All language subtitles for [Chinese Simplified] [DownSub.com]

af Afrikaans
ak Akan
sq Albanian
am Amharic
ar Arabic
hy Armenian
az Azerbaijani
eu Basque
be Belarusian
bem Bemba
bn Bengali
bh Bihari
bs Bosnian
br Breton
bg Bulgarian
km Cambodian
ca Catalan
ceb Cebuano
chr Cherokee
ny Chichewa
zh-CN Chinese (Simplified)
zh-TW Chinese (Traditional)
co Corsican
hr Croatian
cs Czech
da Danish
nl Dutch
en English Download
eo Esperanto
et Estonian
ee Ewe
fo Faroese
tl Filipino
fi Finnish
fr French
fy Frisian
gaa Ga
gl Galician
ka Georgian
de German
el Greek
gn Guarani
gu Gujarati
ht Haitian Creole
ha Hausa
haw Hawaiian
iw Hebrew
hi Hindi
hmn Hmong
hu Hungarian
is Icelandic
ig Igbo
id Indonesian
ia Interlingua
ga Irish
it Italian
ja Japanese
jw Javanese
kn Kannada
kk Kazakh
rw Kinyarwanda
rn Kirundi
kg Kongo
ko Korean
kri Krio (Sierra Leone)
ku Kurdish
ckb Kurdish (Soranî)
ky Kyrgyz
lo Laothian
la Latin
lv Latvian
ln Lingala
lt Lithuanian
loz Lozi
lg Luganda
ach Luo
lb Luxembourgish
mk Macedonian
mg Malagasy
ms Malay
ml Malayalam
mt Maltese
mi Maori
mr Marathi
mfe Mauritian Creole
mo Moldavian
mn Mongolian
my Myanmar (Burmese)
sr-ME Montenegrin
ne Nepali
pcm Nigerian Pidgin
nso Northern Sotho
no Norwegian
nn Norwegian (Nynorsk)
oc Occitan
or Oriya
om Oromo
ps Pashto
fa Persian
pl Polish
pt-BR Portuguese (Brazil)
pt Portuguese (Portugal)
pa Punjabi
qu Quechua
ro Romanian
rm Romansh
nyn Runyakitara
ru Russian
sm Samoan
gd Scots Gaelic
sr Serbian
sh Serbo-Croatian
st Sesotho
tn Setswana
crs Seychellois Creole
sn Shona
sd Sindhi
si Sinhalese
sk Slovak
sl Slovenian
so Somali
es Spanish
es-419 Spanish (Latin American)
su Sundanese
sw Swahili
sv Swedish
tg Tajik
ta Tamil
tt Tatar
te Telugu
th Thai
ti Tigrinya
to Tonga
lua Tshiluba
tum Tumbuka
tr Turkish
tk Turkmen
tw Twi
ug Uighur
uk Ukrainian
ur Urdu
uz Uzbek
vi Vietnamese
cy Welsh
wo Wolof
xh Xhosa
yi Yiddish
yo Yoruba
zu Zulu
Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated: 1 00:00:00,200 --> 00:00:01,866 这才不到三个月 2 00:00:01,866 --> 00:00:05,333 那个号称要碾压台积电的晶圆厂“Terafab” 3 00:00:05,333 --> 00:00:06,566 怎么就熄火了？ 4 00:00:06,566 --> 00:00:08,866 前脚马斯克亲自泼冷水 5 00:00:08,866 --> 00:00:10,866 后脚它的技术后盾英特尔 6 00:00:10,866 --> 00:00:13,066 自己都跑去台积电抢产能了 7 00:00:13,066 --> 00:00:14,200 甚至 8 00:00:14,200 --> 00:00:18,266 马斯克自己手里的特斯拉都随时准备好了退出跑路 9 00:00:18,266 --> 00:00:20,033 明明是强强联手 10 00:00:20,100 --> 00:00:22,833 为什么就是造不出足够多的AI芯片？ 11 00:00:22,833 --> 00:00:24,633 朋友们，就在最近 12 00:00:24,633 --> 00:00:26,766 出现了两个非常诡异的信号 13 00:00:26,766 --> 00:00:28,600 一个发生在上周 14 00:00:28,733 --> 00:00:32,266 在SpaceX向SEC提交了S-1上市申请文件里 15 00:00:32,366 --> 00:00:34,733 马斯克向全世界投资人坦白 16 00:00:34,733 --> 00:00:37,100 “我可能搞不定芯片供应了” 17 00:00:37,100 --> 00:00:38,433 原话是这么说的： 18 00:00:38,433 --> 00:00:42,700 “我们寄望于通过建设 Terafab 解决芯片供应难题 19 00:00:42,833 --> 00:00:44,933 但该项目仍有可能失败” 20 00:00:44,933 --> 00:00:48,633 “我们无法保证 Terafab 能在预期周期内达成目标 21 00:00:48,733 --> 00:00:51,066 甚至无法保证项目能够落地” 22 00:00:51,066 --> 00:00:52,200 更扎心的是 23 00:00:52,233 --> 00:00:54,266 在另一份文件中还补了一句： 24 00:00:54,266 --> 00:00:57,066 “英特尔和特斯拉均有可能退出该项目” 25 00:00:57,066 --> 00:00:59,633 一旦这两个核心伙伴撤资离场 26 00:00:59,800 --> 00:01:03,500 Terafab将同时失去关键客户和制程技术支持 27 00:01:03,600 --> 00:01:06,133 那这个项目不就剩个空架子了吗？ 