BV1T5XWB1EdK · 浑水摸鱼清源1 00:00:00,120 --> 00:00:01,879 我刚从GTC回来 2 00:00:01,879 --> 00:00:05,759 我确信华尔街并不理解英伟达 3 00:00:05,759 --> 00:00:10,769 这是因为最好的长期投资来自理解公司的产品 4 00:00:10,769 --> 00:00:12,330 而不仅仅是他们的利润 5 00:00:12,330 --> 00:00:13,890 经过我所见的一切 6 00:00:13,890 --> 00:00:18,050 我相信英伟达将成为地球上首个市值突破 7 00:00:18,050 --> 00:00:20,010 一万亿美元的公司 8 00:00:20,010 --> 00:00:22,410 让我展示为什么你的时间有价值 9 00:00:22,410 --> 00:00:24,129 那我们直接进入正题 10 00:00:24,129 --> 00:00:24,690 看 11 00:00:24,690 --> 00:00:27,129 我不是来复述黄仁勋的演讲的 12 00:00:27,129 --> 00:00:27,670 相反 13 00:00:27,670 --> 00:00:31,149 我想分享我在参加GTC时的所学所得 14 00:00:31,149 --> 00:00:31,719 我自己 15 00:00:31,719 --> 00:00:33,799 采访英伟达高管 16 00:00:33,799 --> 00:00:35,659 试用原型机器人 17 00:00:35,659 --> 00:00:37,340 乘坐自动驾驶汽车 18 00:00:37,340 --> 00:00:38,920 触摸量子计算机 19 00:00:38,920 --> 00:00:41,560 甚至还和黄仁勋本人交谈 20 00:00:41,560 --> 00:00:43,179 在所有重大宣布之后 21 00:00:43,179 --> 00:00:46,340 主流媒体和华尔街分析师聚焦于 22 00:00:46,340 --> 00:00:48,359 英伟达的新Rubin GPU 23 00:00:48,359 --> 00:00:50,560 但我认为他们忽略了更大的图景 24 00:00:50,560 --> 00:00:53,359 Vera Rubin 不仅仅关乎更快的芯片 25 00:00:53,359 --> 00:00:59,289 这是整个AI革命的蓝图，对数据中心支出影响深远 26 00:00:59,289 --> 00:01:01,829 未来AI系统的设计方向 27 00:01:01,829 --> 00:01:05,228 当然还有哪些股票将因此获利 28 00:01:05,228 --> 00:01:08,688 让我分解我在GTC学到的最重要内容 29 00:01:08,688 --> 00:01:10,069 以及让我感到意外的地方 30 00:01:10,069 --> 00:01:13,448 英伟达的Vera Rubin平台和新Grok 31 00:01:13,448 --> 00:01:14,680 三款推理芯片 32 00:01:14,680 --> 00:01:19,079 这些硬件如何融入英伟达的AI战略，特别是代理系统 33 00:01:19,079 --> 00:01:20,150 比如OpenClaw 34 00:01:20,150 --> 00:01:24,349 在自动驾驶汽车和人形机器人中看到的惊人之处 35 00:01:24,349 --> 00:01:27,709 以及这对我看英伟达股票前景的启示 36 00:01:27,709 --> 00:01:29,549 让我感到意外的一点是 37 00:01:29,549 --> 00:01:33,109 英伟达的Vera Rubin平台与Blackwell有何不同 38 00:01:33,109 --> 00:01:34,939 这不仅仅是一个更快的系统 39 00:01:34,939 --> 00:01:40,400 许多标题暗示Rubin在网络、内存甚至计算方面采用了根本不同的方法 40 00:01:40,400 --> 00:01:42,239 内存和计算 41 00:01:42,239 --> 00:01:45,640 这出于两个重要原因 42 00:01:45,640 --> 00:01:49,079 首先AI模型不再只训练一次 43 00:01:49,079 --> 00:01:52,539 而是通过强化学习持续微调 44 00:01:52,539 --> 00:01:58,698 其次AI工作负载正从人类编写的短提示转向自主代理 45 00:01:58,698 --> 00:01:59,879 比如OpenClaw 46 00:01:59,879 --> 00:02:01,859 Perplexity计算机和Claude 47 00:02:01,859 --> 00:02:03,500 这些代理会调用工具 48 00:02:03,500 --> 00:02:04,920 浏览网站 49 00:02:04,920 --> 00:02:08,610 编写代码并一次性处理数百万个标记 50 00:02:08,610 --> 00:02:12,710 这消耗的标记量是普通聊天提示的数千倍 51 00:02:12,710 --> 00:02:14,110 这使得功耗效率更高 52 00:02:14,110 --> 00:02:15,449 低延迟推理 53 00:02:15,449 --> 00:02:17,800 人工智能的新主要成本驱动因素 54 00:02:17,800 --> 00:02:22,099 这就是为什么我预计数据中心支出将加速增长 55 00:02:22,099 --> 00:02:23,159 而非放缓 56 00:02:23,159 --> 00:02:24,879 就像大多数分析师预测的那样 57 00:02:24,879 --> 00:02:29,099 这也是vera rubin系统与blackwell的根本不同之处 58 00:02:29,099 --> 00:02:32,930 旨在尽可能生成更多有效token 59 00:02:32,930 --> 00:02:33,759 每RAC 60 00:02:33,759 --> 00:02:35,379 每瓦特和每美元 61 00:02:35,379 --> 00:02:40,449 因此这些开源布料风格代理实际上可以大规模部署 62 00:02:40,449 --> 00:02:44,229 英伟达在ruin平台发布了七款新芯片 63 00:02:44,229 --> 00:02:45,819 我想尊重您的时间 64 00:02:45,819 --> 00:02:47,280 所以我将全部列出 65 00:02:47,280 --> 00:02:50,439 但随后我会聚焦对投资者真正重要的两款 66 00:02:50,439 --> 00:02:52,900 ruin GPU是核心AI芯片 67 00:02:52,900 --> 00:02:57,860 配备新Transformer引擎，token吞吐量远超blackwell 68 00:02:57,860 --> 00:03:00,180 推理性能提升约五倍 69 00:03:00,180 --> 00:03:02,719 训练性能提高三倍五成 70 00:03:02,719 --> 00:03:05,830 并降低token成本超90% 71 00:03:05,830 --> 00:03:08,810 媒体标题正确强调了这些惊人改进 72 00:03:08,810 --> 00:03:11,639 但稍后我会展示它们对整体局势的意义 73 00:03:11,639 --> 00:03:16,360 vera rubin CPU是基于ARM的处理器，含88个定制核心 74 00:03:16,360 --> 00:03:20,319 专为处理GPU不擅长的复杂任务而设计 75 00:03:20,319 --> 00:03:22,099 如任务调度与控制 76 00:03:22,099 --> 00:03:22,740 分支处理 77 00:03:22,740 --> 00:03:24,379 逻辑运算和数据预处理 78 00:03:24,379 --> 00:03:28,149 vera CPU调度并协调多代理工作负载 79 00:03:28,149 --> 00:03:30,508 处理API和工具调用 80 00:03:30,508 --> 00:03:35,110 并运行同一机架上安装的其他软件和服务 81 00:03:35,110 --> 00:03:36,909 例如数据日志记录 82 00:03:36,909 --> 00:03:38,629 监控安全服务 83 00:03:38,629 --> 00:03:39,710 等等 84 00:03:39,710 --> 00:03:42,370 vera内存容量约是三倍 85 00:03:42,370 --> 00:03:44,509 每个核心内存带宽翻倍 86 00:03:44,509 --> 00:03:48,259 与grace相比GPU连接速度加倍 87 00:03:48,259 --> 00:03:51,159 还可实现全机密计算 88 00:03:51,159 --> 00:03:53,620 而grace CPU不具备此功能 89 00:03:53,620 --> 00:03:54,259 所以是的 90 00:03:54,259 --> 00:03:57,259 CPU在AI领域仍至关重要 91 00:03:57,259 --> 00:04:00,460 只是与传统CPU差异显著 92 00:04:00,460 --> 00:04:03,539 我们熟悉nvlink六交换芯片 93 00:04:03,539 --> 00:04:06,840 在机架层面连接所有72颗GPU 94 00:04:06,840 --> 00:04:11,580 nvlink六将带宽提升至前代约三倍 95 00:04:11,580 --> 00:04:13,099 六太字节每秒 96 00:04:13,099 --> 00:04:16,639 足以每秒传输约250部全长度 97 00:04:16,639 --> 00:04:19,928 4K电影在芯片间 98 00:04:19,928 --> 00:04:23,288 kinectx九超级尼克是网络接口卡 99 00:04:23,288 --> 00:04:27,970 位于每个计算托盘，用于在网路与GPU内存间传输数据 100 00:04:27,970 --> 00:04:29,908 同时加密通信流量 101 00:04:29,908 --> 00:04:31,608 因此网络保持快速 102 00:04:31,608 --> 00:04:32,988 可预测且安全 103 00:04:32,988 --> 00:04:34,939 随着更多机架被添加 104 00:04:34,939 --> 00:04:39,240 Spectrum6以太网交换机提供连接Ruben机架的骨干网络 