semi·news
Headlines要闻 / Research研究 /
Sunday, May 31, 2026 2026年5月31日 星期日

Samsung Snaps HBM4E First; Nvidia Bets $6.5B on US Photonics 三星抢发HBM4E,英伟达豪赌65亿美元美国硅光

Samsung shipped industry-first HBM4E samples six months ahead of SK hynix while Nvidia committed roughly $6.5B to onshore photonics manufacturing — and Nikon undercut ASML on ArF pricing in the same week the Supermicro co-founder was arrested over $2.5B in alleged Nvidia GPU smuggling. 三星比SK海力士提前六个月送出业界首批HBM4E样品;英伟达约65亿美元押注美国本土硅光制造;同一周尼康以低价突袭ASML的ArF光刻业务,而美超微联合创始人因涉嫌25亿美元英伟达GPU走私被捕。

Memory 存储

Samsung ships industry-first HBM4E samples — 3.6 TB/s, six months ahead of SK hynix 三星送出业界首批HBM4E样品:3.6 TB/s带宽,较SK海力士领先六个月

Tech Times · 16h ago Tech Times · 16小时前

Samsung confirmed shipment of the first HBM4E samples to lead customers at 3.6 TB/s per stack — roughly 30% above HBM4 and a meaningful lead window over SK hynix, whose HBM4E samples weren't expected until late 2026 or early 2027. The combination of P4-line conversion (covered last week as a 2027 DRAM-supply risk) and a real bandwidth lead reframes Samsung's position: from the memory vendor visibly behind SK hynix in HBM3E to the credibly-first vendor on HBM4E. For Nvidia's Rubin generation roadmap — and especially the Rubin CPX variant whose substrate/memory orders failed to materialize earlier this month — this changes which memory supplier is the gating partner. 三星确认向头部客户出货首批HBM4E样品,单堆叠带宽达3.6 TB/s——较HBM4高出约30%,且明显领先SK海力士(其HBM4E样品原计划要到2026年底或2027年初才出)。叠加P4厂的产能转换(上周报道的2027年DRAM供应风险)与真实带宽领先,三星在存储格局中的地位被改写:从在HBM3E上明显落后SK海力士,变成HBM4E首发的可信厂商。对英伟达Rubin代路线图——尤其是本月早些时候基板/存储订单未能落地的Rubin CPX变体——这改变了关键存储供应商的人选。

Samsung overtakes Micron to claim the top spot in automotive memory 三星超越美光,登顶车用存储芯片市场

Seoul Economic Daily (via Google News) · 5h ago Seoul Economic Daily (via Google News) · 5小时前

New 2025 share data shows Samsung edging Micron out of the lead in the automotive memory market — a quietly-important category where qualification cycles run multi-year and vendor switching is rare. The shift matters more for what it implies than for the dollar revenue: Samsung's automotive HBM and LPDDR5X qualifications give it a foothold for next-gen ADAS / autonomous-driving compute platforms where on-vehicle memory bandwidth is rising sharply. Micron is now playing defense in a segment it dominated for the better part of a decade. 2025年最新份额数据显示三星在车用存储市场以微弱优势超越美光——这是一个低调但重要的品类:车规认证周期以年计、客户更换供应商极少。这一份额变化对营收数字本身的意义有限,但其象征更重要:三星在车规HBM与LPDDR5X上的认证为下一代ADAS/自动驾驶计算平台(车载存储带宽正快速攀升)取得立足点。美光则在其曾主导近十年的这一细分市场上转入防守。

AI & Accelerators AI与加速器

Nvidia commits ~$6.5B to American photonics manufacturing — signals CPO is the next AI battleground 英伟达约65亿美元押注美国硅光制造——共封装光学成下一AI战场的信号

