AI memory stocks fall after supply chain report on NVIDIA Rubin

A supply chain report on NVIDIA's Rubin cabinets caused the AI memory sector to decline first.

The report noted that memory capacity per rack may decrease from approximately 55 TB to around 28 TB. Following this, Micron's stock fell about 7.7% in a single day, and SK Hynix dropped over 8% in early trading the next day. More subtly, the report's author, Dylan Patel, later clarified that many shares had only circulated the most sensational excerpts, and this was not a "catastrophic bearish" report.

This event has triggered such a strong reaction because it strikes at the most sensitive point in the AI hardware market. Over the past period, the market has not been trading on ordinary memory cycles, but rather on the expectation that, after mass production of the Rubin platform, AI server racks will continue to drive demand for HBM and associated memory, thereby lifting memory suppliers’ revenues and pricing power. Since this year’s GTC, key market themes have repeatedly centered on HBM4, SK Hynix’s market share, and Micron’s pursuit of AI memory.

But the phrase "memory cut" is too crude.

The adjustments disclosed by SemiAnalysis primarily refer to changes in the CPU-side SOCAMM and LPDDR configurations within the Rubin NVL72 racks. Most systems are likely to use 96GB modules instead of higher-capacity 192GB modules, reducing the memory capacity per rack from a planned ~55TB to ~28TB. This change affects the memory value per rack, but it does not directly imply a corresponding reduction in HBM4 demand on the GPU side.

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

The market decline reflects position adjustments following negative keywords affecting high-position themes.

So far, it has been confirmed that market reactions have been significant, but the event remains at the level of supply chain reports. SemiAnalysis disclosed that NVIDIA may reduce the CPU-side SOCAMM configuration to ensure the delivery schedule of the Rubin NVL72. The report cites figures including a reduction in memory capacity per rack from approximately 55 TB to about 28 TB, and a decrease in rack cost from approximately $7.6 million to about $6.8 million. These figures should be understood as reported by SemiAnalysis and are not yet the official final BOM (bill of materials) confirmation from NVIDIA.

Over the past few quarters, AI memory stocks rose based on a straightforward narrative: the more AI racks deployed, the greater the shortage of advanced memory, and the higher the suppliers' profits.

The simpler the story, the greater the impact of the negative headline. Once “memory capacity halved” appears, the market will first downgrade the value of memory per cabinet, rarely distinguishing immediately which type of memory is being adjusted.

Micron's response best illustrates the point.

It is both a traditional DRAM supplier and a beneficiary of AI server memory upgrades. Much of the market’s previous resilience toward it stemmed from the revaluation that “AI memory is no longer just a cyclical product.” If Rubin’s per-rack memory capacity declines, investors will immediately worry whether Micron’s revenue expectations per rack for SOCAMM and LPDDR have been overestimated.

SK Hynix has also declined, indicating that this impact extends beyond a single supplier.

It is stronger in the HBM space, and there were previously reports that it secured the majority of orders related to Vera Rubin. However, when AI memory trading becomes crowded, capital does not wait for all details to be verified. The simultaneous decline in memory stocks reflects a contraction in sector-wide risk appetite, rather than each company being impacted by the same fundamental factors.

Dylan Patel’s subsequent clarification also points to this. He stated that the report did not intend to create a “doomsday” narrative, and many people overlooked the context.

In market terms, rather than executing a full supply chain analysis, traders quickly reduced positions after encountering negative keywords associated with a high-performing sector.

This time, the system memory on the CPU side was primarily reduced, not the HBM4 next to the GPU.

The memory in the Rubin cabinet cannot be summarized with just one word. The simplest breakdown is two layers:

The first layer is HBM4 on the GPU side, serving the acceleration chip itself;

The second layer consists of the CPU-side SOCAMM and LPDDR, which function 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 some workload performance.

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

What could change is the number, capacity, and procurement value of SOCAMM modules in each Rubin NVL72 cabinet. If most systems shift from 192GB modules to 96GB modules, the per-unit value of high-capacity SOCAMMs will indeed decrease, putting pressure on the revenue elasticity of related suppliers.

However, HBM4 on the GPU side is a different line.

The Rubin platform continues to revolve around the Rubin GPU and Vera CPU, with HBM4 remaining the core memory component for GPU packaging and computational power delivery. Current information does not indicate a simultaneous reduction in HBM4 capacity or Rubin GPU shipments. Previously, multiple parties predicted that HBM would still be regarded as one of the most scarce and highest-pricing-power components in AI servers, with SK Hynsi being viewed by the market as a primary beneficiary.

You can think of an AI cabinet as an extremely expensive, high-performance server.

HBM is more like high-speed memory mounted next to the GPU, while SOCAMM is more like system memory that can be replaced across the entire system. This update primarily focuses on the latter.

