Reports of NVIDIA cutting the memory configuration for its Vera Rubin racks have caused market turbulence, but Morgan Stanley argues this actually underscores the severity of a memory shortage, not a signal of cooling AI demand.
Last week, research firm SemiAnalysis published a report stating that NVIDIA (NVDA) plans to reduce LPDDR5 memory capacity for Vera Rubin racks from 55 TB to 28 TB, and shrink the SOCAMM memory module specification from 192 GB to 96 GB. This news was interpreted by the market as a sign of weakening AI demand, directly causing shares of Micron Technology (MU) to plunge 13% on the same day it received NVIDIA's HBM4 certification, marking its largest single-day drop since April 2025. Morgan Stanley semiconductor analyst Joseph Moore promptly issued a research report refuting this interpretation.
According to the report, Morgan Stanley explicitly states that it has verified some racks will indeed ship with lower configurations, but emphasizes this adjustment is entirely due to supply-side constraints, not demand weakness. Concurrently, the firm raised its 2026 global semiconductor industry revenue forecast from $8.07 trillion to $8.8 trillion, and maintained its Overweight ratings on MU and SNDK with price targets of $1,050 and $1,750, respectively.
Rumor Confirmed, But Logic Misread by Market
Morgan Stanley's report confirms that NVIDIA's reduction of the Vera Rubin rack memory configuration is indeed happening, but stresses the market has interpreted the event in the opposite direction.
The report notes that NVIDIA and its cloud computing customers are willing to purchase every available GB of SOCAMM memory, and will revert to higher configurations as soon as supply catches up. Morgan Stanley believes the sole purpose of this adjustment is to minimize the impact of DRAM shortages on GPU rack sales, which itself is evidence of a genuine shortage—not a product of double-ordering windows.
In terms of scale, assuming 53,000 to 70,000 racks are built next year, under the 55 TB configuration, rack SOCAMM demand would approach 5% of total global DRAM demand. A full halving could reduce demand by approximately 1.4 million TB in the most extreme scenario, equivalent to over 2% of a 6.2 million TB market, affecting a higher-value segment. Morgan Stanley expects higher configurations to return to supply shortly after shipments begin.
April SIA Data: No Quick Fix for Memory Shortage
Morgan Stanley also commented on the April Semiconductor Industry Association (SIA) billings data released on June 5, viewing it as further reinforcing the view of persistent tight memory supply.
Overall semiconductor sales in April fell 2.2% month-over-month, significantly better than Morgan Stanley's estimate of -12.1% and the 10-year seasonal average of -10.6%, with year-over-year growth accelerating to 106.4%. Memory performance was particularly notable: DRAM sales declined only 3.7% month-over-month, far better than the estimated -24.2% and 5-year average of -27.6%, with a year-over-year surge of 375.3%. The 3-month rolling year-over-year growth rate reached 298.5%, a historical high since 2001, with average selling price growth accelerating for nine consecutive quarters. NAND sales fell 4.2% month-over-month, also significantly better than the estimated -11.2%, with a year-over-year increase of 366.0%; the 3-month rolling year-over-year revenue growth of 307.0% and average selling price growth of 213.4% both set historical records.
Morgan Stanley points out that the April data confirms the judgment that there is no quick solution to the memory shortage. DRAM has become the primary bottleneck for AI infrastructure build-out, with constraints including dense HBM wafer capacity, cleanroom and EUV equipment, and limited incremental NAND capacity, all supporting a "higher for longer" pricing environment. The firm views long-term agreements (LTAs) as a symptom of supply tightness and hyperscale cloud vendors scrambling for capacity, not a cause of this cycle.
Raising Forecasts, Bullish on Memory and AI Supply Chain
Based on the above assessment, Morgan Stanley significantly raised its industry forecasts. The 2026 global semiconductor revenue forecast is increased from $9.1 trillion to $16.07 trillion, with the year-over-year growth expectation raised from +91% to +103%, primarily driven by memory. The DRAM forecast jumps from $150.3 billion to $581.9 billion, and NAND from $67.7 billion to $293.5 billion. For 2027, the firm expects industry revenue to grow a further 22% to $19.6 trillion, mainly driven by continued pricing pass-through in memory.
At the individual stock level, Morgan Stanley maintains its Overweight rating on MU with a base case price target of $1,050 (current price $864, implying 22% upside potential); it maintains its Overweight rating on SNDK with a base case price target of $1,750 (current price $1,559, implying 12% upside potential). Additionally, the firm continues to favor AI infrastructure beneficiaries NVDA (Overweight, target $288), Broadcom (AVGO) (Overweight, target $502), and ALAB (Overweight, target $240), as well as capital equipment/supply chain names LRCX, KLAC, and MKSI.
Broad-Based Market Also Improves, Cycle Expansion Signals Strengthen
Morgan Stanley notes that the highlights of the April SIA data are not limited to memory; the broad-based market's overall outperformance is equally noteworthy, indicating this semiconductor upcycle is broadening from AI-driven to more widespread areas.
By category, discrete devices, analog, MCU, and MPU all exceeded Morgan Stanley's estimates and the 10-year seasonal average. The industrial sector is re-accelerating from cycle lows, with demand expanding into non-AI related areas, and inventory digestion continues to progress. The firm states that while supply chains have seen price increases, the primary driver is cost pass-through rather than value recapture, with a few exceptions like Analog Devices (ADI).
Morgan Stanley concludes that the April data further proves the current semiconductor upcycle is no longer a narrow, purely AI-driven story but is increasingly evolving into a broader, supply-constrained upcycle. In this context, the firm also maintains a positive view on analog/MCU suppliers ADI, NXP, and ALGM.
Comments