28 00:01:06,133 --> 00:01:08,600 但这还没有结束，就在几天后 29 00:01:08,733 --> 00:01:11,833 6月1日，COMPUTEX 2026的会场上 30 00:01:11,833 --> 00:01:13,666 另一个不该认输的人 31 00:01:13,666 --> 00:01:15,266 英特尔CEO陈立武 32 00:01:15,266 --> 00:01:18,400 亲口说了一句让所有分析师始料未及的话： 33 00:01:18,400 --> 00:01:20,800 “英特尔的转型需要5到10年 34 00:01:20,800 --> 00:01:23,533 我们非常乐意成为台积电的前几大客户” 35 00:01:23,533 --> 00:01:24,433 你没听错 36 00:01:24,433 --> 00:01:27,666 那个曾经把“自己设计、自己制造、自己定义标准” 37 00:01:27,666 --> 00:01:29,133 写进骨子里的英特尔 38 00:01:29,133 --> 00:01:33,900 那个前不久还在用18A和台积电疯狂打擂台的英特尔 39 00:01:33,900 --> 00:01:37,766 现在公开承认要把一部分关键生产交给台积电 40 00:01:37,900 --> 00:01:41,000 甚至大肆谈论和台积电的友情往事 41 00:01:41,000 --> 00:01:42,633 他为什么要这么做？ 42 00:01:42,633 --> 00:01:43,700 是疯了？ 43 00:01:43,700 --> 00:01:45,200 还是清醒了？ 44 00:01:45,200 --> 00:01:47,666 大家好，我是Olivia，之前两集 45 00:01:47,666 --> 00:01:50,400 我们聊了黄仁勋砸下的1500亿 46 00:01:50,400 --> 00:01:52,400 还有苏姿丰飞过来的100亿 47 00:01:52,400 --> 00:01:56,266 他们全都在做同一件事——拼命抢台积电产能 48 00:01:56,266 --> 00:01:57,100 而现在 49 00:01:57,100 --> 00:01:58,300 陈立武也来了 50 00:01:58,300 --> 00:02:01,633 想绕开台积电的马斯克呢，也低头了 51 00:02:01,633 --> 00:02:02,333 今天 52 00:02:02,333 --> 00:02:04,933 我们就把这一块最重要的拼图放进去 53 00:02:04,933 --> 00:02:07,133 把这局战争的全貌理清楚 54 00:02:07,133 --> 00:02:09,166 为什么Terafab可能要失败了？ 55 00:02:09,166 --> 00:02:10,966 为什么陈立武认输了？ 56 00:02:10,966 --> 00:02:13,566 为什么这些AI时代最重要的巨头 57 00:02:13,566 --> 00:02:15,333 最近都在往台积电跑？ 58 00:02:15,333 --> 00:02:16,400 更重要的是 59 00:02:16,433 --> 00:02:18,366 在这场无声的权力转移中 60 00:02:18,366 --> 00:02:20,700 作为投资者，我们应该看到什么？ 61 00:02:20,700 --> 00:02:22,466 这不是孤立的事件 62 00:02:22,466 --> 00:02:25,666 而是这轮AI战争在对市场进行重估 63 00:02:25,666 --> 00:02:27,700 上游材料、先进制程 64 00:02:27,700 --> 00:02:31,133 先进封装、光通信、存储、电力…… 65 00:02:31,133 --> 00:02:32,500 都在这条线上 66 00:02:32,500 --> 00:02:34,666 其实咱们之前聊过的一些公司 67 00:02:34,666 --> 00:02:36,800 都在这一轮狂欢的便车 68 00:02:36,800 --> 00:02:39,266 感谢大家一路以来的支持和相伴 69 00:02:39,266 --> 00:02:43,500 我把这次黄仁勋、苏姿丰、陈立武共同押注下 70 00:02:43,500 --> 00:02:47,033 完整的产业链名单，还有最有价值的三个企业 71 00:02:47,033 --> 00:02:51,700 从核心逻辑，到财报重点，再到市场分歧，关键催化 72 00:02:51,700 --> 00:02:53,466 整理成一份完整清单 73 00:02:53,466 --> 00:02:56,300 想要的朋友可以在评论区打出888 74 00:02:56,300 --> 00:02:58,300 如果留言人数多 75 00:02:58,300 --> 00:03:02,733 我会花两天时间准备好，统一发给真正需要的朋友 76 00:03:02,733 --> 00:03:04,466 我们先来看英特尔 77 00:03:04,500 --> 00:03:06,766 为什么突然不坚持要自己造了 78 00:03:06,766 --> 00:03:09,300 反而要跑到台积电去抢产能呢？ 79 00:03:09,300 --> 00:03:11,566 不少朋友都感觉挺反常的吧？ 