semi·news
Headlines要闻 / Research研究 /
Wednesday, May 27, 2026 2026年5月27日 星期三

Memory Crosses the $1 Trillion Line; Taiwan Cracks the Smuggling Pipe 存储跨入万亿俱乐部,台湾撬开走私管道

Micron and SK hynix both crossed $1 trillion in market value on AI memory demand, TSMC's CEO defused the bonus revolt with a 30%-plus pledge, and Taiwan made its first arrests in a network smuggling Nvidia GPUs to China via Japan. 美光与SK海力士同日跨过万亿美元市值门槛——由AI存储需求驱动;台积电CEO以30%+奖金涨幅承诺化解员工不满;台湾警方对经日本转运英伟达GPU至中国的走私网络展开首次抓捕。

Memory 存储

Micron and SK hynix both cross $1 trillion market cap on AI memory demand 美光与SK海力士同日跨入万亿美元市值俱乐部,AI存储需求驱动

Yahoo Finance · 10h ago Yahoo Finance · 10小时前

Two memory chipmakers — Micron and SK hynix — became the first pure-play DRAM/HBM vendors to cross a $1 trillion market capitalisation, joining Nvidia, TSMC and a handful of others. SK hynix is up ~250% year-to-date, Micron's HBM order book is reportedly sold out through 2026, and Micron's data-center revenue grew 150% YoY in its most recent quarter. The crossing reframes a long-running argument: in this AI cycle, memory is no longer the cyclical commodity sidecar to logic — it is co-equal infrastructure, and the buyers are pricing it that way. The risk angle: the same digitimes-cited analysts who validate the move are also asking when (not whether) cyclicality returns. 存储芯片厂美光和SK海力士同日跨入万亿美元市值门槛,成为首批进入万亿俱乐部的纯DRAM/HBM厂商,与英伟达、台积电等少数公司并列。SK海力士年初至今涨幅约250%,美光的HBM订单据称已售罄至2026年,最近一季数据中心收入同比增长150%。这一里程碑改写了一个延续多年的争论:在本轮AI周期中,存储不再是逻辑芯片的周期性附属品,而是同等重要的基础设施,市场也在按此定价。风险面:肯定本轮的分析师同时在问——周期性回归是什么时候、而非是否会发生。

Rambus launches DDR5-9600 client chipset aimed at next-gen AI PCs Rambus推出面向下一代AI PC的DDR5-9600客户端芯片组

Embedded.com · 24m ago Embedded.com · 24分钟前

Rambus unveiled a DDR5-9600 client chipset — the fastest mainstream-PC memory interface part shipping to date — explicitly positioned for on-device AI workloads where bandwidth-per-watt at the client edge has become a real constraint. The launch matters because client memory has lagged datacenter HBM by ~2 generations on bandwidth; closing that gap is what lets a NPU-equipped consumer PC actually run multi-billion-parameter models locally without thrashing against DRAM. Watch for AMD's and Intel's 2026 client roadmaps to validate the spec as a default rather than premium tier. Rambus发布DDR5-9600客户端芯片组——目前出货的主流PC存储接口部件中速率最高——明确瞄准设备端AI负载,其中边缘带宽每瓦已成为真实瓶颈。意义在于:客户端存储在带宽上已落后数据中心HBM约2个代际,弥合这一差距是让搭载NPU的消费PC在本地运行多十亿参数模型而不被DRAM拖死的关键。后续关注AMD与英特尔2026年客户端路线图是否将此规格作为默认配置而非高端选项。

AI & Accelerators AI与加速器

Broadcom adds South Korea's FuriosaAI to its custom-ASIC empire 博通将韩国FuriosaAI纳入其定制ASIC版图

