Morgan Stanley comments on Nvidia "memory halving" rumors: This is a real shortage; supply constraints have become the biggest obstacle for AI.

Rumors about Nvidia cutting the memory configuration of Vera Rubin racks caused a stir in the market, but Morgan Stanley believes this precisely proves the severity of memory shortages, rather than being a signal of cooling AI demand.

Last week, research firm SemiAnalysis released a report stating that Nvidia plans to reduce the LPDDR5 memory capacity of Vera Rubin racks from 55 TB to 28 TB, and the specification of SOCAMM memory modules from 192 GB to 96 GB. This news was interpreted by the market as weakening AI demand, directly leading to a 13% plunge in Micron (MU) stock on the same day it received Nvidia HBM4 certification, marking the largest single-day drop since April 2025. Morgan Stanley semiconductor analyst Joseph Moore immediately issued a research report refuting the above interpretation.

According to Chase Wind Trading Desk, Morgan Stanley clearly stated in its latest report that it has verified that some racks will ship with lower configurations, but emphasized that the adjustment is entirely due to supply-side constraints, not weak demand. Meanwhile, the firm raised its 2026 global semiconductor industry revenue forecast from $807 billion to $880 billion, and maintained overweight ratings on MU and SNDK, with target prices of $1,050 and $1,750 respectively.

Rumors are true, but the market misunderstood the logic

Morgan Stanley confirmed in its report that the reduction in Nvidia Vera Rubin rack memory configuration is indeed happening, but emphasized that the market's interpretation of this event is exactly the opposite.

The report points out that both Nvidia and cloud computing clients are willing to buy every GB of SOCAMM memory available, and once supply catches up, higher configurations will be restored immediately. Morgan Stanley believes the sole purpose of this adjustment is to minimize the impact of DRAM shortages on GPU rack sales, and that this itself is evidence of real shortages—not a result of cyclical double ordering.

In terms of scale impact, assuming 53,000 to 70,000 racks are built next year, at the 55 TB configuration, SOCAMM demand for racks would approach 5% of global DRAM demand. If fully halved, in the most extreme scenario, it would reduce demand by about 1.4 million TB, accounting for over 2% of a 6.2 million TB market size, and impacting the higher-value segments of the market. Morgan Stanley expects higher configurations to resume delivery soon after shipments begin.

April SIA data: No quick solution for memory shortages

Morgan Stanley also commented on the Semiconductor Industry Association (SIA) billing data released on June 5 for April, believing it further reinforces the continuous tightness in memory supply.

Overall semiconductor sales in April fell 2.2% month-on-month, significantly better than Morgan Stanley’s forecast of -12.1% and the 10-year seasonal average of -10.6%, with year-on-year growth accelerating to 106.4%. Among them, memory was particularly outstanding: DRAM sales fell only 3.7% MoM, far better than the forecast of -24.2% and the 5-year average of -27.6%, year-on-year growth soared to 375.3%, 3-month rolling YoY growth reached 298.5%, the highest since 2001, and average price YoY growth has accelerated for nine consecutive quarters; NAND sales fell 4.2% MoM, again much better than the forecast of -11.2%, YoY growth was 366.0%, 3-month rolling revenue YoY growth 307.0% and average price growth 213.4% both hit historical records.

Morgan Stanley pointed out that April data confirms the judgment that there is no rapid solution to memory shortages. DRAM has become the main bottleneck in AI infrastructure construction, with HBM wafer capacity, cleanroom, and EUV equipment constraints, as well as limited incremental NAND capacity, together supporting a “higher for longer” pricing environment. The firm believes that long-term supply agreements (LTA) are symptoms of tight supply and hyperscale cloud providers grabbing capacity, not the cause of this cycle.

Upgraded forecast, optimistic about memory and AI supply chain

Based on the above judgments, Morgan Stanley significantly raised its industry forecasts. The 2026 global semiconductor revenue forecast was raised from $910 billion to $1.607 trillion, with expected YoY growth upgraded from +91% to +103%, mainly driven by memory, where DRAM forecast jumped from $150.3 billion to $581.9 billion, and NAND from $67.7 billion to $293.5 billion. In 2027, the firm expects industry revenue to further grow 22% to $1.96 trillion, mainly driven by continued memory price transmission.

On the individual stock level, Morgan Stanley maintained an overweight rating for MU, with a baseline target price of $1,050 (current price $864, potential upside of 22%); maintained an overweight rating for SNDK, with a baseline target price of $1,750 (current price $1,559, potential upside of 12%). Additionally, the firm continues to favor AI infrastructure beneficiaries NVDA (overweight, target price $288), AVGO (overweight, target price $502) and ALAB (overweight, target price $240), as well as capital equipment/supply chain names LRCX, KLAC, and MKSI.

Broad market improves simultaneously, cycle expansion signals strengthen

Morgan Stanley pointed out that the April SIA data's highlights are not limited to memory; the broad market’s overall outperformance is also noteworthy, indicating that the current semiconductor upcycle is expanding from AI-driven to broader fields.

By category, discrete devices, analog chips, MCUs and MPUs all exceeded Morgan Stanley's estimates and the 10-year seasonal average. The industrial sector is re-accelerating from the cyclical low, with demand expanding to non-AI related fields, and inventory digestion continues. The firm stated that although supply chain price increases have emerged, the main driver is cost pass-through rather than value recapture, with a few companies like ADI being exceptions.

Morgan Stanley concluded that April data further proves that the current semiconductor upcycle is no longer a purely AI-driven narrow market, but is evolving into a broader supply-constrained upcycle. Against this backdrop, the firm also maintains a positive outlook for analog/MCU suppliers ADI, NXP, and ALGM.

~~~~~~~~~~~~~~~~~~~~~~~~

The above highlights are from Chase Wind Trading Desk.

For more detailed interpretations, including real-time commentary and frontline research, please join [Chase Wind Trading Desk Annual Member]

Risk Warning and DisclaimerThe market has risks, investment needs to be cautious. This article does not constitute individual investment advice, nor does it take into account the unique investment goals, financial situations, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article suit their specific circumstances. Investing based on this, responsibility is at your own risk.