A source-checked analysis of U.S. data-center delays across grid interconnection, transformers, turbines, and local opposition; why ERCOT lowered risk with 40.3 GW of solar, 22.0 GW of batteries, and 5.1 GW of demand response; and what the bottleneck means for Big Tech, GPUs, HBM, memory, and power equipment stocks.
On the day KOSPI crashed 9.65% and tripped the year's eighth circuit breaker, the triggers were a report that China has begun mass-producing domestic immersion DUV lithography tools and CXMT's 466% listing-day surge. We steelman both the major-negative case and the noise case, then judge across four axes: substance of the information, tense, stock-level channels, and what was already priced. The verdict: a genuine terminal-value repricing for equipment names (right direction, excessive speed), a confirmation of an already-priced post-2028 bear axis for Korean memory, and for today's crash itself a liquidity shock that is roughly 20% information and 80% positioning. Five measurable discriminators and this week's earnings-and-FOMC reads close it out.
The folk version of the comparison, fake demand then versus real demand now, is wrong. End demand was real in 1999 too. What cut Cisco down 89% was not absent demand but sequencing: financing that ran five to ten years ahead of it. The monetization gap ratio, infrastructure spend divided by end AI revenue, has narrowed from 8x to 4.6x and reads like 1996, while OpenAI's $1.4 trillion commitment stack and the arrival of NVIDIA vendor financing read like 1999. In a system where both signals are lit, the bridge that decides rupture is the capital market. We audit the agent-replaces-labor evidence and map the implications for Korean memory.
We break the San Francisco AI Declaration and the $950 billion of Korea-US AI cooperation into contracts versus intent, then calculate what is already in Korean share prices and what is genuinely new. Roughly 20-25% of every AI capex dollar accrues to Korea, and almost all of it is memory. The real weight of this week lies in locking that share through 2030, in Samsung Foundry winning Broadcom as an anchor tenant, and in Korea joining NVIDIA's circular financing loop. That none of the declaration's four pillars claims the model layer is what caps the multiple.
KOSPI surged 3.8% on heavy foreign buying while the meta screener flags APR, Gigavis, and Samsung Electro-Mechanics as top quality re-rating candidates.
Alphabet's print reconfirmed AI demand fundamentals. What remains is rates, and this summer's trigger for rates is oil. Of Brent at $94, $16-20 is war premium rather than physical supply-demand, and that fear component travels through the August CPI and the September FOMC into AI multiples. We decompose the oil price, dissect the three engines behind 4.7% yields, map the depreciation wave rolling toward cloud profit cycles, and add the four channels through which this chain reaches Korean memory, ending with the triple verdict of July 30. One conclusion stands out: selling Samsung Electronics and SK Hynix on rates is a category error.
A current ranking of 19 Korean AI data-center names by direct revenue exposure, foreign and real-money flow, valuation and entry point. LG Electronics, Iljin Electric, Samsung SDS, GST and SK Gas lead the risk-adjusted list.
Alphabet's Q2 sharply lowered the odds of a post-2028 AI demand cliff: Cloud grew 82%, backlog reached $514bn, and existing customers are consuming 50% above their commitments. The same print delivered the first negative quarterly FCF on record, a 2026 CAPEX guide raised to up to $205bn, and Alphabet renting compute from SpaceX at roughly $920M a month because its own capacity ran out. We read the end of the demand debate and the start of the cash debate in one print, translate it into HBM, server DRAM and eSSD, and set up the Microsoft and Amazon scorecards for late July.
FADU's Q2 2026 revenue of KRW 73.16bn and operating profit of KRW 16.27bn are assessed alongside Gen5 commercialization, the Gen6 FC6161, OCP/FDP, QLC, customer concentration, valuation and July 22 flows.