AI hardware supply chains are being redrawn AI硬件供应链正在重塑

Reported Chinese DUV production, an Arizona-made GB300 milestone, and new accelerator economics point to a more regionalized AI hardware stack. The remaining constraints extend beyond wafers to packaging, memory, and power delivery. 中国国产DUV量产的报道、亚利桑那制造GB300的里程碑和新的加速器经济性，都显示AI硬件栈正趋向区域化。剩余的约束已从晶圆延伸至封装、存储和供电。

AI & Accelerators AI与加速器

AMD positions Helios racks on tokens per dollar AMD以每美元Token产出定位Helios机架

EE Times · 9h ago EE Times · 9小时前

AMD unveiled new data-center GPU systems at Advancing AI and said its Helios racks deliver 30% more tokens per dollar than comparable Nvidia systems. The claim shifts the comparison from peak accelerator specifications to delivered inference economics, where rack integration, networking, and software all matter. AMD在Advancing AI活动上发布新的数据中心GPU系统，并称其Helios机架的每美元Token产出比可比Nvidia系统高30%。这一主张将比较重点从加速器峰值规格转向实际推理经济性，机架集成、网络与软件都将成为关键。

Report says Google's Frozen V2 TPU puts SRAM on-die 报道称Google的Frozen V2 TPU将SRAM集成到芯片上

Wccftech · 8h ago Wccftech · 8小时前

A report cited by Wccftech says Google is considering a Frozen V2 TPU that integrates SRAM on silicon and reduces dependence on TSMC's CoWoS packaging for Gemini workloads. The design is unconfirmed, but it illustrates why hyperscalers may exchange packaging complexity for chip-level memory integration and different cost-power trade-offs. Wccftech援引的报道称，Google正考虑一种Frozen V2 TPU，将SRAM集成到芯片上，以降低Gemini工作负载对TSMC CoWoS封装的依赖。该设计尚未确认，但它说明超大规模云厂商为何可能以芯片级存储集成换取封装复杂度下降及不同的成本—功耗取舍。

Foundry & Manufacturing 晶圆代工与制造

Report: TSMC Arizona fabricates GB300 GPUs, packaging stays in Taiwan 报道称TSMC亚利桑那厂制造GB300 GPU，封装仍在台湾进行

Wccftech · 10h ago Wccftech · 10小时前

Wccftech, citing Commercial Times, reports that TSMC has fabricated some Nvidia GB300 AI GPUs on its 4nm process at Arizona Fab 21. The chips reportedly still travel to Taiwan for advanced packaging, showing that US wafer production does not yet make the accelerator supply chain domestic end to end. Wccftech援引Commercial Times报道称，TSMC已在亚利桑那Fab 21采用4nm工艺制造部分Nvidia GB300 AI GPU。据称这些芯片仍需运往台湾进行先进封装，说明美国晶圆制造尚未使加速器供应链实现端到端本土化。

Memory 存储

China distributor price moves lift RTX 50-series GPU prices 中国渠道调价推高RTX 50系列GPU价格

Tom's Hardware · 6h ago Tom's Hardware · 6小时前

MSI and Colorful raised prices for Nvidia's RTX 50-series cards in China, with the article describing increases of up to 59% across the range and as much as 75% over MSRP at the top end. Broad changes in board pricing can signal pressure in GPU, memory, and channel costs, though consumer-card pricing remains a noisy supply indicator. MSI与七彩虹上调了Nvidia RTX 50系列显卡在中国的价格；报道指全系列涨幅最高达59%，顶级型号较建议零售价最高高出75%。板卡价格的大范围变化可能反映GPU、存储与渠道成本压力，但消费级显卡价格仍是带有噪声的供给指标。

Equipment & Materials 设备与材料

Report: China starts producing domestic immersion DUV tools 报道称中国开始生产国产浸没式DUV设备

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

Tom's Hardware reports that a Shanghai state-backed company has begun mass production of immersion DUV lithography machines and plans first deliveries this year to SMIC, Hua Hong, and CXMT. If confirmed, it would advance substitution of restricted lithography equipment, but it does not imply a domestic EUV capability; process qualification will determine its manufacturing impact. Tom's Hardware报道称，一家上海国资背景企业已开始量产浸没式DUV光刻机，并计划今年向SMIC、华虹和CXMT交付首批设备。若获证实，这将推动受限光刻设备的替代，但不代表已具备国产EUV能力；工艺验证将决定其制造影响。

Policy & Geopolitics 政策与地缘政治

South Korea's $518 billion semiconductor push focuses attention on capacity policy 韩国5180亿美元半导体计划聚焦产能政策

Reuters Technology (Google News) · 4h ago Reuters Technology (Google News) · 4小时前

Foley & Lardner examines South Korea's $518 billion semiconductor investment push and its implications for global technology companies. The scale reinforces how governments are using policy to secure domestic production, memory leadership, and AI infrastructure; the test is whether incentives become usable power, water, talent, and fab capacity. Foley & Lardner分析了韩国5180亿美元半导体投资计划及其对全球科技公司的影响。该计划的规模表明各国正以政策保障本土生产、存储领先地位和AI基础设施；检验标准是激励能否转化为可用的电力、用水、人才与晶圆厂产能。

Challengers 挑战者

Lattice completes $1.6 billion acquisition of AMI Lattice完成16亿美元收购AMI

Reuters Technology (Google News) · 6h ago Reuters Technology (Google News) · 6小时前

Lattice Semiconductor has completed its $1.6 billion acquisition of AMI, according to The Business Journals. The deal broadens Lattice's AI data-center infrastructure platform beyond programmable logic into adjacent system-management capabilities, with integration determining the potential for cross-selling. 据The Business Journals报道，Lattice Semiconductor已完成对AMI的16亿美元收购。该交易扩大了Lattice的AI数据中心基础设施平台，使其从可编程逻辑延伸到相邻的系统管理能力；整合效果将决定交叉销售潜力。