本年7月天门锚索价格，历久唱空AI的知名科技行业评述东说念主EdZitron近期发布1.5万字重磅著作《LetAIBurn》，抛出了个迄今具冲击力的判断：

着实的AI泡沫，试验上即是“OpenAI泡沫”。

以下是中枢内容顾忌：

空头中枢结论：OpenAI是“系统要紧机构”

EdZitron认为OpenAI是扫数这个词AI投资周期的“信用锚”——自ChatGPT问世以来，投资者对AI更变寰宇、GPU需求增、大模子盈利的扫数信仰都建立在OpenAI不息成长的前提上。旦它倒下，冲击远角兽自己，类比2008年雷曼兄弟流露金融体系杠杆脆弱，OpenAI失败将终结AI狂热期间。

营救论据：触目惊心的财务与买卖模式纰谬

财务黑洞：征引所谓OpenAI审计财报数据，2025年营收130.7亿好意思元、总开销340亿好意思元、净亏本385亿好意思元（2024年营收37亿、亏50.9亿，亏本增速远收入）；2026年Q1营收57亿、钱37亿，展望全年170亿（2025年为90亿），到2030年底累计虚耗或8520亿好意思元，离盈利越来越远。

买卖模式两大硬伤（称“史上大成本错配”）：

理成本过：用户发问不息虚耗GPU、电力、服务器成本，巨额廉价/用度户致限度越大亏本越重；面前AI服务订价远低于试验成本（“1好意思元40好意思元的服务”），企业收入跟不上成本增长，且不少企业已在削减AI开销。

成本开支远快于现款流：AI行业掀开销已从覆按转向理算力、GPU采购、数据中心诞生，OpenAI2026年仅算力开销展望500亿好意思元，四大科技巨头2026年景本开销7650亿好意思元；且AI行业70营收试验是OpenAI/Anthropic的运筹帷幄开销（风投→AI实验室→云厂商→英伟达的资金闭环），非真实健康需求。

雷曼时刻的传逻辑

OpenAI崩溃不会遽然发生，而是先融资变难→数据中心运用率下降→干系企业估值重估→传至科技股市集；Oracle、CoreWeave等依赖AI基建需求的企业当其冲。此外OpenAI、Anthropic、SpaceX的IPO若发达欠安，会激励流动转动（基金接盘、投资者抛售其他股）的四百四病

让AI行业自生自灭（汉文全文）

本周我发布了《软银黑粉全指南》，讲透了科技圈狂赌徒的辩护旧事。当初几笔幸运的早期投资让它表象限，如今却为AI泡沫史上限度大、也解气的崩盘埋下伏笔。本周五，我还会度剖析存储芯片行业，解释为什么咫尺配台能游戏PC贵到普通东说念主难以承受。

订阅我的付费栏价比，也恰是靠付费读者的营救，我本领每周产出这篇万字、度调研的费长文。

不要补救、不要补贴、不要独特待、不要税收减、不要《芯片与科学法案》歪斜政策，不成建立主权金钱基金托底AI产业。是期间让扫数这个词AI行业自食遵守了——它对待社会的式，本就配不上任何坦护。这个行业是处，是告贷无门的硅谷编造出来的骗局，虚构催生了数万亿好意思元的公论造势，还埋下多场本可避的金融危险；论若何，咱们都不该动手保全它。

只须你听到有东说念主拿起“行业补救”“大而不成倒”或是“主权金钱基金托市”，就要明显：这是业内东说念主士在想设法营造“行业不成倒下”的假象。但事实上，这里每企业都和普通初创公司样脆弱不胜。

另外，媒体乃至全社会都减轻给与了“泡沫离散后政府会动手补救”的预设，只因2008年金融危险的始作俑者后皆备逃走追责。但我须厘清：AI产业和金融业有着试验离别。它对国民经济并非刚需，所谓行业体量，靠劈头盖脸的营销炒作撑起来。

这是群失败者堆出来的泡沫。它之是以能越吹越大，非是硅谷、主流媒体和歪曲的股市达成了默契——股市浪漫圈钱与轮回融资。本年季度，OpenAI营收57亿好意思元，Anthropic接近50亿好意思元，可这些收入大多来自不计成本虚耗AI算力代币的企业客户，而脚下扫数企业都在狂削减干系开支。

生成式东说念主工智能历久实现不了通用东说念主工智能（AGI），它能作念到的事，也和行家领会里的东说念主工智能相去甚远。它莫得自主行动本领，谈不上任何“智能”，不存在自我意志与着实的学问储备。论给大模子套上些许层研讨框架、编写些出嫁套剧本，按照OpenAI我方的官说法，它在数学层面注定会产滋事实幻觉。据我估算，全行业至少七成收入都仅仅OpenAI与Anthropic虚耗算力产生的活水。而这两企业亏本其严重，也即是说扫数这个词AI产业的试验，非是风投把资金运送给云厂商，云厂商再把钱流向英伟达或是数据中心基建。

倘若这套软件真有实实的价值，它不错闻鸡起舞，根柢不需要轮回融资和个东说念主顾惜式宣传来续命。要是它真实二，也不会有巨额狂热信徒，凡是有东说念主质疑这套居品，就群起而攻之。东说念主类历史上，从来莫得哪款器具或商品，需要靠如斯歇斯底里、排他强的宣传去兜销，后施展不是场骗局。不少东说念主对大语言模子、以及背后斥地公司产生了病态的执念，驱动他们留神AI产业的不是任何训练时候或AGI落地的实证，仅仅这种歪曲的情怀依附。

这套软件自己自带阴霾属：面，它会放大东说念主中的坏心；另面，背后运营企业的底也不胜入目。每隔几周，Anthropic就会出批新发言东说念主，个个神气麻痹、毫灵气，言语越来越不端、充满邪教式狂热，与现实寰宇脱节。硅谷自重神论，但Anthropic里面以及粉丝群体中宽裕着令东说念主不安的遵从氛围。你不错联想款游戏端、及其的饭圈，再叠加金融投契、割韭菜、芳华期式内斗，扫数东说念主只为款线上软件狂。

交付诸位，现实里没东说念主存眷什么“智能轮回”，没东说念主钻研分词、词元化。你在街上已然拉个路东说念主说“理算力”，对只会认为你该去看精神科大夫。没东说念主在乎OpenClaw这类小众技俩。通晓点，多外出走走，你咫尺的款式和子异。你父母知说念你注册了几十个Claude账号吗？这种千里迷早已病态。

长篇大套，AI之是以能在经济中占据席之地，根源独一个：微软、谷歌、Meta、亚马逊主义2026年参加7650亿好意思元成本开支，2027年再万多亿。它们再也拿不出新的速增长赛说念，即便生成式AI于今没能创造具备限度、可不息的营收（别说盈利）。我上周在CNBC节目上也点明了这点：这些巨头从未单泄露地说念的AI业务收入。

