semi·news
Headlines要闻 / Research研究 / /
Friday, August 21, 2026 2026年8月21日 星期五

AI Demand Pushes Capacity Into New Constraints AI需求将产能推向新的约束

Leading-edge logic, memory research, and custom accelerators keep advancing, but export rules, power-grid reviews, and materials requirements increasingly determine where that capacity can land. The bottleneck is spreading from the chip itself to the infrastructure and policy around it. 先进逻辑制程、存储研发与定制加速器继续推进,但出口管制、电网审查和材料要求正日益决定这些产能能在何处落地。瓶颈正在从芯片本身扩散到其周边的基础设施与政策环境。

AI & Accelerators AI与加速器

Waymo Designs a 5nm Robocar Accelerator Above 1,000 TOPS Waymo设计5nm自动驾驶加速器,算力超过1,000 TOPS

The Register · 5h ago The Register · 5小时前

Waymo has designed a 5nm machine-learning accelerator for its autonomous-driving stack with more than 1,000 TOPS and an emphasis on very low latency. Bringing the chip in-house gives Waymo tighter control over the timing, power, and sensor-processing requirements that general-purpose accelerators do not optimize for, while adding another large systems company to the custom-silicon trend. Waymo为其自动驾驶系统设计了一款5nm机器学习加速器,算力超过1,000 TOPS,并重点压低延迟。自研芯片让Waymo能更紧密地控制通用加速器难以同时优化的时序、功耗和传感器处理需求,也使采用定制硅方案的大型系统公司再添一员。

Foundry & Manufacturing 晶圆代工与制造

TSMC's 1.6nm Production Could Start in Late 2026 台积电1.6nm或于2026年底量产

Android Authority · 4h ago Android Authority · 4小时前

TSMC could begin mass production of 1.6nm-class chips as early as the fourth quarter of 2026, according to a new report. If that schedule holds, it would accelerate the foundry's leading-edge ramp and give high-performance-computing customers an earlier path beyond its 2nm generation; the date remains a reported target rather than confirmed output. 据最新报道,台积电最早可能在2026年第四季度开始量产1.6nm级芯片。若进度兑现,这将加快其先进制程爬坡,并让高性能计算客户更早获得2nm世代之后的工艺选择;目前这一时间点仍是报道中的目标,并非已确认的产出。

Memory 存储

Micron Plans a $10 Billion Research Push as AI Lifts Memory Demand AI推升存储需求,Micron计划投入100亿美元研发

Outlook Business · 8h ago Outlook Business · 8小时前

Micron plans a $10 billion research push as AI systems increase demand for higher-bandwidth and higher-capacity memory. The scale of the commitment shows that the memory race is shifting from a cyclical capacity response toward longer-horizon process and product development, although the candidate report does not specify how quickly the spending will be deployed. 随着AI系统提高对高带宽、大容量存储的需求,Micron计划投入100亿美元推进研发。如此规模的承诺表明,存储竞争正从周期性的扩产反应转向更长期的制程与产品开发,不过候选报道并未说明这笔资金将以多快速度投入。

CXMT Supplies 11.3% of 2026 DRAM Growth, but Its Share Is Forecast to Ease CXMT贡献2026年DRAM增量的11.3%,后续占比或回落

Wccftech · 8h ago Wccftech · 8小时前

CXMT is estimated to account for 11.3% of global DRAM production growth in 2026, while a Hanwha Securities forecast puts its contribution at 6.9% in 2027. The figures suggest China now has meaningful incremental DRAM capacity, but also that expansion alone may not translate into a rapid challenge to the three incumbent suppliers across products and customers. 据估算,CXMT将贡献2026年全球DRAM产量增量的11.3%,而韩华证券预测其2027年的增量贡献占比将降至6.9%。这些数字说明中国已形成不可忽视的DRAM新增产能,但也表明单纯扩产未必能让其迅速在产品与客户层面挑战三大现有供应商。

Equipment & Materials 设备与材料

An $85 Billion TSMC Spending Forecast Lifts ASML 台积电850亿美元支出预测提振ASML

Yahoo Finance UK · 6h ago Yahoo Finance UK · 6小时前

ASML shares rose after a forecast put TSMC spending at $85 billion, reinforcing expectations for sustained leading-edge tool demand. The figure is a market forecast rather than reported TSMC guidance, but it highlights how rapidly an aggressive foundry buildout would flow into EUV and other equipment orders. 一项预测将台积电支出估至850亿美元后,ASML股价上涨,市场对先进制程设备需求持续的预期随之增强。该数字属于市场预测,并非台积电已公布的业绩指引,但它凸显了激进的晶圆厂扩建将如何迅速转化为EUV及其他设备订单。

