Over the past 72 hours, a wave of sell-side reports hit my terminal: AI-driven demand for HBM will keep the memory market tight until 2028. High-end DRAM shortages, persistent pricing power, a new dawn for Samsung, SK Hynix, and Micron.
I’ve seen this movie before. It ends with a hangover.
Context: Why Now
Memory is the least sexy, most cyclical corner of semiconductors – except when it isn’t. Right now, it’s the hottest. AI training clusters consume HBM3e like a dehydrated runner at a desert oasis. Every H100 GPU needs 80GB of the stuff, and next-gen Blackwell will ask for more. The story is clean: AI demand > supply → shortage → fat margins → stock rerating.
But the crypto world lives on the edge of the same compute stack. Those same GPUs power AI inference and – when the market turns – mining. When I was testing AI-driven trading bots during the 2025–2026 convergence, I watched H100 lead times stretch from 8 weeks to 52. The bottleneck wasn’t just wafer starts – it was HBM packaging. And now the narrative says we’re stuck with that bottleneck for four more years.
Core: The Numbers (and What They Miss)
The core thesis rests on three facts. First, HBM is structurally undersupplied. SK Hynix runs ~50% market share, Samsung ~40%, Micron ~10%. Entry barriers are absurdly high – you need TSV (through-silicon via) packaging, CoWoS integration, and years of customer certification. Second, the price premium is real: HBM ASP is 3–5x that of standard DDR5, pushing gross margins above 50% for those who can deliver. Third, the capex bonanza has begun. Samsung and SK Hynix are spending billions on new HBM-specific fabs, with capital intensity (capex/revenue) hitting 50% or more.
From the front lines of the hype cycle, these are legitimate signals. But here’s what the reports conveniently sidestep: the lead time for a new fab is 2–3 years. Memory is a commodity, not a miracle. When three giants all place the same bet simultaneously, oversupply is the most likely outcome. The 2028 shortage prediction may already be pricing in a supply wave that arrives in 2026–2027.
Contrarian: The Prophecy That Eats Itself
This is the unreported angle. The “shortage until 2028” narrative is not a forecast – it’s a weapon. It drives every major memory maker to accelerate capacity expansion. It gives SK Hynix leverage to sign long-term agreements at sky-high prices. It justifies capex that would otherwise destroy returns. But if all three succeed, the market flips from scarcity to surplus faster than anyone expects. Memory cycles last 2–3 years. This time is different only until it isn’t.
Moreover, the geopolitical heat is missing from the mainstream analysis. China’s export controls on gallium and germanium can choke the supply of advanced materials used in memory production. US sanctions on AI chips to China indirectly reduce HBM demand from that region. Meanwhile, Chinese memory maker CXMT is inching toward HBM-like products – a long shot, but if the gap narrows, the triopoly cracks. The bullish case assumes the status quo holds for five years. That’s a bet I wouldn’t size large.
Takeaway: What to Watch
The rush for HBM will reshape the entire tech hardware landscape – including crypto mining rigs, datacenter costs, and even the price of GPUs that double as validation hardware. But instead of buying the shortage narrative, track the real inflection points:
- CoWoS capacity at TSMC: if it expands faster than HBM fab output, the bottleneck shifts and HBM pricing rolls over.
- HBM4 certification windows: when Samsung or Micron wins a next-gen Nvidia spot, watch for a pre-earnings run.
- Hyperscaler capex guidance: a single “we are pausing AI spend” from Meta or Google could reset the narrative overnight.
Chasing the alpha, one block at a time. The sprint never stops – only the pace. And right now, the pace is about memory.
Turning red candles into green lessons: be early, be skeptical, and never trust a prophecy that benefits the prophet.
