 projected above $1 trillion — and all four majors report being supply-constrained, not demand-constrained.
The bear case (respect it): Michael Burry disclosed a short on MU (plus NVDA, AMAT, SOXX) in his July 2 Substack, shorting at $1,051.87 and writing “Micron defines cyclical like no other” — citing 34 drawdowns >30% in 42 years, a median ROIC of 4% and ROE of 7% he called “frankly terrible,” and “one quarter in every three, Micron is a destroyer of capital.” His broader thesis targets AI-server depreciation math. The genuine risks: (a) double-ordering by memory-hungry customers that could reverse fast, (b) a hyperscaler capex plateau, and (c) all three incumbents plus China adding capacity simultaneously — the classic setup for oversupply 2–3 years out. These are 2027–2028 risks, not this-week risks.
Valuation check
MU: ~12x FY26 / ~8–9x FY27 EPS ($98.52 consensus); EV/EBITDA ~14.5–15x trailing (falling fast as EBITDA compounds); P/B ~10.3x (the one stretched metric, and a lagging one). At the 2018 peak MU traded ~4.5x forward earnings and still fell 57% — but that was because earnings collapsed, which take-or-pay contracts now cushion. DRAM ASPs (~$1.17) are above the prior 2018 peak (~$0.87), and Mehrotra has said trough-cycle gross margins under the new LTAs would be “well beyond the peaks” of prior cycles. Recent targets: KeyBanc $1,750 (July 14), BofA $1,550 (July 21), UBS $1,625, Morgan Stanley $1,200 — Joseph Moore called the July selloff “a buying opportunity.”
SNDK: Forward P/E ~8x; EV/EBITDA elevated on trailing but collapsing forward as quarterly EPS ramps from $23.41 toward $30–33 guided. Analyst targets span a wide $2,000–$3,250 band (Goldman ~$2,200, Bernstein $3,000, Susquehanna $3,250) — these are forward projections, not current value, and the width signals genuine uncertainty about NAND durability. Bernstein estimates SNDK’s contracts carry a ~$0.29/GB floor, “broadly in line with Q2-2026 average selling prices,” and argues that even in a worst-case collapse worse than 2010 the LTAs “should significantly mute earnings downside in 2029 and 2030.”
SK Hynix Q2 earnings — the live catalyst
SK Hynix reports Q2 2026 results July 28 (some sources say July 29) after the U.S. close. Consensus (14 brokerages via Yonhap Infomax): revenue ~84.1 trillion KRW (~$57.6B) and operating profit ~64.1 trillion KRW — which “would surpass the company’s previous annual operating profit record of 47.2 trillion won in 2025,” at a margin near 75–77%, all record highs. A more cautious whisper (Korea Investment & Securities) sits lower at 80.9T revenue / 60.4T operating profit. What confirms the thesis: a beat, reaffirmed HBM sold-out status through 2027, HBM4/Rubin allocation commentary, and disciplined capex. What breaks it: soft HBM pricing, weak 2027 guidance, or a capex surge signaling a supply war.
Recommendations
Micron — Accumulate; treat $850–880 as a gift entry.
On the cash-secured puts at $850–880 expiring this week: these align perfectly with your hold-to-2027 thesis. Assignment at $850–880 means owning MU at ~10–11x FY26 / ~8.5x FY27 EPS — an excellent scale-in level. Do not panic-close them for a fundamental reason; the only reason to act is this week’s binary event stack (SK Hynix + FOMC + PCE all landing while the puts expire). If you are comfortable owning shares at the strike, let assignment happen. If you want to dodge this week’s gap risk, roll to next week/next month at the same or slightly lower strikes to collect more premium and buy time past the events.
Scale-out plan to early 2027: stay the course. Add on any flush toward $800, which has been strong support twice in July.
Continue selling puts on SNDK and SKHY, but keep SNDK size below MU given the higher beta and YMTC exposure.
SanDisk — Buy the dip, smaller size.
Constructive at ~$1,270; the datacenter/eSSD inflection and NBM floors ($42B contracted / $11B guaranteed) are real. But treat it as the higher-risk leg — YMTC is a more direct NAND threat than CXMT is to HBM, and the parabolic YTD run means sharper drawdowns.
Prefer selling puts at strikes well below spot (a bigger cushion) over chasing shares here.
Benchmarks that would flip the calls from buy to reduce:
Contract prices turning negative QoQ (vs merely decelerating) — the single most important signal.
A hyperscaler cutting or pausing 2027 capex guidance on the earnings calls this week/next.
Any customer canceling or renegotiating a Micron take-or-pay LTA, or HBM pricing cracking.
MU sustained break below $800 on fundamentals; SNDK break below its 50-day with negative NAND data.
The 2–3 data points to watch this week
SK Hynix Q2 earnings (July 28/29, after U.S. close) — the highest-signal event. Record numbers with HBM sold out into 2027 and capex discipline = green light to add. Cautious HBM/2027 guidance = trim and wait.
FOMC decision (July 28–29) — a hold is expected under Chair Kevin Warsh, but the tone is hawkish (~25% July-hike odds, ~82% September-hike odds priced). A hawkish surprise or hike lifts discount rates and pressures high-beta memory valuations regardless of fundamentals.
PCE inflation — a hot print raises hike odds and compounds valuation pressure on MU/SNDK; a cool print (June expected ~3.7–3.8%) relieves it. Also watch hyperscaler earnings (MSFT, META, AMZN, GOOGL, AMD) for capex confirmation — the demand backbone of the entire thesis.
Caveats
Forward-looking figures flagged: analyst price targets ($1,200–$3,250) and 2027–2028 capacity/share projections are estimates, not facts. FY27 EPS consensus for MU varies by source — $98.52 is the best-sourced current figure; some cite $102–$149.
Date/figure conflicts: sources disagree on whether SK Hynix reports July 28 or 29, and on exact PCE release timing. One source referenced SK Hynix guidance disappointing and triggering a rout — that appears to reference a prior July episode, not the upcoming print; treat the July 28 report as the live catalyst.
The bear case is not zero. This is still a cyclical industry. The structural arguments (take-or-pay, HBM scarcity) genuinely reduce downside vs prior cycles, but do not eliminate it. Burry has been early-but-right on cyclical turns before, and shorted MU well above current prices ($1,051.87).
Concentration risk: a concentrated MU position via short puts into a triple-event week is high-variance. The calls above assume you can withstand a 15–20% gap and would welcome assignment; if that is not true, reduce size before the events.
