NVIDIA Rubin SOCAMM Memory Cut Triggers Market Panic and MU Decline

Article by: Tide Research

On the morning of June 4, SemiAnalysis, the most influential independent research firm in the semiconductor industry, released a morning brief.

The key message is this: NVIDIA’s Vera Rubin NVL72 may reduce rack-level SOCAMM DRAM capacity from the previously expected ~55TB to ~28TB, with most Rubin systems adopting 96GB SOCAMM modules instead of the previously anticipated 192GB.

Once the news spread, the market's reaction was swift and harsh: memory demand was halved, a negative for Micron. MU's stock plunged more than 10% intraday, crashing from its previous day's all-time high of $1,089 to $971, wiping out over $100 billion in market value in a single day.

Fear is real, but is the direction of that fear correct?

Vera Rubin NVL72 is NVIDIA's next-generation flagship AI supercomputer rack. Each rack contains 72 Rubin GPUs and 36 Vera CPUs. The GPUs are equipped with HBM4, offering 288 GB per chip, resulting in a total of approximately 20.7 TB per rack—this remains unchanged. The change lies in the CPU side.

Each Vera CPU features eight SOCAMM slots, each capable of accommodating modules of varying capacities. NVIDIA's official specifications announced at CES 2026 state that "each Vera CPU supports up to 1.5TB of LPDDR5X," corresponding to a configuration with eight 192GB modules fully populated. With 36 CPUs, this totals 54TB.

SemiAnalysis's report states that the actual shipping configurations are unlikely to be fully populated. Most systems will use 96GB modules, resulting in 8 × 96GB = 768GB per CPU, and with 36 CPUs, this amounts to approximately 28TB.

Reduced from 55TB to 28TB, a capacity cut of nearly half—clickbait headlines would call it "memory requirements halved."

But the market got one key variable wrong.

First, SOCAMM is designed as a slot-based module, not soldered in place.

This is the most easily overlooked technical detail in the entire story. Unlike the GB300 Blackwell Ultra, which features LPDDR soldered directly onto the motherboard, the Vera Rubin platform uses the JEDEC-standardized SOCAMM2 module—hot-swappable, removable, and upgradable. Today, you can install 96GB; tomorrow, if the customer needs more, simply remove it and replace it with 192GB or even 256GB, just like upgrading standard memory modules.

NVIDIA highlighted this design at CES 2026: the entire compute tray assembly time has been reduced from 2 hours to just 5 minutes. Modularity, maintainability, and upgradability represent one of the most significant architectural advancements of Vera Rubin over Blackwell.

Reducing the initial sell-off configuration does not mean demand has disappeared permanently; it’s more like a "board first, pay later" strategy.

Second, the reason for reducing capacity is not because it's unnecessary, but because it's "not enough."

Dylan Patel, founder of SemiAnalysis, said something insightful on Twitter: “I really like one thing—people who share our reports often leave out most of the content. This happens all the time.”

Reader comments on Digg are also telling: 77.8% of comments believe the secondary dissemination is clickbait with misleading headlines.

What was overlooked? The context.

Global LPDDR5X supply will be extremely tight in 2026. Micron explicitly stated at the Wolfe conference at the end of May that memory demand significantly exceeds supply capacity, a situation expected to persist beyond 2026. Micron’s full-year HBM capacity for fiscal year 2026 is already sold out, with average DRAM prices up more than 110% year-over-year and gross margins surging to 74%. Samsung and SK Hynix are also operating at full capacity with full sales.

In this context, the issue NVIDIA faces is not that customers don’t want more memory, but rather, “I can’t get enough LPDDR5X chips to fill every slot.”

Reducing the default SOCAMM configuration per rack is essentially an engineering-level supply chain management strategy: rather than delaying the full rack delivery due to memory shortages, ship with a lower configuration to get computing power online as soon as possible.

This is not a signal of contracting demand; on the contrary, it is a signal of demand overwhelming supply.