80 00:03:11,566 --> 00:03:14,466 其实，它是被时间逼到了墙角 81 00:03:14,466 --> 00:03:16,900 陈立武不是不知道自家制造重要 82 00:03:17,000 --> 00:03:19,500 恰恰相反，他比谁都知道 83 00:03:19,533 --> 00:03:23,633 18A、14A是他重建芯片制造帝国的命门 84 00:03:23,633 --> 00:03:25,233 可是问题在于 85 00:03:25,333 --> 00:03:28,300 先进制程不是你喊一句“我要回来了” 86 00:03:28,300 --> 00:03:30,666 良率就能立刻回到王者水平 87 00:03:30,800 --> 00:03:33,833 Intel 18A虽然已经有良率改善的迹象 88 00:03:33,833 --> 00:03:37,233 但它仍然处在爬坡和客户验证的关键阶段 89 00:03:37,333 --> 00:03:40,300 14A更是下一步重注投入的节点 90 00:03:40,300 --> 00:03:41,266 换句话说 91 00:03:41,300 --> 00:03:43,633 英特尔的制造复兴正在路上 92 00:03:43,633 --> 00:03:45,500 但还没完全到站 93 00:03:45,500 --> 00:03:47,700 可问题是，时间不等人 94 00:03:47,700 --> 00:03:48,533 另一边 95 00:03:48,533 --> 00:03:50,733 CPU的复兴大潮已经来了 96 00:03:50,733 --> 00:03:53,600 AI PC的战火也已经烧到家门口了 97 00:03:53,633 --> 00:03:56,100 AI正在从训练，走向推理 98 00:03:56,133 --> 00:03:58,666 从聊天框，走向Agentic AI 99 00:03:58,866 --> 00:04:00,333 从云端大模型 100 00:04:00,400 --> 00:04:03,900 走向企业系统、边缘设备和AI PC 101 00:04:03,900 --> 00:04:06,500 一旦AI Agent开始真正干活 102 00:04:06,500 --> 00:04:08,466 它就不是简单回答一句话 103 00:04:08,466 --> 00:04:11,566 而是要拆任务、调工具、读文件 104 00:04:11,566 --> 00:04:15,700 跑代码、访问数据库、管理权限、处理网络通信 105 00:04:15,700 --> 00:04:18,566 还要同时协调多个模型和多个应用 106 00:04:18,566 --> 00:04:21,333 这些活，GPU一个人干不了 107 00:04:21,666 --> 00:04:24,233 GPU擅长的是大规模矩阵计算 108 00:04:24,233 --> 00:04:25,466 是暴力算力 109 00:04:25,466 --> 00:04:29,266 但CPU擅长的是系统调度、复杂逻辑、I/O 110 00:04:29,266 --> 00:04:33,033 内存管理、网络通信和企业软件生态 111 00:04:33,033 --> 00:04:35,333 换句话说，AI训练时代 112 00:04:35,333 --> 00:04:37,100 GPU是绝对主角 113 00:04:37,100 --> 00:04:40,033 但到了AI Agent大规模落地的时候 114 00:04:40,200 --> 00:04:42,966 CPU又重新从后台走到了前台 115 00:04:42,966 --> 00:04:43,700 而这 116 00:04:43,700 --> 00:04:45,266 正是英特尔的领域 117 00:04:45,266 --> 00:04:47,266 这不是我替英特尔圆场 118 00:04:47,266 --> 00:04:49,666 而是数据已经开始反映出来了 119 00:04:49,666 --> 00:04:52,666 Intel 2026年第一季度财报里 120 00:04:52,666 --> 00:04:54,333 Client Computing Group 121 00:04:54,466 --> 00:04:57,100 也就是PC芯片所在的客户端业务 122 00:04:57,100 --> 00:04:58,866 收入是77亿美元 123 00:04:58,900 --> 00:05:00,400 Data Center and AI 124 00:05:00,400 --> 00:05:02,700 也就是数据中心和AI业务 125 00:05:02,700 --> 00:05:06,433 收入是51亿美元，同比增长22% 126 00:05:06,433 --> 00:05:07,966 这两个业务加起来 127 00:05:07,966 --> 00:05:11,466 就是英特尔产品业务最核心的收入来源 128 00:05:11,466 --> 00:05:15,333 Intel Q1 2026总净收入是136亿美元 129 00:05:15,500 --> 00:05:19,966 光CCG和DCAI加起来就是128亿美元 130 00:05:19,966 --> 00:05:20,866 也就是说 131 00:05:20,866 --> 00:05:24,466 英特尔真正赚钱、真正有市场存在感的地方 132 00:05:24,500 --> 00:05:28,233 依然是CPU和围绕CPU展开的产品线 133 00:05:28,400 --> 00:05:31,033 这对英特尔来说，当然是好消息 134 00:05:31,100 --> 00:05:33,566 可问题是，好消息来得太快了 135 00:05:33,566 --> 00:05:37,100 它需要时间证明自己重新掌握先进制造节奏 136 00:05:37,100 --> 00:05:38,933 但客户现在就要产品 137 00:05:38,966 --> 00:05:40,733 云厂商现在就要CPU 138 00:05:40,833 --> 00:05:42,600 PC厂商现在就要方案 139 00:05:42,600 --> 00:05:45,900 而互相攻防了这么多年的老对手AMD 140 00:05:45,900 --> 00:05:49,866 也不会站在原地等英特尔慢慢把良率爬上来再开战 141 00:05:49,866 --> 00:05:50,666 这些年 142 00:05:50,766 --> 00:05:53,700 AMD的CPU市占率已经超过了英特尔 143 00:05:53,700 --> 00:05:55,500 实力当然不可小觑 144 00:05:55,500 --> 00:05:59,800 现在AMD又宣布超过100亿美元投资台湾生态 145 00:05:59,800 --> 00:06:04,500 重点扩大战略合作、先进封装和AI基础设施能力 146 00:06:04,500 --> 00:06:09,366 还把6代EPYC、MI450X和Helios机柜级AI平台 147 00:06:09,366 --> 00:06:11,066 绑定到台湾供应链上 148 00:06:11,266 --> 00:06:12,400 这说明什么？ 