Global semi news — Korea, China, Taiwan, the US, and Japan. Government policy, export controls, capex moves, supply-chain shifts, and macro events. AI-classified and tagged with affected tickers. All headlines link back to the originating publisher.
Original: 【狙击龙虎榜午盘】科技股反弹带动指数再度回升 国产算力与人形机器人逆势突围
A-shares outperformed a weak overnight US tech tape, with domestic-compute supply chain (optical interconnect, PCB, liquid cooling) leading while semiconductor upstream stayed under pressure; article flags widening China-vs-overseas AI token-usage gap as key watch item. Reinforces the domestic-AI-infra decoupling narrative — modestly bearish for foreign AI compute exposure into China (Nvidia) while China compute names take share.
Why it matters: China domestic-compute leadership and token-usage decoupling narrative is a sector-wide overhang for Nvidia's China exposure without a hard catalyst.
Original: 千元手机明年或许买不到了
Chinese media flags DRAM prices up ~700% since 2022 with the Big Three memory makers accused of driving the surge — memory now ~60% of a budget phone's BOM, and sub-RMB1,500 handsets may disappear by 2027 as tightness persists into 2H27, even Apple can't dodge hikes. Bullish read-through for Samsung/SK Hynix/Micron pricing power; the Chinese framing (blaming foreign oligopoly) also strengthens the case for accelerated CXMT domestic-substitution push, a medium-term overhang.
Why it matters: Chinese framing of memory-oligopoly pricing directly reinforces the bull case for Samsung/SK Hynix/Micron while accelerating CN domestic-substitution risk.
Open source articleOriginal: 英伟达推出“算力换分成”新模式 AI初创企业可获得GPU资源支持
Nvidia introduced a revenue-share and credit-support scheme letting AI startups access GPUs without upfront capex, with initial partners planning to deploy up to 210,000 GPUs — 170,000 landing in Batam, Indonesia. Chinese press flags this as Nvidia locking in recurring cloud-service revenue beyond hardware, effectively extending its ecosystem lock-in and expanding sovereign-AI demand outside China.
Why it matters: Expands Nvidia's demand runway and CoreWeave-style neocloud model, positive for GPU/HBM supply chain though not a new export-control shock.
Open source articleOriginal: 韩国政府趁热打铁：斥资312万亿韩元在东南地区打造物理AI中心
Korea's deputy PM announced a KRW 312 trillion private-led investment plan (Samsung, SK, Hanwha) to build a physical-AI, semiconductor and aerospace cluster in the southeastern Yeongnam region, one day after unveiling a KRW 392 trillion semiconductor mega-hub. Chinese coverage frames this as Korea doubling down on national-scale AI/chip capex, reinforcing the medium-term capex cycle for Korean memory, foundry and equipment names.
Why it matters: Sector-wide capex tailwind for Korean semi ecosystem but announcements are aspirational and multi-year, not an immediate catalyst.
Original: 2nm代工再爆单！三星拿下Meta巨额ASIC订单 积压订单冲刺50万亿韩元
Chinese press highlights Samsung Foundry emerging as the go-to 2nm ASIC partner for Tesla, Anthropic and now Meta's MTIA accelerator, pushing its backlog toward KRW 50 trillion. Framed as a direct challenge to TSMC's advanced-node dominance and a validation of Samsung's 2nm yield ramp — clearly bullish for Samsung and its HBM/packaging ecosystem, incrementally bearish for TSMC's ASIC pipeline.
Why it matters: Concrete 2nm ASIC win reshaping the Samsung-vs-TSMC advanced-node race directly touches multiple tracked names.
Open source articleOriginal: 群聯潘健成斥資約8666萬元 今年二度加碼買進自家股票35張
Phison Electronics (8299-TW) CEO Pan Jian-cheng purchased 35,000 shares (35 lots) in June through his holding vehicle Chengshuo Investment at an average of NT$2,476/share, deploying approximately NT$86.7M (~US$2.7M). This follows a February purchase of 25,000 shares at NT$1,960/share (~NT$49M), bringing total H1 2026 insider buying to 60,000 shares at over NT$135M (~US$4.2M). Management's action signals positive operational confidence despite GigaDevice's warning that record-high memory prices could temper downstream NAND demand; industry consensus holds that AI-driven NAND demand and tight supply keep short-term fundamentals intact.
Why it matters: Significant insider accumulation (NT$135M in H1) by the controlling CEO signals management confidence and is a tradeable sentiment indicator, but lacks a hard catalyst such as earnings revision, capex announcement, or named contract.
Original: 秋声 | 袁进辉新公司冲港股IPO，成立不到三年
SiliconFlow, China's largest independent AI-inference token supplier, filed for HKEX listing under Chapter 18C at a $7.7B valuation. It aggregates Nvidia and AMD GPUs alongside domestic Ascend/Moore Threads/Metax silicon into a unified inference stack — a dual signal of continued CN demand for foreign accelerators AND accelerating domestic-substitution readiness. 2025 gross margin was −24%, highlighting price war in CN inference.
Why it matters: China inference demand touches Nvidia/AMD but the domestic-heterogeneous stack signals substitution pressure — sector theme not a single-name catalyst.
Open source articleOriginal: 江波龙：上半年净利润预计为92亿元—110亿元，同比增62204%—74394%
Chinese memory-module maker Longsys guided H1 2026 net profit of RMB 9.2–11.0B, up 62,204%–74,394% YoY, citing tight global wafer supply, downstream demand strength, and newly-signed LTA/MOU wafer supply deals with 'multiple major global wafer OEMs.' Strongly confirms the memory upcycle and validates SK Hynix/Samsung/Micron pricing power — the LTAs also suggest wafer OEMs are locking in module-maker offtake at premium terms.
Why it matters: Downstream Chinese module maker's blowout guide + explicit LTA signings with global wafer OEMs is direct positive confirmation of the memory supercycle for SK Hynix/Samsung/Micron.
Open source articleOriginal: 엔비디아, VR-NVL BMC 지원을 OpenBMC 프로젝트에 기여
NVIDIA is contributing device tree support for its VR-NVL data center platform to the OpenBMC open-source project, enhancing systems management capabilities. This reflects NVIDIA's strategic investment in data center infrastructure and ecosystem development for AI deployments.
Why it matters: NVIDIA's OpenBMC contribution advances AI data center infrastructure and ecosystem development, aligning with broader AI infrastructure trends.
Original: 氪星晚报｜西贝餐饮集团退出小女当家；FF洛杉矶总部人去楼空？公司回应：不实
Deutsche Bank estimates Meta's plan to sell external AI compute/model access could bring 8–11.5GW of AI capacity by end-2027, with 1.2–2.7GW sellable, unlocking $9–30B in incremental 2027 revenue (base case ~$17.5B). The re-rating narrative shifts Meta from 'high capex, unclear returns' to a monetized AI cloud optionality — reinforcing sustained hyperscaler AI capex through 2027.
Why it matters: Reinforces hyperscaler AI capex thesis — supportive for accelerator/HBM/networking supply chain but no new capex figure of its own.
Open source articleJul 10, 2026 close · day-over-day
WOLF
$35
-5.26%