谷歌拿不出二个谷歌搜索，微软造不出下代Office，Meta复刻不了Facebook，亚马逊也莫得二个AWS。正因如斯，它们要全社会默许AI是划期间机遇，却从不拿出试验营收数据佐证，只可靠狂钱基建制作秀象。可要是剔除OpenAI和Anthropic两企业（头部AI公司89的活水都来自它们的算力采购），全球AI行业年营收多也就200亿好意思元。

行业从业者总把句话挂嘴边：“咫尺还处于早期阶段”，拿往常互联网泡沫类比，只为劝服行家摄取AI，或是把东说念主类史上限度空前的成本错配包装成“搭建灵验的数字基建”。

别再强行把当下AI泡沫对标互联网泡沫

我直白说：AI显卡只适配生成式AI，险些莫得其他用途。大语言模子每次所谓“翻新”，都仅仅靠数十亿资金堆东说念主力、堆算力换来。全行业顶东说念主才、一起媒体流量、大多数投融资，皆备押注在单赛说念，后产出的大模子成本远于东说念主力，实用价值却大幅不足真东说念主。

AI从业者总泛论虚缥缈的畴昔遐想，只因把居品落地后对比其天价成本，试验果粗造到不值提。仅2026年，Anthropic和OpenAI融资总和（若一起交割完成）就3000亿好意思元，占据全球大多数AI算力资源。他们只可不停画远期大饼，因为咫尺落地的扫数成果，连其参加成本的头都法施展理。

旦AI泡沫离散，如今搭建的海量算力基建险些莫得复用价值。我在付费栏里写过，此次泡沫的后果远比互联网泡沫严重：显卡和光纤不是回事。显卡启动耗电惊东说念主，数据中心运维成本居不下，适用场景其单，且度连合。光纤泡沫离散后不错大领域二次运用、廉价折价计帐，但显卡作念不到。普通好者力邻接，不息运营成本居不下，烂尾的数据中心也很难完工周转。

互联网泡沫时期搭建光纤，往常有套广为流传的坏话：集中流量每90天翻番。各大券商研报、企业财报电话会、投资东说念主路演、扫数科技媒体都在访佛这个说法，但这套数据是捏造的。AT&T集中照顾员安德鲁・奥德里兹科核查真实流量数据后发现，好意思国骨干网流量试验年才翻倍——增速如实可不雅，但远够不上分析师饱读励的离谱数字。运营商靠面目财务作秀覆盖数据缺口，技能堪比中叶纪真金不怕火金方士：出售数十年光纤使用权，次全额计入营收；运营商之间彼此交换带宽，虚增账面收入。

另外，互联网泡沫时期行业如实处于起步阶段，网速差。2000年，仅有52好意思国成年东说念主使用互联网，2003年这比例也只涨到61。寰宇银行数据骄傲，2005年全球网民仅占总东说念主口16，到2024年才擢升至71。往常上网靠56K拨号，按分钟计费，速率慢；2000至2002年，好意思国平均网速仅400千比特/秒，换算成下载速率约50KB/s。如今好意思国平均网速200兆比特/秒，下载速率可达25MB/s。泡沫离散后，光纤依旧保留残值，存活下来的运营商不息受益。比如2004至2010年，威瑞森斥资230亿好意思元铺设庭光纤宽带，依托的恰是泡沫时期留传的光纤资源。

反不雅生成式AI，如今劈头盖脸浸透扫数场景：险些每款App都强制内置AI，脸书、谷歌、微软扫数账号默许开启AI，2023年以来扫数主流媒体反复报说念AI观念。OpenAI、Anthropic嘴上说着算力不足，可“扩建数据中心”除了托举英伟达销量、给云厂商提供费钱渠说念，看不出任何试验价值。算力短缺根柢莫得崎岖行家使用AI居品，也莫得拦住创业公司发布新址品。没东说念主能说簇新建海量机房到底要承载什么刚需，只可辩护解释“给OpenAI、Anthropic虚耗算力”。

Anthropic所谓算力紧缺，涓滴莫得盘桓它覆按、发布Mythos、Fable系列大模子；它从SpaceX采购数百兆瓦算力后，大的新闻仅仅上调用户额度，让客户每月花200好意思元，就能虚耗价值8000好意思元的代币。数据中心诞生迟缓，莫得归天任何AI企业发展。扩建算力的一起逻辑，不外是“造多算力，等着客户费钱虚耗”，莫得骄傲的市集刚需营救。

英伟达宣称2027年底显卡销售额将破损万亿好意思元，为邻接这笔天量需求，行业要参加至少4350亿好意思元搭建算力基建。咱们到底需要多大限度的配套身手？按英伟达商量，到2027年末要售出承载30吉瓦以上算力的显卡，配套数据中心总供电容量需要达到40吉瓦（电源使用率PUE取1.35）。仅机房内AI算力硬件（不含制冷、输电损耗）每兆瓦造价1200万好意思元，仅IT硬件参加就达4350亿好意思元，地皮、土建、东说念主力、输电、制冷等成本还要另算。

OpenAI预估2026年算力开销达500亿好意思元，Anthropic参加限度与之左近。除微软、谷歌、亚马逊为两企业兜底、租出算力外，仅有Meta领有可不雅算力参加（旗下Nebius、CoreWeave邻接算力业务）。而彭博社音书骄傲，Meta主义对外出售闲置算力，这又是大佐证：市集根柢不存在大限度真实刚需。

追捧AI的业内东说念主士会辩解“自建算力是为了发债融资”，但这套说辞站不住脚。当下全球三分之二风投资金涌向AI赛说念，云厂商手合手数千亿未践约算力订单，真相却不胜入目：云厂商五成待践约订单来自OpenAI与Anthropic。剔除这两企业，微软算力订单同比增长停滞，亚马逊增速仅20；谷歌数据虽杂沓，但举座趋势致。比及市集不再追捧AI策略、没东说念主再泛论“自主可控AI”时，扫数东说念主都会看清真相：AI真实市集需求其薄弱。

咫尺莫得可靠字据施展，理算力服务能够盈利。也即是说，即便开源大模子熬过前沿实验室泡沫，其单元经济模子也存在根柢纰谬。当下所谓AI需求，试验是公论裹带、软件强制紧缚AI逼出来的伪需求。如实有小数用户得意付费使用ChatGPT、Claude，但大多数算力虚耗来私费，或是售价低、严重亏本的套餐。用户每月花20、200好意思元订阅服务，其虚耗算力的真实成本远售价。借用科里・多克托罗的话顾忌：行业在用价值40好意思元的算力，好意思元廉价出售。这根柢算不上健康买卖，谈不上原生市集需求。