149 00:06:12,400 --> 00:06:15,733 说明苏姿丰不是只想在GPU上追英伟达 150 00:06:16,000 --> 00:06:18,800 她还野心勃勃的要在CPU、AI服务器 151 00:06:18,800 --> 00:06:22,300 先进封装和系统平台上继续夹击英特尔 152 00:06:22,366 --> 00:06:23,233 与此同时 153 00:06:23,233 --> 00:06:25,733 黄仁勋也在一边虎视眈眈 154 00:06:25,733 --> 00:06:27,500 昨天咱们聊了这件事 155 00:06:27,500 --> 00:06:30,633 他正打算带着他专门为新一波AI agents 156 00:06:30,633 --> 00:06:33,333 打造的超级芯片RTX Spark 157 00:06:33,333 --> 00:06:35,533 重新定义PC战场 158 00:06:35,533 --> 00:06:39,200 更扎心的是，RTX Spark还是Arm架构 159 00:06:39,200 --> 00:06:40,933 黄仁勋一只手拉住微软 160 00:06:40,933 --> 00:06:42,500 一只手绕开x86 161 00:06:42,500 --> 00:06:45,733 直接冲进了英特尔最传统、最核心的地盘 162 00:06:45,733 --> 00:06:48,366 所以陈立武现在面临的不是一个敌人 163 00:06:48,366 --> 00:06:50,466 而是几个战场同时开打 164 00:06:50,466 --> 00:06:51,200 这时候 165 00:06:51,200 --> 00:06:52,666 如果英特尔还坚持 166 00:06:52,666 --> 00:06:55,833 “所有东西必须等我自家18A完全成熟再说” 167 00:06:55,833 --> 00:06:58,466 那它可能等来的不是制造复兴 168 00:06:58,466 --> 00:07:01,933 而是PC和AI服务器机会窗口被别人抢走 169 00:07:01,933 --> 00:07:04,166 所以陈立武必须现实一点 170 00:07:04,166 --> 00:07:06,533 这就是他为什么要向台积电要产能 171 00:07:06,766 --> 00:07:09,400 因为台积电能给他的，不只是晶圆 172 00:07:09,500 --> 00:07:10,500 而是时间 173 00:07:10,800 --> 00:07:11,766 而他的转向 174 00:07:11,766 --> 00:07:14,433 无疑是对“Terafab”的一次公开拆台 175 00:07:14,666 --> 00:07:17,433 虽然不是最沉重、最关键的那一拳 176 00:07:17,433 --> 00:07:20,966 但它直接撕开了这个项目最尴尬的现实： 177 00:07:20,966 --> 00:07:23,933 连负责提供先进制程想象力的英特尔 178 00:07:23,933 --> 00:07:26,100 自己都还需要台积电托底 179 00:07:26,100 --> 00:07:29,766 那Terafab又凭什么在短时间内复制一个台积电？ 180 00:07:29,766 --> 00:07:32,666 说起“Terafab”，咱们之前也聊过 181 00:07:32,666 --> 00:07:35,366 虽然我个人不看好，但是不得不说 182 00:07:35,366 --> 00:07:38,166 很多人对它的追捧也不是空穴来风 183 00:07:38,166 --> 00:07:39,900 如果只看账面实力 184 00:07:39,900 --> 00:07:42,700 Terafab甚至可以说是过去十年里 185 00:07:42,700 --> 00:07:45,633 最接近“硬造一个台积电”的疯狂尝试 186 00:07:45,633 --> 00:07:47,633 你想想，它背后站着谁？ 187 00:07:47,633 --> 00:07:48,733 需求端 188 00:07:48,733 --> 00:07:51,100 直接就是马斯克自己的AI帝国 189 00:07:51,100 --> 00:07:53,666 特斯拉要FSD，要Robotaxi 190 00:07:53,666 --> 00:07:55,433 要Optimus机器人 191 00:07:55,433 --> 00:07:57,466 xAI要训练和部署Grok 192 00:07:57,466 --> 00:08:00,166 SpaceX更夸张，不只要做Starlink 193 00:08:00,166 --> 00:08:02,533 还想把AI算力、卫星网络 194 00:08:02,533 --> 00:08:05,066 甚至未来的太空数据中心串起来 195 00:08:05,066 --> 00:08:05,933 也就是说 196 00:08:05,966 --> 00:08:08,500 Terafab不是先建厂再找客户 197 00:08:08,500 --> 00:08:10,333 而含着金汤匙出生 198 00:08:10,366 --> 00:08:14,133 直接绑定马斯克体系里最烧芯片、最烧算力 199 00:08:14,133 --> 00:08:16,333 最会讲增长故事的几家公司 200 00:08:16,333 --> 00:08:17,600 那技术端呢？ 