“Memory halving” is just a misinterpretation? The true positives and negatives behind the AI memory stock crash

TL;DR

• Memory configuration downgrades in the Rubin rack system triggered a broad correction in the AI memory sector.
• The market is not truly reassessing AI memory demand, but rather 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 rack system sent the AI memory sector into a downturn.

The report indicated that the memory capacity per rack might decrease from approximately 55TB to about 28TB. Subsequently, Micron fell around 7.7% in a single day, and SK Hynix's opening price dropped over 8% the following day. Adding nuance, the report's author, Dylan Patel, later clarified that many shares only captured the most alarming excerpts, emphasizing it was not a "catastrophic negative" report.

The significant reaction stems from hitting the most sensitive nerve in the AI hardware market recently. Lately, the market has not been trading based on typical memory cycles, but on the expectation that volume production of the Rubin platform would continue to drive demand for HBM and supporting memory, thereby re-elevating memory suppliers' revenue and pricing power. Since this year's GTC, themes like HBM4, SK Hynix's market share, and Micron catching up in AI memory have been recurring market narratives.

However, the phrasing "memory being cut" is too simplistic.

The adjustments disclosed by SemiAnalysis primarily refer to configuration changes in the CPU-side SOCAMM and LPDDR within the Rubin NVL72 rack. Most systems might adopt 96GB modules instead of the higher-capacity 192GB modules, reducing the per-rack memory capacity from the planned ~55TB to ~28TB. This change impacts the system memory value per rack, but does not directly imply a synchronous reduction in GPU-side HBM4 demand.

The key is to dissect exactly which profit pool this adjustment impacts and which expectations the market is currently pricing in.

Why the Collective Plunge in AI Memory Stocks?

The market's decline reflects position adjustment in high-valuation themes upon encountering negative keywords.

What's confirmed is that the market reaction was significant, but the event itself remained 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 figures cited include the per-rack memory capacity dropping from ~55TB to ~28TB and the rack cost decreasing from ~$7.6 million to ~$6.8 million. These numbers should be understood as SemiAnalysis's estimates, not Nvidia's final confirmed BOM.

In recent quarters, the rise of AI memory stocks has been driven by a straightforward narrative: more AI racks lead to greater scarcity of advanced memory, resulting in fatter margins for suppliers.

The simpler the story, the more damaging a negative headline can be. Once "memory capacity halved" appears, the market first revises down the per-rack memory value, seldom distinguishing which type of memory is being adjusted initially.

Micron's reaction is most telling.

It is both a traditional DRAM supplier and a beneficiary of AI server memory upgrades. The elasticity the market previously assigned to it largely came from the repricing of "AI memory as no longer just a cyclical product." If the memory capacity per Rubin rack decreases, capital immediately worries that expectations for Micron's revenue per rack in SOCAMM and LPDDR segments were too high.

SK Hynix also declined, indicating the shock extended beyond a single supplier.

It has a stronger position in HBM, with the market previously abuzz that it had secured a major share of the 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 sector risk appetite rather than each company facing the same fundamental impact.

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

In market parlance, capital didn't fully trade a supply chain analysis; it executed a rapid position reduction in a high-valuation sector upon encountering a negative keyword.

AI Memory Begins to Redefine Profit Pools

The primary downgrade in this instance pertains to CPU-side system memory, not the GPU-adjacent HBM4.

Memory within the Rubin rack cannot be summarized with a single term. The simplest breakdown involves two layers:

The first layer is GPU-side HBM4, serving the accelerator chips themselves;

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

The former determines the speed of data feeding the GPU, while the latter impacts overall system scheduling, maintenance, and performance in certain workloads.

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

It potentially alters the quantity, capacity, and procurement cost of SOCAMM modules per Rubin NVL72 rack. If most systems shift from 192GB to 96GB modules, the per-rack value of high-capacity SOCAMM indeed decreases, pressuring the revenue elasticity of related suppliers.

But GPU-side HBM4 is a different story.

The Rubin platform still centers around the Rubin GPU and Vera CPU, with HBM4 remaining the core memory segment for GPU packaging and computational power release. Current information does not indicate a synchronous reduction in HBM4 capacity or Rubin GPU shipments. Multiple prior forecasts still view HBM as one of the tightest and most pricing-powerful segments in AI servers, with SK Hynix perceived as a primary beneficiary.

Think of an AI rack 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 replaceable system memory of the whole machine. This adjustment mainly pertains to the latter.

