Capital Chases AI's Next Bottlenecks 资本追逐AI的下一道瓶颈

Nvidia pushes into power infrastructure and possible accelerator M&A as memory, optical I/O and manufacturing capacity become strategic constraints. Capital is flowing toward the physical stack, from higher TSMC spending to YMTC's $4.9 billion IPO target. 随着存储、光学I/O和制造产能成为战略约束，Nvidia正把触角伸向电力基础设施和潜在的加速器并购。资本继续涌向AI的实体技术栈，从台积电提高资本开支到长江存储母公司寻求规模达49亿美元的IPO。

AI & Accelerators AI与加速器

Nvidia's Cloverleaf Deal Targets the Data-Center Power Constraint Nvidia借Cloverleaf交易瞄准数据中心电力约束

The Register · 2h ago The Register · 2小时前

The Register reports that Nvidia's Cloverleaf deal extends the chipmaker's spending into the infrastructure needed to deploy its next generation of accelerators. The move matters because power and data-center construction, rather than GPU supply alone, increasingly determine how quickly new AI capacity can come online. The Register报道称，Nvidia通过Cloverleaf交易把资金投入延伸到部署下一代加速器所需的基础设施。此举的重要性在于，决定新增AI算力上线速度的因素正从GPU供给本身扩展到电力和数据中心建设。

Nvidia May Reduce Memory on Some AI Chips; AMD Plays Down the Risk Nvidia部分AI芯片或缩减存储配置，AMD淡化影响

Yahoo Finance · 4h ago Yahoo Finance · 4小时前

A report says Nvidia may ship AI chips with less memory, while AMD says it is not concerned about the prospect. Any reduction would test how customers trade model capacity and bandwidth against price and availability, and could alter expected HBM content per accelerator. 报道称，Nvidia可能推出存储容量更低的AI芯片，而AMD表示并不担忧这一动向。若配置确实下调，客户将在模型容量、带宽、价格与供货之间重新权衡，也可能改变单颗加速器的HBM搭载预期。

Alibaba's RISC-V Chip Reportedly Runs a 27B Model Without a GPU 阿里巴巴RISC-V芯片据称无需GPU即可运行27B模型

Intelligent Living · 4h ago Intelligent Living · 4小时前

Alibaba's RISC-V processor reportedly ran a 27-billion-parameter AI model without relying on a GPU. The claim points to a broader search for domestically controllable inference platforms, although the candidate report provides no independent performance, power or cost comparison. 报道称，阿里巴巴的RISC-V处理器无需GPU即可运行一个270亿参数的AI模型。这一结果反映了对自主可控推理平台的持续探索，但候选报道没有提供独立的性能、功耗或成本对比。

Foundry & Manufacturing 晶圆代工与制造

China's Share of Mature-Node Chips Rises, CRS Reports 美国国会研究处称中国成熟制程芯片份额上升

Legis1 · 8h ago Legis1 · 8小时前

A report citing the U.S. Congressional Research Service says China's share of mature-node chip production is increasing. The shift matters because legacy-node capacity underpins automotive, industrial and power semiconductors, making it a supply-chain and policy issue even without leading-edge technology. 援引美国国会研究处的报道称，中国在成熟制程芯片生产中的份额正在上升。成熟制程支撑汽车、工业和功率半导体，因此即便不涉及先进制程，这一变化仍会影响供应链与政策判断。

Memory 存储

Micron's Memory Pipeline Is Reportedly Sold Out Through 2028 Micron存储产品据称已售罄至2028年

Ad-hoc-news.de · 5h ago Ad-hoc-news.de · 5小时前

Ad-hoc-news reports that Micron's memory pipeline is sold out through 2028. If that order coverage holds, it signals unusually long customer commitments and gives Micron visibility through several product cycles, though the candidate item does not break out HBM from conventional DRAM or NAND. Ad-hoc-news报道称，Micron的存储产品产能已售罄至2028年。若订单覆盖确实如此，意味着客户承诺周期异常延长，并让Micron获得跨越多个产品周期的可见度；不过候选信息未区分HBM、传统DRAM与NAND。

SK hynix Maps Co-Packaged Optics to 100 Tb/s Nodes SK hynix规划面向100 Tb/s节点的共封装光学路线

