Custom silicon, high-bandwidth memory, cooling, and software are becoming co-equal constraints on AI deployment. Policy proposals are increasingly following the compute wherever it is accessed. 定制芯片、高带宽存储、散热和软件正成为AI部署中同等重要的约束。政策方案也开始追随算力的实际访问路径而延伸。
Google’s expanded relationship with Marvell points to custom silicon work beyond the TPU itself, including memory, networking, storage, and data movement. Those surrounding subsystems can determine how much useful work an AI cluster extracts from its accelerators. Google扩大与Marvell的合作，显示定制硅的范围正从TPU本身延伸到存储、网络、存储设备和数据搬运。这些外围子系统会决定AI集群能从加速器中释放多少实际有效算力。
A benchmarking study examines two-phase versus single-phase direct-to-chip cooling on an NVIDIA B200 server, including warm-water operation. As accelerator densities rise, the cooling choice affects both sustained performance and the electricity and water economics of the data center. 一项基准研究在NVIDIA B200服务器上比较两相和单相直触芯片冷却，并考察温水运行条件。随着加速器功率密度提升，散热方案同时影响持续性能以及数据中心的电力和用水经济性。
AMD has released ROCm 10, its latest open compute and AI software stack, and says it delivers up to 3.3× inference uplift over ROCm 7 in its cited comparisons. The release puts software maturity alongside hardware availability as a competitive variable for AMD’s AI accelerator users. AMD发布了最新的开放计算与AI软件栈ROCm 10，并在其引用的对比中宣称相对ROCm 7最高可实现3.3倍推理提升。该版本表明，对AMD AI加速器用户而言，软件成熟度与硬件供给同样是竞争变量。
SK hynix has broken ground on a $4 billion HBM plant in Indiana, extending advanced-memory production and packaging investment into the United States. The new site adds a geographically distinct supply-chain node for AI memory, a component that increasingly gates accelerator-system deliveries. SK hynix已为其位于印第安纳州、投资40亿美元的HBM工厂破土动工，将先进存储生产与封装投资延伸至美国。该项目为AI存储增加了一个地域上独立的供应链节点，而这一环节正日益限制加速器系统交付。
Qualcomm’s reported HBC initiative is drawing Samsung and SK hynix into a new memory race beyond conventional HBM. The development underscores how AI-memory differentiation is moving toward system-level bandwidth, packaging, and interface choices rather than capacity alone. 报道称，Qualcomm的HBC布局正将Samsung和SK hynix带入传统HBM之外的新一轮存储竞争。这表明AI存储的差异化正从单纯容量转向系统级带宽、封装和接口选择。
TSMC paid NT$36 billion in second-quarter bonuses, a 50.6% year-on-year increase, while revenue rose 36% to NT$1.2 trillion, according to its SEC filings cited in the report. The divergence highlights workforce retention as a rising cost in the advanced-node expansion serving AI demand. 据报道援引的SEC文件，台积电第二季度发放360亿新台币奖金，同比增长50.6%，同期营收增长36%至1.2万亿新台币。这一增速差异凸显出，在服务AI需求的先进制程扩张中，留住人才正成为不断上升的成本。
Lam Research has broken ground on an AI semiconductor laboratory in Oregon. For process-equipment suppliers, closer work on AI-oriented device and process problems can help translate new architectures and materials requirements into manufacturable process flows. Lam Research已为其位于俄勒冈州的AI半导体实验室破土动工。对于制程设备供应商而言，围绕AI导向器件和工艺问题开展更紧密的研究，有助于将新架构与材料需求转化为可制造的工艺流程。
The Trump administration is reportedly drafting an export-control rule aimed at Chinese remote access to advanced AI computing, with an industry consultation possible as early as September. Extending controls from physical chip shipments to remote use would broaden the compliance perimeter for cloud providers and AI infrastructure operators. 报道称，特朗普政府正在起草一项出口管制规则，旨在限制中国远程访问先进AI算力，并可能最早于9月向行业征询意见。若管制从实体芯片出货延伸至远程使用，云服务商和AI基础设施运营商的合规边界将随之扩大。
Nvidia has formed a federal political action committee for employees, according to the report. The move formalizes the company’s policy engagement as export rules, competition scrutiny, and infrastructure incentives increasingly affect the market for AI systems. 据报道，Nvidia已成立面向员工的联邦政治行动委员会。随着出口规则、竞争审查和基础设施激励日益影响AI系统市场，这一举措将公司的政策参与进一步制度化。
The administration is considering additional semiconductor tariffs that could extend to data-center servers, according to the report. Including finished systems would make trade policy relevant not only to component sourcing but also to the location and cost structure of AI infrastructure deployment. 据报道，政府正在考虑新增半导体关税，范围可能延伸至数据中心服务器。若将整机纳入，贸易政策将不仅影响零部件采购，也会影响AI基础设施部署的地点选择和成本结构。