Memory and Links Become the AI System 存储与互连成为AI系统本体

Computex's second wave shifts from accelerator announcements to the pieces that decide whether those accelerators ship: HBM capacity, optical links, PCIe scale-up, 2nm ramps and export-control routing. Computex第二波重点从加速器发布转向决定加速器能否出货的部件：HBM产能、光互连、PCIe扩展、2nm爬坡以及出口管制路径。

Memory 存储

SK hynix plans to double wafer production within five years SK hynix计划五年内将晶圆产量翻倍

Korea JoongAng Daily via Google News · 21m ago Korea JoongAng Daily via Google News · 21分钟前

SK hynix's chairman said the company plans to double wafer production within five years as AI memory demand keeps pulling capacity forward. The statement matters because HBM supply is no longer a quarterly allocation problem; it is becoming a multi-year fab, materials and equipment plan. The risk for buyers is that capacity added for HBM also reshapes conventional DRAM supply, keeping PC and consumer memory exposed to AI-driven pricing. SK hynix董事长表示，公司计划在五年内将晶圆产量翻倍，以应对AI存储需求持续提前拉动产能。这个表态重要，因为HBM供应已不再只是季度配额问题，而是转变为跨越数年的晶圆厂、材料和设备计划。对买方而言，风险在于为HBM增加的产能同样会重塑传统DRAM供给，让PC和消费存储继续暴露在AI驱动的价格压力下。

SK hynix previews 48GB HBM4E with 4 TB/s bandwidth SK hynix预览48GB HBM4E，带宽达4 TB/s

Wccftech · 46m ago Wccftech · 46分钟前

SK hynix previewed HBM4E at Computex with a 12-high, 48GB stack and a claimed 4 TB/s of bandwidth. The numbers push the memory roadmap beyond today's HBM3E allocation fight and toward the Rubin-generation training and inference systems now being designed. Thermal stack design, known-good-die discipline and packaging slots will be as important as the headline bandwidth. SK hynix在Computex预览HBM4E，展示12层、48GB堆叠，并宣称带宽达到4 TB/s。这些数字把存储路线图推过当前HBM3E配额战，指向正在设计中的Rubin世代训练与推理系统。热堆栈设计、已知良品裸片管理和封装槽位，将与标题带宽同样重要。

Air Liquide invests in Korea to support SK hynix AI memory production Air Liquide在韩国投资，支撑SK hynix AI存储项目

Air Liquide via Google News · 2h ago Air Liquide via Google News · 2小时前

Air Liquide announced a milestone investment in South Korea tied to SK hynix's advanced AI memory project. Industrial gas capacity is easy to miss in the HBM story, but high-volume memory ramps require reliable supply of process gases and on-site infrastructure. The announcement is another reminder that AI memory is pulling the upstream materials chain into named, long-horizon commitments. Air Liquide宣布在韩国进行一项里程碑式投资，服务于SK hynix先进AI存储项目。工业气体产能在HBM叙事中容易被忽略，但高量级存储爬坡依赖稳定的工艺气体和现场基础设施。该公告再次说明，AI存储正在把上游材料链拉入具名、长期的产能承诺。

Micron's HBM4 push signals a larger Nvidia supply role 美光HBM4推进暗示其在Nvidia供应链中角色扩大

DigiTimes via Google News · 3h ago DigiTimes via Google News · 3小时前

DigiTimes reports Micron's HBM4 push is pointing to a bigger role in Nvidia's supply chain. That would not displace SK hynix overnight, but it would give Nvidia more room to diversify HBM sourcing as demand moves to next-generation stacks. For Micron, HBM4 is the route from cyclical DRAM recovery to a more strategic AI-infrastructure supplier position. DigiTimes报道称，美光推进HBM4，显示其在Nvidia供应链中的角色可能扩大。这不会一夜之间取代SK hynix，但会让Nvidia在下一代堆叠需求到来时拥有更多HBM来源分散空间。对美光而言，HBM4是从周期性DRAM复苏走向更具战略性的AI基础设施供应商地位的路径。

Interconnects & Scale-Up 互连与扩展

Ayar Labs joins Nvidia's NVLink ecosystem for optical infrastructure Ayar Labs加入Nvidia NVLink生态，推进光互连基础设施

Scientific Computing World via Google News · 30m ago Scientific Computing World via Google News · 30分钟前

Ayar Labs joined Nvidia's NVLink ecosystem to bring co-packaged optics into rack-scale AI infrastructure. The significance is architectural: optical I/O is no longer only a future datacenter slide, it is being wired into the scale-up fabric that decides how many accelerators can behave like one system. Packaging, thermals and serviceability still decide how fast CPO becomes deployable hardware. Ayar Labs加入Nvidia NVLink生态，将共封装光学带入机架级AI基础设施。其意义在架构层面：光I/O不再只是未来数据中心路线图，而是被接入决定多颗加速器能否像一个系统工作的scale-up fabric。封装、散热和可维护性仍会决定CPO多快成为可部署硬件。

