semi·news
Headlines要闻 / Research研究 / /
Thursday, July 30, 2026 2026年7月30日 星期四

AI Demand Rewires the Silicon Stack AI需求重塑芯片全栈

Cloud operators are widening their accelerator bets while memory, packaging and wafer tools absorb capacity pressure. Public funding is moving toward photonics and specialized compute as supply-chain resilience becomes a design constraint. 云服务商正扩大加速器押注，存储、先进封装与晶圆设备承接产能压力。随着供应链韧性成为设计约束，公共资金也流向硅光子和专用计算。

AI & Accelerators AI与加速器

Microsoft plans AMD Helios and Nvidia Vera Rubin rack-scale deployments 微软计划部署AMD Helios与NVIDIA Vera Rubin机架级系统

Wccftech · 2h ago Wccftech · 2小时前

Microsoft says Azure will be among the first cloud platforms to deploy rack-scale AI infrastructure based on AMD Helios and Nvidia Vera Rubin. The report says Helios may cost roughly 40% more than Vera Rubin, showing that Microsoft is preserving supplier optionality even as system costs rise. The competition is moving from individual GPUs to integrated racks, networking and power delivery. 微软表示，Azure将成为最早部署基于AMD Helios和NVIDIA Vera Rubin的机架级AI基础设施的云平台之一。报道称，Helios成本可能比Vera Rubin高约40%，显示微软即使面对系统成本上升仍保留多供应商选择。竞争正从单颗GPU转向集成机架、网络与供电系统。

Foundry & Manufacturing 晶圆代工与制造

Intel targets advanced packages beyond 24× reticle size Intel瞄准超过24倍光罩尺寸的先进封装

Wccftech · 3h ago Wccftech · 3小时前

Intel says advances in New Mexico remove an encapsulation constraint that had limited very large multi-die systems, with the report pointing to packages beyond 24 times reticle size. Such scaling is central to assembling AI systems from chiplets when a monolithic die is no longer economical or manufacturable. Yield, substrate availability and thermal design remain the practical tests at these dimensions. Intel称其在新墨西哥州的技术进展消除了限制超大型多芯粒系统的一项封装约束，报道指向超过24倍光罩尺寸的封装。当天然单片大芯片不再具备经济性或可制造性时，这种扩展能力是以chiplet构建AI系统的关键。良率、封装基板供给和散热设计仍是这一尺寸下的实际考验。

Memory 存储

Samsung posts record profit on device and memory sales 三星凭借器件与存储销售实现创纪录利润

Nikkei Asia · 31m ago Nikkei Asia · 31分钟前

Samsung Electronics reported an all-time-high profit, with device and memory sales driving the result. It adds evidence that AI infrastructure demand is feeding through to memory suppliers as well as accelerator vendors. Strong earnings give Samsung more room to fund HBM capacity, though execution against established suppliers remains a separate question. 三星电子公布创纪录利润，器件和存储销售是主要驱动力。这进一步表明，AI基础设施需求不仅利好加速器厂商，也正在传导至存储供应商。强劲盈利让三星有更多空间投入HBM产能，但能否赶上既有供应商仍是另一项考验。

Apacer chief warns commodity DRAM supply could fall 70% in 2027 Apacer负责人警告2027年通用DRAM供给或下降70%

Wccftech · 7h ago Wccftech · 7小时前

Apacer CEO C.K. Chang said commodity DRAM supply could fall by as much as 70% in 2027 as HBM consumes more wafer capacity, echoing concerns from SK hynix. The claim is a forecast, not an industry plan, but it captures the allocation tension between AI-focused HBM and mainstream DRAM. If sustained, that trade-off would widen price and availability differences across memory segments. Apacer首席执行官张国森表示，随着HBM占用更多晶圆产能，2027年通用DRAM供给可能最多下降70%，呼应了SK hynix的供给担忧。该说法是预测而非行业量产计划，但它揭示了AI用HBM与主流DRAM之间的产能分配矛盾。若趋势持续，不同存储细分市场的价格和可得性差异将扩大。

Equipment & Materials 设备与材料

Lam Research guidance exceeds analyst expectations by $1 billion Lam Research业绩指引较分析师预期高出10亿美元

techtimes.com · 3h ago techtimes.com · 3小时前

Lam Research issued guidance reported to be $1 billion above analysts' expectations as AI-related chip capacity demand remains strong. For wafer-fab equipment suppliers, guidance is an early signal that customers are converting AI demand into process-tool orders. It must still be read alongside customer capital-spending plans because memory and logic cycles can diverge sharply. 据报道，Lam Research给出的业绩指引比分析师预期高出10亿美元，反映AI相关芯片产能需求依然强劲。对晶圆制造设备供应商而言，指引是客户将AI需求转化为工艺设备订单的早期信号。仍需结合客户资本开支计划解读，因为存储和逻辑的投资周期可能明显分化。

Policy & Geopolitics 政策与地缘政治

US awards GlobalFoundries $300 million for silicon photonics and takes 1% stake 美国向GlobalFoundries提供3亿美元硅光子资金并持有1%股份

The Register · 6h ago The Register · 6小时前

The US government will provide GlobalFoundries $300 million to pursue silicon photonics and take a 1% equity stake, according to the report. Combining grant funding and ownership aligns public support with optical interconnect technology used to move data among AI systems. It extends industrial policy beyond leading-edge logic into links that increasingly constrain large clusters. 据报道，美国政府将向GlobalFoundries提供3亿美元用于推进硅光子技术，并取得其1%的股权。将拨款与持股结合，是把公共支持与AI系统数据互连技术直接绑定的少见做法。此举将产业政策从先进逻辑扩展到日益限制大型集群扩展的光互连。

Aeluma set to receive $30 million in CHIPS funding Aeluma将获得3000万美元CHIPS资金

Compound Semiconductor · 9h ago Compound Semiconductor · 9小时前

Aeluma is set to receive $30 million in CHIPS funding, according to the report. The award points to continued use of US incentives for compound-semiconductor capability rather than only high-volume digital logic. Smaller allocations can be material for specialized suppliers because pilot lines, packaging and qualification costs are large relative to revenue. 据报道，Aeluma将获得3000万美元CHIPS资金。该奖励表明，美国激励措施仍在支持化合物半导体能力，而不仅仅是大规模数字逻辑制造。对专业供应商而言，较小的资金规模也可能意义重大，因为试产线、封装和认证成本相对于收入很高。

Challengers 新兴挑战者

Disaggregated inference pairs AMD Helios with Cerebras 解耦推理将AMD Helios与Cerebras组合

SiliconANGLE · 3h ago SiliconANGLE · 3小时前

A disaggregated-inference design pairs AMD's Helios platform with Cerebras hardware, assigning different parts of the serving stack to different compute systems. It seeks to balance throughput and latency rather than asking one accelerator architecture to optimize every stage. The approach shows how alternative AI-chip suppliers can enter deployments through composable systems instead of wholesale platform replacement. 一种解耦推理设计将AMD的Helios平台与Cerebras硬件结合，把服务栈的不同部分分配给不同计算系统。它试图平衡吞吐量和延迟，而不是要求一种加速器架构优化所有环节。该方法表明，替代型AI芯片供应商可通过可组合系统进入部署，而无需整体替换平台。

Newsletter 邮件订阅

Daily semiconductor briefing. 每日半导体简报。