For positions, the distinction is straightforward: if Micron has greater exposure in the SOCAMM segment, a reduction in per-unit value will hit its expectations first; SK Hynix’s HBM business is relatively independent, but it will still be dragged down by sector-wide sentiment in a crowded trade.

Reducing system memory and directly extrapolating it as a breakdown in HBM4 demand is not yet sufficiently supported by evidence.

A more reasonable breakdown is that the CPU side’s profit pool indeed faces downward pressure, while the GPU side’s HBM demand still depends on Rubin’s total shipments and the pacing of HBM4 orders.

The AI memory market can no longer be generalized by a single line stating “all memories are strong.” Micron, SK Hynix, and Samsung Electronics have different exposures across HBM, SOCAMM, traditional DRAM, and NAND. Different types of memory within the same server rack also correspond to distinct pricing, gross margins, and supply-demand constraints.

The optimistic outlook stems from cost and delivery timing.

SemiAnalysis's estimates show that the cost of the Rubin NVL72 rack may decrease from approximately $7.6 million to about $6.8 million, a reduction of around $800,000.

For cloud providers like Microsoft, Google, Amazon, and Meta, AI racks are not simply about purchasing hardware—they involve calculating hourly compute costs, lead times, and large-scale deployment stability.

If downsizing enables Rubin to deliver faster, the reduced value per unit may be offset by a greater number of cabinets.

The logic is not complicated. If there is a shortage of high-capacity SOCAMMs, NVIDIA can choose easier-to-deliver configurations, reducing the BOM cost per cabinet and minimizing the risk of a single component delaying overall unit delivery.

For buyers, if a lower memory configuration does not significantly impact core workloads, receiving the cabinet sooner may be more appealing than waiting for the fully configured version.

The issue is that this step is still theoretical.

Lower costs do not automatically translate to increased orders. For the reduced value per unit to be offset by a rise in total cabinet volume, NVIDIA needs to deliver more Rubin NVL72 units, and cloud providers must place additional or accelerated orders.

There is currently no publicly available information on orders, quarterly guidance, or actual shipment data to substantiate this.

To understand this with a simple scenario: if the capacity of a certain type of SOCAMM in a single rack approaches halving, the total number of racks shipped must significantly increase to bring the overall Bit demand for this stage back to the original target.

Even with a cost reduction of approximately 10%, it cannot be directly assumed that customers will purchase enough additional cabinets. Large cloud providers' procurement is also influenced by power availability, data center construction, GPU supply, advanced packaging, and networking equipment; a single BOM reduction is just one of many variables.

HBM's situation is relatively more stable, but not entirely immune.

If Rubin’s overall shipment volume remains strong, HBM4 will still be one of the most directly benefited segments; if subsequent evidence shows that whole-system deliveries are hindered by other bottlenecks, HBM will also be affected by the platform’s shipment pace.

The difference is that this report did not directly reduce the HBM4 configuration; the market is waiting for total cabinet shipment volumes, not just focusing on SOCAMM capacity figures.

The greatest current risk is that the market revalues based on profit pool allocations, but subsequent data fails to support an optimistic interpretation.

If NVIDIA or its supply chain ultimately confirms that the Rubin NVL72 will consistently use a lower SOCAMM configuration, with no significant upward revision in total cabinet shipments, CPU-side system memory suppliers will face a more prolonged compression of revenue expectations.

For Micron, the key is not just the overall label of "AI memory benefits," but the revenue breakdown across different products.

In future earnings reports and conference calls, monitor whether management discloses the growth trajectory of AI server-related DRAM, SOCAMM, and HBM, and whether gross margins have been affected by changes in specifications, pricing, or customer negotiation power.

If the company only provides an optimistic view of total demand but fails to explain the impact of SOCAMM configuration adjustments, the market may continue to apply a discount.

For SK Hynix, the verification focus is more on HBM.

If its HBM4 order share, shipment pace, and pricing remain strong, this pullback is more likely a sector sentiment fluctuation; only if Rubin’s total shipments or HBM delivery cadence are subsequently revised downward will the market extend the impact from SOCAMM to the HBM theme.

This is also a typical change after the AI memory theme reaches its midpoint.

In the early market, buyers bet on the trend: AI server racks were being built in increasing numbers, while advanced memory became increasingly scarce.

The underlying asset has now accumulated significant gains, and capital is beginning to scrutinize whether each profit has been truly realized. A single supply chain detail can trigger a 7%-8% daily swing, indicating that trading in this sector has become overcrowded, making negative information more likely to be amplified.

It’s still too early to label this pullback as “bad news priced in” or “collapse in AI demand” before actual shipments and financial results are released.

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

What will most significantly influence the assessment going forward is whether NVIDIA confirms the final BOM for Rubin NVL72, whether the actual shipment plan for Rubin cabinets can be increased, and the revenue exposure and gross margin changes for Micron, SK Hynix, and Samsung Electronics in HBM and SOCAMM/LPDDR.