105 00:04:39,240 --> 00:04:42,810 并与存储舱及共封装光模块协同工作 106 00:04:42,810 --> 00:04:46,350 这是系统中NVIDIA投入20亿美元研发的部分 107 00:04:46,350 --> 00:04:50,649 非相干票据代码CHR和动量票据代码 108 00:04:50,649 --> 00:04:52,769 LITE符号开始发挥作用 109 00:04:52,769 --> 00:04:53,569 请留言 110 00:04:53,569 --> 00:04:56,750 如果你想让我制作关于光网络的完整视频 111 00:04:56,750 --> 00:05:00,490 因为这项技术旨在提升网络韧性 112 00:05:00,490 --> 00:05:03,009 更少错误且更节能 113 00:05:03,009 --> 00:05:04,350 接下来这两款芯片 114 00:05:04,350 --> 00:05:07,850 Grock3 LPU和BlueField4 DPU 115 00:05:07,850 --> 00:05:10,579 我认为NVIDIA在这里创新最多 116 00:05:10,579 --> 00:05:11,800 而GPU 117 00:05:11,800 --> 00:05:15,040 CPU和网络芯片则明显升级 118 00:05:15,040 --> 00:05:16,199 Grock3重新定义 119 00:05:16,199 --> 00:05:18,360 令牌生成机制的工作原理 120 00:05:18,360 --> 00:05:23,000 BlueField4为AI代理新增上下文内存层 121 00:05:23,000 --> 00:05:23,620 顺便说一句 122 00:05:23,620 --> 00:05:24,920 最新研究显示 123 00:05:24,920 --> 00:05:29,779 每天使用AI的美国员工收入比未使用者高40% 124 00:05:29,779 --> 00:05:31,980 这意味着AI不再是可选 125 00:05:31,980 --> 00:05:35,918 这是你拥有或被他人超越的优势 126 00:05:35,918 --> 00:05:37,639 这就是Outskill的作用 127 00:05:37,639 --> 00:05:39,079 本视频的赞助商 128 00:05:39,079 --> 00:05:41,038 Outskill正在举办为期两天的 129 00:05:41,038 --> 00:05:43,139 AI战略研讨会本周末 130 00:05:43,139 --> 00:05:44,399 十六小时培训 131 00:05:44,399 --> 00:05:50,050 汇聚来自NVIDIA、微软等百位专家知识 132 00:05:50,050 --> 00:05:53,329 助你更自信地自主使用AI 133 00:05:53,329 --> 00:05:57,879 紧跟工具快速迭代并将其技能转化为更高价值 134 00:05:57,879 --> 00:05:58,939 更高薪酬的工作 135 00:05:58,939 --> 00:06:03,310 前一千名通过我的链接注册者获免费C 136 00:06:03,310 --> 00:06:06,709 无论你在科技、销售、管理或营销领域 137 00:06:06,709 --> 00:06:08,569 将学习使用AI代理 138 00:06:08,569 --> 00:06:10,310 创建自动化流程 139 00:06:10,310 --> 00:06:14,430 并连接到日常使用的软件和电子表格 140 00:06:14,430 --> 00:06:17,170 这是提升AI知识的绝佳途径 141 00:06:17,170 --> 00:06:21,680 获得真实竞争优势并理解技术背后的逻辑 142 00:06:21,680 --> 00:06:25,879 全球已有超过1000万人参加 143 00:06:25,879 --> 00:06:28,939 本次名额正以更快速度被填满 144 00:06:28,939 --> 00:06:33,069 请立即通过下方链接注册获取免费席位 145 00:06:33,069 --> 00:06:33,810 好的 146 00:06:33,810 --> 00:06:35,029 我们先从Grock开始 147 00:06:35,029 --> 00:06:39,040 这是投资者最需理解的重要收购 148 00:06:39,040 --> 00:06:43,279 Grock3芯片似乎正在取代Ruin Cpx GPU 149 00:06:43,279 --> 00:06:46,139 NVIDIA最初为此设计用于推理 150 00:06:46,139 --> 00:06:49,139 但这款Grock芯片并非GPU 151 00:06:49,139 --> 00:06:50,500 这是一个LPU 152 00:06:50,500 --> 00:06:52,339 语言处理单元 153 00:06:52,339 --> 00:06:53,279 这简直疯狂 154 00:06:53,279 --> 00:06:56,180 仅仅NVIDIA整合它的速度 155 00:06:56,180 --> 00:07:00,319 NVIDIA宣布以200亿美元授权Gro技术 156 