The Register · 36m ago The Register · 36分钟前

Broadcom announced a development partnership with Seoul-based FuriosaAI for next-generation AI inference chips — the latest addition to a stable that already includes hyperscaler XPU work for Google, Meta and OpenAI. For Broadcom, FuriosaAI brings a non-US partner with a usable RNGD architecture and a credible LPDDR-anchored inference cost story; for FuriosaAI, it brings access to Broadcom's CoWoS allocation and Tomahawk fabric stack. The move continues the broader pattern of 2026: every credible non-Nvidia inference startup needs a foundry-services or networking gatekeeper as a co-pilot. 博通宣布与韩国首尔的FuriosaAI在下一代AI推理芯片上建立开发合作——这是博通定制ASIC业务的最新一站,此前已包括谷歌、Meta、OpenAI的XPU项目。对博通而言,FuriosaAI带来了一家拥有可用RNGD架构、以LPDDR为锚的推理成本叙事可信的非美系合作方;对FuriosaAI而言,则换来了博通CoWoS产能与Tomahawk网络栈的入场券。这延续了2026年的更大模式:每一家可信的非英伟达推理初创,都需要一家代工服务或网络厂商作为「副驾驶」。

Nvidia's Rubin CPX launch wobbles as memory and substrate orders fail to materialize 英伟达Rubin CPX发布生变:存储与基板订单未如期落地

TheElec · 2h ago TheElec · 2小时前

Korea's TheElec reports that the launch of Nvidia's Rubin CPX — the compute-heavy variant of the Rubin generation targeted at the prefill stage of LLM inference — has become uncertain because the bill-of-materials orders for its memory and advanced substrate haven't arrived at the expected suppliers. The signal cuts against the broader narrative that Nvidia's roadmap is on autopilot: at the bleeding edge, BOM placement is now the practical schedule constraint, not silicon design. If true, the slip would also reshuffle the implied 2026-Q4 / 2027-Q1 hyperscaler accelerator mix. 韩国TheElec报道,英伟达Rubin CPX——Rubin代号面向LLM推理预填充阶段、计算密度更高的变体——的发布出现不确定性,原因是其存储与先进基板的BOM订单未按预期到达对应供应商。这一信号与「英伟达路线图自动驾驶」的主流叙事相悖:在尖端节点,BOM下单已实质成为日程瓶颈,而非硅设计本身。若属实,延后还将重新洗牌2026年Q4 / 2027年Q1超大规模客户的加速器组合。

Foundry & Manufacturing 晶圆代工与制造

TSMC CEO Wei pledges 30%-plus bonus increase, Samsung workers vote yes on $340k deal 台积电CEO魏哲家承诺奖金涨幅30%+,三星员工以多数票通过约34万美元奖金方案

Bloomberg · 2h ago Bloomberg · 2小时前

Twin labor outcomes resolved this week's flashpoints. TSMC CEO C.C. Wei publicly pledged employee bonus growth above 30% in 2026 if performance holds — directly responding to the strike-threat reports from earlier in the week. Separately, Samsung's chip-division union voted 74% in favor of an average bonus equivalent to ~10.5% of operating profit (about $340-400k per senior engineer), formally ending months of strike-threat tension. The outcomes matter because the Asian leading-edge fabs' single biggest 2026 risk — beyond geopolitics — was talent flight; both companies have now bought another year of stability at a known cost. 本周两大劳资紧张同时落地。台积电CEO魏哲家公开承诺:若业绩维持,2026年员工奖金涨幅将超过30%——直接回应本周早些时候的罢工威胁报道。另一边,三星芯片部门工会以74%多数票通过平均约相当于营业利润10.5%的奖金方案(资深工程师人均约34-40万美元),正式结束持续数月的罢工威胁。意义在于:除地缘政治外,2026年亚洲先进制程代工的最大单一风险就是人才流失;两家公司均以已知代价为自己又买到了一年稳定。

TSMC to raise 3nm wafer prices by up to 15% in H2 2026; further 5-10% in 2027 台积电2026年下半年3nm晶圆涨价至多15%,2027年再涨5-10%

