[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"content-doc-dcioa8f12c3d":3},{"user":4,"document":8,"mainDocument":27,"columnUrl":29,"subscription":30,"footer":42,"text":80},{"isAuthenticated":5,"isAdmin":5,"displayName":6,"avatarUrl":6,"nid":6,"groupLevel":7},false,"",-10,{"id":9,"fullTitle":10,"subTitle":6,"url":11,"columnId":12,"columnName":13,"columnUrl":14,"summary":6,"contentHtml":15,"mainContentHtml":6,"posterUrl":16,"createDate":17,"displayDate":18,"displayDateSlash":19,"pageviews":20,"tags":21,"hidden":5,"isSubContent":5,"replyDocOrTargetId":6,"contentType":23,"videoId":6,"liveVideoUrl":6,"useContentVideo":5,"duration":24,"price":24,"priceText":25,"priceBadgeText":25,"priceBadgeClass":26,"freeForMinGroupLevel":24,"redirectUrl":6,"readyToStream":5},"dcioa8f12c3d","大摩：AI-内存短缺将持续数年-三大路径破局-存储与异构算力迎结构性机会","\u002Fdoc\u002Fdcioa8f12c3d","col18178739ee","美股资讯","\u002Fcol\u002Fcol18178739ee","\u003Cp> \u003C\u002Fp>\n\u003Cp>\u003Cspan>美股投资网获悉，摩根士丹利发布半导体行业研报指出，DRAM内存短缺将贯穿本轮AI产业周期，而AI算力扩张不会等待新晶圆厂建成投产。行业正通过硬件降配、推理架构解耦、CXL内存池化三大技术路径绕过内存瓶颈，在供给约束下持续推进算力建设。该行继续看好美光(MU)、闪迪(SNDK)等存储龙头，同时看好CXL互联、异构推理赛道的增量机遇，推荐Astera Labs(ALAB)、Marvell(MRVL)、Cerebras(CBRS)及英伟达(NVDA)。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>大摩强调，当前 AI 内存短缺并非短期扰动，而是算力性能增长远超内存供给速度的结构性矛盾。前沿大模型规模每 6 个月翻一番，头部模型上下文窗口年增速达 5-6 倍，叠加推理并发需求持续提升，内存容量与带宽需求持续爆发。而 DRAM 晶圆厂建设周期长达数年，供给弹性极低，短缺格局将持续数年。英伟达CEO黄仁勋也公开表示，行业需要转变思路，通过架构创新应对内存约束，而非单纯等待产能扩张。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>面对 HBM 与主存供给紧张，行业最直接的应对方式是选择性降低单设备内存规格(De-speccing)。以英伟达 Rubin 架构为例，单机架 LPDDR5 容量从原计划 54TB 降至 28TB，单 GPU HBM 容量从 288GB 下调至 192GB，通过降低内存堆叠高度保障整机出货量。大摩指出，降配并非消除内存瓶颈，而是将压力在内存层级间转移削减高速本地内存后，KV 缓存等非高频数据向 NAND 存储下沉，同时跨 GPU 数据交互增加，带动对网络互联带宽的需求，本质上是以更高速的网络互联、更低成本的存储资源，补充稀缺的高带宽内存。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>第二条破局路径是推理架构解耦(Disaggregation)。AI推理包含预填充(Prefill)与解码(Decode)两个差异显著的阶段Prefill 以算力消耗为主，Decode则高度依赖内存带宽。传统架构用同一加速器兼顾两类任务，资源利用率偏低。当前行业正加速走向异构推理，将两个阶段拆分至不同硬件算力密集型Prefill由通用GPU承担，内存带宽密集型 Decode 则由搭载大容量片上 SRAM 的专用加速芯片处理。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>典型代表包括Cerebras的晶圆级引擎、英伟达收购的 Groq LPU 架构，均能在解码阶段提供远超传统 GPU 的内存带宽效率。大摩认为，异构推理将成为 AI 基础设施的重要演进方向，专精解码环节的算力厂商将获得明确的增量市场。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>第三条路径是 CXL 技术重构内存体系。CXL(计算高速互联)通过高速互联实现内存的扩展、共享与池化，让内存不再绑定单一处理器，成为突破内存容量约束的核心技术路径。大摩测算，AI需求将推动 CXL 及相关内存附加芯片市场规模在 2030 年达到约 60 亿美元，远超传统 CPU 内存扩展市场体量。其核心价值在于构建分层内存架构最高频数据保留在HBM，次高频数据放入 CXL 内存池，低频冷数据下沉至 NAND，用成本梯度匹配数据访问频率。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>投资主线方面，大摩重申三大方向一是继续超配美光、闪迪，降配源于供给短缺而非需求走弱，AI内存需求长期向上，供给缺口将持续消化产能；二是布局 CXL 与 scale-up 互联龙头 Astera Labs、Marvell；三是关注异构推理受益者Cerebras，以及通过收购Groq完善异构布局的英伟达。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>风险提示AI 算力需求增速不及预期；CXL 技术落地进度慢于预期；内存产能扩张超预期。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>最专业的美股资讯,推荐美股大数据 \u003Ca href=\"https:\u002F\u002FStockwe.com\u002F\">https:\u002F\u002FStockwe.com\u002F\u003C\u002Fa>\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>\u003Cspan>如何识别美股市场异常波动？美国机构主力资金买卖情况，出货和吸筹，使用美股投资网VIP会员，2008年成立于美国硅谷，由前纽约证券交易所分析师Ken创立，联合多位摩根斯坦利分析师，谷歌 Meta工程师利用AI和大数据，配合十多年美股实战经验和业内量化模型，建立了一个股市数据库 \u003Ca href=\"https:\u002F\u002FStockWe.com\u002F\">https:\u002F\u002FStockWe.com\u002F\u003C\u002Fa> 每天处理千万级股票数据：捕捉期权大单，实时主力资金流向、机构持仓变化、川普突发新闻，精准交易信号第一时间发到您手机APP！