semi·news
Headlines要闻 / Research研究 /
Monday, May 25, 2026 2026年5月25日 星期一

Huawei Bypasses EUV (On Paper); Micron Says the Memory Crunch Has Years To Run 华为宣称'绕开EUV';美光称存储紧张还要持续多年

Huawei unveils a LogicFolding architecture plus a 'Tau scaling law' it says replaces Moore's, claiming 1.4nm-class chips by 2031 without EUV. Micron tells JPMorgan the HBM/DRAM/NAND shortage will outlast 2026. Samsung gets to a 900-layer V-NAND prototype, Intel pushes glass substrates at Rio Rancho, and ASML's High-NA pivot to Asia takes the stock toward €1,900. 华为发布LogicFolding架构,提出意在替代摩尔定律的'Tau Scaling Law',宣称无需EUV即可在2031年量产1.4nm级芯片。美光向JPMorgan表示HBM/DRAM/NAND短缺将持续到2026年之后。三星完成首个900层V-NAND原型,英特尔在Rio Rancho推进玻璃基板,ASML的High-NA亚洲转向令股价指向1900欧元。

Policy & Geopolitics 政策与地缘

Huawei claims 1.4nm-class chips by 2031 via 'LogicFolding' — no EUV required 华为宣称2031年实现1.4nm级芯片——通过'LogicFolding'架构,无需EUV

Tom's Hardware · 50m ago Tom's Hardware · 50分钟前

Huawei unveiled a chip-design framework it calls 'LogicFolding,' built around what it markets as the 'Tau scaling law' (a replacement for Moore's Law). The pitch: 1.4nm-equivalent transistor density by 2031 with 55% higher density and dramatically lower power, **without** access to EUV lithography. The technical claim is unverified — no working silicon — but the *political* claim is unambiguous: China is asserting it can route around the export-control architecture entirely rather than chase it. Western analysts will spend the next quarter trying to decide whether this is real engineering progress or pre-Computex theater. 华为发布名为'LogicFolding'的芯片设计框架,核心是其所称替代摩尔定律的'Tau Scaling Law'。主张:无需EUV即可在2031年达到1.4nm等效晶体管密度,密度提高55%、功耗显著降低。技术主张尚无实测硅片佐证,但政治主张明确:中国正在表态——不再追赶出口管制架构,而是宣称可以完全绕开。西方分析师未来一个季度的核心争论将是:这是真实的工程突破,还是Computex前的舞台动作。

Huawei also says it can build 1.4nm chips without ASML's most advanced machines 华为另称:无需ASML最先进设备即可量产1.4nm芯片

TechSpot · 1h ago TechSpot · 1小时前

Companion statement out of Shenzhen — Huawei is explicitly framing the LogicFolding claim as 'no High-NA EUV needed.' Whether the design framework actually delivers the claimed density on DUV-only tooling is the open engineering question, but the messaging is the point: Huawei wants the international audience (and the U.S. policy audience) to hear 'sanctions did not work.' Watch for ASML to respond at Computex; the company's $400M+ High-NA tools are explicitly the moat that this announcement is targeting. 深圳同步发声:华为明确将LogicFresh主张定位为'无需High-NA EUV'。该设计框架能否在仅有DUV设备的条件下达到所述密度,是真正的工程未决问题;但叙事意图清晰——华为希望国际舆论(以及美国政策圈)听到'制裁未奏效'。下周Computex期间ASML的回应值得关注;其单台4亿美元以上的High-NA设备,正是这一表态明确瞄准的护城河。

Memory 存储器

Micron warns memory crunch will outlast 2026 as AI demand outpaces HBM/DRAM/NAND supply 美光警告:AI需求超出HBM/DRAM/NAND供给,存储紧张将延续至2026年后