MSN · 7h ago MSN · 7小时前

Reports converge on Nvidia making a multi-billion-dollar onshore commitment to silicon-photonics manufacturing — figures range from $4B (initial onshore plant) to $6.5B (full multi-site program), with the Tekedia coverage framing it as Nvidia explicitly designating co-packaged optics as the next AI battleground beyond accelerator silicon. The bet lines up with Lightmatter's COUPE partnership, Tower Semi's $1.3B photonics order book, and the Marvell DSP ramp covered earlier this week. Net read: every credible 2027 AI rack now has a CPO story, and Nvidia is buying out the gating supplier risk rather than just consuming it. 多家媒体证实英伟达正向硅光子在美制造做出数十亿美元级投入——金额版本从40亿美元(首座在美工厂)到65亿美元(完整多站项目)不等。其中Tekedia的报道将其定义为:英伟达明确把共封装光学(CPO)定为加速器硅片之外的下一AI战场。这一押注与Lightmatter的COUPE合作、Tower半导体13亿美元硅光订单簿,以及本周早些时候报道的Marvell DSP爬坡相互呼应。结论:每一个可信的2027年AI机架都有了CPO故事,而英伟达正在直接买入「关键供应商风险」,而非仅仅消费它。

Groq closes $650M round from existing backers for AI Inference Cloud expansion Groq完成6.5亿美元融资(老股东参与),用于扩张AI推理云

MLQ.ai (via Google News) · 12h ago MLQ.ai (via Google News) · 12小时前

MLQ.ai confirmed Groq's $650M round is now closed, led by existing investors, with proceeds explicitly earmarked for AI Inference Cloud capacity rather than R&D into a successor LPU architecture. The round lands alongside the previously-disclosed Nvidia $20B partnership — a curious juxtaposition that suggests Groq is positioning itself as a complement, not a replacement, for Nvidia inference on hyperscaler stacks. Watch for how the new Groq Cloud regions land: that's where the inference-economics narrative either holds up or falls apart against incumbent GPU pricing. MLQ.ai确认Groq 6.5亿美元融资已经落地,由老股东主导,资金明确用于AI推理云的产能扩张,而非用于下一代LPU架构的研发。本轮融资与此前披露的英伟达200亿美元合作并列出现,这一组合耐人寻味:Groq似乎正定位为「英伟达推理在超大规模客户栈上的补充」,而非替代。后续需关注Groq Cloud新区域的落地情况——推理经济学叙事正是在那里与现有GPU定价正面对决,要么撑住、要么破裂。

Equipment & Materials 设备与材料

Nikon undercuts ASML on ArF lithography pricing — in-house manufacturing as the lever 尼康以低价突袭ASML的ArF光刻业务——以自有产线作为价格杠杆

Tom's Hardware · 1d ago Tom's Hardware · 1天前

Nikon confirmed an aggressive pricing posture on ArF immersion lithography tools, explicitly leveraging its in-house manufacturing (vs. ASML's reliance on a wider supplier ecosystem) to cut prices and target customers ASML had effectively monopolized. ArF immersion is the workhorse for mature-to-leading-edge nodes outside EUV territory — exactly where the China-domestic fabs, Powerchip, GlobalFoundries, and the new Tata Gujarat fab buy heavily. Nikon's move re-opens a competitive front ASML had largely closed off in the late 2010s; ASML's stock weakness this week (covered in the parallel Seeking Alpha downgrade) tracks the narrative. 尼康确认对其ArF浸没光刻设备采取激进定价策略,明确以自有产线(相对于ASML依赖更广供应商生态)作为价格杠杆,瞄准此前几乎被ASML垄断的客户群。ArF浸没光刻是EUV以外、从成熟到先进节点的主力工艺——正是中国本土晶圆厂、力积电、GlobalFoundries及塔塔古吉拉特新厂的高强度需求所在。尼康此举重新打开了ASML在2010年代后期已基本关闭的一条竞争战线;本周ASML股价的疲软(与Seeking Alpha同步下调评级的报道一致)也印证了这一叙事。

Korean prosecutors raid semiconductor packaging-material suppliers over alleged price collusion 韩国检方突查半导体封装材料供应商,涉嫌价格串通

TheElec (via Google News) · 4h ago TheElec (via Google News) · 4小时前

TheElec reports that Korean prosecutors executed search-and-seizure operations at multiple semiconductor packaging-material suppliers on suspicion of coordinated price increases — affecting underfill, encapsulants, and the resin classes that move into HBM and CoWoS-class packaging. The raids matter because materials-side price discipline (or the lack of it) flows directly into the gross-margin trajectory of Samsung and SK hynix, both at a moment when HBM4E pricing power is being established. A formal cartel finding would compress margins for the package material vendors and incrementally help the memory makers. TheElec报道,韩国检方对多家半导体封装材料供应商执行搜查扣押,涉嫌联合涨价——影响范围涵盖底部填充胶、封装胶以及进入HBM和CoWoS级封装的树脂材料。意义在于:材料端的价格纪律(或缺失)直接传导至三星与SK海力士的毛利率走势——而当前正值HBM4E的定价权确立之际。若最终认定卡特尔行为,将压缩封装材料厂利润,并边际利好存储厂。

