Nvidia reaches into the model ecosystem while HBM4 base dies absorb Samsung's 4nm capacity and TSMC's 3nm output value overtakes 5nm. AI demand is now reshaping ownership, foundry allocation, and memory pricing at the same time. Nvidia将触角伸向模型生态，HBM4基础裸片占用Samsung的4nm产能，而TSMC 3nm产值首次超过5nm。AI需求正同时重塑产业所有权、晶圆代工产能分配与存储价格。
Nvidia is acquiring open-source model hub Hugging Face for $12.9 billion, according to EE Times. The deal would extend Nvidia beyond compute hardware into a distribution and collaboration layer used across the AI software ecosystem, tightening its connection to developers and model deployment. 据EE Times报道，Nvidia将以129亿美元收购开源模型平台Hugging Face。此举将Nvidia的版图从计算硬件延伸到AI软件生态广泛使用的模型分发与协作层，并强化其与开发者及模型部署流程的连接。
Dell projects that AI will account for 75% of datacenter demand by 2030. The forecast signals that servers, accelerators, networking, and power infrastructure are converging around AI workloads, but it remains a long-range vendor estimate rather than booked demand. Dell预计，到2030年AI将占数据中心需求的75%。这一预测表明服务器、加速器、网络与供电基础设施正围绕AI工作负载汇聚，但它仍是厂商的长期估算，并非已经锁定的订单。
TSMC's 3nm quarterly output value is estimated to exceed NT$400 billion in the third quarter, surpassing 5nm for the first time. The crossover indicates that the newest leading-edge node is becoming the foundry's primary value engine, even though output value is not the same as wafer volume or unit shipments. TSMC第三季度3nm产值预计将超过新台币4,000亿元，并首次超越5nm。这一交叉点表明最新先进制程正在成为其主要产值引擎，但产值并不等同于晶圆投片量或芯片出货量。
Samsung has reportedly allocated more than half—and as much as 60%—of its 4nm foundry output to HBM4 base dies as memory demand outruns supply. Because the base die uses a logic process rather than the DRAM process used for the stacked memory layers, HBM4 is now competing directly for leading-edge foundry capacity. 据报道，随着HBM需求超过供给，Samsung已将超过一半、最高达60%的4nm晶圆代工产出分配给HBM4基础裸片。基础裸片采用逻辑制程，而非堆叠存储层所用的DRAM制程，因此HBM4正在直接争夺先进逻辑代工产能。
DRAM contract prices are forecast to rise 13–18% in the third quarter, a slower increase as PC customers limit purchases to essential refreshes. The moderation suggests that AI-driven memory tightness is still lifting prices, but downstream buyers outside the datacenter are increasingly resisting the pass-through. 第三季度DRAM合约价格预计上涨13%至18%，涨幅有所放缓，PC客户正将采购限制在必要的换机需求。价格仍受到AI带来的存储紧张支撑，但数据中心以外的下游买家正更明显地抵制成本转嫁。
Samsung has adopted a photonics test platform that was previously validated by TSMC. Reuse of a qualified test stack can reduce integration friction as foundries and customers move silicon photonics toward higher-volume manufacturing, while also making test compatibility a competitive part of the ecosystem. Samsung已采用一套此前通过TSMC验证的光子测试平台。随着晶圆代工厂与客户推动硅光子进入更大规模制造，复用已验证的测试体系可降低集成摩擦，也使测试兼容性成为生态竞争的一部分。
Malaysia is considering Huawei AI chips for an AI project despite a warning from the United States. The decision tests how far U.S. technology controls can shape procurement in third countries and gives Huawei another potential route to build an overseas accelerator footprint. 尽管受到美国警告，马来西亚仍在考虑为一个AI项目采用Huawei AI芯片。这一选择将检验美国技术管制对第三国采购决策的影响力，也可能为Huawei拓展海外加速器部署提供新路径。
Belgian authorities have detained a dual national on suspicion of semiconductor espionage. The case shows that protection of chip know-how is moving beyond export licensing into criminal enforcement, raising compliance and personnel-security stakes for European semiconductor organizations. 比利时当局已拘留一名双重国籍人士，其涉嫌从事半导体间谍活动。此案显示芯片技术保护正从出口许可延伸至刑事执法，并提高欧洲半导体机构在合规与人员安全方面的风险。
Broadcom reported a sharp rise in third-quarter profit as demand grew for its custom chips. The result reinforces the shift toward hyperscaler-designed accelerators and interconnect silicon, where Broadcom supplies design expertise and networking components alongside the dominant merchant-GPU market. 随着定制芯片需求增长，Broadcom第三季度利润大幅上升。该业绩进一步印证超大规模云厂商正转向自研加速器与互连芯片；Broadcom在主流商用GPU市场之外提供定制设计能力和网络组件。