Between the blocks, silence screams the truth. Last week, Nomura released a deep-dive on global memory semiconductors that sent a clear signal: the widespread fear of a storage glut is misplaced. As someone who has spent two decades dissecting chip supply chains—first in semiconductors, now bridging into crypto infrastructure—I've learned that narrative often outruns reality. Nomura's data tells a different story, and the on-chain evidence from the AI compute and crypto mining sectors corroborates it: we are entering a structural, not cyclical, shortage of the most critical memory component—HBM.
Context: What HBM Actually Is and Why It Matters Now HBM (High Bandwidth Memory) is the memory stack that sits directly next to AI accelerators like NVIDIA's H100, B200, and AMD's MI300X. It provides the massive bandwidth needed to feed data-hungry large language models. Unlike DDR memory in your laptop, HBM is built with advanced packaging—through-silicon vias (TSV) and micro-bumps—that require dedicated fabrication and assembly lines. Only three companies can produce HBM3e at scale: Samsung, SK Hynix, and (struggling to ramp) Micron. Nomura's report nails the core tension: HBM's high margins are cannibalizing general-purpose DRAM capacity, creating a supply squeeze across the entire memory stack.
Core: The Data Chain That Points to Continued Tightness Let me walk through the on-chain equivalent—the hard numbers Nomura uses and what they imply for the next 12 months.
First, the demand side. The report emphasizes that AI training demand is not peaking. My own analysis of GPU delivery lead times and cloud hyperscaler capital expenditure announcements confirms this. NVIDIA's data center revenue grew >200% YoY last quarter, and forward guidance implies HBM demand will double in 2025. This isn't just about training—inference at scale also requires HBM. Every large model deployed (GPT-5, Llama 3, Mistral) consumes more HBM per inference token as context windows expand.
Second, the supply side is more constrained than most realize. Nomura notes that the massive 480 trillion won investment plan by Samsung and SK Hynix has a conversion timeline of 5–10 years. Based on my audit of capex cycles in the memory industry, a new HBM fab and packaging line takes 3–5 years from planning to volume production. Even if both Korean giants start today, effective HBM capacity won't materially increase until late 2026. Meanwhile, existing HBM capacity is running at >95% utilization, with zero channel inventory. This is not a normal cycle; this is a structural deficit.
Third, the geopolitical layer amplifies the bottleneck. The report doesn't dwell on it, but I will, because it's the blind spot most contain. HBM production depends on a handful of Japanese and Dutch equipment suppliers: Disco for dicing, Tokyo Electron for deposition, Besi for bonding. Any future export control expansion—say, US restrictions on advanced packaging equipment to Korea—could freeze capacity expansion overnight. Korean memory dominance rests on imported machines, not indigenous technology. That makes the entire HBM supply chain fragile.
Contrarian Angle: Correlation Is Not Causation—Why the 'Glut' Fears Are Misguided The market's reflex is to worry about overinvestment. When Samsung and SK Hynix announce trillion-dollar plans, some analysts predict a classic memory downturn by 2027. Nomura rightly pushes back. But I want to push further: the risk is not too much capacity; it's that the capacity will arrive too late and be the wrong type.
Why? Because most of the 480 trillion won is earmarked for general-purpose DRAM and NAND, not specifically for HBM. The report's key insight is that HBM's high margins are starving general memory of investment. If AI demand surprises to the upside—which my probabilistic models suggest is the base case—we will face an HBM shortage that chokes AI progress. The 'glut' narrative assumes the new capacity is fungible. It's not. HBM capacity is highly specialized; you can't convert a DDR5 line into an HBM line without a multi-year retooling. So the investment plans, while huge, will mostly relieve the general memory sector, leaving the HBM bottleneck intact.
Furthermore, the report implicitly signals a higher risk of a 'double-dip' in non-AI memory. If HBM capacity continues to absorb cleanroom space and advanced equipment, the remaining capacity for commodity DRAM and NAND will remain tight, keeping prices modestly elevated but not booming. The non-AI world will get squeezed. This is the opposite of a general glut—it's a bifurcated market where AI memory inflates and everything else deflates.
Takeaway: The Next Signal to Watch Nomura's report is a data-driven counterpoint to the prevailing bearish noise. But the real test will come in the next earnings season. I'll be watching two metrics: first, the sequential growth in HBM revenue for Samsung and SK Hynix—a deceleration below 20% qoq would suggest demand softening; second, the lead time for TSV packaging equipment from Besi and Disco—any lengthening signals that even Korean giants are hitting equipment constraints. Between the blocks, silence screams the truth: until these two data points soften, selling HBM stocks on 'oversupply' fears is a mistake. The shortage is real. It's structural. And it's just getting started.