AI饱读励者宣称扩大限度就能摊薄成本，但现实恰好相悖：限度扩张从未镌汰AI使用成本。新代显卡、博通自研芯片、亚马逊Trainium、谷歌TPU，钢绞线皆备没能实现降价。就算畴昔成本古迹式下落，市面上也曾上架数万机柜的H100、H200、B100、B200、B300、AMD显卡，又该若何措置？

每次刷到有东说念主在外交平台说“行业尚早，大部分东说念主还没用过智能体”，我都忍不住想反驳。智能体观念早已充斥扫数媒体，凡是点基础网页检索、代码生成，从业者都强行包装成“智能体”。行家对AI好奇平平，中枢原因很简便：现存居品作念不到行家期待的自主智能——需巨额东说念主工辅导，自动完成种种事务；扫数东说念主都了了模子会编造信息；数据中心虚耗巨量电力、水资源，还能拿到大额税收补贴，背后掌权的却是凯文・奥利里这类投契富豪，或是奥特曼、阿莫迪这类脱离底层的硅谷精英。

那些天天心焦“AI会越好意思国”的东说念主，活在孩童式幻想里。先说事实：Anthropic我方都承认，廉价开源模子（包括自ClaudeHaiku4.5、月之暗面KimiK2.7），在安全罅隙检测上果和旗舰Fable模子不相崎岖。所谓大国科技竞争论调，仅仅游说政府披发补贴、减税收、廉价出让地皮给数据中心斥地商与地产投契商的器具。AI机房造价昂，泡沫离散后，企业想筹资完工、周转资产会难如登天。

我明显扫数东说念主都期盼泡沫崩盘后能迎来圆满结局，风尚套用熟习的历史案例（哪怕纳斯达克暴跌77）天门锚索价格，只因互联网泡沫后如实出身了批质企业。有东说念主拿步、AWS例如，施展AI行业终能盈利，但两类类比不成立。记者、券商分析师刻意忽略扫数风险信号，只会访佛“这就像往常的步”“对标AWS”。2003至2015年，亚马逊为AWS累计成本开支297亿好意思元，和当下涌向生成式AI的万亿参加不在个量；何况早在大模子问世前，云服务器托管就也曾领有实实、全企业层面的刚需求。

AI产业非“大而不成倒”，政府也力补救

OpenAI掌舵东说念主山姆・奥特曼忽视，把公司5股权划拨给好意思国政府，这笔股权估值约4200亿好意思元。奥特曼与OpenAI其他管还建议，好意思国头部AI企业各拿出5股权，组建类似阿拉斯加基金的国有载体。可OpenAI预估到2030年末累计亏本将达8520亿好意思元，戋戋5国有股权，只可减慢崩盘，法更变终结局。

营救这场泡沫的，是四大彼此交汇、法不息的危险：

数据中心投契泡沫

：市集盲目扩建AI显卡机房，预期每年能产生4500亿好意思元以上机房收入。可剔除两靠风投输的头部实验室，全行业原生刚需仅有几十亿好意思元限度。

AI初创公司泡沫

：大多数AI初创企业估值虚，看不到被收购、上市的可行旅途。扫数创业公司都依赖采购OpenAI、Anthropic代币，现款流虚耗速率快，吞吃全球大多数风投资金。

私东说念主信贷泡沫

：资管机构把万亿待业金、保障资金参加亏本、未完工的数据中心诞生技俩。

半体泡沫

：数据中心催生的需求填满芯片供应链，内存、存储价钱，扫数消费、企业电子居品加价，连机房私用硬件成本也同步高潮，酿成恶轮回。短短10个月，千兆瓦数据中心举座造价从500亿翻倍至1000亿好意思元。

对比2008年金融危险，就能认识AI产业不存在系统风险。往常若AIG歇业，全球数十亿人人储蓄、养老账户、市政基金、保障保单会一起清；持有AIG货币基金的普通投资者会本归。往常补救案中，好意思联储出往来商信贷器具、按期证券假贷器具，每个往来日向金融机构注入上千亿流动，才避市集停摆。

反不雅OpenAI与Anthropic，连同扫数这个词生成式AI行业，对宏不雅经济系统影响。它们仅仅营救好意思股“七大巨头”估值的符号，抛开2026年约7500亿至万亿算力开销、600亿计营收，对实体经济影响一丁点儿。剥离这两企业，全球剩余AI企业年收入险些不错忽略不计。

就算政府动手补救，也莫得明确补救处所。除非限期注入全球资金，等企业摸索出盈利模式——试验等同于兜底。两端部实验室莫得巨额表内债务；阿莫迪旗下Anthropic价值350亿好意思元的TPU硬件采购订单，有阿波罗全球经管、博通兜底。它们的股权价值只和风投基金挂钩：旦法上市，对应基金整期收益率将崩盘。

门聊聊软银

唯度绑定OpenAI、存在系统风险的大型企业，独一软银。我在《软银黑粉全指南》中详备阐发：软银押上扫数这个词公司畴昔，为奥特曼间断的算力钱提供400亿好意思元短期过桥贷款。要是OpenAI法按虚的私募估值上市，软银将遭受致命流动危险。

即便软银出现危险，风险也远不足往常全球金融崩溃。但软银是日本股市中枢上市公司，大鼓励是限度1.6万亿好意思元的日本政府待业金GPIF，后续或然能拿到日本政府定向扶植。但这仅仅单区域企业风险，和生成式AI产业自己关。

除非到2030年AI算力原生买卖需求暴涨几十上百倍，覆盖全球软件行业7790亿好意思元年总收入，不然泡沫崩盘带来的大限度损失可避。

旦云厂商大幅削减成本开支，英伟达显卡需求会一会儿挥发；随之崩盘的是代工AI服务器的台湾ODM厂商；内存、存储芯片企业收入暴跌，整条科技硬件供应链堕入历久零碎。这切的泉源，仅仅群只会招东说念主、裁人、钱在未教会证软件上的行业有策画者。

我反复强调：投资者、政策制定者浑浊了两件事——靠举债、机房投契、巨头寻求新增长故事催生的巨额成本开支，不等于多元、可不息的真实买卖需求。企业估值靠市集心思营救，莫得塌实的单元营收托底。旦市集预期转向下行，股价莫得任何基本面营救。

补充私东说念主信贷风险：我度担忧限度数万亿好意思元、流动差的浮动利率私东说念主信贷市集。咫尺法统计它在AI泡沫中的竣工敞口，但数百亿待业金也曾参加烂尾、盈利远景的数据中心技俩。公立待业金律例谢却奏凯重仓亏本AI基建企业，可监管罅隙让它们通过私东说念主信贷基金波折完成这场豪赌。