201 00:08:17,600 --> 00:08:20,433 也不缺，直接牵手了英特尔 202 00:08:20,433 --> 00:08:22,000 按照此前的设想 203 00:08:22,033 --> 00:08:24,800 Terafab并不是从零摸黑造芯片 204 00:08:24,866 --> 00:08:28,400 而是计划借用Intel 14A这样的先进工艺路线 205 00:08:28,400 --> 00:08:30,966 英特尔虽然这几年被台积电压着打 206 00:08:30,966 --> 00:08:32,766 但它毕竟底子还在 207 00:08:32,766 --> 00:08:34,633 而且从基辛格时代开始 208 00:08:34,633 --> 00:08:37,933 它们就在用四年5节点激进计划拼命追赶 209 00:08:37,933 --> 00:08:40,500 从落后台积电2代，追到现在 210 00:08:40,566 --> 00:08:44,066 这些年的技术积累、工程师体系都是现成的 211 00:08:44,133 --> 00:08:45,166 更重要的是 212 00:08:45,166 --> 00:08:48,133 还有美国本土先进制造的政策光环 213 00:08:48,133 --> 00:08:51,166 Terafab不是找了一个草台班子来造芯片 214 00:08:51,166 --> 00:08:54,533 它至少试图把自己绑到全球少数几个 215 00:08:54,533 --> 00:08:57,300 还敢冲击最先进制程的玩家身上 216 00:08:57,300 --> 00:08:58,433 而且这样一来 217 00:08:58,566 --> 00:09:00,466 设备端的问题也就解决了 218 00:09:00,633 --> 00:09:05,200 Terafab直接能靠英特尔供应链绑定落地设备配额 219 00:09:05,200 --> 00:09:06,500 至于人才端呢 220 00:09:06,500 --> 00:09:10,200 直接挥舞钞票从台湾新竹科学园区挖人 221 00:09:10,200 --> 00:09:11,566 对于马斯克来说 222 00:09:11,566 --> 00:09:15,433 TeraFab的目标是2027年量产2纳米芯片 223 00:09:15,666 --> 00:09:18,933 他没有十年时间慢慢培养自己的工程师团队 224 00:09:18,966 --> 00:09:20,733 他需要在最短的时间内 225 00:09:20,733 --> 00:09:23,400 把一群知道“怎么做”的人聚在一起 226 00:09:23,400 --> 00:09:25,366 而台积电的资深工程师们 227 00:09:25,366 --> 00:09:28,600 当然就是全世界最知道“怎么做”的那群人 228 00:09:28,600 --> 00:09:29,333 为此 229 00:09:29,333 --> 00:09:32,933 他甚至拿出了特斯拉和SpaceX的双重股权 230 00:09:32,933 --> 00:09:33,666 钱呢？ 231 00:09:33,666 --> 00:09:34,600 更不用说 232 00:09:34,800 --> 00:09:36,600 Terafab如果真的落地 233 00:09:36,666 --> 00:09:38,800 背后是SpaceX上市融资 234 00:09:38,866 --> 00:09:42,200 是特斯拉的现金流，是xAI的融资能力 235 00:09:42,200 --> 00:09:44,700 是美国制造回流的政策叙事 236 00:09:44,700 --> 00:09:47,033 别的公司建厂最怕没钱 237 00:09:47,300 --> 00:09:48,500 马斯克最不缺的就是 238 00:09:48,500 --> 00:09:51,566 让资本市场相信他能改变世界的能力 239 00:09:51,566 --> 00:09:54,433 所以你看，按理说这盘棋很豪华 240 00:09:54,433 --> 00:09:57,866 客户有了，技术有了，设备有了 241 00:09:57,933 --> 00:10:00,200 人才有了，钱也有了 242 00:10:00,200 --> 00:10:02,100 可为什么就是不行呢？ 243 00:10:02,100 --> 00:10:04,600 从3 月 21 日正式对外发布 244 00:10:04,600 --> 00:10:06,966 到今天为止也不足三个月 245 00:10:06,966 --> 00:10:09,666 为什么合作双方就都开始悲观了呢？ 246 00:10:09,666 --> 00:10:11,266 到底差的是什么？ 247 00:10:11,266 --> 00:10:12,500 其实归根结底 248 00:10:12,500 --> 00:10:15,466 还是马斯克小看的晶圆制造的难度 249 00:10:15,466 --> 00:10:17,933 马斯克喜欢快，喜欢颠覆 250 00:10:17,933 --> 00:10:19,300 喜欢“move fast” 251 00:10:19,300 --> 00:10:22,200 可晶圆制造偏偏是世界上最慢、最重 252 00:10:22,200 --> 00:10:24,066 最讲秩序的行业之一 253 00:10:24,066 --> 00:10:27,000 它不是你把最好的零件买回来装在一起 254 00:10:27,000 --> 00:10:28,866 就能变成一个台积电的 255 00:10:28,866 --> 00:10:30,000 首先一开始 256 00:10:30,000 --> 00:10:32,300 这个合作本身就没有真正锁死 257 00:10:32,300 --> 00:10:36,200 Terafab表面看是SpaceX、Tesla、Intel的铁三角 258 00:10:36,200 --> 00:10:39,800 但文件里它还不是最终、有约束力的合资协议 259 00:10:39,800 --> 00:10:42,100 万一哪天英特尔不想合作了 260 00:10:42,100 --> 00:10:46,366 那Terafab将直接失去Intel 14A工艺、工程团队 261 00:10:46,366 --> 00:10:48,366 甚至阿斯麦光刻机配额 262 00:10:48,366 --> 00:10:49,400 也就是说 263 00:10:49,400 --> 00:10:52,500 咱们刚刚提到的Terafab的两个重要优势 264 00:10:52,500 --> 00:10:53,766 技术和设备 265 00:10:53,766 --> 00:10:55,633 根本就是镜花水月 266 00:10:55,633 --> 00:10:57,000 一旦它们消失了 267 00:10:57,000 --> 00:10:59,033 那这座所谓超级晶圆厂 268 00:10:59,100 --> 00:11:01,766 瞬间就会从“AI芯片自救计划” 269 00:11:01,766 --> 00:11:04,700 变成一块还没打地基的空地 270 00:11:04,700 --> 00:11:06,800 而且即使英特尔不离不弃 271 00:11:06,800 --> 00:11:08,633 设备问题也依旧存在 272 00:11:08,766 --> 00:11:11,566 依靠英特尔供应链绑定一部分设备配额 273 00:11:11,566 --> 00:11:14,333 不等于“能按自己的节奏拿到足够设备” 274 00:11:14,333 --> 00:11:16,433 更不等于“拿到设备就能量产” 275 00:11:16,433 --> 00:11:19,333 High-NA EUV产能本身非常稀缺 276 00:11:19,333 --> 00:11:22,433 而且前面已经排着台积电、三星、英特尔 277 00:11:22,433 --> 00:11:24,600 你说我马斯克加钱行不行？ 