Third, less memory does not mean fewer racks.

The market did a simple calculation: halving the memory per rack → halving total demand. But there’s another variable in this equation: shipment volume.

If each rack's SOCAMM is reduced from 55TB to 28TB, NVIDIA can install significantly more racks under the same LPDDR5X supply constraints. Previously, a batch of memory was sufficient for only 100 racks; now, it can support nearly 200.

The total LPDDR5X consumption has not decreased; it has simply been distributed across more racks. For NVIDIA, this is a pragmatic choice to bring Rubin to market faster; for memory manufacturers, the total order volume may not decline.

Moreover, the memory requirements on the CPU side for inference scenarios vary significantly. Not all workloads require 1.5TB of LPDDR5X. While large model training is indeed memory-intensive, many inference tasks—especially agentic AI and long-context inference—can flexibly schedule KV caches between HBM and LPDDR via NVLink-C2C. For many customers, 768GB of CPU-side memory is already sufficient.

Because SemiAnalysis was merely the second straw to break the camel's back.

The final straw was Broadcom. Before the U.S. market opened on June 4, Broadcom released its Q2 earnings. The numbers themselves were strong: revenue of $22.19 billion, a 48% year-over-year increase, and non-GAAP EPS of $2.44, exceeding expectations. However, CEO Hock Tan did not raise the full-year AI chip revenue guidance of $100 billion, and the market deemed it "insufficient." Broadcom’s stock plunged 15%, dragging down the entire semiconductor sector in a broad sell-off.

Micron did not have any negative company-specific news that day. Multiple outlets, including TipRanks, Motley Fool, and 24/7 Wall St., explicitly stated that this was a case of "collateral damage" leading to a correlated decline. As a core player in the AI memory supply chain, Micron’s performance is closely tied to sentiment around AI capital expenditures; Broadcom’s guidance prompted the market to reassess the projected growth rate for the entire AI chip supply chain.

On the same day, SemiAnalysis’s report spread widely, giving traders who were already looking for reasons to sell a perfect narrative: not only was overall AI sentiment weakening, but the specific figures for memory demand were also shrinking.

A stock with a $1 trillion market cap has surged 900% over the past year and just set a new all-time high yesterday. At this level, any negative headline serves as a catalyst for profit-taking. Panic doesn’t need to be right—it just needs an excuse.

Three judgments.

First, SemiAnalysis’s report itself is accurate, but the market has misinterpreted it. Rubin’s NVL72 default SOCAMM configuration is likely indeed below its theoretical maximum, determined by supply chain realities and customer demand elasticity. However, there is a significant gap between a “lower default configuration” and “shrinking memory demand”—a gap defined by a modular, plug-in upgrade architecture and an industry reality where demand far exceeds supply.

Second, Micron’s core risk currently lies not in SOCAMM, but in HBM4. As SemiAnalysis reported in February this year, Micron holds zero share in NVIDIA’s Rubin platform HBM4 orders, with SK Hynix taking 70% and Samsung 30%. Although Micron announced mass production and shipment of HBM4 in March, its market share is expected to be only 18%. In contrast, Micron’s position in SOCAMM remains very strong: it was the first company to launch a 256GB SOCAMM2 and has been NVIDIA’s core SOCAMM partner for the past five years. The impact on Micron from reduced SOCAMM configurations is far smaller than the marginalization of its HBM4 market share.

Third, this decline was a profit-taking event in a stock with a trillion-dollar market cap following an all-time high, amplified by two independent catalysts. Broadcom provided an emotional shock, while SemiAnalysis supplied narrative fuel. Together, they triggered a 10% pullback in a stock that had risen ninefold over the past 12 months. From a trading perspective, this isn't called "panic"—it's called "normal."

Dylan Patel’s tweet was right: most people who shared his report indeed missed the most important part of it.

The most dangerous mistake in semiconductor investing isn't misjudging the direction—it's getting the headline right but miscalculating the formula.