数千AI初创企业估值暴跌，联邦政府根柢力兜底。除非好意思国国会划拨数千亿征税东说念主资金，地说念收购毫价值的创业公司股权，只为稳住风投基金账面收益。如斯大限度补救法案，需要参众两院同期通过，政层面行欠亨：朝上派选民早已厌倦科技巨头间断拿补贴，保守派选民贯主张财政谨慎、留神工薪阶级征税东说念主利益。

别忘了，2008年轮银行补救法案轮投票奏凯未通过，两党议员反对与营救东说念主数各占半。其时全好意思金融体系正在及时坍塌，才拼集动后续补救。AI产业不存在同等别、山水相连的系统危险，不可能催生两党快速达成共鸣的蹙迫法案。

单看数据中心泡沫：政府向来放任未完工、消灭工业资产闲置多年。2008年四季度，全好意思11住宅空置，计入度假房后空置率达15。地皮自己具备残值，哪怕莫得建起堆满亏本显卡的巨型机房。政府莫得任何政策、经济动天真手补救，也不会出台项法案。金融危险后数千建筑企业奏凯歇业，2007至2012年好意思国建筑公司总量奏凯腰斩。

有东说念主会辩称，畴昔总统会出台定向科技补贴、项纾困政策，但这种泛论算不上严谨金融分析。要是扫数对AI泡沫风险的反驳，都只浑沌预计政坛腐化，那仅仅稠浊视听，而非客不雅基本面分析。

关键点：泡沫离散后，大型科技巨头短期不会垮台。英伟达概况率丢掉全球市值宝座，七大科技企业市值大幅缩水；除非出现端恶财务操作，微软、谷歌、Meta、亚马逊多作念数百亿资产减值；若查实英伟达违章向运送端显卡，还可能濒临SEC监管处罚。

但这不代表散户投资者、科技从业者、芯片供应链工东说念主、待业金持有东说念主、普通人人不会承受雄伟损失。泡沫离散后，AI行业大限度裁人、纸面金钱挥发数万亿好意思元、民用电子居品加价、私东说念主信贷基金连锁亏本，会让和硅谷投契圈层毫干系的普通东说念主常常蒙受经济损失。

这亦然我写下本文中枢不雅点的原因。

泡沫来临时，放任AI行业自生自灭

我重申态度：不要补救、不要补贴、不要独特待、不要税收减、不要《芯片法案》定向歪斜，不要建立主权金钱基金托举生成式AI产业。是期间让扫数这个词AI行业自食遵守——它对待全社会的式，不配得回丝坦护。要是这套恶运、法盈利的买卖模式脱离外部不息输就难以为继，这些企业须立承担后果，体面歇业。

列国政府以来对科技行业味退缩，被全球顶富豪集体蒙蔽，误以为奥特曼、阿莫迪正在造划期间居品，可试验上，他们仅仅作念出了东说念主类史上盈利本领差的行家软件。

咱们不需要所谓“自主AI国策略”，不需要AI主权金钱基金，不执着“好意思国跑大模子赛说念”——ChatGPT、Claude背后的大语言模子，是东说念主类史上宣传夸大、营销具诈欺的软件之。

就算大模子只可行动小众办公器具，也关紧要。着实的问题在于：AI行业提取限地皮、资金、稀缺当然资源，只为不息进套历久亏本、历久看不到盈利旅途的时候。它能出圈，仅仅扫数科技企业抱团造的转动视野器具，覆盖个狂暴真相：行业再也拿不出能驱动增长、颠覆消费与企业服务的全新址品。大模子能实现的切，都配不上万亿成本参加。

这场泡沫能彭胀于今，离不开被成本俘获的财经、科技媒体：媒体刻意夸大模子本领，对事实失真、经济崩盘隐患浮光掠影。巨额记者、分析师、评述东说念主减轻被富豪首创东说念主吸引，轻信他们在造有自我意志的东说念主工智能。比及市集万亿市值化为泡影，这群媒体从业者须承担饱读励投契狂热的包袱。2022年末以来大多数AI报说念，试验是服务顶富豪、吹起熬煎资产泡沫的宣传器具，终全球数亿普通东说念主将历久承受经济损失；哪怕中枢公司倒闭，奥特曼、阿莫迪依旧稳坐亿万富豪席位。

泡沫来临之时，须放任扫数这个词AI行业崩盘，竣工承受市集出清。生成式AI也曾拿到过多资金、媒体曝光、政策待与地球稀缺资源；要是脱离风投不息输、媒体脑吹捧就法存活，它不配社会保护，理当直面普通东说念主做生意失败时要承受的冰冷市集章程。

不存在神奇技能挽救这些买卖模式崩坏的企业。给OpenAI、Anthropic划拨4200亿全球资金，更变不了其底层单元成本失衡的硬伤，也虚构造不出英伟达2027年前显卡销售所需的40000亿年度原生买卖收入。

这群行业掌舵东说念主不是在造普惠变革畴昔，而是创造新机制固化金钱分化，给微软、谷歌、亚马逊、Meta找新借口扩大recurring收入，借“翻新”之名连合全球算力基础身手。

要是政策制定者读到这篇著作，请认清：你们被AI行业游说团体系统诈欺了。他们刻意渲染自身经济不可或缺，只为泡沫离散后拿到政府补救，或是劝服征税东说念主出资，在各州修建税数据中心园区。他们研发的大语言模子架构，历久终了不了自主通用东说念主工智能的数十年甘心。

我不幻想泡沫始作俑者会承担个东说念主法律包袱，但崩盘后出台监管、政策修订时，须追责到底：山姆・奥特曼、达里奥・阿莫迪、萨提亚・纳德拉、桑达尔・皮查伊、安迪・贾西、黄仁勋、马克・扎克伯格，以及扫数刻意制造全民AI共鸣、埋下下场全球金融危险隐患的管。

科技行业不完成层结构修订，硅谷历久只会制造靠共鸣炒作的泡沫，产出庄重现存金钱分层的器具。

因此天门锚索价格，凡是政客、说客、企业管示意定向补救AI行业，奏凯拒；他们索要新税收减、产业补贴的诉求概视。条款政策制定者安适演：要是行业对生成式AI的一起中枢假定一起造作，历久经济代价是什么。AI高潮终结之后，咱们须竣工复盘整场崩盘，杜同类投契狂热重演：查清每造势媒体、每支跟风风投、每位夸大宣传的企业管与网红，若何联手编造大模子更变寰宇的叙事，制造全民投契。

历史上每轮大型资产泡沫，尘埃落定后风险制造者都少被追责。我担忧本次崩盘带来的经济冲击会常常涉及普通庭。咱们须尽全力竣工复盘泡沫发源，落地长监管机制杜历史重演，这意味着全社会要直面几个千里重议题：零落拘谨的私东说念主金融体系、被成本裹带的主流媒体、科技翻新的投融资、估值、并购、补贴章程。