StorageReview · 8h ago StorageReview · 8小时前

SK hynix has outlined a co-packaged-optics roadmap targeting 100 Tb/s nodes, energy below 1 pJ/bit and latency under 10 ns. Those targets address the interconnect wall around memory-rich AI systems, where moving data can consume more energy and time than arithmetic. SK hynix公布共封装光学路线图，目标包括100 Tb/s节点、低于1 pJ/bit的能耗和低于10 ns的延迟。这些指标直指高存储带宽AI系统的互连瓶颈，因为数据搬运消耗的能量与时间可能超过计算本身。

Equipment & Materials 设备与材料

ASML's Order Book Is Reportedly Full Through 2028 ASML订单据称已排满至2028年

Ad-hoc-news.de · 9h ago Ad-hoc-news.de · 9小时前

Ad-hoc-news reports that ASML's order book is full through 2028 even as large investors remain split on the stock. A multi-year backlog would preserve demand visibility for lithography tools, but it also raises the execution bar for shipment timing, service capacity and customer fab readiness. Ad-hoc-news报道称，尽管大型投资者对ASML股票看法分化，其订单已排满至2028年。多年积压订单提高了光刻设备需求的可见度，同时也对交付节奏、服务能力和客户晶圆厂准备度提出更高要求。

Policy & Geopolitics 政策与地缘政治

U.S. Lawmakers Question the UAE Export-Control Decision 美国议员质询对阿联酋的出口管制决定

Quiver Quantitative · 4h ago Quiver Quantitative · 4小时前

Representatives Gregory Meeks and Sydney Kamlager-Dove are seeking answers about a U.S. export-control decision involving the UAE. Their intervention keeps scrutiny on how Washington balances AI-chip access for partners against diversion and national-security risks. 美国众议员Gregory Meeks和Sydney Kamlager-Dove正要求政府说明一项涉及阿联酋的出口管制决定。此举使华盛顿如何在伙伴获得AI芯片与防范转运和国家安全风险之间取舍，再次受到审视。

Earnings & Markets 财报与市场

TSMC Raises 2026 Capital Spending While Increasing Dividends 台积电提高2026年资本开支并上调股息

Yahoo Finance · 9h ago Yahoo Finance · 9小时前

TSMC is committing to higher 2026 capital spending while also raising dividend payouts, according to Yahoo Finance. Funding both expansion and shareholder returns signals confidence in cash generation, while suppliers will watch how quickly the larger budget converts into tool orders and installed capacity. 据Yahoo Finance报道，台积电承诺提高2026年资本开支，同时增加股息派发。扩产与股东回报并举显示其对现金创造能力的信心，供应商则会关注新增预算多快转化为设备订单和实际产能。

YMTC's Parent Targets a $4.9 Billion Shanghai IPO 长江存储母公司拟在上海IPO募资49亿美元

Reuters · 10h ago Reuters · 10小时前

The parent of Chinese NAND producer YMTC is targeting a $4.9 billion Shanghai initial public offering, Reuters reports. The financing would give China's flash-memory champion a large domestic capital pool for expansion while international technology restrictions continue to shape its equipment options. 据Reuters报道，中国NAND厂商长江存储的母公司计划在上海IPO，目标募资49亿美元。这笔融资将为中国闪存龙头提供大规模境内扩张资金，而国际技术限制仍在影响其设备选择。

Challengers & New Silicon 新势力与新硅

Nvidia Enters Early Talks With $2.3 Billion Korean AI-Chip Startup Rebellions Nvidia与估值23亿美元的韩国AI芯片初创公司Rebellions展开初步洽谈

BigGo News · 6h ago BigGo News · 6小时前

Nvidia is reportedly in early talks with Korean AI-accelerator startup Rebellions, which the report values at $2.3 billion. A partnership or transaction would give Nvidia a route into a local challenger while testing whether the incumbent will increasingly buy or align with alternative inference architectures. 报道称，Nvidia正与韩国AI加速器初创公司Rebellions进行初步洽谈，后者估值为23亿美元。若形成合作或交易，Nvidia将借此切入本土新势力，也将检验行业龙头是否会更多收购或结盟于不同的推理架构。