A four-quarter review of Amazon, Microsoft, and Alphabet tests how long AI demand can outgrow expectations, whether $570-580 billion of 2026 capital spending can earn an adequate return, and what it means for HBM, server DRAM, and enterprise SSD demand beyond 2028.
The question that never went away in the AI CAPEX debate was who would actually use all that infrastructure. On NVIDIA's books, non-hyperscale revenue has reached parity with hyperscale and its quarterly growth rate has already flipped ahead. Sovereign AI revenue tripled in a year, Meritz's semiconductor team pushed back on the market's narrow frame with a re-tightening server DRAM market, and Codex added three million users in three days. We cross-verify the demand-side evidence that landed right after Kioxia's limit-down and the Nasdaq slide, and map what it means for the 45/35/20 demand scenarios.
As NVIDIA expands Vera CPU into NVL72 and standalone servers, we verify whether SoCAMM2 capacity reduction is negative for Simmtech. Connecting NVIDIA, TrendForce, SK Hynix, and Simmtech official materials with July 16th stock price, supply-demand, and consensus to establish module-unit-based investment logic and entry conditions.
A weekly synthesis connecting TSMC and ASML earnings, DRAM and NAND contract prices, HBM long-term supply agreements, Chinese memory and Kimi K3, and the sharp sell-off in Korean semiconductor stocks onto a single timeline from July 13–19, 2026. Earnings in 2026–2027 are strong, but the market is already discounting 2028 supply expansion, efficiency improvements, and customer profitability.
We decompose two months of remarks and actions by SK Hynix Chairman Chey Tae-won through the lens of information edge and vested interest. His words strongly support an AI memory volume supercycle, but a roughly KRW 40 trillion new-share issuance, a doubling of capacity and a token-cost-cut goal point to a volume-conversion strategy rather than perpetuating excess margins. We organize that separation between a strengthening company thesis and a weakening margin thesis through a remarks timeline, his repeated mental models, scenarios and buy conditions.
AI memory demand is likely to meet or exceed today's high expectations, but a large upside beat is not yet the base case. We probability-weight demand into base, beat and miss scenarios (45/35/20), lay out six upside variables and six downside risks, then connect the memory-layer supply bottlenecks, the site-by-site expansion maps of Samsung, SK Hynix and Micron, and the substance of Chairman Chey's expansion argument. The conclusion is a shortage thesis through 2027 and re-verification from 2028.