这套计帐相通适用于集中上狂热的AI信徒。数百万网民酿成端排外心态，凡是有东说念主不把买卖科技公司的营销话术当成客不雅科学说念理，就恶语相向。科技圈层里这种邪教式遵从，是根蒂固的文化弊病，须。

接下来的市集重置可避，AI泡沫终结会激励全科技行业重估，给依附华尔街成本、视行家福祉的硅谷文化记要的现实警钟。行业内个东说念主顾惜文化根柢不在乎普通作事者，他们顾惜富豪精英，幻想活在顶层把控的分层社会。

我不给与他们这种局促、自私的畴昔叙事，不承认这是东说念主类唯归宿。

几周前我写过段话：公论把AI泡沫包装成东说念主类可避的科技畴昔，试验却是中期硅谷旺盛又漫长的终结。独一个被纸面金钱、缺乏价值创造主的科技行业，才会容忍万亿成本浪费在套表面、落地均实证的时候上；也独一想想缺乏、成本成瘾的硅谷圈层，会减轻被阿莫迪、奥特曼这类擅长挑动东说念主心的投契者收割。

AI高潮须终结，扫数亏本AI企业，应当在全球资金兜底的前提下竣工出清。

让AI行业自生自灭。

LetAIBurn

Thisweek,IpublishedtheHater’sGuidetoSoftbank—asordidtaleoftech’smostdegenerategambler,who,thankstoacoupleofearlyluckywins,hasmanagedtosetthefoundationsfortheAIbubble’sbiggest(andpossiblymostgratifying)downfall.And,onFriday,I’mgoingtotakeadeepdiveintothememoryindustry—andthereasonwhyyoucan’taffordanewgamingPC.

Subscribingtopremiumisbothgreatvalueandmakesitpossibletowritetheselarge,deeply-researchedfreepieceseveryweek.

Nobailouts,nohandouts,nospecialtreatment,notaxbreaks,noCHIPSact,andnosovereignwealthfund.ItistimetotelltheAIindustrytogofuckitself,becauseit’seffectivelydonethesametotherestofsociety.Thisindustryisunworthy—ashamconjuredupbyatechindustrythat’srunoutofideas,atrillion-dollars’worthofmanufacturedconsentandentirely-avoidablefinancialcrises—andshouldnotbeprotectedunderanycircumstance.

Everysingletimeyouhearsomebodydiscuss“bailout”or“toobigtofail”or“sovereignwealthfunds,”knowthatthisistheindustry,onsomelevel,attemptingtocreatetheairthatitcannotdie,wheninfacteveryoneofthesecompaniesisjustasweakandbrittleasanyotherstartup.

Ialsothinkthatthemedia—andtheworldatlarge—istooreadytoaccepttheprospectofabailoutafterwatchingthosewhodrovetheworldintoaditchin2008escapeblame,andImustbeclear:theAIindustryisverydifferenttothefinancialindustry.Itisinessentialtotheeconomy,anditsrelevanceisonlyaslargeasthehypecampaignthatsitsbehindit.

ThisisanindustryoflosersthathasinflatedonlybecauseofthejointmanufacturedconsentofSiliconValley,themainstreammedia,andanenshittifiedstockmarketthatrewardsgriftingandcircularfinancing.OpenAIhad$5.7billionandAnthropicalittleunder$5billioninthefirstquarterofthisyear—andthoserevenuesmostlycamefromcompaniesthatwereburningAItokensatahorrendousratebecausethey’djustbeenforcedtopaytheactualcostofAI—andnoweverybody’spullingbackonthatspend.

GenerativeAIwillnotbringusAGI,nordoesitdomuchofwhatweassociatewithartificialintelligence.Itisnotautonomous.Itisnot“intelligent.”Itdoesnothavethoughts,or“knowledge,”andnomatterhowmanylayersofharnessesandscriptsyouputontopofit,itisstill(perOpenAI)mathematicallycertaintohallucinate.Iestimatethatatleast70oftheentireAIindustry’srevenuesaremadeupofOpenAIandAnthropic’scomputespend,andasbothcompaniesarehorrendouslyunprofitable,thismeansthattheAIindustryis,forthemostpart,venturecapitalistsfunnellingmoneytohyperscalerssothattheycanfunnelthatmoneytoNVIDIAordatacentercapex.

Ifthissoftwarewereworthy,itwouldstandonitsowntwofeet.Itwouldn’tneedcircularfinancingandacultofpersonalitytopropitup,either.Ifitweretrulyspecial,therewouldn’tneedtobeanarmyofcrazedacolytesthatattackyoufornotpledgingyourselftothegraveyardsmash.Therehasneverbeenatoolorproductinhistorysoldwithsuchhysteriaandaggressivemonoculturalforcethathaseverturnedouttobeanythingmorethanagrift.Somepeoplehavedevelopedunhealthyrelationshipswithlargelanguagemodels(LLMs)andthecompaniesthatmakethem,andthat,notanycertaintyorproofofArtificialGeneralIntelligence(AGI),iswhatmotivatesthem.

Thissoftwareisuniquelydark,bothinwhatitunlocksinsomepeoplethroughitsuseandinthesenseoftheentitiesthatsellit.Somepeopleareingenuineaweofeachoftherotationofclammy,soullesspod-peoplethatsaunteroutofAnthropiceveryfewweeks.Eachonesoundsalittleweirder,morecultish,moredisconnectedfromtherealworld.SiliconValleymaybelieveitselfatheistic,butAnthropichasaworryingsenseoffanaticism,bothinthepeoplethatworkthereanditsfanbase.Imaginetheabsoluteworstfanbaseofavideogamepossible,andthenaddlayersoffinancialization,griftingandhighschooldramalacedwithpseudo-religiousattachment.Allforafuckingapp!

Please,people.Nobodyintherealworldcaresabout“loops.”Nobodyisthinkingabouttokenization.Ifyousaidinferencetoaguyonthestreetthey’dtakeyoutoseeadoctor.Nobodygivesashit.Theydon’tknowwhatOpenClawiseither.Growup.Gooutside.Yousoundlikealunatic.DoesyourmotherknowhowmanyClaude20xaccountsyouhave?It’sobsessive!

Anyway,theonlyreasonthatAIhasanypresenceinoureconomyisthatMicrosoft,Google,Meta,andAmazonareintentonspendingmorethan$765billionincapitalexpendituresin2026andatrillionmorein2027becausetheyhavenootherhypergrowthideas,eventhoughgenerativeAIhasyettoshowanyrealpotentialassomethingthatcandrivemeaningfulrevenues(letaloneprofits),asevidencedbythefactthatnoneofthesecompaniesbreakouttheiractualAIrevenues,apointImadeonCNBClatelastweek.

