“Memory Halved” is Just a Misreading? The Real and False Bearish Factors Behind the AI Memory Stock Plunge

TL;DR

• Rubin cabinet system memory configuration downgraded, triggering a broad pullback in AI memory stocks.
• What the market is truly revaluing is not the demand for AI memory, but the profit distribution across different memory segments.
• Related tickers: MU (US), NVDA (US), 000660.KS (Korea), 005930.KS (Korea), SMH (US ETF), SOXX (US ETF)

A supply chain report regarding Nvidia's Rubin cabinet sent AI memory stocks tumbling.

The report mentioned that the memory capacity per cabinet might drop from approximately 55TB to around 28TB. Subsequently, Micron fell about 7.7% in a single day, and SK Hynix's stock opened down over 8% the following day. More subtly, the report's author, Dylan Patel, later clarified that many shared excerpts only highlighted the most alarming parts, stating this was not a "catastrophic negative" report.

The reason this event triggered such a strong reaction is that it struck the most sensitive nerve in the current AI hardware market. Recently, the market has not been trading on the typical memory cycle, but rather on the expectation that after the mass production of the Rubin platform, AI cabinets will continue to drive demand for HBM and supporting memory, thereby re-elevating the revenue and pricing power of memory suppliers. Since this year's GTC, HBM4, SK Hynix's market share, and Micron's efforts to catch up in AI memory have been the main themes of repeated market trading.

But saying "memory was cut" is too simplistic.

The adjustment disclosed by SemiAnalysis mainly refers to configuration changes in the CPU-side SOCAMM and LPDDR within the Rubin NVL72 cabinet. Most systems might use 96GB modules instead of the higher capacity 192GB modules, reducing the memory capacity per cabinet from the planned ~55TB to ~28TB. This change impacts the system memory value per cabinet but cannot be directly extrapolated to conclude a simultaneous downward adjustment in GPU-side HBM4 demand.

What truly needs to be dissected is which profit pool this adjustment affects and which expectations the market is currently pricing in.

Why Did AI Memory Stocks Plunge Collectively?

The market is reacting to position adjustments in a high-valuation theme encountering negative keywords.

What is confirmed so far is that the market reaction has been severe, but the event itself remains at the level of a supply chain report. SemiAnalysis disclosed that Nvidia, to ensure the delivery schedule for the Rubin NVL72, might downgrade the CPU-side SOCAMM configuration. The numbers cited in the report include cabinet memory capacity dropping from ~55TB to ~28TB and cabinet cost decreasing from approximately $7.6 million to about $6.8 million. These figures should be understood as SemiAnalysis's estimates and have not yet been confirmed as the final official BOM (Bill of Materials) by Nvidia.

Over the past few quarters, the rise of AI memory stocks was driven by a straightforward narrative: more AI cabinets mean more demand for advanced memory, leading to fatter profit margins for suppliers.

The simpler the story, the greater the damage from a negative headline. Once "memory capacity halved" appears, the market first revises down the memory value per cabinet, rarely taking the time to distinguish which type of memory was adjusted.

Micron's reaction is the most telling example.

It is both a traditional DRAM supplier and a beneficiary of the memory upgrade cycle for AI servers. A significant portion of the premium the market previously assigned to it came from the re-pricing narrative that "AI memory is no longer just a cyclical commodity." If the system memory capacity per Rubin cabinet decreases, capital immediately worries whether the per-cabinet revenue expectations for Micron in the SOCAMM and LPDDR segments were set too high.

SK Hynix also fell, indicating the impact has spread beyond a single supplier.

It has a stronger position in HBM, and the market had previously heard rumors that it secured the majority share of HBM orders related to Vera Rubin. However, when AI memory trades become crowded, capital doesn't wait for all details to be verified before acting. The synchronized decline in memory stocks reflects a contraction in risk appetite for the sector, rather than each company facing the same fundamental shock.

Dylan Patel's subsequent clarification actually points to this. He stated the report was not intended to create a "catastrophe" narrative and that many people ignored the context.

In market parlance, this means capital wasn't fully trading a supply chain analysis; it was executing rapid position reduction when a high-valuation sector encountered a negative keyword.

AI Memory Starts to Redivide Profit Pools

This adjustment primarily targeted CPU-side system memory, not the HBM4 next to the GPU.

Memory in the Rubin cabinet cannot be summarized with a single term. The simplest breakdown is two layers:

The first layer is GPU-side HBM4, serving the accelerator chip itself;

The second layer is CPU-side SOCAMM and LPDDR, functioning more like the system's main memory.

The former determines the speed at which data is fed to the GPU, while the latter affects overall system scheduling, maintenance, and performance for certain workloads.

The "55TB to 28TB" mentioned by SemiAnalysis primarily pertains to the CPU-side system memory.

This could change the number of SOCAMM modules, their capacity, and the total procurement cost per Rubin NVL72 cabinet. If most systems switch from 192GB modules to 96GB modules, the per-cabinet value of high-capacity SOCAMM indeed decreases, putting pressure on the revenue elasticity of related suppliers.

But GPU-side HBM4 is a different story.

The Rubin platform still revolves around the Rubin GPU and the Vera CPU, with HBM4 remaining the core memory component for GPU packaging and computing power. Current information does not indicate a simultaneous downward revision in HBM4 capacity or Rubin GPU shipments. Previous forecasts still regard HBM as one of the tightest and most pricing-powerful segments in AI servers, with SK Hynix seen as a primary beneficiary.