Lightmatter joins NVLink Fusion with photonic interconnects Lightmatter以光子互连加入NVLink Fusion

Business Wire via Google News · 10h ago Business Wire via Google News · 10小时前

Lightmatter also joined Nvidia's NVLink Fusion ecosystem, positioning its photonic interconnect technology for next-generation AI infrastructure. Having both Lightmatter and Ayar around the NVLink conversation suggests Nvidia is opening the scale-up perimeter without giving up the software and fabric control point. For optics startups, the prize is not a generic transceiver socket; it is being designed into the accelerator fabric itself. Lightmatter也加入Nvidia NVLink Fusion生态，将其光子互连技术定位到下一代AI基础设施中。Lightmatter和Ayar同时围绕NVLink出现，说明Nvidia正在打开scale-up边界，但不放弃软件和互连控制点。对光互连初创而言，真正的奖品不是通用收发器插槽，而是被设计进加速器fabric本身。

Astera Labs shows a 320-lane PCIe 6.0 switch for scale-up AI Astera Labs展示320通道PCIe 6.0交换芯片，用于AI扩展

Tom's Hardware · 1h ago Tom's Hardware · 1小时前

Astera Labs showcased a 320-lane PCIe 6.0 switch that can scale up to 80 accelerators using PCIe alone. That is not a replacement for NVLink-class fabrics in every training cluster, but it matters for vendor-agnostic systems and mixed accelerator deployments. The announcement keeps PCIe scale-up relevant as customers look for ways to avoid total dependence on a single closed interconnect stack. Astera Labs展示了一款320通道PCIe 6.0交换芯片，可仅用PCIe扩展多达80个加速器。这不会在所有训练集群中取代NVLink级互连，但对厂商中立系统和混合加速器部署很重要。该发布让PCIe scale-up继续保持相关性，因为客户正在寻找避免完全依赖单一封闭互连堆栈的方法。

Foundry & Manufacturing 晶圆代工与制造

AMD starts Venice production on TSMC's 2nm process AMD Venice开始在TSMC 2nm制程生产

IT Brief Australia via Google News · 7h ago IT Brief Australia via Google News · 7小时前

AMD has reportedly started Venice production on TSMC's 2nm process, a concrete signal that leading-edge server CPU roadmaps are moving from tape-out talk to manufacturing allocation. The story matters for TSMC because 2nm demand is not only mobile or AI accelerators; server CPUs are competing for the same early process capacity. For AMD, the timing sets up a sharper platform fight against Intel's 18A recovery narrative. 报道称AMD Venice已开始在TSMC 2nm制程生产，这是领先服务器CPU路线图从流片叙事进入制造配额的具体信号。对TSMC而言，2nm需求不只是手机或AI加速器；服务器CPU也在争夺同一批早期先进制程产能。对AMD而言，这个时间点会让其与Intel 18A复苏叙事的服务器平台竞争更直接。

Intel Foundry cites advanced-process progress while Tan calls TSMC a key partner Intel Foundry称先进制程取得进展，陈立武称TSMC是关键伙伴

DigiTimes via Google News · 12h ago DigiTimes via Google News · 12小时前

Intel Foundry used Computex messaging to emphasize progress in advanced processes, while CEO Lip-Bu Tan also described TSMC as a key partner. That dual framing is revealing: Intel wants foundry credibility without pretending the industry can abruptly detach from TSMC. The practical read is that Intel's recovery still depends on landing real external volume while maintaining pragmatic relationships across the existing supply chain. Intel Foundry在Computex信息中强调先进制程进展，同时CEO陈立武也称TSMC是关键伙伴。这种双重叙事很有信息量：Intel需要代工可信度，但并不假装行业能突然摆脱TSMC。实际解读是，Intel复苏仍取决于拿到真实外部量产，同时与现有供应链保持务实关系。

Synopsys and Samsung Foundry expand AI-driven design, test and 3DIC enablement Synopsys与Samsung Foundry扩展AI驱动设计、测试与3DIC支持

EE Times Asia · 12h ago EE Times Asia · 12小时前

Synopsys and Samsung Foundry expanded their collaboration around AI-driven design, test and 3DIC enablement for advanced nodes. The content is EDA-heavy, but the strategic point is manufacturing readiness: advanced packaging and 3D integration need signoff, test and yield workflows as much as new substrates. Samsung's foundry pitch becomes more credible when its toolchain partners can show packaged-system support, not just transistor-level PDK claims. Synopsys与Samsung Foundry扩展围绕先进节点的AI驱动设计、测试和3DIC支持。内容偏EDA，但战略重点是制造就绪度：先进封装和3D集成需要签核、测试和良率流程，和新载板同样重要。当工具链伙伴能展示封装系统级支持，而不只是晶体管级PDK说法时，Samsung Foundry的代工叙事才更可信。