SandboxAQ Searches for PFAS-Free Semiconductor Materials SandboxAQ寻找无PFAS半导体材料替代方案

Manufacturing Dive · 7h ago Manufacturing Dive · 7小时前

SandboxAQ is working to identify PFAS-free alternatives for semiconductor manufacturing, where fluorinated chemicals are embedded in multiple process steps. Viable substitutes matter because tighter PFAS regulation could create qualification and supply risks for fabs; any replacement will still need to meet unusually strict purity, performance, and process-compatibility requirements. SandboxAQ正尝试为半导体制造寻找无PFAS替代材料,而含氟化学品目前嵌入多道工艺环节。可行替代品之所以重要,是因为更严格的PFAS监管可能给晶圆厂带来认证与供应风险;任何替代方案仍须满足极严苛的纯度、性能和工艺兼容性要求。

Policy & Geopolitics 政策与地缘政治

Nvidia Denies a China-Specific AI Chip Launch by Year-End Nvidia否认将在年底前推出中国特供AI芯片

Reuters · 4h ago Reuters · 4小时前

Nvidia denied a report that it plans to roll out a China-specific AI chip by the end of 2026. The denial leaves its China roadmap unresolved as the company tries to serve a large inference market without crossing U.S. export-control thresholds or running into Beijing's push for domestic alternatives. Nvidia否认了有关其计划在2026年底前推出中国特供AI芯片的报道。该回应意味着其中国市场路线图仍未明朗,公司既要争取庞大的推理市场,又不能越过美国出口管制门槛,还要面对北京推动国产替代的压力。

Texas Pauses 1,800 Data-Center Projects for a Grid Audit 得州暂停1,800个数据中心项目以进行电网审查

Wccftech · 3h ago Wccftech · 3小时前

Texas has reportedly paused roughly 1,800 data-center projects while ERCOT conducts a comprehensive audit of the development pipeline. The action turns electricity access into an immediate gating factor for AI infrastructure, with consequences for the timing of server, accelerator, memory, and networking demand even when chips themselves are available. 据报道,得州已暂停约1,800个数据中心项目,等待ERCOT对项目储备进行全面审查。此举使电力接入成为AI基础设施的直接约束因素;即便芯片供应充足,服务器、加速器、存储和网络设备的需求兑现时间也可能受到影响。

Court Hears That Former Samsung Engineers Took 18nm DRAM Process Recipes to CXMT 法院审理前Samsung工程师向CXMT泄露18nm DRAM工艺配方一案

Tom's Hardware · 9h ago Tom's Hardware · 9小时前

A Korean court heard allegations that former Samsung engineers stole recipes for the company's 18nm-class DRAM process and intended them for CXMT; one former engineer is now imprisoned. The case puts process know-how, rather than only equipment access, at the center of China's memory expansion and could intensify scrutiny of hiring and technology transfer across the regional supply chain. 韩国法院审理的指控称,前Samsung工程师窃取了该公司18nm级DRAM制程配方并拟提供给CXMT,其中一名前工程师现已入狱。此案表明,制程诀窍与设备获取同样是中国存储扩张的核心,并可能加剧区域供应链对人才流动和技术转移的审查。

Challengers & New Silicon 新势力与新硅

KT Unveils an On-Premise NPU Server for Sovereign AI KT发布面向主权AI的本地部署NPU服务器

RCR Wireless News · 8h ago RCR Wireless News · 8小时前

KT unveiled NPU LLM Station, a fully on-premise server aimed at sovereign-AI deployments. The product gives organizations an alternative to cloud-hosted GPU inference when data residency and operational control matter, although the candidate report provides no throughput, power, or customer-deployment figures for comparing the NPU platform with incumbent systems. KT发布了NPU LLM Station,这是一款面向主权AI部署的全本地服务器。当数据驻留和运营控制至关重要时,该产品为机构提供了云端GPU推理之外的选择;不过候选报道没有给出吞吐量、功耗或客户部署数据,尚难与主流系统直接比较。

Newsletter 邮件订阅

Daily semiconductor briefing. 每日半导体简报。