00:07:00,319 --> 00:07:04,019 并在12月24日雇佣了核心工程团队 157 00:07:04,019 --> 00:07:06,778 2025年黄仁勋展示了首个Grock 158 00:07:06,778 --> 00:07:09,848 在GTC演讲中展示了三LPU机架 159 00:07:09,848 --> 00:07:14,348 从收购宣布到首次公开演示仅三个月 160 00:07:14,348 --> 00:07:17,790 从交易开始到首颗芯片仅九个月 161 00:07:17,790 --> 00:07:20,870 这比大多数初创公司还要快 162 00:07:20,870 --> 00:07:22,449 NVIDIA行动如此迅速 163 00:07:22,449 --> 00:07:23,610 因为每个Croc 164 00:07:23,610 --> 00:07:27,509 LPU内置512MB片上SRAM 165 00:07:27,509 --> 00:07:28,970 存储模型权重 166 00:07:28,970 --> 00:07:31,259 激活值和KV缓存 167 00:07:31,259 --> 00:07:34,220 而非分散到外部DRAM 168 00:07:34,220 --> 00:07:36,060 我知道这听起来像字母组合 169 00:07:36,060 --> 00:07:37,689 让我用英文解释 170 00:07:37,689 --> 00:07:39,110 静态随机存取 171 00:07:39,110 --> 00:07:41,149 内存或SRAM容量小 172 00:07:41,149 --> 00:07:42,459 但速度极快 173 00:07:42,459 --> 00:07:44,220 直接集成在芯片上 174 00:07:44,220 --> 00:07:46,420 每比特成本高且耗电 175 00:07:46,420 --> 00:07:51,470 但内存访问可预测且低延迟 176 00:07:51,470 --> 00:07:53,370 DRAM容量大 177 00:07:53,370 --> 00:07:54,839 但外部内存较慢 178 00:07:54,839 --> 00:07:56,459 每比特成本低 179 00:07:56,459 --> 00:07:58,800 适合大容量存储 180 00:07:58,800 --> 00:08:02,980 但访问耗时更多 181 00:08:02,980 --> 00:08:04,300 在演讲中 182 00:08:04,300 --> 00:08:07,269 黄仁勋专门讲解了这点差异 183 00:08:07,269 --> 00:08:09,829 Ruin GPU拥有288GB高带宽内存 184 00:08:09,829 --> 00:08:12,209 这是DRAM 185 00:08:12,209 --> 00:08:13,509 而一个Grock LPU有512MB内存 186 00:08:13,509 --> 00:08:17,910 容量几乎少600倍 187 00:08:17,910 --> 00:08:20,610 我的观点是这些是根本不同的芯片 188 00:08:20,610 --> 00:08:24,170 服务于AI链条的不同环节 189 00:08:24,170 --> 00:08:27,069 甚至位于独立机架 190 00:08:27,069 --> 00:08:29,199 NVIDIA的LPU机架连接256个Grock 191 00:08:29,199 --> 00:08:33,559 构建专用超低延迟解码路径 192 00:08:33,559 --> 00:08:38,490 而Ruin GPU专注于训练 193 00:08:38,490 --> 00:08:41,220 预填充和注意力机制 194 00:08:41,220 --> 00:08:42,600 如果回顾NVIDIA原计划 195 00:08:42,600 --> 00:08:45,240 曾有Ruin CPX GPU专用于长上下文推理 196 00:08:45,240 --> 00:08:50,929 我甚至专门制作过视频介绍 197 00:08:50,929 --> 00:08:52,909 该芯片已从最新资料中消失 198 00:08:52,909 --> 00:08:55,519 Grock系统已取代其位置 199 00:08:55,519 --> 00:08:58,700 换句话说 200 00:08:58,700 --> 00:09:00,139 所以说 201 00:09:00,139 --> 00:09:02,419 英伟达投入了二十亿美元 202 00:09:02,419 --> 00:09:06,429 他们在不到一年时间内将Gro的LPU架构整合到系统中 203 00:09:06,429 --> 00:09:09,889 并悄然用自家的Ruin CPX加速器进行替换 204 00:09:09,889 --> 00:09:11,990 采用的方案每瓦能提供最高三倍 205 00:09:11,990 --> 00:09:14,970 五倍更高的推理吞吐量 206 00:09:14,970 --> 00:09:18,870 在服务大型模型时每操作每秒收入提升十倍 207 00:09:18,870 --> 00:09:20,090 我真心认为 208 00:09:20,090 --> 00:09:22,330 