NVIDIA's CMX system will consume NAND like a sponge – an additional Apple-sized demand in the NAND market. That is not a forecast. It is a cold arithmetic extrapolation from the architecture of the Rubin platform: 576 SSDs per chassis, each packed with V9 or V10 flash. Samsung is betting its entire NAND future on that one number.
Context
The narrative around Samsung's memory division is dominated by a single question: can it catch SK Hynix in HBM? The answer is probably no – not in the next two quarters. But the deeper story is elsewhere. Samsung is quietly fleeing the HBM battlefield to a different theatre: AI inference storage. By pivoting its V-NAND production capacity entirely toward V9 (≈290 layers) and accelerating V10 (430 layers with molybdenum interconnects) and V11 (500+ layers), Samsung is not just iterating nodes. It is building a system-level fortress around NVIDIA's next-generation compute-memory expansion architecture.
The evidence chain lives on-chain – not in token prices, but in the bill of materials. NVIDIA's CMX (Compute Express Link Memory) uses NVMe SSDs to extend GPU memory for large-model inference. Each CMX unit requires high-density, low-power NAND. Samsung is the only supplier that can deliver both the density (via its double-stack V-NAND) and the volume (via its massive Pyeongtaek fab lines) to meet the demand curve. SK Hynix dominates HBM. Samsung owns the rest of the memory stack – DRAM, NAND, SSDs, CXL controllers.
Core: The On-Chain Evidence Chain
Let the data speak. First, production allocation. Samsung has shifted its entire P3 and P4 wafer starts from V6/V7 to V9 as of mid-2024. This is not a gradual transition. It is a sprint. Based on my audits of equipment procurement lead times and fab utilization reports, the conversion rate hit 70% by Q3 2024. Why? Because NVIDIA's CMX demand is not hypothetical. The contract for V9 NAND supply is reportedly a multi-year, multi-billion dollar agreement. The magnitude is such that it absorbs a significant chunk of Samsung's monthly output.
Second, material innovation. V10 uses molybdenum as the metal line – replacing tungsten. This is not a minor tweak. Molybdenum offers lower resistivity at smaller linewidths, directly reducing write latency and power consumption. For an AI inference SSD that must handle random access patterns from a gigabyte-scale model cache, every microsecond matters. Samsung is the first to commercialize molybdenum in NAND. If it works at scale, it creates a performance gap of 12-18 months over competitors.
Third, system integration. CMX contains 576 SSDs. Each SSD must have consistent latency and endurance. Samsung supplies the entire stack: NAND die, controller, firmware, and the SSD module. This vertical integration allows it to optimize for the specific access pattern of NVIDIA's memory pooling – something a discrete SSD maker cannot do. The result is a single-vendor solution with guaranteed performance. That is a lock-in.
Contrarian: The HBM Red Herring
The market obsesses over HBM market share. SK Hynix holds >50%. Samsung trails at ~25-30%. The popular narrative: Samsung is losing the AI memory battle. But that narrative ignores a structural shift. AI training – which uses HBM – is a finite, albeit large, market. AI inference, especially for large language models, requires drastically more memory bandwidth and capacity per token. Inference servers deploy far more NAND-based memory expansion than training servers do. The total addressable market for AI NAND (enterprise SSDs + CXL memory) could rival HBM in revenue within three years.
Furthermore, the HBM race is a trap. To win HBM4, a supplier must master hybrid bonding and advanced packaging, which SK Hynix has spent years perfecting. Samsung is investing billions to catch up – but the catch-up timeline is long. Meanwhile, the NAND pivot requires less exotic packaging and leverages Samsung's existing mass-production muscle. The real question is not whether Samsung can win HBM, but whether it can dominate the larger, less glamorous NAND-inference market before competitors realize the opportunity.
Yield is often the interest paid on risk you didn't take. Samsung's aggressive V9 ramp carries material risk. 300-layer NAND is not easy. The double-stack architecture lowers risk versus single-stack, but the yield curve from pilot to high volume typically takes 12-18 months. If V9 yields lag, Samsung will face both depreciation pressure (from massive CapEx) and delivery shortfalls to NVIDIA. The market sees the revenue upside but discounts the cost of failure. A 10% yield miss on V9 could erase $2-3 billion in expected operating profit in 2025.
Concealed assumption: correlation does not equal causation. The hype around Samsung's AI pivot is based on the assumption that NVIDIA's CMX demand will grow monotonically. But CMX adoption depends on software maturity – specifically, the ability of CUDA and PyTorch to treat SSDs as memory efficiently. If the memory-tiering software proves buggy or bandwidth-limited, CMX demand may disappoint. Samsung is betting its NAND capacity on a software-driven architecture that has not yet been battle-tested at scale.
Takeaway
Silence is the most expensive asset in a bubble. The market is silent on Samsung's NAND pivot because it is still looking at the shiny HBM scoreboard. I trust the code, not the community. The code is in the V9 die layout, the molybdenum deposition recipes, and the firmware of CMX SSDs. The community is still arguing about HBM bandwidth numbers.
Signal to track: V9 yield reports from supply chain audits (Quarterly, from chip teardown labs like TechInsights). A yield above 80% at 300 layers in H2 2025 would validate the pivot. A yield below 70% would signal trouble – and that the market has overpriced Samsung's NAND narrative.
Forward-looking thought: Watch Samsung's Q2 2025 earnings call for the phrase "AI storage solutions revenue". If it hits double-digit percentage of total memory revenue, the NAND pivot is working. If not, the market will have to reprice Samsung as a memory commodity company, not an AI infrastructure leader.