GoogledoesnothavethenextGoogleSearch,MicrosoftdoesnothavethenextMicrosoftOffice,MetadoesnothavethenextFacebook,andAmazondoesnothavethenewAWS.That’swhytheyneedyoutobelievethatAIisabigdealwithoutthemeverhavingtoprovewhyoutsideofcapitalexpenditures.Theywantyoutoassumethatallthismoneycan’tbewrong,eventhoughwhenyouremoveOpenAIandAnthropic(whorepresent89oftherevenuesofthelargestAIcompanies)theAIindustryis,atbest,pullingin$20billioninannualrevenue.

Andlorddotheywantyoutosay“it’searly,”andthatit’sjustliketheDotComBubble,allsothatyou’lleitheracceptAIasyourlordandsavioror,alternatively,helpjustifyoneofthelargestmisallocationsofcapitalinhistoryas“buildingusefulinfrastructure.”

StopPretendingThisAIIsLikeTheDotComBubble

Newsflash!AIGPUsareusefulforgenerativeAIandnotmuchelse.Every“innovation”inLLMshasonlybeenmadepossiblebythrowingbillionsofdollarsattheproblemeitherinheadcountorcomputecosts—everyounceoftalentinthetechindustry,everybitofmediaattention,everydollarofcapitalexpenditures,allfocusedononeindustrythathassuccessfullycreatedLLMsthataremoreexpensiveandsignificantlylessusefulthanhumanbeings.

ThereasoneveryAIpersonspeaksinpie-in-the-skyhypotheticalsisthattheactualoutcomesaredecidedlymediocrewhenyoucomparethemtotheirruinouscosts.AnthropicandOpenAIraised(assumingtheroundscompletelyclose)over$300billionin2026alone,andtakeupthevastmajorityofavailableAIcompute.Theyneedyoutospeakinthefuturetense,becausenothing—absolutelynothing—aboutwhat’sbeencreatedsofarjustifiesevenafractionofitsfinancialandinfrastructuralcost.

WhentheAIbubblebursts,noneofthisinfrastructurewillbeparticularlyuseful.AsIsaidinmypremiumabouthowthisisworsethantheDotComBubble,GPUsarenotfiberopticcable.Theycostvastlymorepowertorun,carryfarhigherongoingoperationalexpenses(ToPowerAndRunTheDataCenters),havesignificantlylessutility,andaresignificantlymorecentralized.Unlikefiber,theycannotberepurposedbroadly,liquidatedcheaply,orauctionedatasteepdiscountbycreditors.Thesearenotgoingtobeusefulforhobbyists,norwilltheybecheapertorun,norwillincompletedatacentersbecheapertofinish.

Thedot-combubble’sinfrastructurebuildoutwasjustifiedbyanotherpowerfulmyth—thatinternettrafficwasdoublingevery90days.Theclaimspreadthroughanalystreports,earningscalls,investorroadshows,andeverytechmediaoutletalive.Butthemathematicswerefiction.NetworkresearcherslikeAndrewOdlyzko(atAT&T),lookingatactualtrafficdata,foundthatU.S.backbonetrafficwasdoublingroughlyonceayear—rapidgrowth,butnothingclosetothewildclaimsanalystspushed.Carriersburiedthediscrepancyunderlayersofcreativeaccountingthatwouldhaveimpressedmedievalalchemists.Theysold“indefeasiblerightsofuse”—essentiallydecades-longleasesonfibercapacity—andbookedtheentirevalueimmediatelyasrevenue.Theyengagedinelaborate“capacityswaps”betweencarrierstoinflatetop-linenumbers.

Wewerealsofairlyearly,andinternetspeedswereatrocious.In2000,only52ofAmericanadultswereusingtheinternet,andby2003,thatnumberhadonlyincreasedto61.PertheWorldBank,in2005only16oftheworldusedtheinternet,andin2024,thatnumberhadincreasedto71.Whentheinternetwasconnectedtoviaa56kmodem,accesswaschargedby-the-minute,andobviouslymuchslower.Backin2000,2001,or2002,theaverageUSinternetspeedwas,atbest,400Kilobits/s,orroughly50kilobytesasecond,comparedtotheaverageUSinternetspeedofover200Megabitspersecond,or25megabytesasecond,today.Fiberopticcableretainedenormousresidualvalueafterthebubbleburst,bothfortheoperatorsthatsurvivedandtheproviders.Verizonspent$23billiononbringingFiOStopeople’shomesbetween2004and2010,forexample,buildingonthatleftoverfiber.

GenerativeAI,ontheotherhand,isfuckingeverywhere,andanyonewithaninternetconnectionexperiencesitineffectivelythesameway.It’snon-consensuallyavailableineffectivelyeveryapp—everyFacebook,GoogleandMicrosoftaccount,forexample—andeverymediaoutletknowntomanhasmentionedAImultipletimessince2023.OpenAIandAnthropicmightclaimtheyneedmoredatacenters,butit’sunclearwhat“moredatacenters”actuallyachievesotherthanproppingupNVIDIAandgivinghyperscalerssomethingtoinvestin.

Alackofdatacentercapacityisn’tholdingbackpeoplefromusinggenerativeAI,norisitstoppinganybodyfromlaunchingaproduct,norcananyoneactuallyexpresswhatitisthatthey’rebeingbuiltforotherthan“reasonsforAnthropicandOpenAItospendmoney.”Anthropic’ssupposedlackofcomputedidnotstopittrainingorlaunchingMythosorFable,andwhenitboughthundredsofmegawattsofcomputefromSpaceX,thebiggestnewswasthatitexpandedratelimitstoallowuserstoburn$8,000worthoftokensfor$200amonth.NothingaboutthepainfullyslowpaceofdatacenterdevelopmentappearstoberestrainingasingleAIcompany,theentireargumentformoredatacentersappearstobe“weneedmorecomputesothatpeoplecanbuyit”farmorethananycogentpositionaroundwhatthesecapacityshortagesactuallymean.

Howmuchinfrastructuredoweneedtospend$435billionormoretojustifythe$1trillioninGPUsalesthatNVIDIAclaimsit’llhavebytheendof2027?That’showmuchdemandwe’llneed.AsNVIDIAintendstoselloveratrilliondollarsofBlackwellandVeraRubinGPUsbytheendof2027,itneedstohavearound(assumingaPUEof1.35)40GWofdatacentercapacitybuilttosupportthe30GW+ofGPUsitwillhavesold.Atabout$12amegawattofcriticalIT(IE:thestuffinthedatacenterthatrunsAIcompute,andnoteverythingelse,likethecoolingsystemsandanytransmissionloss),that’s$435billioninpurelyITcapitalexpenditurebeforepower,land,construction,labor,transmission,cooling,oranyothercosts.