我们将回顾英伟达的Gro克交易 209 00:09:22,330 --> 00:09:24,370 视为他们最重要的收购 210 00:09:24,370 --> 00:09:25,480 自Mellanox以来 211 00:09:25,480 --> 00:09:31,110 Mellanox正是让英伟达掌控了GPU全系列的网络技术 212 00:09:31,110 --> 00:09:31,970 X以太网 213 00:09:31,970 --> 00:09:33,289 量子和FinBand 214 00:09:33,289 --> 00:09:35,450 以及他们的BlueField GPU 215 00:09:35,450 --> 00:09:37,940 到目前为止我们讨论了Ruin GPU 216 00:09:37,940 --> 00:09:40,919 Vera CPU和Grok LPU 217 00:09:40,919 --> 00:09:44,659 但BlueField四代才是真正将所有组件串联起来的关键 218 00:09:44,659 --> 00:09:48,328 BlueField四代是数据处理单元或DPU 219 00:09:48,328 --> 00:09:50,749 位于Vera Rubin计算机架内部 220 00:09:50,749 --> 00:09:52,408 Rock LPU X射线 221 00:09:52,408 --> 00:09:55,289 以及独立的上下文内存和存储机架 222 00:09:55,289 --> 00:09:57,529 每个机架的LPU负责 223 00:09:57,529 --> 00:09:58,330 网络通信 224 00:09:58,330 --> 00:10:00,809 内存访问和数据控制 225 00:10:00,809 --> 00:10:05,029 使GPU和LPU能专注于生成token 226 00:10:05,029 --> 00:10:06,450 在存储方面 227 00:10:06,450 --> 00:10:11,740 BlueField是英伟达新TX上下文内存机架的核心处理器 228 00:10:11,740 --> 00:10:17,940 这些机架将长期代理上下文存储在独立硬盘而非廉价GPU内存 229 00:10:17,940 --> 00:10:22,370 并在需要时及时将数据拉回GPU 230 00:10:22,370 --> 00:10:24,720 这就是Rubin保持token速度高效的原因 231 00:10:24,720 --> 00:10:29,360 同时为长上下文窗口的代理降低五倍功耗 232 00:10:29,360 --> 00:10:31,299 这意味着性能上的提升 233 00:10:31,299 --> 00:10:32,360 在机架层面 234 00:10:32,360 --> 00:10:34,720 搭配Vera Rubin机架与Grok 235 00:10:34,720 --> 00:10:37,340 三LPU机架可生成最高三 236 00:10:37,340 --> 00:10:39,789 五倍的每瓦推理token 237 00:10:39,789 --> 00:10:45,169 单个TX上下文内存机架每秒token输出提升五倍 238 00:10:45,169 --> 00:10:49,350 长上下文任务功耗效率提升五倍 239 00:10:49,350 --> 00:10:50,929 综合来看 240 00:10:50,929 --> 00:10:52,289 Ruin GPU 241 00:10:52,289 --> 00:10:53,730 Vera CPU 242 00:10:53,730 --> 00:10:54,389 Grok 243 00:10:54,389 --> 00:10:56,190 三LPU和BlueField 244 00:10:56,190 --> 00:10:57,200 四DPU 245 00:10:57,200 --> 00:11:00,100 以及上下文内存和网络架构 246 00:11:00,100 --> 00:11:04,019 这标志着英伟达硬件组合的全面革新 247 00:11:04,019 --> 00:11:06,190 数据中心可灵活组合配置 248 00:11:06,190 --> 00:11:06,889 例如 249 00:11:06,889 --> 00:11:11,960 专注于训练和批量推理的数据中心主要部署Vera Rubin 250 00:11:11,960 --> 00:11:13,679 NVL72机架 251 00:11:13,679 --> 00:11:14,620 但为了实时性 252 00:11:14,620 --> 00:11:17,339 代理型工作负载或延迟真的至关重要 253 00:11:17,339 --> 00:11:22,259 詹森建议数据中心约25%的算力可转向新架构 254 00:11:22,259 --> 00:11:23,360 Lpx机架 255 00:11:23,360 --> 00:11:26,500 这是投资者需要关注的另一个洞察点 256 00:11:26,500 --> 00:11:28,019 华尔街分析师正在讨论 257 00:11:28,019 --> 00:11:32,350 我们应该关注英伟达数据中心营收的两个关键原因 258 00:11:32,350 --> 00:11:32,870 首先 259 00:11:32,870 --> 