For holdings, the distinction is direct: if Micron has greater exposure in the SOCAMM segment, the downward revision in per-rack value impacts its expectations first; SK Hynix's HBM thesis is relatively independent but can still be dragged down by sector sentiment in crowded trades.

Extrapolating the system memory downgrade directly to a collapse in HBM4 demand lacks sufficient evidence.

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

The AI memory market can no longer be covered by a single "all memory is strong" narrative for all suppliers. Micron, SK Hynix, and Samsung Electronics have different exposures to HBM, SOCAMM, traditional DRAM, and NAND, and different memory types within the same rack correspond to different prices, margins, and supply-demand constraints.

Can Cost Reduction Translate into More Rack Shipments?

The optimistic interpretation hinges on cost and delivery cadence.

Calculations by SemiAnalysis suggest the cost of a Rubin NVL72 rack could drop from approximately $7.6 million to around $6.8 million, a reduction of about $800,000.

For cloud giants like Microsoft, Google, Amazon, and Meta, AI racks are not just hardware purchases but involve calculating per-hour compute costs, delivery timelines, and large-scale deployment stability.

If the downgrade allows faster delivery of Rubin systems, the decrease in per-rack value might be offset by a higher number of racks shipped.

The logic is not overly complex. If high-capacity SOCAMM supply is tight, Nvidia opting for more easily deliverable configurations can lower the BOM per rack and reduce the risk of a single component bottlenecking the entire system delivery.

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

The issue is that this step remains speculative for now.

A cost reduction does not automatically equate to increased orders. For the "decline in per-rack value" to be offset by a "rise in total racks," Nvidia needs to deliver more Rubin NVL72 units, and cloud providers need to place additional or accelerated orders.

Current available materials lack public order data, quarterly guidance, or actual shipment figures to confirm this.

To use a simple scenario: if the SOCAMM capacity per rack is nearly halved, the total rack shipments would need to increase significantly for the total Bit demand in this segment to return to original expectations.

Even with a cost reduction of about 10%, one cannot directly conclude that customers will purchase sufficiently more racks. Procurement by large cloud providers is also influenced by power availability, data center construction, GPU supply, advanced packaging, and networking equipment; a single BOM reduction is just one variable.

The situation for HBM is comparatively 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 system delivery is constrained by other bottlenecks, HBM will also be affected by the platform's shipment pace.

The difference is that this specific report did not directly downgrade HBM4 configuration. The market needs to wait for total rack shipment figures, not just focus on the SOCAMM capacity number.

Shipment Data is the True Pricing Anchor

The biggest risk currently is that the market revalues based on profit pool segmentation first, only for subsequent data to fail to support the optimistic narrative.

If Nvidia or the supply chain ultimately confirms a long-term adoption of the lower SOCAMM configuration for the Rubin NVL72, without a significant upward revision in total rack shipments, CPU-side system memory suppliers would face more sustained compression of revenue expectations.

For Micron, the key is not just the overarching "AI memory beneficiary" label, but the revenue breakdown by product.

In upcoming earnings reports and calls, the focus should be on whether management discloses the growth trajectory for AI server-related DRAM, SOCAMM, and HBM, and whether margins are changing due to specifications, pricing, or customer negotiations.

If the company only provides optimistic statements about total demand but cannot explain the impact of the SOCAMM configuration change, the market may continue to apply a discount.

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

If its HBM4 order share, shipment cadence, and pricing remain strong, this correction looks more like sector sentiment volatility. If subsequent total Rubin shipments or HBM delivery schedules are also revised downward, the market would then extend the impact from SOCAMM to the HBM main thesis.

This also represents a typical shift for the AI memory theme as it enters a middle phase.

In the early stages, the market bought the direction: more AI racks built, more advanced memory scarcity.

Now, with representative stocks having accumulated significant gains, capital starts scrutinizing whether each profit pool is actually materializing. A single supply chain detail can trigger 7%-8% daily swings, indicating the sector is already quite crowded, making negative information easier to amplify.

Until actual shipment data and earnings breakdowns are available, it is premature to definitively label this correction as either "all bad news priced in" or "AI demand collapse."

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

The factors most likely to alter the assessment going forward remain Nvidia's confirmation of the final Rubin NVL72 BOM, whether the actual shipment plan for Rubin racks can be upgraded, and the revenue exposure and margin changes for Micron, SK Hynix, and Samsung Electronics in HBM versus SOCAMM/LPDDR.