Nvidia tests leaner Rubin Ultra memory configurations while TSMC reports mature yields on selected CoWoS products. Capital is following the bottlenecks, from Sony's Japan investment to Meta's reuse of DDR4 through CXL. Nvidia测试更精简的Rubin Ultra存储配置，台积电则披露部分CoWoS产品已达到成熟良率。资本与系统设计正追随瓶颈移动，从Sony在日本的投资，到Meta借助CXL复用DDR4。
AMD CEO Lisa Su laid out a full-stack AI strategy spanning GPUs, CPUs, networking, software, and rack-scale systems for agentic workloads. The framing matters because AMD is competing for the architecture of the entire AI rack, where system integration and software can be as decisive as accelerator specifications. AMD CEO苏姿丰提出覆盖GPU、CPU、网络、软件与机架级系统的全栈AI战略，重点面向智能体工作负载。这一定位意味着AMD争夺的不只是加速器份额，而是整套AI机架架构；系统集成与软件的重要性正接近单颗芯片规格。
GPU cloud provider Nebius is pursuing a rapid expansion toward 1 GW of powered capacity and expects to use multiple financing routes to fund it. The plan is another measure of how quickly accelerator demand is translating into power, construction, and balance-sheet requirements rather than chip orders alone. GPU云服务商Nebius计划快速扩张至1 GW供电容量，并准备通过多种融资方式支撑建设。该计划再次表明，加速器需求正迅速转化为电力、基础设施与资产负债表压力，而不只是芯片采购订单。
A Sony-TSMC agreement brings total overseas investment in Japan's chipmaking sector to $37 billion, according to Nikkei Asia. The total shows how Japan is pairing domestic industrial policy with foreign capital and customer commitments to rebuild manufacturing depth. 据Nikkei Asia报道，Sony与台积电达成交易后，日本芯片制造领域吸引的海外投资总额达到370亿美元。这一规模显示，日本正把国内产业政策、海外资本与客户承诺结合起来，重建半导体制造纵深。
TSMC says selected AI products using a 5.5-reticle-size CoWoS implementation have reached 98% packaging yield. The company also points to a 14-reticle-size CoWoS roadmap for 2029, underscoring that larger packages—not only leading-edge lithography—are becoming central to accelerator scaling. 台积电表示，采用5.5倍光罩尺寸CoWoS方案的部分AI产品，封装良率已达到98%。公司还提出到2029年扩展至14倍光罩尺寸的路线图，说明加速器扩展越来越依赖更大封装，而不只是先进光刻制程。
Nvidia is reportedly testing lower-memory Rubin Ultra designs, including configurations with 192 GB and a move back to HBM4, as memory supply tightens. These are tests rather than a final product decision, but they show that HBM availability can reach far enough upstream to reshape accelerator configurations. 据报道，在存储供应趋紧之际，Nvidia正在测试更低容量的Rubin Ultra设计，其中包括192 GB配置以及回退至HBM4的方案。这些仍是测试而非最终产品决定，但已表明HBM供应足以向上游传导，直接改变加速器配置。
Meta cut the server count for a workload by 25% by pooling and reusing older DDR4 memory through CXL. The result gives CXL a concrete utilization case, though mismatched DIMMs, power overhead, and telemetry complexity make the approach harder to reproduce outside hyperscale fleets. Meta通过CXL池化并复用旧DDR4，使某项工作负载所需的服务器数量减少25%。这一结果为CXL提供了具体的资源利用案例，但DIMM规格混杂、额外功耗与遥测复杂度，会让非超大规模运营商更难复制。
Hanmi Semiconductor is defending profitability while rival Hanwha Semitek faces financial strain in Korea's TC bonder market. The divergence is a reminder that strong HBM packaging demand does not translate uniformly into margins, particularly as suppliers invest to qualify tools and expand capacity. 在韩国TC Bonder市场，Hanmi Semiconductor维持盈利能力，而竞争对手Hanwha Semitek面临财务压力。这种分化说明，旺盛的HBM封装需求不会自动转化为一致的利润率，尤其是在供应商仍需投入设备验证与扩产的阶段。
Valens Semiconductor raised its 2026 revenue outlook, sending its shares higher. The revision improves demand visibility for the chip supplier and matters more than the immediate stock move because guidance resets expectations for the remaining quarters. Valens Semiconductor上调2026年营收展望，推动股价上涨。相较于短期市场反应，业绩指引的调整更重要，因为它提高了需求能见度，并重新设定市场对余下季度的预期。
Cerebras reported a 103% year-over-year increase in core revenue for the second quarter. The growth is a meaningful commercialization signal for wafer-scale computing, although one quarter of expansion does not by itself establish the durability or breadth of customer demand. Cerebras报告第二季度核心营收同比增长103%。这一增速为晶圆级计算的商业化提供了实质信号，但单季扩张仍不足以证明客户需求具备长期持续性与足够广度。
Tenstorrent posted GCC compiler patches for its Ascalon XG processor core. Upstream compiler work is not proof of a shipping product, but it is a necessary step toward a usable software ecosystem and gives developers an earlier view of the core's support requirements. Tenstorrent为其Ascalon XG处理器核心提交了GCC编译器补丁。上游编译器工作并不等同于产品已经出货，但它是构建可用软件生态的必要步骤，也让开发者更早看到该核心的支持需求。