OpenAIestimatesit’llspend$50billiononcomputein2026,andAnthropicwilllikelyspendcomparableamounts.OutsideofMicrosoft,Google,AmazonrentingorbackstoppingcapacityforAnthropicandOpenAI—theonlyothertechfirmwithanymeaningfulcomputespendisMeta(withNebiusandCoreWeave)...andBloombergisreportingthatMetaisplanningtostartsellingitscomputebecauseitdoesn’tneedallofit,yetanothersignthatactual,realdemanddoesnotexistatscale.

AIboosterswithblackmoldproblemswillsay“thisisjusttohelpthemraisedebt,”butthatdoesn’tholdup,consideringthedemanddoesn’tappeartobethereatscalewithtwo-thirdsofallventurecapitalfundingfocusedonAI,andhyperscalersaresittingonmassiveremainingperformanceobligations—hundredsofbillionsofdollars’worth—ratherthanthegrimtruththat50ofhyperscalerRPOsarefromAnthropicandOpenAI,hidingthefactthatMicrosoft’sRPOgrowthisflatyear-over-yearandAmazon’sisonlygrowingatamodest20whenyouremoveAnthropicandOpenAI’shundredsofbillionsofdollars’ofcomputespend.Google’sisalittlemessier,butthepatternholds.WheneveryonestopsaskingaboutyourAIstrategyorramblingonabout“sovereignAI,”it’llbecomeblatantlyobviousthattheactualdemandforAIwasnotparticularlystrong.

Wehavelittlecompellingevidencethatprovidinganyinference-basedservicesisprofitable,whichmeansthatevenifopensourceAIoutlivesthefrontierAIlabs,theuniteconomicsarefundamentallybroken.AIdemandis,atthispoint,adirectresultofsocietalpressureandnon-consensuallyoverwhelmingcustomerswithAIfeatures.WhiletherearepeoplethatlikeandpayforChatGPTorClaude,vastmajorityofAIcomputedemandisfromservicesprovidedtopeopleeitherforfreeorsoldatsuchamassivediscountthatit’simpossiblethatanyoneona$20or$200-a-monthplancouldevenaffordtheseserviceshadtheypaidtheiractualtokencost.ToparaphraseCoryDoctorow,yourdemandisbasedonselling$40worthofcomputeforadollar.That’snotarealbusiness,noristhatorganicdemand.

AIevangelistsclaimscalingwilldrivecostsdown,butthatwouldrequirethemto…becomecheaper.Morecomputeisn’t(andhasn’t)loweringthecostofAI.NewerGPUsaren’tloweringthecost.Barely-testedBroadcomGPUs,AmazonTrainiumXPUs,andGoogleTPUsaren’tloweringthecosts.Eveniftheyweretosomehowmagicallydosointhefuture,whatdowedowiththemountainofH100,H200,B100,B200,B300orAMDGPUsalreadydeployedacrosstensofthousandsofdatacenterracks?

EverytimeIreadsomebodyonTwittersaythat“we’reearly”orthat“mostpeoplehaven’teventriedagents”Ifeellikescreaming.Motherfucker,everyoneistalkingaboutagentsineverysinglemediapropertyallthetime.AIboosterswillrefertoliterallyanyAIfeatureasanagent,evenifit’sabasicwebsearchorgeneratingcode.Thereasonthatmostpeoplearekindof“meh”aboutAIisthatitdoesn’tdothingsthattheyassociatewithAI(autonomouslyandautomaticallytakingcareofthethingstheyneedwithlittlepromptingorcoaxing),everybodyknowsithallucinates,andAIdatacentersarehorrifyingmonolithsofcapitalthatgetmassivetaxbreaks,useatonofwater,andarefrontedbyultra-wealthytecheliteslikeKevinO’Leary,orcharmlesslydisconnectedValleyelitistslikeAltmanandAmodei.

Everysinglepersonfreakingoutabout“whatifChinadoesAIbetterthanAmerica”islivinginachild’sfantasy.Ohno!Anthropicitselfalreadyadmittedthatcheaperopen-sourcemodels—includingClaudeHaiku4.5andKimiK2.7—wereabletoidentifytheverysamesafetyvulnerabilitiesastheirflagshipFablemodel.Nationalcompetitionrhetoricisjustatooltojustifypublicsubsidies,taxhandouts,andlandgiveawaystodatacenterandlandspeculators.AIdatacentersaremassive,expensiveoperations,andraisingmoneytofinish(orfurnish)oneafterthebubbleburstswillbevery,verydifficult.

Irealizethateverybodywantstheretobeahappyendingafterallofthiscollapses.Igetthatit’seasiertothinkofthingsinfamiliarterms—evenifsaidtermsinvolveda77dropintheNASDAQ—becausetherewassomethinggoodandniceattheend.PeoplepointtoUberorAWSasproofAIwilleventuallyturnprofitable,butthoseanalogiesdonothold.Themarketcapturedtechandbusinessjournalistsandsell-sideanalyststhatinsistedonignoringeverywarningsignandwavingawayproblemsbysayingitwas“justlikeUber(nope!)”or“justlikeAmazonWebServices.”Between2003and2015,Amazonspent$29.7billiononcapexforAWS,afractionofthetrillionsbeingthrownatgenerativeAItoday,andAWShadtangible,enterprise-widedemandforcloudhostinglongbeforeLLMsexisted.

AIIsNotTooBigToFail—AndYouCan’tBailItOutAnyway

Capital-hogSamAltmanhasfloatedtheideaofhanding5ofOpenAItotheUSgovernment,astakewortharound$42billion.AltmanandotherOpenAIexecutiveshavesuggestedthateachofAmerica’sleadingAIdevelopersallot5percentoftheirequitytoavehicleliketheAlaskaPermanentFund,andconsideringOpenAIhasprojectedtoburn$852billionthroughtheendof2030,that5stakewouldonlyexisttoprolongtheinevitablecollapse.

Thisbubbleisbuiltonfourinterlockingunsustainablecrises:

Adatacenterspeculationbubble,wherewe’rebuildingAIGPUcapacityinexpectationof$450billionormoreinannualdatacenterrevenueforanindustrythat,withouttwounsustainableventure-backedlabs,hasonlyafewbilliondollars’worthoforganicdemand.

AnAIstartupbubble,wherethevastmajorityofAIstartupsarebothover-valuedandhavenoforeseeablepathtoacquisitionorapublicoffering.ThesestartupsalsorelyentirelyonbuyingtokensfromOpenAIandAnthropic,makingthemfarmorecash-intensive,soakingupthemajorityofglobalventurecapitalfunding.

Aprivatecreditbubble,whereassetmanagershavesunkbillionsofdollarsofpensionandinsurancefundcapitalintounprofitableAIdatacenterconstructionprojects.

Asemiconductorbubble,wheresupplychainshavebecomesaturatedwithartificialdemandfromthosebuildingAIdatacenters,inflatingthecostofRAMandstorage,makingallconsumerandenterpriseelectronicsmoreexpensive,includingthehardwareinsideAIdatacentersthemselves—creatingaviciouscyclethathasdoubledthecostofagigawattdatacenterfrom$50billionto$100billioninalittleunder10months.

