Does 55TB become 28TB? The rumors and panic behind Rubin's memory being cut in half

Writing: Trend Research

Early morning on June 4th, the most influential independent research institution in the semiconductor industry, SemiAnalysis, released a morning report.

The core message is just one sentence: The SOCAMM DRAM capacity per rack in NVIDIA's Vera Rubin NVL72 may have dropped from the previously expected approximately 55TB to about 28TB. Most Rubin systems will use 96GB SOCAMM modules, rather than the 192GB modules that the market had widely anticipated.

Once the news spread, the market's reaction was simple and brutal: memory demand halved, bearish for Micron. MU temporarily plummeted over 10% during trading, from the previous day's all-time high of $1,089 to $971, evaporating over $100 billion in market value in a single day.

The panic is real, but the question is, is the direction of the panic correct?

Let's do the math clearly first

Vera Rubin NVL72 is NVIDIA's next-generation flagship AI supercomputing rack. Each rack contains 72 Rubin GPUs and 36 Vera CPUs. The GPUs use HBM4, each with 288GB, totaling about 20.7TB per rack, which remains unchanged. The change is on the CPU side.

Each Vera CPU has 8 SOCAMM slots, each capable of holding 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 configuration with 8 modules of 192GB each fully populated. With 36 CPUs, that totals 54TB.

SemiAnalysis's report states that the actual shipped configurations are unlikely to be fully populated. Most systems will use 96GB modules, so 8×96GB=768GB per CPU, and 36 CPUs total about 28TB.

From 55TB down to 28TB, the capacity shrinks by nearly half, and headlines call it "memory demand halved."

But this calculation misses a key variable.

The logical flaw in the panic

First, SOCAMM is a plug-in design, not a soldered one.

This is the most overlooked technical detail in the entire story. Unlike LPDDR soldered onto the motherboard in GB300 Blackwell Ultra, Vera Rubin platforms use JEDEC-standardized SOCAMM2 modules, which are pluggable, hot-swappable, and upgradable. Today, with 96GB modules, if customer needs increase tomorrow, they can swap out for 192GB or even 256GB modules, just like changing memory sticks.

NVIDIA emphasized this design at CES 2026: the entire compute tray assembly time was reduced from 2 hours to 5 minutes. Modular, maintainable, upgradable—this is one of the biggest architectural evolutions of Vera Rubin compared to Blackwell.

Lowering initial shipment configurations does not mean permanent demand disappears. It’s more like a "buy now, upgrade later" strategy.

Second, the reason for reducing capacity is not "not needed," but "not enough."

SemiAnalysis founder Dylan Patel tweeted a meaningful comment: "I like one thing—the people who repost our report mostly miss most of the content. This kind of thing happens often."

Reader comments on Digg about this news also reveal the background: 77.8% believe the secondary dissemination is cherry-picking headlines.

What was missed? The background.

In 2026, global LPDDR5X supply is extremely tight. Micron explicitly stated at the Wolfe conference at the end of May that memory demand significantly exceeds supply capacity, and this situation is expected to continue beyond 2026. Micron's entire HBM capacity for FY2026 is sold out, with DRAM prices up over 110% year-over-year, gross margins soaring to 74%. Samsung and SK Hynix are also operating at full capacity and sales.

Against this backdrop, NVIDIA's problem isn't that customers don't want more memory, but that "they can't get enough LPDDR5X chips to fill every slot."

Reducing the default SOCAMM configuration per rack is essentially an engineering supply chain management decision: rather than delaying entire rack deliveries due to memory shortages, it’s better to ship with lower configurations so compute power can go online faster.

This is not a signal of demand shrinking; quite the opposite, it’s a sign that demand is overwhelming supply.

Third, fewer memory modules ≠ fewer racks.

The market made a simple multiplication: halving memory per rack → halving total demand. But there's another variable: shipment volume.

If SOCAMM per rack drops from 55TB to 28TB, under the same LPDDR5X supply constraints, NVIDIA can assemble more racks. Previously, a batch of memory was enough for 100 racks; now, nearly 200 racks.

Total LPDDR5X consumption hasn't decreased; it’s just redistributed across more racks. For NVIDIA, this is a pragmatic choice to bring Rubin to market faster; for memory manufacturers, total order volume may not decline.

Moreover, the inference scenario has high elasticity in CPU-side memory demand. Not all workloads require 1.5TB of LPDDR5X. Large model training indeed consumes a lot of 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 fall 10%?

Because SemiAnalysis is only the second straw that broke the camel’s back.

The first was Broadcom. Before the US stock market opened on June 4, Broadcom released its Q2 earnings report. The numbers were decent: revenue of $22.19 billion, up 48% year-over-year, and Non-GAAP EPS of $2.44, beating expectations. But CEO Hock Tan did not raise the full-year AI chip revenue guidance to $100 billion, and the market felt it was "not enough." Broadcom's stock plunged 15%, dragging down the entire semiconductor sector.

On that same day, Micron had no company-specific negative news. Media outlets like TipRanks, Motley Fool, 24/7 Wall St. clearly pointed out this was a "contagion" decline. As a core player in the AI memory chain and highly tied to AI capex sentiment, Broadcom’s guidance caused the market to reassess the expected growth rate of the entire AI chip industry.

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

A stock valued at over a trillion dollars, which had risen 900% in the past year and just hit a record high the day before, now faces a negative headline as a profit-taking catalyst. Panic doesn’t need to be correct, only an excuse.

Trend Interpretation

Three judgments.

First, SemiAnalysis’s report itself is accurate, but the market’s interpretation is wrong. The default SOCAMM configuration in Rubin NVL72 likely does fall below the theoretical maximum, which is jointly determined by supply chain realities and customer demand elasticity. But "default configuration lowered" and "memory demand shrank" are separated by a modular, pluggable architecture and an industry where demand far exceeds supply.

Second, Micron’s current core risk isn’t SOCAMM but HBM4. SemiAnalysis already reported in February that Micron’s share of HBM4 orders for the Rubin platform is zero, with SK Hynix taking 70% and Samsung 30%. Although Micron announced mass production of HBM4 in March, market share is expected to be only 18%. Conversely, Micron’s position in SOCAMM is very solid: it was the first to launch 256GB SOCAMM2 and has been NVIDIA’s key partner for SOCAMM solutions over the past five years. The reduction in SOCAMM configuration has a much smaller actual impact on Micron than the marginalization of HBM4 share.

Third, the nature of this decline is profit-taking after a trillion-dollar stock hit a new high, amplified by two independent catalysts. Broadcom provided the emotional shock, SemiAnalysis supplied the narrative ammunition. The combination caused a stock that had risen 9-fold over the past 12 months to pull back 10%. From a trading perspective, this isn’t "panic," it’s "normal."

Dylan Patel’s tweet is correct: most people reposting his report indeed missed the most important part of it.

The most dangerous thing in semiconductor investing isn’t misjudging the direction but seeing the right headline and getting the formula wrong.