278 00:11:24,633 --> 00:11:26,900 行，可能能往前挤一点 279 00:11:26,900 --> 00:11:28,033 但问题是 280 00:11:28,033 --> 00:11:31,433 人家整个供应链早就按季度、按年份 281 00:11:31,433 --> 00:11:33,666 按客户验证节奏排好了 282 00:11:33,666 --> 00:11:36,433 不是你今天喊一句“我要改变世界” 283 00:11:36,566 --> 00:11:39,933 明天ASML就能把机器空投到德州 284 00:11:39,933 --> 00:11:43,200 更何况，先进制程不是只有光刻机 285 00:11:43,533 --> 00:11:45,766 光刻只是最耀眼的那颗明珠 286 00:11:45,766 --> 00:11:50,933 后面还有沉积、蚀刻、离子注入、CMP、清洗 287 00:11:50,933 --> 00:11:54,533 检测、量测、缺陷分析、自动搬运 288 00:11:54,533 --> 00:11:56,833 洁净系统、化学品供应 289 00:11:56,833 --> 00:11:59,966 任何一环掉链子，整条产线都得等 290 00:11:59,966 --> 00:12:02,666 所以设备的问题，不是“买不买得起” 291 00:12:02,700 --> 00:12:05,700 而是“能不能按时、成套、稳定 292 00:12:05,700 --> 00:12:07,733 优先地进入你的产线” 293 00:12:07,733 --> 00:12:09,666 这背后拼的不是现金 294 00:12:09,700 --> 00:12:12,233 而是行业信用和供应链地位 295 00:12:12,233 --> 00:12:15,233 同样，人才也不是你想挖就能挖 296 00:12:15,233 --> 00:12:18,533 马斯克确实可以高薪挖人，但问题是 297 00:12:18,533 --> 00:12:22,033 一座先进晶圆厂不是挖几个高手就能跑起来 298 00:12:22,033 --> 00:12:26,233 它需要几千名工程师、设备技术员、制程整合团队 299 00:12:26,233 --> 00:12:30,366 良率团队、封装团队和供应链团队长期磨合 300 00:12:30,366 --> 00:12:31,933 挖一个工程师 301 00:12:31,933 --> 00:12:33,900 可以带走他的个人经验 302 00:12:33,900 --> 00:12:36,166 但带不走台积电那个“平台 303 00:12:36,166 --> 00:12:38,433 那套把几千人拧成一股绳 304 00:12:38,433 --> 00:12:43,100 把几十年的经验沉淀成标准作业程序的组织能力 305 00:12:43,100 --> 00:12:45,733 更可怕的，是良率的地狱 306 00:12:45,733 --> 00:12:47,933 Intel 14A听起来先进 307 00:12:47,933 --> 00:12:49,700 但先进不等于成熟 308 00:12:49,700 --> 00:12:51,833 成熟不等于稳定量产 309 00:12:51,833 --> 00:12:53,966 芯片制造最狠的地方在于 310 00:12:53,966 --> 00:12:57,033 做出一颗芯片和稳定量产一批芯片 311 00:12:57,033 --> 00:12:58,533 完全是两种能力 312 00:12:58,533 --> 00:13:01,666 你在实验室里做出样片，叫技术突破 313 00:13:01,833 --> 00:13:04,566 你每个月稳定出货几十万片晶圆 314 00:13:04,600 --> 00:13:06,266 客户拿去做成产品 315 00:13:06,433 --> 00:13:09,333 良率稳定、成本可控、交期可靠 316 00:13:09,400 --> 00:13:10,833 这才叫商业制程 317 00:13:10,833 --> 00:13:12,233 但这可不容易 318 00:13:12,233 --> 00:13:13,433 一个粉尘颗粒 319 00:13:13,433 --> 00:13:14,700 一次温度波动 320 00:13:14,700 --> 00:13:17,733 一道蚀刻深一点，一层薄膜厚一点 321 00:13:17,766 --> 00:13:19,400 一个金属互连缺陷 322 00:13:19,400 --> 00:13:21,133 一个封装应力变化 323 00:13:21,133 --> 00:13:22,966 都可能让一批芯片报废 324 00:13:23,100 --> 00:13:27,866 良率从20%爬到50%，从50%爬到70% 325 00:13:28,133 --> 00:13:30,266 从70%爬到85% 326 00:13:30,266 --> 00:13:34,066 每一个百分点都是无数工程师、设备商 327 00:13:34,066 --> 00:13:37,133 材料商、客户设计团队一起熬出来的 328 00:13:37,133 --> 00:13:39,733 台积电可怕的地方，不只是技术 329 00:13:39,733 --> 00:13:42,366 而是它那种近乎偏执的制造文化： 330 00:13:42,366 --> 00:13:44,966 客户要求不能掉、良率每天抠 331 