2026-11-17 15:36
Lead: Morgan Stanley downgrades cloud-side semiconductors and legacy memory, citing a shift of approximately 5GW of capacity deployment from 2028 to 2029—equivalent to 20% of original 2028 forecast.
November 15, Morgan Stanley revised down ratings for a group of stocks—cloud-side semiconductors and legacy memory. The report opens by acknowledging investor pushback against these downgrades.
Its response is direct: demand remains strong. But by the time "cloud capex deceleration" becomes visible in data, pricing has already priced it in.
This is a Q&A-style report addressing four concentrated investor concerns. The sector view remains "attractive". The overarching stance: industry outlook is positive, but individual names should be rotated out ahead of macro shifts.
The report is authored by four analysts.
Daniel Yen, Charlie Chan.
Daisy Dai, Ethan Jia.
What Investors Are Asking
Four questions, each more pointed than the last.
First: Is the downgrade of Montage’s earnings forecast justified? Second: What underpins the multiple compression, and what catalysts are driving this?
Third: Have server and CPU shipment forecasts been revised? Why act now given such strong demand? Fourth: Why retain CXMT and Nanya?
The report answers each in turn. Together, the four responses form the core thesis: demand is intact, but timing is off.
An Early Exit Downgrade
Previously, two reports had been issued—one forecasting peak growth in cloud-side semiconductors, another signaling peak pricing power in legacy memory. This report serves as a follow-up clarification.
The downgraded firms fall into two categories: those with heavy exposure to cloud-side semiconductors and legacy memory. The former’s revenue tracks server deployments; the latter’s hinges on memory pricing.
In the cloud-side segment, Aspeed (Xin Hua) produces server management chips; Montage (Lanqian) manufactures memory interface chips. The report now adopts a more conservative assumption for RCD shipments in 2027.
RCD refers to Register Clock Driver chips on memory modules—components that contribute to per-server content value. The report notes this opportunity remains intact.
In the legacy memory cohort, Phison (Qunlian) received a “reduce” rating. Winbond (Hua Bang), Macronix (Wang Hong), and others were similarly adjusted, with timing concentrated around November 10–11.
The report does not frame the downgrade as bearish on the sector. It repeatedly emphasizes robust demand fundamentals, confirmed by Q3 earnings. What has changed is the timing of the assessment.
In one sentence: By the time capex deceleration becomes evident in data, valuation adjustments have already occurred. Therefore, position reduction should happen preemptively—not reactively.
Demand Unchanged, Timing Shifted
The two most concrete figures on the page both relate to timing.
The first is 5GW. The U.S. team has shifted roughly 5GW of new capacity deployment from 2028 to 2029. Per the report’s framework, this represents 20% of the original 2028 forecast.
The second is the capital expenditure projection. Spending expectations from hyperscalers remain unchanged: $1.6 trillion in 2027, $1.8 trillion in 2028.
Combined, these figures convey a clear message: deployment intent persists, but the timing of equipment procurement and deployment is now harder to pin down. Orders may be delayed—but they will not vanish.
The report attributes the downward revision to three sources of uncertainty: procurement timing, platform transition, and supply chain execution. It also acknowledges that the 2027 RCD shipment assumption may be overly conservative.
This sentence deserves isolation: a conservative assumption implies upside potential. If customer orders and deployments proceed more smoothly than current assumptions allow, this upside could materialize.
The Q3 earnings season has just concluded. The report argues this moment is appropriate for incorporating greater risk, citing post-midterm election execution uncertainty.
It also reminds readers: intention to spend and actual spending are two distinct concepts. The former is guided by forward-looking statements; the latter is measured by equipment delivery and deployment.
Pricing Power in Legacy Memory Weakening
The issue with legacy memory isn’t demand—it’s the sustainability of pricing and returns.
During upcycles, fixed-price long-term contracts enabled rapid price increases. As demand weakens, such contracts offer diminishing protection.
The report’s exact wording: when shortages narrow, pricing and return sustainability must be reassessed. Its updated supply-demand model points to normalization by 2028.
The conclusion lands on valuation multiples: lower multiples should be applied before normalization, not after peak profitability rolls over.
A supporting observation: the breadth of earnings upgrades is peaking. Visibility into 2028 server production is weakening.
