Author: Chaoxiang Research
At 4 AM on June 4, SemiAnalysis, the most influential independent research institution in the semiconductor industry, published a morning report.
The core information was just one sentence: The SOCAMM DRAM capacity per rack for NVIDIA's Vera Rubin NVL72 might drop from the previously expected ~55TB to ~28TB. Most Rubin systems will use 96GB SOCAMM modules, not the 192GB modules widely anticipated by the market.
After the news spread, the market's reaction was simple and brutal: Memory demand is halved, bearish for Micron. MU plummeted over 10% intraday, sharply falling from the all-time high of $1,089 set just the day before to $971, wiping out over a hundred billion dollars in market cap in a single day.
The panic was real, but the question is, was the panic directed correctly?
First, Let's Do the Math
The Vera Rubin NVL72 is NVIDIA's next-generation flagship AI supercomputing rack. Each rack packs 72 Rubin GPUs and 36 Vera CPUs. The GPUs use HBM4, 288GB per GPU, totaling about 20.7TB per rack; this part remains unchanged. The change is on the CPU side.
Each Vera CPU has 8 SOCAMM slots, which can accommodate modules of different capacities. The official specification NVIDIA announced at CES 2026 is "up to 1.5TB LPDDR5X per Vera CPU," corresponding to a configuration with all 8 slots filled with 192GB modules. With 36 CPUs, that's 54TB.
What this SemiAnalysis report says is: Actual shipping configurations will most likely not be fully populated. Most systems will use 96GB modules, 8×96GB=768GB per CPU, and 36 CPUs equals about 28TB.
From 55TB to 28TB, a reduction of nearly half—headline writers could easily call it "memory demand slashed in half."
But the market miscalculated a key variable in this equation.
The Logical Flaws in the Panic
First, SOCAMM is a socketed, pluggable design, not soldered on.
This is the most easily overlooked technical detail in the whole story. Unlike the LPDDR soldered onto the motherboard in the GB300 Blackwell Ultra, the Vera Rubin platform adopts the JEDEC-standardized SOCAMM2 module—pluggable, hot-swappable, and upgradeable later. You can plug in 96GB today; tomorrow, if the customer needs more, pull them out and replace them with 192GB or even 256GB, as simple as swapping RAM sticks.
NVIDIA specifically emphasized this design at CES 2026: The entire compute tray 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 the initial shipping configuration does not equate to permanent demand disappearance. It's more like a "board the train first, pay the fare later" strategy.
Second, the reason for the capacity reduction is not "not needed," but "not enough available."
SemiAnalysis founder Dylan Patel made a meaningful remark on Twitter: "I love how most people who amplify our reports leave out most of the content of the report. This happens all the time."
Reader comments on Digg regarding this news are also telling: 77.8% of the comments considered the secondary amplification to be sensationalized and out-of-context headlines.
What was left out? The context.
Global LPDDR5X supply is extremely tight in 2026. Micron explicitly stated at the Wolfe conference in late 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, DRAM average selling prices are up over 110% year-over-year, and gross margins have soared to 74%. Samsung and SK Hynix are similarly running at full capacity with strong sales.
Against this backdrop, NVIDIA's problem is not that customers don't want more memory, but that "I can't get enough LPDDR5X chips to fill every slot."
Reducing the default SOCAMM configuration per rack is essentially supply chain management at the engineering level: Rather than delaying the delivery of entire racks due to memory shortages, it's better to ship with a lower configuration first, getting the computing power online faster.
This is not a signal of contracting demand; on the contrary, it's a signal of demand overwhelming supply.
Third, less memory per rack ≠ fewer racks overall.
The market did a simple multiplication: Memory per rack halved → total demand halved. But there's another variable in this equation: shipment volume.
If SOCAMM per rack drops from 55TB to 28TB, NVIDIA, under the same LPDDR5X supply constraints, could actually assemble more racks. The same batch of memory that could only build 100 racks before might now build close to 200.
The total consumption of LPDDR5X hasn't decreased; it's just distributed across more racks. For NVIDIA, this is a pragmatic choice to get Rubin to market faster; for memory manufacturers, total order volume may not decline.
Furthermore, the demand for CPU-side memory in inference scenarios is highly elastic. Not all workloads require 1.5TB of LPDDR5X. Large model training indeed consumes memory, but for many inference tasks, especially agentic AI and long-context reasoning, KV cache can be flexibly scheduled between HBM and LPDDR via NVLink-C2C. For many customers, 768GB of CPU-side memory is already sufficient.
Then 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. The numbers themselves weren't bad: revenue of $22.19 billion, up 48% year-over-year, Non-GAAP EPS of $2.44 beating expectations. But CEO Hock Tan did not raise the full-year AI chip revenue guidance of $100 billion, which the market deemed "not enough." Broadcom's stock plunged 15%, dragging down the entire semiconductor sector.
There was no company-specific negative news from Micron that day. Multiple media outlets including TipRanks, Motley Fool, and 24/7 Wall St. explicitly pointed out that this was a "collateral damage" co-movement decline. As a core AI memory chain stock, Micron is highly tied to AI capital expenditure sentiment; Broadcom's guidance led the market to re-evaluate the expected growth rate of the entire AI chip supply chain.
The dissemination of SemiAnalysis's report on the same day gave traders already looking for selling reasons a perfect narrative: not only is overall AI sentiment weakening, but even the specific numbers for memory demand are shrinking.
A trillion-dollar market cap stock, up 900% over the past year, having just hit an all-time high the previous day. At this level, any negative headline is a catalyst for profit-taking. Panic doesn't need to be correct; it just needs an excuse.
Chaoxiang Interpretation
Three judgments.
First, the SemiAnalysis report itself is accurate, but the market's interpretation of it is wrong. The default SOCAMM configuration for Rubin NVL72 will most likely indeed be lower than the theoretical maximum, determined jointly by supply chain realities and customer demand elasticity. However, there's a gap between "default configuration reduction" and "memory demand contraction"—a gap filled by a pluggable, upgradeable modular architecture and an industry reality where demand far outstrips supply.
Second, Micron's core risk currently lies not in SOCAMM, but in HBM4. SemiAnalysis reported back in February that Micron's share in NVIDIA's Rubin platform HBM4 orders was zero, with SK Hynix taking 70% and Samsung 30%. Although Micron announced HBM4 mass production shipments in March, its market share is projected to be only around 18%. Conversely, Micron's position in the SOCAMM field is very solid: it was the first to launch the 256GB SOCAMM2 module and has been a core partner for NVIDIA's SOCAMM solution for five years. The actual impact of SOCAMM configuration reduction on Micron is far less than the marginalization of its HBM4 share.
Third, the nature of this decline is profit-taking in a trillion-dollar stock after hitting an all-time high, amplified by two independent catalysts. Broadcom provided the sentiment shock; SemiAnalysis provided the narrative ammunition. Combined, they triggered a 10% pullback in a stock that had risen 9-fold over the past 12 months. From a trading perspective, this isn't called "panic"; it's called "normal."
Dylan Patel's tweet was correct: Most people who amplified his report indeed left out the most important part.
The most dangerous thing in semiconductor investing isn't getting the direction wrong; it's reading the headline correctly but calculating the formula incorrectly.