TrendForce (via Google News) · 6h ago TrendForce (via Google News) · 6小时前

TrendForce reports TSMC will hike 3nm wafer prices by as much as 15% in 2H 2026 and pencil in another 5-10% in 2027, driven by AI and custom-ASIC demand that has now saturated Fab 18 and the early N3-family capacity. The increase will flow straight into Apple, Qualcomm, MediaTek, AMD and Nvidia bill-of-materials — and the news lands on the same day TSMC's stock hit a 52-week high. The structural read: TSMC is finally collecting the pricing premium that node leadership theoretically commanded since 2020, now that competitive 3nm alternatives are at least 12-18 months behind. TrendForce报道,受AI与定制ASIC需求拉动、Fab 18及N3家族早期产能已饱和,台积电将于2026年下半年对3nm晶圆涨价至多15%,并预留2027年再涨5-10%。涨幅将直接转入苹果、高通、联发科、AMD与英伟达的物料成本;新闻发布当日台积电股价同步创下52周新高。结构性解读:在自2020年以来理论上属于工艺领先者的「溢价」,终于由台积电真正收取——前提是竞争性3nm替代方案至少落后12-18个月。

Policy & Geopolitics 政策与地缘

Taiwan arrests three over Nvidia chip smuggling network — Japan used as transshipment 台湾首次依英伟达走私案抓捕3人,日本作为转运地

Tom's Hardware · 2h ago Tom's Hardware · 2小时前

Taiwan authorities arrested three individuals in connection with smuggling Nvidia AI GPUs to China via Japan — the first formal criminal crackdown on this trade route. The case builds on the Supermicro-channel allegations from earlier in the month, but adds a new wrinkle: routing through Japan to obscure the final destination. Practically, it means the BIS and Taiwan's MOEA now have a documented network rather than just theory, which raises the compliance bar for OEM partners and increases the cost of grey-market diversion across the entire region. 台湾警方依涉嫌将英伟达AI GPU经日本走私至中国的案件抓捕3名嫌疑人——这是该走私路径首次正式刑事追诉。该案在本月早些时候涉Supermicro渠道的指控之上又添新层:通过日本转运以掩盖最终目的地。实质意义:美国商务部BIS与台湾经济部如今手中是有据可查的走私网络而非理论模型,这提高了OEM合作方的合规门槛,并增加了整个亚太区灰色市场转运的成本。

China adds nine homegrown AI chips to 'secure and reliable' procurement list for the first time 中国首次将9款国产AI芯片纳入「安全可靠」采购目录

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

Nine domestically designed AI accelerators — spanning Huawei Ascend, Cambricon, Hygon and others — have been formally added to China's 'secure and reliable' procurement list, a status that effectively mandates them as the default choice for state-owned enterprises, central banks and government IT buyers. This is the first time AI accelerators have been included in that list, and is a direct policy mirror to the US Commerce Department's export-control regime. The practical effect: a large baseline order book gets locked in for Huawei et al. regardless of what happens with US H200 licensing. 9款国产AI加速器——涵盖华为昇腾、寒武纪、海光等——正式进入中国「安全可靠」采购目录;该地位实质上把它们设定为国企、央行与政府IT采购的默认选项。这是AI加速器首次进入该目录,与美国商务部出口管制制度形成镜像。实质效果:无论美国H200许可走向如何,华为等国产芯片厂都将获得一笔较大的基线订单。

Equipment & Materials 设备与材料

Entegris and JSR/Inpria sign EUV photoresist cross-licensing deal Entegris与JSR/Inpria签署EUV光刻胶交叉许可协议

EE Times Asia · 11h ago EE Times Asia · 11小时前

Entegris and JSR's Inpria subsidiary signed a cross-licensing agreement covering metal-oxide and chemically-amplified EUV photoresist IP — the two technology families that together cover essentially all leading-edge EUV resist designs in production today. The deal removes a long-running IP overhang on Inpria's metal-oxide resist roadmap (which TSMC, Samsung and Intel are increasingly relying on for High-NA EUV at A14/A10) and gives Entegris a clear path to scale its material-supply business into the same node window. ASML's High-NA install base is the customer that wins from cleaner IP plumbing. Entegris与JSR子公司Inpria签署交叉许可协议,覆盖金属氧化物与化学放大型EUV光刻胶IP——这两大技术家族几乎涵盖了目前量产中的所有先进EUV光刻胶设计。协议消除了Inpria金属氧化物光刻胶路线长期存在的IP风险(台积电、三星、英特尔在A14/A10的High-NA EUV节点上对其依赖加深),同时让Entegris获得了将材料供应业务扩展到同代节点的清晰路径。最终受益者是ASML的High-NA装机客户群——更清晰的IP水管。