Foundry & Manufacturing 晶圆代工与制造

TSMC says energy efficiency has overtaken raw performance as the top AI-chip priority 台积电表示:能效已超越裸性能,成为AI芯片客户的首要优先项

TweakTown (via Google News) · 12h ago TweakTown (via Google News) · 12小时前

TSMC executives told customers at a Hsinchu briefing this week that energy efficiency — perf/W — has displaced raw transistor performance as the dominant figure of merit AI customers ask for at A16 and A14. The message tracks the GW-scale power constraints documented in arXiv's recent +100 MW datacenter paper and the GridPilot grid-response work. Practically, this reshapes node-selection economics: customers who would have gravitated to leading-edge performance are now weighing backside-power-delivery and BSPDN-enabled efficiency wins against a node lag at the same energy budget. Roadmap implication: A16-with-BSPDN beats N2 for many AI workloads. 台积电高管本周在新竹的客户简报中表示:能效(perf/W)已取代裸晶体管性能,成为AI客户在A16与A14节点上最看重的指标。该信号与近期arXiv上「+100 MW数据中心论文」与「GridPilot电网响应」研究中所记录的吉瓦级电力约束完全一致。实质意义:节点选择经济学被重塑——原本会奔向最先进性能节点的客户,如今正在权衡背面供电(BSPDN)所带来的能效红利与对应的节点滞后。路线图含义:在很多AI负载下,「带BSPDN的A16」会跑赢N2。

Apple gives Intel's foundry turnaround real momentum 苹果给英特尔代工业务的复苏注入实质动能

Taipei Times (via Google News) · 21h ago Taipei Times (via Google News) · 21小时前

Taipei Times reports that Apple's ramping use of Intel Foundry for selected modem and connectivity silicon is materially helping Intel's foundry-revenue trajectory — a meaningful shift after years of speculation about whether Apple would ever commit non-trivial volume outside TSMC. The piece notes the new volume is on Intel 18A, not the most advanced node, but the legitimization effect on Intel Foundry's customer-pipeline credibility may be more valuable than the silicon-area dollars. Couple this with Intel's $3.3B India glass-substrate deal announced the same week, and Intel Foundry's 2H26 narrative is finally constructive. Taipei Times报道:苹果对英特尔代工的选择性采用——主要在调制解调器与连接性硅片上——正实质性帮助英特尔代工业务的收入曲线,这是多年来关于「苹果是否会向台积电以外厂商投入非琐碎产能」的猜测之后的一次实质转折。文章指出新增产能落在Intel 18A而非最先进节点,但对英特尔代工客户管线可信度的「正当化」效应,可能比硅面积本身带来的美元更宝贵。叠加同一周宣布的英特尔33亿美元印度玻璃基板交易,英特尔代工业务在2026年下半年的叙事终于由防守转为建设性。

Policy & Geopolitics 政策与地缘

Supermicro co-founder arrested in $2.5B Nvidia GPU China-smuggling case 美超微联合创始人涉嫌25亿美元英伟达GPU走私中国被捕

abhs.in (via Google News) · 8h ago abhs.in (via Google News) · 8小时前

The Taiwan Nvidia-smuggling probe (covered last week with the arrest of three intermediaries) escalated to a Supermicro co-founder being taken into custody, with the case now formally implicating an OEM founder rather than just channel intermediaries. The scope — $2.5B in alleged diverted GPUs — and the personal involvement of a recognizable industry figure both raise the political stakes for ongoing US-China export-control negotiations. It also forces a reckoning at every US-headquartered server OEM about whether downstream channel compliance, not chip-supplier compliance, is the actual bottleneck. 台湾英伟达走私案(上周报道时已抓捕3名中间人)升级——美超微一名联合创始人被收押,案件正式将OEM创始人本人纳入,而不再仅限于渠道中间人。涉案规模约25亿美元的转移GPU加上知名行业人物的个人卷入,使中美出口管制谈判的政治成本进一步抬高。该案也迫使每一家美国总部服务器OEM正视:真正瓶颈或在于下游渠道合规,而非上游芯片供应商合规。

