TL;DR

Reduced memory configuration in the Rubin cabinet system triggered a broad pullback across the AI memory sector.

The market is not truly re-evaluating AI memory demand, but rather reassessing profit allocation across different memory segments.

Related tickers: MU (US), NVDA (US), 000660.KS (Korean), 005930.KS (Korean), SMH (US ETF), SOXX (US ETF)

A supply chain report on NVIDIA’s Rubin cabinets caused an initial selloff in the AI memory space.

The report indicated that single-cabinet memory capacity could drop from approximately 55TB to around 28TB. Subsequently, Micron fell nearly 7.7% in a single day, while SK Hynix opened down over 8% the next day. More subtly, Dylan Patel, the report’s author, later clarified that many shared excerpts only captured the most alarming parts—this was not a “catastrophic negative” report.

The magnitude of the market reaction stems from its timing: it struck at the most sensitive point in the current AI hardware cycle. Over recent periods, the market wasn’t trading a typical memory cycle, but rather betting on continued HBM and companion memory demand driven by post-Rubin platform volume ramp-up. Memory suppliers’ revenue and pricing power were expected to be significantly elevated. Since this year’s GTC, themes such as HBM4, SK Hynix’s market share, and Micron’s push into AI memory have repeatedly dominated market narratives.

But the phrase “memory cut” is overly simplistic.

SemiAnalysis’s disclosed adjustments primarily refer to changes in CPU-side SOCAMM and LPDDR configurations within the Rubin NVL72 cabinet. Most systems may now adopt 96GB modules instead of higher-capacity 192GB modules, reducing per-cabinet memory capacity from an estimated 55TB to about 28TB. This shift affects the system memory value per cabinet but does not directly imply a synchronized reduction in GPU-side HBM4 demand.

What truly needs clarification is which profit pool is being impacted—and what expectation the market is currently pricing in.

Why Did AI Memory Stocks Collapse Collectively?

The market reacted to high-valuation themes encountering negative keywords through rapid position unwinding.

One confirmed aspect is that the market response has been severe, yet the event remains at the supply chain report stage. SemiAnalysis revealed that NVIDIA may have reduced CPU-side SOCAMM configurations to ensure timely delivery of Rubin NVL72 cabinets. The report cited figures such as per-cabinet memory capacity dropping from ~55TB to ~28TB, and cabinet cost decreasing from ~$7.6M to ~$6.8M. These figures should be understood as SemiAnalysis’s analytical estimates—not final, official BOM (Bill of Materials) confirmation from NVIDIA.

Over the past few quarters, gains in AI memory stocks were fueled by a straightforward narrative: more AI cabinets mean tighter supply of advanced memory, leading to thicker margins for suppliers.

The simpler the story, the greater the damage from a negative headline. Once “memory capacity halved” appears, markets first revise downward the per-cabinet memory value, with little immediate distinction between which type of memory is being adjusted.

Micron’s reaction best illustrates the issue.

It is both a traditional DRAM supplier and a beneficiary of AI server memory upgrades. Market expectations for its upside were largely based on the re-pricing of “AI memory is no longer just a cyclical product.” If single-cabinet system memory capacity in Rubin drops, investors immediately worry whether Micron’s per-unit revenue expectations in SOCAMM and LPDDR segments were overestimated.

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

While stronger in HBM, the market had previously speculated SK Hynix secured the majority of Vera Rubin-related HBM orders. However, when AI memory trading becomes crowded, capital doesn’t wait for all details to be verified before acting. Synchronized declines across memory stocks reflect a contraction in sector risk appetite—not that every company faces identical fundamental pressure.

Dylan Patel’s subsequent clarification points to this reality: he emphasized the report did not intend to create a “disaster” narrative, and many overlooked the context.

In market terms, capital didn’t fully digest a supply chain analysis—it reacted rapidly to a high-valuation theme encountering a negative keyword.

AI Memory Profit Pools Are Being Recalibrated

The primary adjustment involves CPU-side system memory, not GPU-side HBM4.

Memory inside Rubin cabinets cannot be summarized by a single term. The simplest breakdown consists of two layers:

First layer: GPU-side HBM4, serving the accelerator chip itself;

Second layer: CPU-side SOCAMM and LPDDR, functioning more like system RAM for the entire machine.

The former determines how fast data is fed to the GPU; the latter influences overall system scheduling, maintenance, and partial workload performance.

The “55TB to 28TB” figure mentioned by SemiAnalysis mainly refers to CPU-side system memory.

This change likely affects the number, capacity, and procurement cost of SOCAMM modules per Rubin NVL72 cabinet. If most systems shift from 192GB to 96GB modules, the per-unit value of high-capacity SOCAMM indeed falls, pressuring revenue elasticity for related suppliers.

But GPU-side HBM4 operates on a separate track.

The Rubin platform still centers on the Rubin GPU and Vera CPU. HBM4 remains the core memory component for GPU packaging and compute release. There is currently no evidence suggesting HBM4 capacity or Rubin GPU shipments are being scaled back in parallel. Previous forecasts continue to view HBM as one of the most constrained and price-powerful components in AI servers, with SK Hynix widely seen as the primary beneficiary.