\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp> \u003C\u002Fp>","https:\u002F\u002Fwww.tradesmax.com\u002Fimages\u002Fa_Stock\u002FA\u002FAI\u002FAI.jpg","2026-10-06T15:55:04","2026.10.06","2026\u002F10\u002F06",40397,[22],"AI","Article",0,"免费","success",{"id":9,"fullTitle":10,"subTitle":6,"url":11,"columnId":12,"columnName":13,"columnUrl":14,"summary":6,"contentHtml":15,"mainContentHtml":6,"posterUrl":16,"createDate":17,"displayDate":18,"displayDateSlash":19,"pageviews":20,"tags":28,"hidden":5,"isSubContent":5,"replyDocOrTargetId":6,"contentType":23,"videoId":6,"liveVideoUrl":6,"useContentVideo":5,"duration":24,"price":24,"priceText":25,"priceBadgeText":25,"priceBadgeClass":26,"freeForMinGroupLevel":24,"redirectUrl":6,"readyToStream":5},[22],"\u002Fcol\u002Fstocknews",{"visible":5,"marketingHtml":31,"services":32,"recentDocuments":41},"\u003Cfigure class=\"image\">\u003Ca href=\"https:\u002F\u002Fstockwe.com\u002Fdoc\u002Fdcio537efad5\" rel=\"noopener noreferrer\">\u003Cimg src=\"\u002Fimg\u002Fstockwewebfiles\u002Fweb-202408-stk\u002F1586109431mceclip0.jpg\">\u003C\u002Fa>\u003C\u002Ffigure>\u003Cdiv class=\"text-center\">\u003Ch2 class=\"card-title mx-auto\">\u003Cbr>\u003Ca rel=\"noopener noreferrer\" href=\"https:\u002F\u002Fstockwe.com\u002Fdoc\u002Fdcio537efad5\">案例介绍：英伟达深度研究报告\u003C\u002Fa>\u003C\u002Fh2>\u003C\u002Fdiv>",[33,37],{"productId":34,"serviceName":35,"priceText":36},"prod_PPxdDdK87QaiLv","月付","$12.95美元",{"productId":38,"serviceName":39,"priceText":40},"prod_PPxeMs3bix1da5","年付","$149.00美元",[],{"links":43,"images":71,"summaryHtml":76,"aboutTitle":77,"aboutHtml":78,"copyrightHtml":79},[44,47,50,53,56,59,62,65,68],{"label":45,"url":46},"深度报告","\u002Fcol\u002FdepthReport",{"label":48,"url":49},"VIP会员","\u002Fvip",{"label":51,"url":52},"期权推荐","\u002FOption",{"label":54,"url":55},"低价暴涨股","\u002FPenny",{"label":57,"url":58},"AI智能体","\u002FAiAgent",{"label":60,"url":61},"常见问题","https:\u002F\u002Fstockwe.com\u002FFAQ",{"label":63,"url":64},"美股课程","\u002Fcol\u002Fvideos",{"label":66,"url":67},"免责声明","\u002Fdisclaimer",{"label":69,"url":70},"联系我们","\u002FContactUs",[72,73,74,75],"\u002Fimages\u002Ffooter\u002Ffooter-1-wechat.gif?v=20260930c","\u002Fimages\u002Ffooter\u002Ffooter-2-telegram.png?v=20260930c","\u002Fimages\u002Ffooter\u002Ffooter-3-whatsapp.png?v=20260930c","\u002Fimages\u002Ffooter\u002Ffooter-4-line.png?v=20260930c","邮箱: buy@TradesMax.com 美国电话 626-378-3637","公司介绍","\u003Cp class=\"MsoNormal\">美股大数据 \u003Ca href=\"https:\u002F\u002Fstockwe.com\" rel=\"noopener\">StockWe.com\u003C\u002Fa> 是一个美国领先的金融和美股信息大数据提供商，紧盯华尔街金融市场和行情，2008年成立于美国硅谷，创始人是前纽约证券交易所资深分析师Ken，联合多位摩根斯坦利分析师，谷歌 Meta工程师利用AI和大数据，配合十多年美股实战经验和业内量化交易模型，每天处理海量股票数据：挖掘潜力大牛股，捕捉期权异动大单，实时主力资金流向、机构持仓变化、川普突发新闻，美股买卖信号第一时间发到您手机APP。\u003C\u002Fp>","专业美股投资者都在这里",{"loading":81,"search":82,"searchPlaceholder":82,"hotContent":83,"draft":84,"noData":85,"searchNoData":86,"courseContent":87,"more":88,"buyNow":89,"subscribeNow":90,"encoding":91,"paidContent":92},"Loading...","搜索","热门内容","草稿","目前没有任何内容公布","当前检索内容没有数据","课程内容","更多","立即购买后观看","- 立即订阅 -","视频编码中...","付费内容"]