AI Demand Rewires Memory and Interconnects AI需求重塑存储与互连

AI infrastructure demand is spreading beyond accelerators into HBM4, dense DDR5, CXL fabrics, high-speed links, and the test equipment needed to power them. Capital is following the same path, from TSMC's 2nm ecosystem to a $230 million-plus challenger round. AI基础设施需求正从加速器外溢至HBM4、高密度DDR5、CXL互连、高速链路以及配套供电测试设备。资本也沿着同一路径配置，从台积电2nm生态延伸至超过2.3亿美元的新势力融资。

AI & Accelerators AI与加速器

AI Networking Startups Target Nvidia's NVLink Stronghold AI网络初创公司瞄准Nvidia NVLink腹地

The Register · 4h ago The Register · 4小时前

Intel spin-off Cornelis Networks and newcomer Delos Data are pitching open interconnect alternatives for scaling AI systems beyond a single rack. The contest matters because cluster performance increasingly depends on the network fabric, giving buyers a reason to seek options outside Nvidia's vertically integrated NVLink stack. Intel分拆公司Cornelis Networks与新创企业Delos Data正推出开放式互连方案，用于将AI系统扩展到单机架之外。随着网络互连日益决定集群性能，客户也更有动力寻找Nvidia垂直整合NVLink体系之外的选择。

Nvidia Details Groq 3 LPX for Deterministic Interactive Inference Nvidia详解Groq 3 LPX确定性低延迟推理

NVIDIA Developer · 8h ago NVIDIA Developer · 8小时前

Nvidia is positioning Groq 3 LPX deterministic execution as a power-efficient engine for highly interactive inference alongside its Vera Rubin platform. The design highlights a growing split between throughput-oriented AI compute and latency-sensitive serving, where predictable token delivery can matter more than peak benchmark performance. Nvidia将Groq 3 LPX的确定性执行定位为Vera Rubin平台旁侧的高能效交互式推理引擎。该设计凸显AI计算正分化为吞吐量导向与延迟敏感型两类负载；在后者中，可预测的token输出往往比峰值基准性能更重要。

RTX 5090 Prices Pass $6,000 as Chinese AI Demand Absorbs Supply 中国AI需求吸收供给，RTX 5090价格突破6000美元

Wccftech · 8h ago Wccftech · 8小时前

GeForce RTX 5090 retail prices now start above $6,000 and some listings exceed $9,000, versus Nvidia's original $1,999 launch price. Reports that Chinese AI companies are buying the cards in volume show how constrained accelerator access is redirecting demand into consumer GPUs and distorting that market. GeForce RTX 5090零售价目前从6000美元以上起步，部分报价超过9000美元，而Nvidia首发定价为1999美元。有报道称中国AI公司正在批量采购这款显卡，显示受限的加速器获取渠道正把需求导向消费级GPU，并扭曲其市场价格。

Foundry & Manufacturing 晶圆代工与制造

MediaTek Accelerates TSMC's 2nm Customer Ramp MediaTek加速进入台积电2nm客户阵列

Yahoo Finance · 3h ago Yahoo Finance · 3小时前

MediaTek is moving into TSMC's 2nm generation as Nvidia and Alphabet also commit funding to the node's ecosystem, according to the report. A broader group of anchor customers lowers utilization risk for TSMC's expensive ramp and indicates that 2nm demand is extending beyond a single flagship buyer. 据报道，MediaTek正加速进入台积电2nm制程，Nvidia与Alphabet也已向相关生态投入资金。更多核心客户参与可降低台积电高成本爬坡阶段的产能利用率风险，也表明2nm需求并非只依赖单一旗舰客户。

Memory 存储

Kioxia Calls a Halt to Data-Center NAND Price Increases Kioxia叫停数据中心NAND涨价

Chosun · 1h ago Chosun · 1小时前

Kioxia's chief executive has ordered an end to further NAND price increases for data-center customers. The decision could temper storage-cost inflation just as AI deployments raise capacity demand, while also testing whether other NAND suppliers continue pushing prices higher. Kioxia首席执行官已下令停止进一步上调数据中心客户的NAND价格。在AI部署推高存储容量需求之际，这一决定可能缓和存储成本通胀，同时也将检验其他NAND供应商是否继续提价。

Micron Demonstrates a 512 GB DDR5 Module at 9,200 MT/s Micron展示512 GB、9200 MT/s DDR5内存条