00:11:37,619 鲁宾为英伟达提供了突破性扩展新途径 260 00:11:37,619 --> 00:11:42,778 通过在更多专用机架上销售高价值组件和服务 261 00:11:42,778 --> 00:11:44,909 比如Grock和内存机架 262 00:11:44,909 --> 00:11:45,590 其次 263 00:11:45,590 --> 00:11:48,070 如果他们像拆分网络业务那样 264 00:11:48,070 --> 00:11:49,899 DPUs和LP 265 00:11:49,899 --> 00:11:51,360 拆分内存业务 266 00:11:51,360 --> 00:11:55,940 产品结构将揭示客户的主要工作负载方向 267 00:11:55,940 --> 00:11:56,500 从头到尾 268 00:11:56,500 --> 00:11:59,669 从经典模型训练到支持AI代理 269 00:11:59,669 --> 00:12:04,590 这能帮助我们在供应链中发现更多潜力股 270 00:12:04,590 --> 00:12:07,599 现在让我们谈谈这些代币的实际使用者 271 00:12:07,599 --> 00:12:09,418 詹森称Open Claw 272 00:12:09,418 --> 00:12:11,889 是个人AI的操作系统 273 00:12:11,889 --> 00:12:12,690 Open Claw 274 00:12:12,690 --> 00:12:16,129 是一个开源代理，可浏览互联网、调用工具 275 00:12:16,129 --> 00:12:19,818 同时运行数百万代币 276 00:12:19,818 --> 00:12:24,418 这就是我认为代币需求将远超多数分析师预期的原因 277 00:12:24,418 --> 00:12:27,558 数据中心不会只为数十亿人服务 278 00:12:27,558 --> 00:12:32,669 而是可能为数十亿持续在线的AI代理消耗代币 279 00:12:32,669 --> 00:12:35,110 完成人们已有的所有任务 280 00:12:35,110 --> 00:12:37,809 但速度更快且持续时间更长 281 00:12:37,809 --> 00:12:41,009 包括启动更多自定义代理 282 00:12:41,009 --> 00:12:43,769 Open Claw的问题在于它是开源 283 00:12:43,769 --> 00:12:47,259 拥有计算机全系统权限的AI代理 284 00:12:47,259 --> 00:12:50,839 这对企业来说是安全合规噩梦 285 00:12:50,839 --> 00:12:52,678 这就是Nemo Claw的作用 286 00:12:52,678 --> 00:12:55,658 英伟达的开源框架，封装Open Claw 287 00:12:55,658 --> 00:12:57,029 加入策略引擎 288 00:12:57,029 --> 00:12:59,990 隐私路由和安全运行环境 289 00:12:59,990 --> 00:13:04,620 使企业可设置代理使用权限边界 290 00:13:04,620 --> 00:13:06,159 决定代理可访问哪些数据 291 00:13:06,159 --> 00:13:11,480 并确保所有操作在本地或云端安全运行 292 00:13:11,480 --> 00:13:13,080 现在我们回到原点 293 00:13:13,080 --> 00:13:16,360 Open Claw将推动代币需求飙升 294 00:13:16,360 --> 00:13:21,100 Nemo Claw是让代理安全落地的控制层 295 00:13:21,100 --> 00:13:24,340 而英伟达的Ruin架构是硬件支撑 296 00:13:24,340 --> 00:13:26,440 专为高效处理海量代币流设计 297 00:13:26,440 --> 00:13:30,519 随着更多企业接入Open Claw 298 00:13:30,519 --> 00:13:33,470 使用Nemo Claw的代理将消耗更多代币 299 00:13:33,470 --> 00:13:36,470 这就是软件故事最终体现的方式 300 00:13:36,470 --> 00:13:38,720 最终体现在英伟达数据中心营收中 301 00:13:38,720 --> 00:13:42,779 这就是为什么理解股票背后的科学如此重要 302 00:13:42,779 --> 00:13:47,049 我们可以在财报数据出现前就捕捉到这些需求信号 303 00:13:47,049 --> 00:13:51,740 但GCC也明确表示英伟达不会止步于软件代理 304 00:13:51,740 --> 00:13:56,740 他们正在进军机器人和自动驾驶领域，将AI带入物理世界 305 00:13:56,740 --> 00:13:58,559 接下来让我们谈谈这个话题 306 00:13:58,559 --> 00:14:00,240 如果你觉得我值得这个 307 00:14:00,240 --> 00:14:02,899 请点赞并订阅频道 308 00:14:02,899 --> 00:14:06,039 甚至分享这个视频真的能帮助我很多 309 00:14:06,039 --> 00:14:07,740 这能让我知道要制作更多内容 310 00:14:07,740 --> 00:14:08,600 就像这样 311 00:14:08,600 --> 00:14:09,519 谢谢现在 312 00:14:09,519 --> 00:14:11,419 让我们讨论物理AI 313 00:14:11,419 --> 00:14:12,740 在GTC上真正让我惊讶的是 314 00:14:12,740 --> 00:14:13,480 GTC 315 00:14:13,480 --> 00:14:17,409 不是英伟达在谈论机器人和自动驾驶 316 00:14:17,409 --> 00:14:20,590 而是这项技术已经发展的程度 317 00:14:20,590 --> 00:14:24,289 以及在机器人领域报道的滞后程度 318 00:14:24,289 --> 00:14:29,860 像敏捷Digit这样的人形机器人已经在GXO仓库执行真实班次 319 00:14:29,860 --> 00:14:33,940 GXO为各大品牌运营超大规模合同物流仓库 320 00:14:33,940 --> 00:14:34,759 比如耐克 321 00:14:34,759 --> 00:14:36,159 亚马逊和苹果 322 00:14:36,159 --> 00:14:38,120 他们设计并运营这些仓库 323 00:14:38,120 --> 00:14:42,360 并越来越多地使用自动化和机器人服务客户 324 00:14:42,360 --> 00:14:45,799 AJO的Digit已在机器人即服务模式下部署 325 00:14:45,799 --> 00:14:47,080 作为服务模式 326 00:14:47,080 --> 00:14:49,409 不仅仅是舞台上的展示 327 00:14:49,409 --> 00:14:50,990 当我参加GTC时 328 00:14:50,990 --> 00:14:55,460 我看到了工业仓库的完整机器人生态系统 329 00:14:55,460 --> 00:14:57,360 医院甚至零售领域 330 00:14:57,360 --> 00:15:02,139 都在使用相同的Isaac和Cosmos世界模型堆栈进行训练 331 00:15:02,139 --> 00:15:03,740 大多数投资者错过的 332 00:15:03,740 --> 00:15:06,080 是这项技术一旦验证 333 00:15:06,080 --> 00:15:09,440 即使少数设计在实地证明有效 334 00:15:09,440 --> 00:15:11,460 由于所有人使用相同堆栈训练 335 00:15:11,460 --> 00:15:15,500 一个仓库或工厂的模拟能力 336 00:15:15,500 --> 00:15:18,639 可以调整并复用于下一个百位客户 337 00:15:18,639 --> 00:15:21,740 无需每次都从头开始 338 00:15:21,740 --> 00:15:24,340 在GTC看到的所有机器人 339 00:15:24,340 --> 00:15:27,019 结合与Spencer Wong的访谈所得 340 00:15:27,019 --> 00:15:28,919 我怀疑机器人革命 341 00:15:28,919 --> 00:15:29,399 我意思是 342 00:15:29,399 --> 00:15:33,960 物理AI革命可能比大多数人意识到的更早到来 343 00:15:33,960 --> 00:15:38,460 在自动驾驶领域我体验了英伟达L2++梅赛德斯 344 00:15:38,460 --> 00:15:42,269 在旧金山市区行驶 345 00:15:42,269 --> 00:15:46,610 事实证明对我们来说的边缘情况 346 00:15:46,610 --> 00:15:48,240 旧金山路况 347 00:15:48,240 --> 00:15:50,759 我们看到有人闯红灯和停车标志 348 00:15:50,759 --> 00:15:53,200 至少被五次加塞 349 00:15:53,200 --> 00:15:58,200 每条街都有双排停车和施工 350 00:15:58,200 --> 00:16:01,059 车辆轻松处理所有这些场景 351 00:16:01,059 --> 00:16:04,080 让我最惊讶的是这一切显得如此自然 352 00:16:04,080 --> 00:16:08,629 人机之间的交接过程在双向操作中都流畅无阻 353 00:16:08,629 --> 00:16:10,429 这个系统简直 354 00:16:10,429 --> 00:16:13,309 比大多数人处理复杂任务更出色 355 00:16:13,309 --> 00:16:15,230 预测其他驾驶员的行为 356 00:16:15,230 --> 00:16:18,009 判断何时有足够的空间切入空隙 357 00:16:18,009 --> 00:16:20,870 绕开静止和移动的障碍物 358 00:16:20,870 --> 00:16:24,049 在我会犹豫的地方找到安全路径 359 00:16:24,049 --> 00:16:25,450 如果是我亲自驾驶的话 360 00:16:25,450 --> 00:16:29,379 我稍后会发布完整未剪辑的行程作为独立视频 361 