Together, these factors erode confidence in forward revenue growth. This explains why the report adjusts valuation before earnings.
Two Companies Left: Different Logic
The report maintains “overweight” ratings for CXMT and Nanya. Their retention reasons differ significantly.
CXMT benefits from three additional drivers: capacity expansion, localization, and process migration. Management confirms customers have pre-booked enterprise-grade capacity through 2027.
DDR5 supply remains tight—this is CXMT’s second rationale for retention.
Nanya leverages process technology and DDR5 migration. DDR5 already contributes about 10% of its revenue. This path differentiates it from other DDR4 suppliers.
The report offers a structural view on niche memory: smaller supply base, higher sensitivity to minor supply-demand imbalances, and lower direct exposure to AI-driven demand.
This duality holds true: elasticity comes from this structure, but so does fragility.
Both companies still face memory cycle risks. The difference lies in CXMT’s added growth drivers and Nanya’s clearer migration path.
Who Else Is Still Being Overweighted?
The report concludes with a coverage list, showing ratings and closing prices as of November 13. No target prices are provided.
Still rated “overweight”: TSMC, MediaTek, Alchip.
Also included: SMIC, ASMPT, Global Unichip.
“Reduce” ratings include: MetaX, Vanguard, Silergy, Phison.
One additional power semiconductor firm: AWSC.
These are closer to legacy memory and mature nodes.
When cross-referenced with the main text, the downgrade distribution becomes clear: cloud-side exposure and legacy memory pricing are the two axes of this adjustment. All else remains unchanged.
The Report’s Own “Face-Saving” Conditions
Bulls can monitor three signals: faster deployment, upward revisions to earnings forecasts, and renewed capex acceleration after Q1 2027.
If any one of these is validated, the cautious stance collapses. The report itself lists faster deployment as a factor capable of challenging its conservative view.
Conversely, weakening order momentum and slower price increases are identified as catalysts by the report. It admits these timing points remain uncertain.
The entire conclusion rests on one assumption: the 5GW delay will indeed translate into delayed equipment orders. If this assumption fails, the downgrade came too early.
The report also leaves two tail risks: first, niche memory exhibits high price elasticity—prices rise fast during upswings and fall rapidly during downturns. Second, direct exposure to AI demand is limited. Thus, when the AI narrative cools, impact is relatively muted.
These two points explain the coexistence of reduction and accumulation within the same report—they are exposed to different cycles.
An additional watch window worth noting: Q1 2027 capex guidance will be the first real test of this judgment. A raised guidance would signal that caution should yield to optimism.
How to Read This Rating Table
Ratings reflect risk-adjusted total returns over the next 12 to 18 months. Target prices span the same horizon—making this a cross-year assessment, not a quarterly view.
The three-tier structure must be clarified: “Overweight” implies performance above the sector average; “Neutral” implies parity; “Underweight” implies underperformance.
Ratings denote relative performance, not absolute price direction. A “favorable” sector view doesn’t imply every stock should be held.
Prices listed are as of close on November 13. They serve as pricing benchmarks, not targets. The report notes historical prices were not adjusted for stock splits.
Final reminder: the value of this report lies not in its conclusion, but in its timing. It simultaneously asserts “strong demand” and “exit now” in the body. Which will play out first—the reader must decide.
Institutional Ledger | Morgan Stanley
Source: TechFlow Column
Disclaimer: Contains third-party opinions, does not constitute financial advice
AI technology stocks face the largest sell-off in a decade
08-06
After technology stocks double, they average a 40% drop the following year
08-05
NVIDIA Races for Speed, AMD Races for Volume: A $170 Billion CPU Battle
07-24
Google AI Doesn't Just Burn Cash: Cloud Revenue Up 82%
07-24
The AlphaGo Moment in Cybersecurity, or Starting from These 466 Agent Matchups
12 days ago
The Trump administration plans to offer Constellation Nuclear a $5 billion loan for capacity expansion, with the additional power mostly pre-allocated to AI giants such as Google
19 days ago
Silicon Valley Executives Spark Controversy Over Using AI for Household Tasks; Survey Reveals Half of American Adults Worry About Impact on Relationships
19 days ago