00:13:44,966 --> 00:13:48,366 问题必须追到根、工程师和供应商一起熬 332 00:13:48,366 --> 00:13:50,433 马斯克擅长快速迭代 333 00:13:50,500 --> 00:13:54,166 可晶圆制造不能像火箭一样一边炸一边学 334 00:13:54,166 --> 00:13:56,866 芯片客户要的不是“我们学到了很多” 335 00:13:56,933 --> 00:13:58,466 而是“你按时交货” 336 00:13:58,466 --> 00:14:00,100 更重要的是 337 00:14:00,100 --> 00:14:02,700 而是它身后站着合作多年的产业链 338 00:14:02,700 --> 00:14:04,966 还有一整个台湾半导体生态 339 00:14:04,966 --> 00:14:07,700 一颗AI芯片从设计到交付 340 00:14:07,700 --> 00:14:09,700 不是台积电一个人完成的 341 00:14:09,700 --> 00:14:12,700 前面有EDA和IP配合设计规则 342 00:14:12,700 --> 00:14:17,600 中间有设备、材料、光刻胶、硅片、气体、化学品 343 00:14:17,600 --> 00:14:22,166 后面有先进封装、测试、基板、HBM集成 344 00:14:22,266 --> 00:14:25,333 再往后还有广达、纬创、英业达 345 00:14:25,333 --> 00:14:27,500 鸿海这些服务器ODM 346 00:14:27,500 --> 00:14:30,966 把芯片变成整机、机柜、AI工厂 347 00:14:30,966 --> 00:14:33,300 这才是马斯克最难复制的地方 348 00:14:33,300 --> 00:14:34,966 他可以挖几个工程师 349 00:14:34,966 --> 00:14:38,533 但挖不走新竹科学园区几十年形成的供应链密度 350 00:14:38,533 --> 00:14:40,066 他可以买设备 351 00:14:40,066 --> 00:14:43,366 但买不走台积电和ASML、应用材料 352 00:14:43,366 --> 00:14:46,900 Lam Research、东京电子之间的长期协同 353 00:14:46,900 --> 00:14:48,533 他可以找英特尔合作 354 00:14:48,533 --> 00:14:52,500 但短期内复制不了苹果、英伟达、AMD、高通 355 00:14:52,500 --> 00:14:56,700 联发科一代代产品帮台积电磨出来的工艺数据库 356 00:14:56,700 --> 00:14:58,566 他可以建一座德州工厂 357 00:14:58,566 --> 00:15:00,100 但建不出台湾那种 358 00:15:00,100 --> 00:15:04,066 一通电话就能把封装、测试、基板、服务器代工 359 00:15:04,066 --> 00:15:07,433 散热、电源全部拉上桌的产业网络 360 00:15:07,433 --> 00:15:10,733 这也是为什么黄仁勋、苏姿丰、陈立武三个人 361 00:15:10,733 --> 00:15:12,500 都在往台湾跑的原因 362 00:15:12,500 --> 00:15:14,166 因为他们太清楚了 363 00:15:14,166 --> 00:15:17,200 AI时代真正稀缺的不是一栋厂房 364 00:15:17,200 --> 00:15:19,266 而是厂房背后的那张网 365 00:15:19,266 --> 00:15:21,733 台积电只是这张网最亮的节点 366 00:15:21,733 --> 00:15:23,000 真正可怕的 367 00:15:23,033 --> 00:15:25,600 是它背后整个台湾半导体产业链 368 00:15:25,600 --> 00:15:28,466 已经变成了AI时代最重要的基础设施 369 00:15:28,466 --> 00:15:31,066 黄仁勋说台湾是AI革命中心 370 00:15:31,066 --> 00:15:32,800 不是因为他喜欢喊口号 371 00:15:32,800 --> 00:15:36,066 而是因为英伟达的AI王座需要一整条供应链： 372 00:15:36,133 --> 00:15:38,933 和联发科的合作搞RTX Spark 373 00:15:38,933 --> 00:15:40,966 和台积电合作造芯片 374 00:15:41,033 --> 00:15:45,333 先进封装把GPU、CPU和高速内存连起来 375 00:15:45,400 --> 00:15:49,366 广达、纬创、鸿海把它们装成服务器和机柜 376 00:15:49,366 --> 00:15:52,233 最后还要靠光通信、电力、液冷 377 00:15:52,233 --> 00:15:54,666 把整座AI工厂真正跑起来 378 00:15:54,666 --> 00:15:57,366 苏姿丰的100亿美元也不是随便撒钱 379 00:15:57,366 --> 00:16:00,766 AMD点名日月光、矽品、力成 380 00:16:00,766 --> 00:16:02,266 就是因为她很清楚 381 00:16:02,266 --> 00:16:05,566 如果永远卡在台积电CoWoS排队队伍后面 382 00:16:05,666 --> 00:16:08,433 AMD就很难真正追上英伟达 383 00:16:08,433 --> 00:16:11,333 所以她要扶持更多先进封装路径 384 00:16:11,333 --> 00:16:15,133 让MI450、EPYC Venice、Helios 385 00:16:15,133 --> 00:16:18,000 这些系统级产品有机会按时交付 386 00:16:18,000 --> 00:16:19,466 陈立武更现实 387 00:16:19,733 --> 00:16:21,600 英特尔要复兴芯片制造 388 00:16:21,600 --> 00:16:24,033 但AI PC窗口已经打开了 389 00:16:24,166 --> 00:16:27,733 他不能等18A、14A完全成熟再开战 390 00:16:27,966 --> 00:16:29,766 所以他要台积电的产能 391 00:16:29,766 --> 