Huawei chairman publicly thanks US for export restrictions — 'supercharged' China's chip industry 华为董事长公开「感谢」美国出口管制——称其「加速催熟」中国芯片产业

Tom's Hardware · 1d ago Tom's Hardware · 1天前

Huawei's chairman publicly credited US export restrictions with accelerating China's semiconductor self-sufficiency push — a deliberately provocative framing that arrived in the same news window as the company's LogicFolding / Tau Scaling Law architecture announcement. The rhetoric is strategic rather than spontaneous: it's targeted at the Beijing audience evaluating subsidy allocation for the next five-year plan, and at the Washington audience considering whether the existing controls are working. The Lisa Su / Jensen Huang Reuters analysis from the same day frames the broader posture difference: Huang is performative in China; Su is understated; both are still showing up. 华为董事长公开把美国出口管制定义为「加速催熟」中国半导体自给体系的因素——这是一个刻意挑衅性的表述,与公司同期发布的LogicFolding/Tau Scaling Law架构信号在同一新闻窗口出现。这种修辞是战略性的、非即兴的:面向北京评估下一个五年规划补贴分配的受众,也面向华盛顿评估现行管制是否奏效的受众。同日Reuters对Lisa Su与黄仁勋的分析勾勒出更大的姿态差异:黄仁勋在中国「表演式」、Lisa Su「低调」——但两人都还在去。

Earnings & Markets 财报与市场

Qualcomm tumbles, dragging semiconductor and AI stocks lower 高通暴跌,拖累半导体与AI股板块走低

Reuters (via Google News) · 2h ago Reuters (via Google News) · 2小时前

Qualcomm's sharp Friday selloff — triggered by guidance commentary tied to the Snapdragon C platform's $300 laptop positioning and uncertainty around Samsung modem business — pulled the broader SOX lower and ended the parabolic May rally on a defensive note. The macro read isn't earnings deterioration; it's a positioning unwind in a sector that's run hard. SOX exposure in the S&P 500 hit 18% (more than 2× the 2000 tech-bubble peak) this week, and Friday looked like the first day where that concentration started to feel uncomfortable to risk allocators. 高通周五大跌——起因是其与Snapdragon C平台「300美元笔记本」定位以及三星调制解调器业务不确定性相关的指引言论——拖累整个SOX指数走低,结束了五月的抛物线式行情,以防守姿态收尾。宏观解读不在业绩恶化,而是涨势凶猛板块的仓位平衡解构。本周费城半导体指数(SOX)在标普500中的权重达到18%,已超过2000年科技泡沫峰值的2倍以上;周五的盘面看起来是这种集中度第一次让风险配置者感到不舒服的一天。

Tabletop 3D printer cuts semiconductor 3D-patterning time from days to minutes 桌面级3D打印机将半导体3D图形化时间从「天」压缩到「分钟」

Tech Xplore (via Google News) · 17h ago Tech Xplore (via Google News) · 17小时前

Tech Xplore covered a tabletop two-photon 3D printer demonstration that compresses certain classes of semiconductor 3D-patterning work — currently done with stitched lithography or multi-step etch flows — from multi-day workflows down to minutes. The relevance isn't volume manufacturing (it isn't); it's R&D and prototyping for chiplet interposers, photonic-package waveguides, and panel-level packaging test vehicles, where iteration speed is currently the binding constraint. If the throughput holds up, this is the kind of tool that quietly reshapes who can prototype advanced packaging — labs and small foundries get capabilities that previously required a full equipment-set commitment. Tech Xplore报道了一台桌面级双光子3D打印机的演示成果,可将某些类别的半导体3D图形化工序——目前依靠拼接光刻或多步刻蚀实现——从多日工作流压缩到「分钟级」。其意义不在量产(确实不是),而在chiplet中介层、光子封装波导、面板级封装测试载体的研发与原型环节——这些环节的瓶颈正是迭代速度。若吞吐量站得住,这是一类悄然重塑「谁能做先进封装原型」的工具:实验室与小型代工厂获得了原先需要整套设备投入才有的能力。