00:16:29,379 --> 00:16:33,919 对投资者来说关键点是自动驾驶已经到来 362 00:16:33,919 --> 00:16:38,250 并准备好在多种车型和车队中推广 363 00:16:38,250 --> 00:16:39,389 不仅仅是芯片 364 00:16:39,389 --> 00:16:43,570 而是包含英伟达AlphaMayo推理模型的完整软件栈 365 00:16:43,570 --> 00:16:47,019 以及英伟达Drive Hyperion平台为核心 366 00:16:47,019 --> 00:16:49,119 这里的重要合作伙伴是Uber 367 00:16:49,119 --> 00:16:50,979 英伟达驱动的自动驾驶出租车 368 00:16:50,979 --> 00:16:56,000 使用Drive Hyperion和AlphaMayo计划部署在Uber网络 369 00:16:56,000 --> 00:16:59,679 如洛杉矶和旧金山等城市明年即可落地 370 00:16:59,679 --> 00:17:03,019 随后将在2028年前扩展到28个城市 371 00:17:03,019 --> 00:17:06,159 英伟达宣布像比亚迪 372 00:17:06,159 --> 00:17:09,358 日产和五十铃正在开发自己的L4级 373 00:17:09,358 --> 00:17:10,638 车辆用于出行 374 00:17:10,638 --> 00:17:12,559 叫车应用和商业车队 375 00:17:12,559 --> 00:17:14,539 这不仅限于自动驾驶出租车 376 00:17:14,539 --> 00:17:18,940 英伟达正瞄准自动驾驶卡车、巴士和工业车辆 377 00:17:18,940 --> 00:17:23,650 所有都将共享相同的软件仿真工具和硬件组件 378 00:17:23,650 --> 00:17:24,809 当我询问埃隆·马斯克 379 00:17:24,809 --> 00:17:28,549 他认为这些系统最大的近期应用场景是什么 380 00:17:28,549 --> 00:17:29,589 比如OpenAI 381 00:17:29,589 --> 00:17:30,150 他说 382 00:17:30,150 --> 00:17:31,450 自动驾驶车辆 383 00:17:31,450 --> 00:17:32,769 他解释道 384 00:17:32,769 --> 00:17:37,269 尽管英伟达的汽车业务仅占总收入不到1% 385 00:17:37,269 --> 00:17:37,630 今天 386 00:17:37,630 --> 00:17:39,640 就像CUDA当初起步时一样 387 00:17:39,640 --> 00:17:42,500 如今英伟达正在交付训练好的AI模型 388 00:17:42,500 --> 00:17:44,640 标准化的仿真环境 389 00:17:44,640 --> 00:17:46,869 以及出行车队的车载大脑 390 00:17:46,869 --> 00:17:47,769 叫车服务 391 00:17:47,769 --> 00:17:48,990 配送货车 392 00:17:48,990 --> 00:17:51,480 卡车和全球汽车 393 00:17:51,480 --> 00:17:56,159 这些将持续反哺他们的薇拉·鲁宾AI工厂 394 00:17:56,159 --> 00:17:58,278 从GTC大会整体来看 395 00:17:58,278 --> 00:17:59,939 趋势非常明显 396 00:17:59,939 --> 00:18:02,919 英伟达不只是销售更快的GPU 397 00:18:02,919 --> 00:18:06,779 他们正在渗透AI经济的每个环节 398 00:18:06,779 --> 00:18:07,799 代币 399 00:18:07,799 --> 00:18:08,559 智能体 400 00:18:08,559 --> 00:18:10,420 机器人和自动驾驶汽车 401 00:18:10,420 --> 00:18:12,619 支撑这一切的数据中心 402 00:18:12,619 --> 00:18:16,710 这就是华尔街分析师大多忽视的更大图景 403 00:18:16,710 --> 00:18:21,170 这就是我认为英伟达将成为全球首家万亿美元公司的原因 404 00:18:21,170 --> 00:18:25,699 这就是理解股票背后科学原理为何如此重要 405 00:18:25,699 --> 00:18:28,759 如果你想看到更多股票背后的科学原理 406 00:18:28,759 --> 00:18:31,269 不妨观看接下来的这个视频 407 00:18:31,269 --> 00:18:31,969 感谢观看 408 00:18:31,969 --> 00:18:34,509 下次再见，这里是股票代码 409 00:18:34,509 --> 00:18:34,888 你 410 00:18:34,888 --> 00:18:35,969 我是亚历克斯 411 00:18:35,969 --> 00:18:39,230 提醒你最好的投资选择 412 00:18:39,230 --> 00:18:40,579 就是你自己
评论采集失败
采集完成，该视频无弹幕