semi·news
Headlines要闻 / Research研究 /
Saturday, May 23, 2026 2026年5月23日 星期六

US Memory Fabs Light Up as Huawei Bypasses NAND Sanctions 美国存储产能上线,华为绕开NAND管制

Micron's Virginia fab begins quadrupling US DDR4 output the same day TSMC approves another $20B for Arizona, while Huawei unveils a 122TB SSD that sidesteps US 3D-NAND sanctions through a new die-on-board packaging route. DRAM is up nearly 98% YTD, and the supply chain is splitting visibly along jurisdictional lines. 美光弗吉尼亚厂当日投产、美国DDR4产能预计将翻两番;台积电同日批准向亚利桑那再追加200亿美元资本开支;华为则发布122TB SSD,通过新的die-on-board封装路径绕开美国对3D NAND的管制。DRAM年内涨幅已近98%,供应链正沿着管辖边界明显裂解。

Memory 存储器

Micron's Virginia fab begins producing America's most advanced DDR4, quadrupling US output 美光弗吉尼亚厂投产美国最先进DDR4,产能将翻两番

Tom's Hardware · 1d ago Tom's Hardware · 1天前

Micron's expanded Manassas, Virginia fab has begun producing what it calls 'America's most advanced DRAM' — DDR4 destined primarily for automotive and defense, both of which were hit hardest by this year's DDR4 squeeze as the big memory makers shifted to DDR5/HBM. The expansion is on track to quadruple Manassas output. The strategic point isn't bleeding-edge nodes; it's that the US now has trailing-edge memory capacity insulated from the China-Taiwan supply axis. 美光位于弗吉尼亚州马纳萨斯的扩建厂已开始生产其所称的'美国最先进DRAM'——主要为汽车与国防领域供货的DDR4;由于大厂转向DDR5/HBM,这两个行业正承受DDR4严重短缺。扩建后产能预计将翻两番。重点不在尖端节点,而在于美国从此拥有不依赖中国—台湾轴心的成熟节点存储产能。

Hyperscalers reportedly bankrolling SK Hynix's $100B fab blitz 传超大规模厂商出资支持SK海力士1000亿美元扩产

Ad Hoc News · 1d ago Ad Hoc News · 1天前

Multiple outlets are reporting that the next wave of SK Hynix capex — roughly $100B across new HBM and DRAM lines through 2030 — is being underwritten in part by direct hyperscaler commitments, not just on-spec orders. The structure (prepayments, capacity reservations, equity stakes) blurs the line between customer and capital partner and locks in the bottleneck on which the current AI cycle most depends. 多家媒体报道,SK海力士下一轮约1000亿美元的资本开支(涵盖HBM与DRAM新线,延至2030年)将部分由超大规模厂商直接出资承担,而非仅通过订单。预付款、产能预留与股权安排等结构,正模糊客户与资本伙伴的边界,把当下AI周期最关键的瓶颈一并锁定。

Spot DRAM nearly doubles YTD as the AI memory shortage filters past HBM DRAM现货年内近翻倍,AI存储短缺已蔓延至非HBM产品

AOL.com · 1d ago AOL.com · 1天前

Spot DRAM prices are now up roughly 98% since the April launch window, with the wave that started in HBM3E now visible in commodity DDR5 and even DDR4. Nanya's CEO said this week that DRAM margins across all segments are currently above HBM margins — a useful reminder that 'the AI memory shortage' is no longer a leading-edge story. 现货DRAM价格自4月发货窗口以来已上涨约98%,起源于HBM3E的供需紧张已传导至DDR5与DDR4。南亚科技首席执行官本周表示,目前全系列DRAM的利润率高于HBM——这提醒市场:'AI存储短缺'已不再只是尖端产品的故事。

Foundry & Manufacturing 晶圆代工与制造

TSMC approves another $20B for Arizona, following first-year Fab 21 profit 台积电批准对亚利桑那再追加200亿美元,Fab 21首年已盈利

Reuters / MSN · 1d ago Reuters / MSN · 1天前

TSMC's board approved a $20B capex increase for its Arizona campus this week, on top of the existing commitment. The move follows Fab 21's surprise $514M Q1 2026 net profit reported last week — once the US site cleared the economic credibility bar, the appetite for further phases visibly hardened. The new tranche is reportedly aimed at accelerating advanced packaging on-site, not just adding more wafer capacity. 台积电董事会本周批准对亚利桑那园区在现有承诺之外再追加200亿美元资本开支。此前Fab 21出乎意料地在2026年一季度录得5.14亿美元净利;美国厂跨越经济性门槛后,推进后续阶段的意愿明显加强。据报道,新一笔投资重点在于在地推进先进封装,而非单纯扩产晶圆产能。

AI & Accelerators AI与加速器

Enthusiast runs Kimi K2.5 (1T params) on 768GB of cheap Intel Optane — about 4 tokens/sec 玩家以768GB廉价Intel Optane运行Kimi K2.5万亿参数模型,约4 token/秒

