A supply chain report on NVIDIA's Rubin rack has caused a downturn in the AI memory sector.
The report mentioned that the memory capacity per rack may decrease from around 55TB to about 28TB. Subsequently, Micron dropped by approximately 7.7% in a single day, and SK Hynix fell over 8% at the open the following day. What's more nuanced is that the report's author, Dylan Patel, later clarified that many reposts only highlighted the most alarming part, and this was not a "disastrous bearish" report.
The reason this incident sparked such a strong reaction is that it hit the most sensitive spot in the AI hardware market. Over the past period, the market has not been trading in an ordinary memory cycle but rather, after the mass production of the Rubin platform, AI racks will continue to drive HBM and complementary memory demand, raising memory suppliers' revenue and pricing power. Since this year's GTC, the main themes traded in the market have been HBM4, SK Hynix's market share, Micron catching up in AI memory, etc.
But to say that "memory has been cut" is too simplistic.
The adjustment disclosed by SemiAnalysis mainly refers to changes in the SOCAMM and LPDDR configuration on the CPU side in the Rubin NVL72 rack. Most systems may adopt 96GB modules instead of higher-capacity 192GB modules, reducing the memory capacity per rack from the planned 55TB to about 28TB. This change will affect the value of system memory in a single rack, but it cannot be directly inferred that GPU-side HBM4 demand has also been reduced in sync.
What really needs to be clarified is which profit pool this adjustment affects and which expectation the market is currently trading.
Why Did AI Memory Stocks Experience a Collective Sell-Off?
The market sold off due to position reactions of high-level themes encountering negative keywords.
What has been confirmed so far is that the market reaction is significant, but the event itself remains at the level of a supply chain report. SemiAnalysis revealed that to ensure the delivery pace of Rubin NVL72, NVIDIA may adjust the CPU-side SOCAMM configuration. The numbers mentioned in the report include a reduction in the memory capacity per rack from around 55TB to about 28TB, and a decrease in rack cost from around $7.6 million to about $6.8 million. These figures should be understood as per SemiAnalysis's report perspective and are not yet NVIDIA's official final Bill of Materials (BOM) confirmation.
Over the past few quarters, the AI memory stock rally has been driven by a rather smooth narrative: the more AI racks, the scarcer advanced memory, and the fatter profits for suppliers.
The simpler the story, the more powerful the impact of negative headlines. Once a "memory capacity cut" occurs, the market will first downgrade the value of memory per server rack, rarely distinguishing in real-time which type of memory is being cut.
Micron's reaction best illustrates the issue.
It is both a traditional DRAM supplier and a beneficiary of AI server memory upgrades. The market's previous optimism toward Micron, to a large extent, stemmed from the repricing that "AI memory is no longer just a commodity." If Rubin cuts server memory capacity, investors will immediately worry whether Micron's revenue expectations from SOCDIMM and LPDDR segments have been set too high.
SK Hynix also followed suit, indicating that this impact has exceeded a single supplier.
It is stronger in the HBM area, and there were previous reports that it had secured a significant portion of Vera Rubin-related HBM orders. However, when AI memory trading becomes crowded, funds will not wait for all details to be confirmed before taking action. The simultaneous decline in memory stocks reflects a contraction in sector risk appetite, rather than every company facing the same fundamental impact.
Dylan Patel's subsequent clarification actually 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, this means that funds did not conduct a thorough supply chain analysis but instead engaged in rapid deleveraging of a high-flying sector upon encountering negative keywords.
AI Memory Begins to Redefine Profit Pool
The main adjustment this time is on the CPU-side system memory, not the HBM4 next to the GPU.
The memory in the Rubin rack cannot be summarized in just one term. The simplest breakdown is in two layers:
The first layer is the GPU-side HBM4, serving the acceleration chip itself;
The second layer is the CPU-side SOCDIMM and LPDDR, more akin to the system's operational memory.
The former determines the speed at which data is fed to the GPU, while the latter impacts overall system scheduling, maintenance, and some workload performance.
The "55TB to 28TB" mentioned by SemiAnalysis primarily affects the CPU-side system memory.
It may change the number, capacity, and procurement amount of SOCAMM modules per Rubin NVL72 cabinet. If most systems switch from 192GB modules to 96GB modules, the per-unit value of high-capacity SOCAMMs does decrease, putting pressure on the revenue elasticity of related suppliers.
However, the GPU-side HBM4 is a different story.
The Rubin platform still revolves around the Rubin GPU and Vera CPU, with HBM4 remaining a core memory element in GPU packaging and computational power release. Current information does not indicate a synchronous reduction in HBM4 capacity or Rubin GPU shipments. Many had previously forecasted HBM to remain one of the most scarce and price-empowered components in AI servers, with SK Hynix being seen as a major beneficiary by the market.
The AI cabinet can be understood as an extremely expensive high-performance server.
HBM is closer to high-speed memory attached to the GPU, while SOCAMM is closer to system memory that is replaceable within the whole machine. This adjustment mainly focuses on the latter.