Wccftech · 1h ago Wccftech · 1小时前

JPMorgan published an investor note today citing Micron management: the shortage across HBM, DRAM, and NAND will *not* be relieved in 2026 as previously hoped — AI demand is outpacing every supply lever the industry can pull (new wafer starts, node migration, packaging capacity, equipment lead times). This reframes the planning horizon for every hyperscaler and OEM whose cost models assumed a return to spot-price normalization next year. Combined with this week's hyperscaler-funded SK Hynix capex story, the picture is consistent: memory is now a multi-year strategic shortage. JPMorgan今日发布投资者备忘录,引用美光管理层观点:HBM、DRAM、NAND的短缺在2026年内不会如此前预期般缓解——AI需求超出了行业能够调动的全部供给杠杆(新晶圆产能、节点迁移、封装产能、设备交期)。这一表态重新校准了所有以'2026年现货价回归正常'为前提的超大规模厂商与OEM的成本模型。与本周'超大规模厂商出资支持SK海力士'的故事叠加,图景一致:存储已进入多年期战略性短缺。

Samsung shows first 900-layer V-NAND prototype, stacking two 450-layer cells 三星首次展示900层V-NAND原型:由两组450层堆叠合并而成

Wccftech · 2h ago Wccftech · 2小时前

Samsung disclosed a prototype 900-layer V-NAND fabricated by stacking two 450-layer chips into a single device. The 1,000-layer threshold long treated as a technical wall is now visibly within reach, though the prototype-to-product gap historically takes 18–24 months in NAND. The strategic question isn't 'can Samsung get there' but 'does the energy and yield cost of stacking past 600 layers earn its keep against TSMC-attached HBM' — the answer matters for whether NAND eats DRAM territory or stays in storage. 三星披露一款900层V-NAND原型:由两枚450层芯片堆叠而成。业内长期视为技术墙的1000层门槛现已肉眼可见,但NAND原型到量产历史上通常需要18–24个月。战略问题不是'三星能否抵达',而是'600层以上的能耗与良率代价是否值得'——答案决定NAND是侵入DRAM领地,还是仍守在存储侧。

AI & Accelerators AI与加速器

AMD Zen 7 'Grimlock' reportedly targets TSMC A14 for a 2028 launch 据报道,AMD Zen 7 'Grimlock'目标台积电A14节点,定档2028年

VideoCardz · 1h ago VideoCardz · 1小时前

Leaked AMD roadmap detail: Zen 7 (codename 'Grimlock') is targeting TSMC's A14 (1.4nm-class) for 2028. The two-generation gap from current Zen 6 / N2 implies AMD is *not* planning a refresh on A16 before A14 — a notable difference from prior cadences where AMD straddled two foundry nodes. If accurate, A14 is the cliff every architecture house will have to scale at roughly the same time, which puts unusual pressure on TSMC's A14 capacity allocation circa 2027. AMD路线图泄露细节:Zen 7(代号'Grimlock')目标台积电A14(1.4nm级)节点、2028年推出。从当前Zen 6 / N2跨两代直达A14意味着AMD不计划在A16上做中间刷新——这与AMD以往跨两代节点同时推出的节奏明显不同。若属实,A14将是所有架构厂商需在相近时间冲击的悬崖,2027年前后台积电A14产能分配将承受异常压力。

Intel preps Nova Lake edge variant: 8 E-cores + 12 Xe3P GPU cores, no P-cores Intel准备边缘版Nova Lake:8个E核+12个Xe3P GPU核,不含P核

Wccftech · 45m ago Wccftech · 45分钟前

Wccftech surfaces an Intel Nova Lake variant aimed specifically at the edge / industrial-AI segment — no Performance cores at all, just 8 Efficiency cores paired with a beefy 12-core Xe3P iGPU. The pattern is consistent with what Apple did at the M-series low end and what hyperscalers have been asking for in inference-only SKUs: GPU bandwidth matters, single-threaded CPU performance does not. Intel previously didn't ship a no-P-core SKU; the fact it's working on one now is a real concession to where edge AI buying is going. Wccftech曝出Intel Nova Lake的边缘/工业AI变体:完全不含Performance核,仅8个Efficiency核搭配较强的12核Xe3P iGPU。该模式与Apple在M系列低端做的事、以及超大规模厂商一直要求的'仅推理SKU'一致:GPU带宽是关键,单线程CPU性能不重要。Intel此前没有过无P核SKU,如今着手准备意味着公司对边缘AI采购走向的真实让步。

Foundry & Manufacturing 晶圆代工与制造

Intel Foundry positions Rio Rancho as its glass-substrate volume site Intel Foundry将Rio Rancho定位为玻璃基板量产基地