Tom's Hardware · 1d ago Tom's Hardware · 1天前

A workstation build using 768GB of decommissioned Intel Optane PMem DIMMs as RAM has been shown running a 1-trillion-parameter Kimi K2.5 model fully locally at roughly 4 tokens/sec with a single consumer GPU. The headline number is slow, but the cost-per-trillion-params is a useful data point as Optane stockpiles get liquidated. It's also a reminder that the byte-addressable persistent-memory tier — for which Intel killed the product line — fits LLM weight residency surprisingly well. 一名玩家在工作站以768GB退役Intel Optane PMem DIMM作为内存,搭配单张消费级GPU,在本地完整运行万亿参数Kimi K2.5,速度约每秒4 token。绝对速度并不快,但在Optane库存被清仓的当下,'每万亿参数的成本'是个有意义的参考。它也提醒业内:Intel已停产的字节寻址持久内存层,恰好与LLM权重常驻的需求高度契合。

Policy & Geopolitics 政策与地缘

Huawei unveils 122TB SSD that bypasses US 3D-NAND sanctions via die-on-board packaging 华为发布122TB SSD,以die-on-board封装绕开美国3D NAND管制

Tom's Hardware · 1d ago Tom's Hardware · 1天前

Huawei has introduced a 122TB enterprise SSD that avoids US-controlled high-layer-count 3D NAND by mounting the NAND dies directly on the PCB (die-on-board) rather than buying packaged chips. The trick gets more silicon area per drive while sidestepping the specific tooling and IP that are export-controlled. It does not solve the underlying density disadvantage, but it ships a real product that the sanctions architecture is supposed to prevent. 华为发布122TB企业级SSD,通过将NAND裸die直接贴装于PCB(die-on-board)的方式,避开美国管制的高层数3D NAND成品芯片。这一手法以更多每盘硅面积换取容量,同时绕开受管制的封装工艺与知识产权。它未解决底层的密度劣势,但实打实地推出了管制体系本应阻断的产品。

Wingtech sues Nexperia for $1.2B, seeks control 闻泰对Nexperia提起12亿美元诉讼,寻求控制权

Tech in Asia · 1d ago Tech in Asia · 1天前

Wingtech, the Chinese parent that bought Nexperia in 2019 and was effectively stripped of control after a Dutch government intervention last year, is suing Nexperia for $1.2B and seeking to reassert ownership. The case is the first major legal test of European-style 'national-security clawback' actions against Chinese-acquired chip assets, and the outcome will shape every similar case in the EU pipeline. 中国母公司闻泰于2019年收购Nexperia,去年遭荷兰政府干预后实质失去控制;闻泰现对Nexperia提起12亿美元诉讼,寻求重新主张所有权。该案是首起对欧式'国家安全追溯'回收中资芯片资产的重大法律检验,其结果将左右欧盟在审中的同类案件。

Nvidia clarifies: $200B CPU forecast includes China — softening last week's 'conceded' line 英伟达澄清:2000亿美元CPU市场预测包含中国——较上周'已让予'表态有所软化

Reuters · 1d ago Reuters · 1天前

Two days after Huang said Nvidia had 'largely conceded' the Chinese AI accelerator market to Huawei, Nvidia clarified to Reuters that its $200B CPU-market forecast for the next cycle 'includes China.' The two statements aren't contradictory — accelerators and CPUs are different addressable markets — but the speed of the clarification and its venue suggest the original line landed harder in Beijing and on the buy-side than the company expected. 黄仁勋两天前称英伟达已'基本让予'中国AI加速器市场,英伟达本周向路透社澄清,其下一周期2000亿美元CPU市场预测'包含中国'。两个表态并不冲突——加速器与CPU是不同的可寻址市场——但澄清的速度与渠道暗示,原话在北京与买方圈引发的反响超出公司预期。

Challengers & New Silicon 新势力与新硅

Lattice Semiconductor acquires AMI to deepen baseboard-controller play in AI datacenters Lattice收购AMI,深化在AI数据中心基板控制层的布局

Yahoo Finance · 1d ago Yahoo Finance · 1天前

Lattice Semiconductor announced an acquisition of AMI (American Megatrends), the dominant firmware/BMC vendor for server motherboards, framed as a way to deepen Lattice's role in AI-datacenter platform control. The bet: as accelerator-heavy servers proliferate, the security/management plane around the GPUs becomes its own real-time problem, and an FPGA vendor that owns the BMC stack can sell into both sides of the board. Price not disclosed in early reporting. Lattice宣布收购AMI(American Megatrends)——服务器主板固件/BMC的主导供应商,意在深化在AI数据中心平台控制层的角色。其逻辑是:随着加速器密集服务器普及,围绕GPU的安全/管理面将成为独立的实时问题,而拥有BMC栈的FPGA厂商可在主板两侧同时切入。早期报道未披露交易金额。