Wccftech · 5h ago Wccftech · 5小时前

Micron has demonstrated a 512 GB DDR5 module running at up to 9,200 MT/s, with next-generation Intel and AMD servers able to reach 12 TB using the design. Denser, faster DIMMs reduce slot pressure for memory-heavy AI and agentic workloads, although the report does not provide production timing or pricing. Micron展示了一款容量512 GB、速率最高9200 MT/s的DDR5内存条，采用该设计的下一代Intel与AMD服务器内存容量可达12 TB。更高密度、更高速的DIMM可缓解AI及智能体工作负载的插槽压力，但报道未给出量产时间与价格。

SK hynix Begins Shipping 16-Layer HBM4 for Nvidia Rubin SK hynix开始为Nvidia Rubin出货16层HBM4

TradingView · 6h ago TradingView · 6小时前

SK hynix is shipping 16-layer HBM4 for Nvidia's Rubin platform, according to the report. Early volume supply strengthens its position in the next accelerator cycle and raises the qualification bar for Micron and Samsung, where stack height, thermals, and yield all compound manufacturing risk. 据报道，SK hynix已开始为Nvidia Rubin平台出货16层HBM4。率先实现批量供货将巩固其在下一轮加速器周期中的地位，也为Micron和Samsung抬高认证门槛，因为堆叠层数、散热与良率会共同放大制造风险。

Astera Labs Extends CXL Memory from Servers to AI Fabrics Astera Labs把CXL内存从服务器扩展至AI互连网络

ServeTheHome · 8h ago ServeTheHome · 8小时前

Astera Labs has launched Leo 2 memory controllers with CXL 3.2 and PCIe Gen6 support, plus Leo X for attaching memory expanders directly to AI-accelerator fabrics. The move pushes pooled memory toward rack scale, where capacity can be allocated across accelerators instead of remaining stranded behind individual host CPUs. Astera Labs发布支持CXL 3.2与PCIe Gen6的Leo 2内存控制器，并推出可将内存扩展器直接接入AI加速器互连网络的Leo X。该方案把池化内存推进到机架级，使容量能够在多个加速器间分配，而不再被锁在单个主机CPU之后。

Equipment & Materials 设备与材料

Tektronix Builds Regenerative Test System for 800-V AI Data Centers Tektronix推出面向800 V AI数据中心的回馈式测试系统

Embedded.com · 4h ago Embedded.com · 4小时前

Tektronix introduced the EA-ELR 21000 Dynamic Test System to validate emerging ±400-V and 800-V DC power architectures in AI data centers. Its auto-ranging electronic load can emulate rapid demand changes and return absorbed energy to the grid, reducing the heat and operating cost of high-power validation. Tektronix推出EA-ELR 21000动态测试系统，用于验证AI数据中心新兴的±400 V与800 V直流供电架构。其自动量程电子负载可模拟快速变化的功率需求，并把吸收的能量回馈电网，从而降低大功率验证过程中的散热与运行成本。

Photoresist Bake Conditions Emerge as a DUV Stochastics Lever 光刻胶烘烤条件成为DUV随机缺陷控制变量

SemiWiki · 8h ago SemiWiki · 8小时前

A new analysis examines how photoresist bake conditions affect stochastic behavior in 193-nm ArF immersion lithography. The work matters because DUV multipatterning remains central to advanced manufacturing, and small process-window changes can translate into defectivity and yield differences across repeated patterning steps. 一项新分析研究了光刻胶烘烤条件如何影响193 nm ArF浸没式光刻中的随机效应。DUV多重图形化仍是先进制造的关键环节，微小的工艺窗口变化会在重复曝光步骤中累积，最终影响缺陷率与良率。

Policy & Geopolitics 政策与地缘政治

Korean Industry Group Quantifies China's Memory Catch-Up 韩国行业协会量化中国存储技术追赶速度

Wccftech · 8h ago Wccftech · 8小时前

The Korea Semiconductor Industry Association estimates that China's YMTC is about one year behind leading NAND technology, while CXMT trails by roughly two years in DRAM and three years in HBM. The assessment, down from an earlier gap of about five years, adds urgency to Korean technology policy and complicates assumptions that controls alone can preserve a durable memory lead. 韩国半导体产业协会估计，中国YMTC在NAND技术上落后领先水平约1年，CXMT在DRAM与HBM上分别落后约2年和3年。相比此前约5年的差距，这一评估将加大韩国技术政策的紧迫感，也削弱了单靠管制即可长期维持存储领先优势的假设。

Challengers & New Silicon 新势力与新硅

Euclyd Raises More Than $230 Million for Agent-Focused AI Chips Euclyd融资超过2.3亿美元，开发面向AI智能体的芯片

SiliconANGLE · 5h ago SiliconANGLE · 5小时前