TATA Electronics signs MoU with ASML for India semiconductor manufacturing 塔塔电子与ASML签署谅解备忘录,推进印度半导体制造

MSN (via Google News) · 11h ago MSN (via Google News) · 11小时前

TATA Electronics signed a memorandum of understanding with ASML covering future-node lithography tools and service support for the company's planned Gujarat fab — India's first announced 28nm/22nm-class facility, currently in equipment-installation phase. The MoU itself is non-binding but materially de-risks tool delivery slots, which had been the principal open question for India's semi build-out. The strategic read: even at trailing-edge nodes, ASML capacity is now a sovereign procurement matter, and India has secured its place in the queue. 塔塔电子与ASML签署谅解备忘录,覆盖其位于古吉拉特邦规划晶圆厂的未来节点光刻设备与服务支持——这是印度首座正式宣布的28nm/22nm级工厂,目前处于设备进场阶段。MoU本身不具法律约束力,但实质上降低了设备交付排期的风险,而这正是印度半导体建设此前最大的悬而未决问题。战略解读:即便是成熟节点,ASML产能如今也是主权采购议题;印度已在排队序列中锁定位置。

Packaging & Advanced Nodes 封装与先进节点

ASE accelerates AI innovation with automated 310mm panel-level packaging 日月光以自动化310mm面板级封装加速AI创新

EE Times Asia · 11h ago EE Times Asia · 11小时前

Outsourced-assembly leader ASE detailed its automated 310×310 mm panel-level packaging (PLP) line — the same panel format TSMC is targeting for its CoPoS pilot. The differentiator is automation: ASE has built a continuous-flow PLP cell that targets >85% utilisation on AI accelerator packaging volumes that previously bottlenecked at hand-tuned wafer-level steps. The signal worth tracking: panel-level packaging is graduating from RD demonstrators to industrial-scale assembly, and the OSAT incumbents are positioning to absorb the AI accelerator volume that TSMC's in-house CoPoS won't take. 封测龙头日月光(ASE)详述其自动化310×310 mm面板级封装(PLP)产线——所采用面板规格与台积电CoPoS试产线相同。差异点在于自动化:ASE构建了连续流式PLP工艺单元,目标AI加速器封装产能利用率超过85%,此前同等步骤需手工调校、效率受限。值得关注的信号:面板级封装正从研发演示阶段毕业,进入工业化量产;而封测厂正在卡位——准备接收台积电自有CoPoS消化不下的AI加速器封装需求。

Samsung plans $1.5 billion chip testing plant in Vietnam 三星计划在越南投资15亿美元建芯片测试厂

Reuters · 1h ago Reuters · 1小时前

Reuters obtained a Samsung internal document outlining a $1.5 billion semiconductor testing plant in Vietnam's Thai Nguyen province — an extension of the company's existing Vietnamese assembly footprint, but a meaningful first step into back-end test that has historically been Korea-anchored. The geographic spread matters: combined with the Vietnam plant, Samsung's HBM+DRAM test workflow now has three regional anchors (Korea, China, Vietnam), reducing single-point exposure if any one of them faces export-control or labor disruption. 路透获取的三星内部文件显示,公司将在越南太原省投资15亿美元建设半导体测试工厂——这是公司既有越南封装业务的延伸,但向后端测试环节迈出了实质性一步(此前后端测试主要锚定在韩国本部)。地理分布意义重大:加上新越南厂后,三星的HBM+DRAM测试流程将拥有韩国、中国、越南三个区域锚点,降低任一节点因出口管制或劳资中断带来的单点风险。