Written by: Trend Research

In the early morning of June 4, the most influential independent research organization in the semiconductor industry, SemiAnalysis, released a morning report.

The key message is just one sentence: The SOCAMM DRAM capacity for Nvidia's Vera Rubin NVL72 per rack may drop from the previously expected about 55TB to about 28TB. Most Rubin systems will use 96GB SOCAMM modules, rather than the widely anticipated 192GB.

After the news spread, the market's reaction was simple and brutal: memory demand was cut in half, negatively impacting Micron. MU at one point plunged over 10% during trading, falling from a historic high of $1089 just the day before to $971, evaporating over $100 billion in market value in a single day.

Panic is real, the question is, is the direction of the panic correct?

First, lets clarify the numbers

The Vera Rubin NVL72 is Nvidia's next-generation flagship AI supercomputer rack. Each rack contains 72 Rubin GPUs and 36 Vera CPUs. The GPUs use HBM4, with 288GB each, totaling about 20.7TB per rack, which has not changed. What has changed is on the CPU side.

Each Vera CPU is equipped with 8 SOCAMM slots, each capable of accepting modules of different capacities. Nvidia's official specs announced at CES 2026 state that "each Vera CPU supports up to 1.5TB LPDDR5X," corresponding to a full configuration of 8 x 192GB modules. With 36 CPUs, that would be 54TB.

SemiAnalysis's report states: The actual shipment configuration is unlikely to be fully populated. Most systems will use 96GB modules, 8x96GB=768GB per CPU, and 36 CPUs will be approximately 28TB.

From 55TB to 28TB, the capacity shrinks by nearly half, leading clickbait headlines to proclaim "memory demand is halved."

However, the market overlooked a key variable in this calculation.

Logical flaws in the panic

First, SOCAMM is a modular design, not soldered.

This is the most easily overlooked technical detail in the entire story. Unlike the LPDDR soldered to the motherboard on the GB300 Blackwell Ultra, the Vera Rubin platform uses JEDEC standardized SOCAMM2 modules, which are pluggable, hot-swappable, and can be upgraded later. Today you install 96GB, tomorrow if the customer needs more, you can swap it out for 192GB or even 256GB, as simple as changing memory sticks.

Nvidia highlighted this design at CES 2026: the entire computing tray's assembly time was compressed from 2 hours to 5 minutes. Modularity, maintainability, and upgradability are among the biggest architectural evolutions of Vera Rubin compared to Blackwell.

Lowering initial shipment configurations does not mean that demand disappears permanently. It's more like a strategy of "get on the bus and pay afterward."

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

Dylan Patel, the founder of SemiAnalysis, made a poignant remark on Twitter: "I love one thing, which is that most people who retweet our report miss the majority of its content. This happens often."

Comments from readers on Digg regarding this news are also quite telling: 77.8% believe the secondary dissemination is selective reporting.

What was missed? The context.

In 2026, global LPDDR5X supply is extremely tight. Micron clearly stated at the Wolfe conference at the end of May that memory demand significantly exceeds supply capacity, a situation expected to continue beyond 2026. Micron's HBM capacity for the entire fiscal year 2026 has already been sold out, and DRAM prices have risen over 110% year on year, with gross margins soaring to 74%. Samsung and SK Hynix have also been running at full capacity with sold-out products.

Against this backdrop, Nvidia's problem is not that customers do not 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 a supply chain management strategy at an engineering level: rather than delaying the delivery of the entire rack due to memory shortages, it's better to ship with a lower configuration first to get the computing power online quickly.

This is not a signal of demand contraction; on the contrary, it signals that demand exceeds supply.

Third, less memory ≠ fewer racks.

The market performed a simple multiplication: halving the memory per rack → total demand is halved. But there is another variable in this equation: shipment volume.

If each rack's SOCAMM drops from 55TB to 28TB, Nvidia can assemble more racks under the same LPDDR5X supply constraints. What originally could accommodate 100 racks of memory now is enough for nearly 200.

The total consumption of LPDDR5X 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, total order volume may not necessarily decrease.

Moreover, inference scenarios have a large demand elasticity for CPU-side memory. Not all workloads require 1.5TB of LPDDR5X. Large model training indeed consumes memory, but many inference tasks, especially agentic AI and long context reasoning, can flexibly schedule KV cache between HBM and LPDDR via NVLink-C2C. For many customers, 768GB of CPU-side memory is already sufficient.

So why did Micron still drop 10%?

Because SemiAnalysis was just the second straw that broke the camel's back.

The first straw was Broadcom. Before the US market opened on June 4, Broadcom released its Q2 earnings report. The numbers themselves were not bad: revenue of $22.19 billion, up 48% year on year, Non-GAAP EPS of $2.44 exceeded expectations. However, CEO Hock Tan did not raise the guidance on full-year AI chip revenues of $100 billion, and the market felt "insufficient." Broadcom's stock plummeted 15%, dragging the entire semiconductor sector down.

On that day, Micron had no negative news at the company level. Various media outlets such as TipRanks, Motley Fool, 24/7 Wall St. clearly pointed out that this was a "collateral damage" type of decline. As a core player in the AI memory chain, Micron’s stock is highly tied to the sentiment of AI capital expenditure; Broadcom's guidance prompted the market to reassess the expected growth rate of the entire AI chip industry chain.

SemiAnalysis's report spread on the same day, providing traders who were already looking for reasons to sell a perfect narrative: not only is overall AI sentiment weakening, but even the specific numbers on memory demand are shrinking.

A stock with a market value of a trillion dollars, which has risen 900% over the past year, just set a historic high the day before. At this level, any negative headline acts as a catalyst for profit-taking. Panic does not need to be correct, it only needs an excuse.

Trend interpretation

Three judgments.

First, SemiAnalysis's report itself is accurate, but the market's interpretation of it is wrong. The default SOCAMM configuration of the Rubin NVL72 will likely indeed be below the theoretical maximum, determined by the realities of the supply chain and the elasticity of customer demand. However, between "default configuration reduced" and "memory demand shrank," lies a modular architecture that is pluggable and upgradable, and an industry reality where demand far exceeds supply.

Second, Micron's current core risk is not SOCAMM, but HBM4. SemiAnalysis reported in February that Micron has a zero share in the HBM4 orders for Nvidia's Rubin platform, with SK Hynix taking 70% and Samsung 30%. Although Micron announced the mass production and shipment of HBM4 in March, its market share is expected to be only 18%. In contrast, Micron's position in the SOCAMM field is very stable: it was the first to launch the 256GB SOCAMM2 and has been Nvidia's core partner in the SOCAMM solution for five years. The actual impact of reduced SOCAMM configuration on Micron is much smaller than being marginalized in HBM4 share.

Third, the nature of this decline is profit-taking for a trillion-dollar stock after hitting an all-time high, amplified by two independent catalysts. Broadcom provided the emotional shock, and SemiAnalysis offered the narrative ammunition. The combination of both allowed a stock that had risen 9 times in the past 12 months to pull back 10%. From a trading perspective, this is not called "panic," this is called "normal."

Dylan Patel's tweet was right: most of those who retweeted his report indeed missed the most important parts of the report.

The most dangerous thing in semiconductor investment is not misreading the direction, but reading the title correctly yet miscalculating the formula.