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

HBM is akin to high-speed memory directly adjacent to the GPU; SOCAMM resembles replaceable system memory for the entire machine. This adjustment primarily impacts the latter.

For holdings, the distinction is clear: if Micron has greater exposure to the SOCAMM segment, a per-unit value decline will hit its expectations first. SK Hynix’s HBM logic remains relatively independent, but it too gets dragged down in crowded trading environments.

Directly extrapolating system memory downgrades into HBM4 demand collapse lacks sufficient evidence.

A more rational interpretation is that the CPU-side profit pool faces downward pressure, while GPU-side HBM4 remains contingent on Rubin total shipments and HBM4 order cadence.

The AI memory market can no longer be captured by a single narrative of “all memory is strong.” Micron, SK Hynix, and Samsung Electronics have varying exposures across HBM, SOCAMM, traditional DRAM, and NAND. Different types of memory within the same cabinet also correspond to distinct pricing, gross margins, and supply-demand constraints.

Can Cost Reduction Unlock Greater Cabinet Deployments?

An optimistic interpretation comes from cost and delivery rhythm.

SemiAnalysis estimates that the Rubin NVL72 cabinet cost could drop from ~$7.6M to ~$6.8M—a reduction of about $800K.

For cloud giants like Microsoft, Google, Amazon, and Meta, AI cabinets aren’t just hardware purchases—they’re calculations of hourly compute cost, supply time, and large-scale deployment stability.

If downgrading allows faster Rubin delivery, the loss in per-unit value may be offset by deploying more cabinets.

The logic is straightforward: if high-capacity SOCAMM supply is constrained, NVIDIA opting for more deliverable configurations reduces per-cabinet BOM cost and mitigates risks of a single component delaying full system delivery.

For buyers, if lower system memory configurations don’t significantly impact core workloads, receiving cabinets earlier may be more attractive than waiting for fully loaded versions.

The issue is that this remains speculative at present.

Cost reduction doesn’t automatically translate into increased orders. To offset “lower per-unit value” with “higher total cabinet volume,” NVIDIA must deliver more Rubin NVL72 cabinets, and cloud providers must increase or accelerate procurement.

No public data—such as orders, quarterly guidance, or actual shipment figures—currently supports this thesis.

To illustrate simply: if a certain SOCAMM capacity is nearly halved per cabinet, total cabinet shipments must rise significantly for the aggregate bit demand in this segment to return to prior expectations.

Even with a ~10% cost reduction, it doesn’t follow that customers will buy enough additional cabinets. Large cloud vendors’ procurement is influenced by power availability, data center construction, GPU supply, advanced packaging, and networking equipment—BOM cost reduction is just one variable.

HBM dynamics are relatively more stable, though not immune.

If Rubin total shipments remain strong, HBM4 remains among the most direct beneficiaries. But if later evidence shows overall system delivery is bottlenecked elsewhere, HBM will still be affected by platform ramp cadence.

The key difference: this report did not directly downgrade HBM4 configuration. What the market awaits is total cabinet shipment volume—not just SOCAMM capacity numbers.

Shipment Data Is the True Pricing Anchor

The biggest risk today is that the market reweights profit pools early, only to find later data fails to support the optimistic outlook.

If NVIDIA or the supply chain ultimately confirms Rubin NVL72 will long-term use lower SOCAMM configurations, and total cabinet shipments do not see a meaningful uptick, CPU-side system memory suppliers will face prolonged revenue expectation compression.

For Micron, the key isn’t just the blanket label “beneficiary of AI memory,” but the revenue breakdown across individual products.

Subsequent earnings reports and conference calls must reveal management’s updates on growth momentum for AI server-related DRAM, SOCAMM, and HBM, as well as whether gross margins have shifted due to specifications, pricing, or customer negotiation dynamics.

If the company offers only optimistic overall demand statements without explaining the impact of SOCAMM configuration changes, the market may continue to apply discounts.

For SK Hynix, the validation point leans more toward HBM.

If its HBM4 order share, shipment cadence, and pricing remain strong, this correction is likely just sector sentiment volatility. But if future Rubin total shipments or HBM delivery rhythms also show downward revision, the impact will spread from SOCAMM to the core HBM storyline.

This reflects a typical evolution as the AI memory theme reaches its mid-stage.

Early on, the market bought direction: more AI cabinets mean scarcer advanced memory.

Now, representative names have already posted substantial gains. Capital is beginning to scrutinize whether each profit segment is actually materializing. A single supply chain detail triggering 7%-8% daily swings shows the trading environment is crowded, and negative signals are easily amplified.

Before actual shipment and financial breakdowns emerge, labeling this pullback as “negative news exhausted” or “AI demand collapse” is premature.

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

What will most decisively shift the narrative next is whether NVIDIA confirms the final BOM for Rubin NVL72, whether Rubin cabinet shipment plans are upgraded, and how revenue exposure and gross margin trends evolve for Micron, SK Hynix, and Samsung Electronics in HBM versus SOCAMM/LPDDR.

Author: BlockBeats

Original: Odaily Planet Daily