Wccftech · 1h ago Wccftech · 1小时前

Intel Foundry confirmed Rio Rancho (NM) will be the company's volume production site for glass substrates — the post-organic-substrate packaging technology Intel has been pushing as a structural advantage over TSMC's organic interposers. Glass substrates have lower CTE mismatch, better dimensional stability at panel scale, and lower loss at high frequencies; they have also been 'two years away' for about five years. Putting a specific fab on the map starts a clock that Intel will be measured against. Intel Foundry确认Rio Rancho(新墨西哥)将作为公司玻璃基板量产基地——这一替代有机基板的封装技术,被Intel视为对台积电有机中介层的结构性优势。玻璃基板在CTE失配、面板级尺寸稳定性、高频损耗上均更优;但'两年后量产'的承诺已说了大约五年。把具体晶圆厂落到地图上,意味着Intel开始接受时间表的检验。

ASU + TSMC Arizona launch a free semiconductor training program 亚利桑那州立大学与台积电亚利桑那联合推出免费半导体培训项目

KTAR News · 1h ago KTAR News · 1小时前

Arizona State University and TSMC Arizona announced a free semiconductor training program aimed at the Phoenix-area workforce. Skill-supply at the U.S. site has been the persistent operational gap behind the headline capex story — Fab 21's Q1 profit number proved economic viability, but staffing the next phases of expansion at U.S. wages is the binding constraint. State-university + foundry partnerships of this shape are the right institutional response; whether the throughput matches the hiring slope of the next two phases is the open question. 亚利桑那州立大学与台积电亚利桑那宣布,面向凤凰城地区劳动力推出免费的半导体培训项目。在头条资本开支故事背后,美国厂运营层面持续的短板一直是技能供给——Fab 21的一季度盈利证明了经济可行性,但以美国工资规模化招聘下一期扩产员工才是真正的约束。州立大学与代工厂的此类合作是合适的制度回应;吞吐能否跟上下两期招聘曲线是未决问题。

Equipment & Materials 设备与材料

ASML's High-NA era and Asia pivot drive record stock; analyst PT reaches €1,900 ASML步入High-NA时代+亚洲转向,创新高股价;分析师目标价指向1900欧元

Ad Hoc News · 1h ago Ad Hoc News · 1小时前

ASML's stock hit a record this week as sell-side analysts marked PTs up to €1,900 on a combination of confirmed High-NA EUV production orders (TSMC, Intel, Samsung) and a visible 'Asia pivot' — meaning a larger share of bookings now coming from Asian foundries as European logic capex stays flat. The Huawei-bypass-EUV story (above) is the bear case the bulls will have to address; sell-side has so far treated it as posturing rather than a real demand risk. ASML本周股价创新高,卖方分析师将目标价上调至1900欧元,叠加因素为:已确认的High-NA EUV量产订单(台积电、Intel、三星)+可见的'亚洲转向'——亚洲代工厂在订单中的比例上升,欧洲逻辑资本开支保持平稳。上文的'华为绕开EUV'是看多方需要回应的看空叙事;卖方目前将其视为姿态而非真实需求风险。

Challengers & New Silicon 新势力与新硅

Multi-kilowatt accelerators make direct-to-chip liquid cooling mandatory 多千瓦级加速器使直接到芯片(D2C)液冷成为强制配置

Electronics360 · 1h ago Electronics360 · 1小时前

Electronics360 piece on what the industry is quietly converging on: with single-package accelerators (Nvidia Rubin, AMD MI455X, Cerebras CS-class, Tenstorrent's roadmap) clearing 3–5 kW per device, **air cooling is functionally obsolete** for any serious deployment. Direct-to-chip (D2C) cold plates are now the floor; immersion is the next-step. The implication for the challenger layer is more interesting than for the incumbents — startups whose differentiator was 'fits in your existing rack' are going to have to rebuild around liquid before customers will look at them. Electronics360的一篇文章,记录了行业正悄悄达成的共识:当单封装加速器(Nvidia Rubin、AMD MI455X、Cerebras CS系列、Tenstorrent路线图)超过3–5 kW后,任何严肃部署的风冷方案功能上已经过时。直接到芯片(D2C)冷板成为下限,浸没式是下一步。对新势力的影响比对巨头更重要——以'兼容现有机柜'为差异点的初创公司,必须围绕液冷重建产品形态,客户才会认真考虑。