The bottleneck of the AI economy migrated again. After chips and power, the binding constraint of 2025–2026 is high-bandwidth memory. Samsung, Micron, and SK hynix are the new central banks of compute.
Video, audio, and a slide deck — for whichever way you metabolise an idea.
Short explainer video on why memory has displaced silicon as the binding constraint of the AI buildout.
→ WATCH ON YOUTUBELong-form audio episode. The thesis, the players, the geopolitics, the investment implication.
→ LISTEN ON SPOTIFYThe deck on the memory stack hierarchy and why each tier is constrained for a different reason.
→ DOWNLOAD PDFFrom chips (2023) to power (2024) to memory (2025–2026). Each constraint priced. Each repriced. The market keeps moving.
In late 2023, the binding constraint on AI was silicon. NVIDIA could not ship Hopper fast enough; lead times stretched past a year; the value of every H100 in inventory tripled. By the end of 2024, NVIDIA had ramped Blackwell production fast enough to clear that bottleneck. The new constraint became power — a 500 MW data centre in Loudoun County, Virginia, could not be powered before 2032 because the grid interconnection queue had collapsed. The brownfield repower thesis priced.
By the time The NVIDIA Innovator's Dilemma went to press in April 2026, the bottleneck had migrated again. Chips were available. Power was being repowered. The new binding constraint was memory — specifically, high-bandwidth memory, and behind it, the entire DRAM stack.
The fight over Samsung, Micron, and SK hynix capacity in 2025 and 2026 — for HBM3, HBM4, GDDR7, and the long tail of LPDDR — is the next chapter of the same story. The vendors got religion in 2024 and are now the most strategically priced semiconductor companies on earth, more strategically priced than NVIDIA itself for some classes of workload.
This appendix exists to track the memory layer of the argument the book makes about silicon. The two are not separate stories. They are the same story with the bottleneck migrating along the value chain. Three bottlenecks in three years. The interesting question is what gets bottlenecked next — and the answer, the moment you write it down, is unembarrassed by ambiguity: result. The token. The unit of useful thought, delivered on time. That is the subject of the companion essay.
Long-form analysis, investor perspective, and the running press digest on chip-and-memory geopolitics.
Comprehensive analysis of the prerequisites, market dynamics, and forward perspectives of the global DRAM shortage driven by AI infrastructure demand.
→ MOBILE-FRIENDLY VERSIONInvestment perspectives on DRAM and memory markets. When to enter, when to hold, when to rotate. For the operator and the allocator.
→ OPEN THE APPCurated articles from CNN, Reuters, Bloomberg, The Verge, Tom's Hardware on GPU shortages, HBM demand, and semiconductor geopolitics.
→ BROWSE COLLECTIONA test of how the same argument lands across linguistic frames, measured against the assistants that already index the literature.
Diskutiere die Speicherknappheit auf Deutsch.
Обсудить тему на русском.
Deep dive in English.
Discuter en français.
用简体中文讨论 AI 内存短缺.
日本語で議論する.
한국어로 논의하기.
Discutir em português.
If you want the upstream literature on chip geopolitics, capital, and the structural rise of computing as a strategic resource.