For investors, the difference is quite straightforward: if Micron is more exposed in the SOCAMM segment, a reduction in per-unit value will impact its expectations first; SK Hynix's HBM logic is relatively independent, but it will also be dragged down by sector sentiment in a crowded trading environment.
Directly extrapolating a reduction in system memory allocation to an HBM4 demand rupture is not yet supported by evidence.
A more reasonable breakdown is that the CPU-side profit pool does face downward revision pressure, while the GPU-side HBM still depends on Rubin's total shipments and the pace of HBM4 orders.
The AI memory market can no longer be covered by a single narrative of "strong memory" for all suppliers. Micron, SK Hynix, and Samsung Electronics have different exposures in HBM, SOCAMM, traditional DRAM, and NAND, and different memories in the same cabinet correspond to different prices, gross margins, and supply constraints.
Will Cost Reduction Lead to Increased Cabinet Shipments?
An optimistic interpretation comes from cost and delivery cadence.
SemiAnalysis' calculations indicate that the Rubin NVL72 cabinet's cost could decrease from approximately $7.6 million to around $6.8 million, representing a reduction of about $800,000.
For cloud providers like Microsoft, Google, Amazon, and Meta, an AI server rack is not just about buying hardware; it's about calculating the hourly cost of compute power, lead times, and the stability of large-scale deployments.
If reducing the specifications can enable Rubin to be delivered more quickly, the decrease in individual server value may be offset by deploying more racks.
The logic is not complicated. If there is a tight supply of high-capacity SOCAMMs, NVIDIA may choose a more readily available configuration to reduce the bill of materials for each rack and mitigate the risk of a component bottleneck delaying the entire system delivery.
For buyers, if a lower system memory configuration does not significantly impact the core workload, receiving the rack sooner may be more attractive than waiting for the fully-loaded version.
The issue is that this step is currently speculative.
A cost reduction does not automatically translate to an increase in orders. For the "decrease in individual server value" to be offset by the "increase in total racks," NVIDIA needs to deliver more Rubin NVL72 units, and cloud providers need to place additional or early orders.
There is currently no publicly available information on orders, quarterly guidance, or actual shipment data to prove this.
As a simple scenario, if a certain type of SOCAMM sees its capacity nearly halved per rack, then the total rack shipments need to increase significantly to bring the total bit demand for this stage back to the original expectation.
Even with a cost reduction of about 10%, it cannot be directly concluded that customers will buy enough additional racks. Large cloud providers' procurement is also influenced by power, data center construction, GPU supply, advanced packaging, and network equipment; a single BOM reduction is just one variable.
The situation with HBM is relatively more stable, but it is not entirely immune.
If Rubin's total shipments remain strong, HBM4 is still one of the most directly benefited segments; if subsequent evidence shows that overall system deliveries are hampered by other bottlenecks, HBM will also be affected by the platform's shipment pace.
The difference is that this report does not directly lower the HBM4 configuration; what the market is waiting for is the total rack shipment volume, not just focusing on the SOCAMM capacity numbers.
Shipment Data Is the True Pricing Anchor
The biggest risk currently is that the market initially revalues based on profit pools but subsequent data does not support an optimistic interpretation.
If NVIDIA or the supply chain ultimately confirm that Rubin NVL72 will adopt a lower long-term SOCAMM configuration, and if there is no significant upward revision in overall cabinet shipments, the CPU-side memory system suppliers will face more prolonged pressure on revenue expectations.
For Micron, the key is not just the overarching label of "AI memory uplift" but rather the revenue split across different products.
In upcoming financial reports and conference calls, it will be crucial to see if the management discloses the growth trajectory of AI server-related DRAM, SOCAMM, and HBM, as well as whether gross margins have changed due to specifications, pricing, or customer negotiations.
If the company only provides an optimistic depiction of total demand but fails to explain the impact of the SOCAMM configuration adjustment, the market may continue to discount the stock.
For SK Hynix, the focus point leans more towards HBM.
If their HBM4 order share, shipment pace, and pricing remain strong, this current pullback would resemble more of a sector sentiment fluctuation; if subsequent Rubin total shipments or HBM delivery pace also see downward revisions, the market will then see the impact transition from SOCAMM to the HBM mainstream.
This is also a typical shift in the AI memory theme as it reaches the mid-stage.
In the early stages, the market bought into the trend: more AI cabinets were being built, and advanced memory was increasingly in short supply.
Now that the representative targets have seen significant gains, funds are starting to scrutinize whether every profit aspect has truly materialized. A single supply chain detail can trigger a 7%-8% daily swing, indicating that sector trading has become somewhat crowded, making negative information more prone to amplification.
Before actual shipments and financial breakdowns emerge, categorizing this pullback as "the end of bearish news" or "AI demand collapse" is premature.
A more cautious view would be to acknowledge the downward pressure on CPU-side unit value volume while separately pricing HBM4 and SOCAMM.
What can still significantly alter the assessment next is whether NVIDIA confirms the final BOM for Rubin NVL72, whether the Rubin cabinet's actual shipment plan can be increased, and the revenue exposure and margin changes for Micron, SK Hynix, and Samsung Electronics between HBM and SOCAMM/LPDDR.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
