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.
Semiconductors were 38.7% of Korea's H1 exports and China is their largest market. With HBM restricted, exports to China concentrate in commodity DRAM and NAND, precisely the segment CXMT and YMTC target. We decompose each company's China exposure from disclosures and show the real risk is not lost China revenue but the second-order effect of displaced volume flowing to other markets and pressuring global ASP. The conclusion is a structural discount variable for 2028 and beyond rather than a 2026 earnings risk.
An analysis of 43 public Facebook posts and 114,907 Korean characters published by South Korea's policy chief Kim Yong-beom between February 23 and July 17, 2026. The article separates personal policy hypotheses from official programs, maps the recurring chain from AI and semiconductors to power, water, fabs, policy finance, regional capacity and social distribution, and translates it into five investor verification gates.
DeepSeek V4-Pro's output costs 34 times less than GPT-5.6 Sol. Is that sustainable pricing, or the road to winner-take-all as in EVs? Hong Kong disclosures from Zhipu and MiniMax split the answer: paid frontier APIs have started clearing marginal inference cost, but neither recovers fully loaded cost including training. China's cost edge is not Huawei silicon or cheap electricity but the stack of model architecture, batching, caching and cloud utilization, with the power gap worth only about 2% of compute price.
Chinese open models are converging on US frontier performance while being served on domestic Chinese inference infrastructure. This is value-chain redistribution rather than an AI demand collapse. The monopoly value of model APIs and leading-edge GPUs falls, while cheaper inference lifts token volume and can raise the value of memory, storage, networking, power and cloud distribution. This post maps how US AI policy and US-China tension reshape that reading, and whether Chinese model APIs can be adopted by Western enterprises.
We verify the July 16, 2026 bipartisan letter from the House Select Committee on China — examining its legal force, product-level impact, the 2027–2028 oversupply risk, and investment implications for Samsung, SK Hynix, Micron, SanDisk, and Kioxia.