00:16:32,166 要台湾生态帮他争取时间 392 00:16:32,433 --> 00:16:34,266 所以这一轮真正利好的 393 00:16:34,266 --> 00:16:35,800 不是一个孤立的概念 394 00:16:35,800 --> 00:16:38,266 而是AI基础设施的几个硬环节： 395 00:16:38,333 --> 00:16:41,733 先进制程、先进封装、AI服务器ODM 396 00:16:41,733 --> 00:16:43,733 高速互连、电力液冷 397 00:16:43,733 --> 00:16:46,133 台积电吃的是确定性交付 398 00:16:46,166 --> 00:16:50,566 日月光、矽品、力成吃的是先进封装外溢 399 00:16:50,566 --> 00:16:55,333 广达、纬创、鸿海吃的是AI服务器和整柜交付 400 00:16:55,333 --> 00:16:58,733 康宁、Marvell、Coherent、Lumentum 401 00:16:58,733 --> 00:17:01,033 吃的是光通信和高速互连 402 00:17:01,033 --> 00:17:04,933 这才是AI投资从上半场进入下半场的变化 403 00:17:04,933 --> 00:17:08,700 上半场，市场只问谁的GPU最强 404 00:17:08,700 --> 00:17:11,400 下半场，市场会开始问： 405 00:17:11,400 --> 00:17:15,300 谁能把GPU变成一整座可运行的AI工厂？ 406 00:17:15,300 --> 00:17:18,300 而答案，不只在英伟达的发布会上 407 00:17:18,300 --> 00:17:20,533 也在晶圆制造的产能表里 408 00:17:20,600 --> 00:17:22,466 在先进封装的产线里 409 00:17:22,466 --> 00:17:23,966 在服务器工厂里 410 00:17:24,200 --> 00:17:26,766 在光通信和液冷公司的订单里 411 00:17:26,766 --> 00:17:28,466 你是怎么布局的呢？ 412 00:17:28,733 --> 00:17:30,200 欢迎聊聊你的看法 413 00:17:30,366 --> 00:17:32,600 我是陪你解开财富密码的Olivia 414 00:17:32,700 --> 00:00:04,400 点赞关注，我们一起同行～ 415 00:00:04,400 --> 00:00:06,566 “Terafab” 416 00:00:34,733 --> 00:00:37,100 “我可能搞不定芯片供应了” 417 00:01:45,200 --> 00:01:47,666 Olivia 418 00:02:53,466 --> 00:03:02,733 888 419 00:03:19,533 --> 00:03:23,633 18A 420 00:04:09,066 --> 00:04:18,566 拆任务、调工具、读文件 421 00:04:52,666 --> 00:05:11,466 Client Computing Group：77亿美元 422 00:06:54,233 --> 00:07:01,933 18A完全成熟 423 00:09:54,433 --> 00:10:00,200 客户 424 00:10:52,500 --> 00:10:55,633 技术 设备 425 00:11:45,766 --> 00:11:59,966 沉积、蚀刻、离子注入 426 00:12:03,533 --> 00:12:12,233 按时、成套、稳定 427 00:12:53,966 --> 00:12:58,533 做出一颗芯片 428 00:13:06,433 --> 00:13:10,833 良率稳定 429 00:13:42,366 --> 00:13:48,366 客户要求不能掉、良率每天抠 430 00:14:09,700 --> 00:14:27,500 前面有EDA和IP配合设计规则 431 00:16:38,333 --> 00:00:05,333 先进制程、先进封装、AI服务器ODM 432 00:00:05,333 --> 00:00:06,566 熄火？ 433 00:03:20,166 --> 00:03:23,633 14A 434 00:04:11,566 --> 00:04:18,566 跑代码、访问数据库、管理权限、处理网络通信 435 00:04:58,900 --> 00:05:11,466 Data Center and AI：51亿美元 436 00:09:55,700 --> 00:10:00,200 技术 437 00:11:49,066 --> 00:11:59,966 CMP、清洗 438 00:12:05,700 --> 00:12:12,233 优先地进入你的产线 439 00:12:55,300 --> 00:12:58,533 稳定量产一批芯片 440 00:13:07,433 --> 00:13:10,833 成本可控 441 00:13:44,966 --> 00:13:48,366 问题必须追到根、工程师和供应商一起熬 442 00:14:12,700 --> 00:14:27,500 中间有设备、材料、光刻胶、硅片、气体、化学品 443 00:16:41,733 --> 00:04:15,700 高速互连、电力液冷 444 00:04:15,700 --> 00:04:18,566 同时协调多个模型和多个应用 445 00:09:56,766 --> 00:10:00,200 设备 446 00:11:50,933 --> 00:11:59,966 检测、量测、缺陷分析、自动搬运 447 00:13:08,333 --> 00:13:10,833 交期可靠 448 00:14:17,600 --> 00:09:57,933 后面有先进封装、测试、基板、HBM集成 449 00:09:57,933 --> 00:10:00,200 人才 450 00:11:54,533 --> 00:11:59,966 洁净系统、化学品供应 451 00:14:22,233 --> 00:09:59,066 再往后还有广达、纬创、英业达、鸿海 452 00:09:59,066 --> 00:10:00,200 钱32231

Can't find what you're looking for?
Get subtitles in any language from opensubtitles.com, and translate them here.