Equipment & Materials 设备与材料

xLight finalizes a $150M CHIPS incentives agreement xLight敲定1.5亿美元CHIPS激励协议

Pulse 2.0 via Google News · 3h ago Pulse 2.0 via Google News · 3小时前

xLight finalized a $150 million CHIPS incentives agreement, adding another photonics-linked project to the US semiconductor policy stack. The materiality is not the dollar amount alone; it is the way advanced light sources and photonics are being treated as strategic manufacturing infrastructure. As AI hardware leans harder on optical links and advanced lithography, photonics policy is moving closer to core chip industrial policy. xLight敲定1.5亿美元CHIPS激励协议，为美国半导体政策体系增加了一个与光子相关的项目。其重要性不只在金额，而在于先进光源和光子技术正被视为战略制造基础设施。随着AI硬件更加依赖光互连和先进光刻，光子政策正在靠近芯片产业政策核心。

Tata Electronics and ASML sign an MoU for semiconductor manufacturing Tata Electronics与ASML签署半导体制造谅解备忘录

MSN via Google News · 12h ago MSN via Google News · 12小时前

Tata Electronics and ASML signed an MoU tied to semiconductor manufacturing, another step in India's attempt to assemble a serious fab ecosystem rather than only a packaging base. An MoU is not a tool delivery schedule, but ASML's involvement matters because lithography access is the credibility bottleneck in any national fab plan. Watch whether this turns into training, service, installed-base commitments or actual scanner allocation. Tata Electronics与ASML签署半导体制造相关谅解备忘录，这是印度试图建立真正晶圆厂生态、而不只是封装基地的又一步。MoU不是设备交付排期，但ASML参与本身重要，因为光刻能力是任何国家晶圆厂计划的可信度瓶颈。后续要看它会变成培训、服务、装机承诺，还是实际光刻机配额。

Policy & Geopolitics 政策与地缘

US moves to halt Nvidia AI-chip shipments to Chinese firms outside China 美国推进阻止Nvidia AI芯片流向中国企业境外实体

MSN via Google News · 3h ago MSN via Google News · 3小时前

The US is moving to halt Nvidia AI-chip shipments to Chinese firms outside China, extending the export-control perimeter beyond nominal destination. This is the operational version of the loophole closure: resellers and OEMs now need to evaluate beneficial ownership and end-use, not simply country-of-delivery. The likely effect is tighter compliance friction across Southeast Asian and offshore purchasing channels. 美国正在推进阻止Nvidia AI芯片流向中国企业境外实体，把出口管制边界从名义目的地扩展出去。这是关闭漏洞后的操作层版本：分销商和OEM现在需要评估实际受益所有人与最终用途，而不只是交付国家。可能结果是东南亚和离岸采购渠道的合规摩擦明显上升。

Huawei proposes a new architectural path to advanced chips under US controls 华为提出在美国管制下推进先进芯片的新架构路径

MSN via Google News · 7h ago MSN via Google News · 7小时前

Huawei is proposing a new architectural path for advanced chips as US controls continue to restrict direct access to leading-edge tools and accelerators. The important point is that China's response is becoming architectural and system-level, not only a search for one-for-one tool substitutes. Whether the path works depends on software stack maturity, packaging capacity and enough domestic silicon volume to absorb inefficiency. 在美国管制持续限制先进工具和加速器直接获取的背景下，华为提出推进先进芯片的新架构路径。重点在于，中国的应对正在变成架构和系统层面的路线，而不只是寻找一对一设备替代。该路径能否成立，取决于软件栈成熟度、封装能力，以及足够大的国内硅片出货量来吸收效率损失。

Challengers & New Silicon 新势力与新硅

Cerebras reportedly raises its IPO range, targeting a $48.8B valuation 报道称Cerebras上调IPO区间，目标估值达488亿美元

MSN via Google News · 12h ago MSN via Google News · 12小时前

Cerebras reportedly raised its IPO range and is targeting a valuation around $48.8 billion. The number is aggressive, but it also reflects how public markets are being asked to price non-GPU AI compute architectures after Nvidia's rack-scale dominance. Cerebras still has to prove durable software and workload pull, yet the IPO range shows challenger silicon remains financeable when it offers a differentiated system story. 报道称Cerebras上调IPO区间，目标估值约488亿美元。这个数字很激进，但也反映出在Nvidia机架级主导之后，公开市场正被要求重新定价非GPU AI计算架构。Cerebras仍需证明软件和工作负载拉力的持续性，但IPO区间说明，只要有差异化系统叙事，新硅势力仍能获得融资窗口。