Think of an AI cabinet as an extremely expensive high-performance server.

HBM is akin to high-speed memory attached directly to the GPU, while SOCAMM is more like the system's replaceable main memory. This adjustment primarily affects the latter.

For investment positions, the distinction is very direct: if Micron has greater exposure to the SOCAMM segment, the reduction in per-cabinet value will impact its expectations first; SK Hynix's HBM logic is relatively independent, but it can still be dragged down by sector sentiment in crowded trades.

There is insufficient evidence to directly extrapolate a downgrade in system memory to a collapse in HBM4 demand.

A more reasonable breakdown is that the CPU-side profit pool faces downward pressure, while GPU-side HBM still depends on the total Rubin shipment volume and the order cadence for HBM4.

The AI memory market can no longer cover all suppliers with a single "memory is strong" narrative. Micron, SK Hynix, and Samsung Electronics have different exposures to HBM, SOCAMM, traditional DRAM, and NAND. Different types of memory within the same cabinet also face different prices, margins, and supply-demand constraints.

Can Cost Reduction Lead to More Cabinet Shipments?

The optimistic interpretation hinges on cost and delivery schedules.

SemiAnalysis's estimates suggest that the cost of a Rubin NVL72 cabinet might fall from approximately $7.6 million to about $6.8 million, a reduction of roughly $800,000.

For cloud providers like Microsoft, Google, Amazon, and Meta, AI cabinets aren't just about buying hardware; it's about calculating the cost per hour of computing power, supply lead times, and large-scale deployment stability.

If the configuration downgrade allows for faster Rubin delivery, the decrease in per-cabinet value could be partially offset by higher shipment volumes.

The logic isn't complex. If high-capacity SOCAMM supply is tight, Nvidia opting for a more easily deliverable configuration can lower the per-cabinet BOM and reduce the risk of a single component bottlenecking the entire system delivery.

For buyers, if a lower system memory configuration doesn't significantly impact core workloads, receiving the cabinets earlier might be more attractive than waiting for the fully-specced version.

The problem is that this step is still speculative.

A cost decrease doesn't automatically translate into increased orders. For the "decline in per-unit value" to be offset by a "rise in total cabinet volume," Nvidia needs to ship more Rubin NVL72 units, and cloud providers need to add or accelerate their procurement.

Current materials do not yet include public order data, quarterly guidance, or actual shipment figures to support this.

To understand with a simple scenario: if the capacity of a certain type of SOCAMM per cabinet is nearly halved, total cabinet shipments would need to increase substantially just to bring the total Bit demand for that segment back to original expectations.

Even a cost reduction of ~10% does not directly imply that customers will buy enough additional cabinets. Procurement by major cloud providers is also constrained by factors like power availability, data center construction, GPU supply, advanced packaging, and networking equipment. A single BOM change is just one variable.

The situation for HBM is relatively more stable, but not entirely immune.

If total Rubin shipments remain strong, HBM4 is still one of the most direct beneficiaries; if it is later proven that overall system delivery is bottlenecked by other issues, HBM will also be affected by the platform's shipment rhythm.

The difference is that this report did not directly downgrade HBM4 configurations; the market now needs to focus on the total cabinet shipment volume, not just the SOCAMM capacity numbers.

Shipment Data is the True Pricing Anchor

The biggest current risk is that the market first revalues based on the profit pool dissection, but subsequent data fails to support the optimistic explanation.

If Nvidia or the supply chain ultimately confirms that the Rubin NVL72 will consistently use lower SOCAMM configurations without a significant upward revision in total cabinet shipments, suppliers of CPU-side system memory will face more persistent revenue expectation compression.

For Micron, the key isn't just the overarching "AI memory beneficiary" label, but the revenue breakdown across different products.

In subsequent earnings reports and calls, one needs to see if management discloses the growth trajectory for AI server-related DRAM, SOCAMM, and HBM, and whether gross margins are changing due to specifications, pricing, or customer negotiations.

If the company only offers optimistic total demand statements without explaining the impact of the SOCAMM configuration adjustment, the market may continue to assign a discount.

For SK Hynix, the verification point leans more towards HBM.

If its HBM4 order share, shipment pace, and pricing remain strong, this pullback looks more like sector sentiment volatility; if subsequent total Rubin shipments or HBM delivery schedules are also revised downwards, the market will then spread the impact from SOCAMM to the core HBM theme.

This is also a typical change during the mid-phase of the AI memory theme.

In the early stage, the market bought the direction: more AI cabinets being built leads to greater shortages of advanced memory.

Now, the representative stocks have already accumulated significant gains, and capital is starting to scrutinize whether each profit pool is being realized. A single supply chain detail triggering a 7%-8% single-day swing indicates the sector is relatively crowded, and negative information is more easily amplified.

Until actual shipment data and earnings breakdowns are available, it is too early to characterize this pullback as either "selling the news" or a "collapse in AI demand."

A more prudent view is to acknowledge the downward pressure on CPU-side per-cabinet value while pricing HBM4 and SOCAMM separately.

The factors most likely to change this assessment going forward remain Nvidia's confirmation of the final Rubin NVL72 BOM, whether the actual Rubin cabinet shipment plans can be raised, and the revenue exposure and gross margin changes for Micron, SK Hynix, and Samsung Electronics across HBM, SOCAMM, and LPDDR.