TocontextualizewhyAIcannotbetreatedlikethe2008financialcrisis,let’srevisitthesystemicriskofthebankingcollapse.IfAIGhadfailedin2008,itwouldhavewipedouthundredsofbillionsinconsumersavings,retirementaccounts,municipalfunds,andinsurancepoliciesacrosstheglobe.Smallinvestors,includinganyonewhoownedmoneymarketfundswithA.I.G.securitiescouldhavebeenwipedoutentirely.Alittle-discussedpartofthescaleofthe2008bailoutweretheemergencyliquiditymechanismscreatedtostopthebleeding—thePrimaryDealerCreditFacilities(PDCF)andTermSecuritiesLendingFacilities(TSLF)thatprovidedasmuchas$100billiondollarstobanksandfinancialinstitutionseverysingletradingdaytopreventtotalmarketseizure.

Bycomparison,OpenAIandAnthropicaresystemicallyirrelevant,muchliketherestofthegenerativeAIindustry.WhiletheirexistencepropsupthesymbolicvaluationoftheUSstockmarket’sMagnificentSeven,theiractualeconomicfootprintistiny,outsideofwhatIestimateisaround$75billionto$100billionof2026computespendandroughly$60billionofcombinedtop-linerevenue.Removethosetwofirms,andtherestoftheglobalAIindustry’sannualrevenuebarelyregisters.

It’salsounclearwhatexactlyyouwouldbailout,unlessthegovernment’splanistofeedthemendlesscapitalforalleternityuntiltheysomehowinventafunctionalprofitablebusinessmodel(so,forever).Neitherofthemcarrymassivebalancesheetdebt—andBroadcomisbackstopping$30billionofAnthropic’s$35billionTPUhardwaredealwithApolloGlobalManagement—andtheirequitypositionsonlymattertoventurecapitalfirmsinthesensethattheirentirefundvintageswillpainfullyunderperformifOpenAIandAnthropiccannotIPO.

WeShouldTalkAboutSoftBank

ThereisexactlyonelargecorporationthatissystemicallydependentonOpenAI:SoftBank.AsIcoveredinthisweek’sHater’sGuidetoSoftbank,SoftBankhaswageredeffectivelyitsentirecorporatefutureon$40billionormoreinshort-termbridgeloanstofundSamAltman’sendlesscomputespendingspree.IfOpenAIcannotcompleteapubliclistingatitsinflatedprivatevaluation,SoftBankwillfaceanexistentialliquiditycrisis.

Thisrisk,again,isnothingcomparedtotheglobalsystemiccollapsethatwouldhaveunfoldedifAIGorLehmanBrotherscollapsedwithoutintervention.Thatbeingsaid,SoftBankisoneofthelargestlistedcompaniesontheJapanesestockmarket,anditssinglelargestinvestorisJapan’s$1.6trilliongovernmentpensioninvestmentfund(GPIF),andthusSoftBankmightsecuresomeformoftargetedJapanesestatesupportdowntheline.Thatisanisolatedregionalcorporaterisk,notaglobaleconomiccollapserisktiedtogenerativeAIitself.

Thereisnoavoidingthecarnagetocome,outsideofamiraculousten-to-onehundredfoldexplosioninorganiccommercialdemandforAIcomputeby2030—ascenariothatwouldrequireAIcomputespendingtooutpacetheentire$779billionannualglobalsoftwareindustry’stotalrevenue.

NobailoutpolicycanreversethedownwardtrendoncehyperscalersslashtheircapitalexpenditurebudgetsandNVIDIA’sGPUdemandevaporates,whichwillinturncollapsetherevenuesofTaiwaneseODMsthatbuildAIserversforhyperscalers,whichwillinturncrushtherevenuesofmemoryandstoragesemiconductorfirms,triggeringaprolongedindustry-widedepressionacrosstheentiretechhardwaresupplychain—allcreatedbyBusinessIdiotsthathavenoideawhattodootherthanhirepeople,firepeopleandburnbillionsofcapitalonunproventools.

AsI’vesaidmanytimes,investorsandpolicymakersareconflatingmassivedebt-fueledcapitalexpenditures—drivenbydatacenterspeculationandhyperscalersdesperateforanewgrowthnarrative—withreal,diverse,sustainablecommercialAIdemand.Valuationsarepumpedupentirelyonmarketsentimentratherthantangibleuniteconomics,whichmeansthatwhenmarketsentimenttakesaviolent,permanentdownwardshift,thereisnounderlyingfundamentalrevenuestreamtopropupstockprices.

Asidenoteonprivatecredit:Iamdeeplyworriedabouttheprivatecreditindustryanditstrillionsofdollarsofilliquidfloating-rateloans,aswedon’treallyhavefullvisibilityonitstotalexposuretotheAIbubble,otherthanthathundredsofbillionsofpensioncapitalhavebeensunkintounfinished,unprofitabledatacenterconstructionprojects.PublicpensionfundsarelegallyrestrictedfrommakingmassivedirectpuntsonunprofitableAIinfrastructurecompanies,butthankstoamassiveregulatoryloophole,theycanmakethesamespeculativebetsbyproxybyshufflingcashtoopaqueprivatecreditfunds.

ThecollapseinvaluationofthousandsofAIstartupswouldnotbemeaningfullysoftenedbyafederalbailoutunlesstheUSgovernmentliterallycommittedhundredsofbillionsoftaxpayerdollarstobuyingworthlessstartupequitypurelytopropupventurecapitalfirms’internalfundreturns.AnysuchmassivefederalrescuepackagewouldhavetopassboththeHouseandtheSenate,andanybailoutoftheAIsectorwouldbeanincredibly-unpopularpoliticaldecision,infuriatingprogressivevoterstiredofBigTech’sendlesssubsidiesandconservativevoterswhoclaimtocareaboutfiscalresponsibilityandworking-classtaxpayers.

Asareminder,theinitial2008bankbailoutbillfaileditsfirstcongressionalvote,withRepublicansandDemocratseachfairlysplitonsupport—andthatrejectionhappenedduringamomentwhentheentireUSfinancialsystemwasactivelyimplodinginrealtime.ThereiszerocomparableemergencysystemicriskfromgenerativeAItoforcerushedbipartisanlegislativeaction.

Asfarasthestandalonedatacenterbubblegoes,thegovernmentisabsolutelywillingtoletunfinishedorabandonedindustrialassetslaydormantforyearsonend.Inthefinalquarterof2008,11ofallUSresidentialhomessatempty,or15ifyouincludevacationproperties.UnlikeAIdatacenterhardware,physicallandretainsresidualvalueevenifyouhaven’tbuiltagiantwarehousefullofloss-makingGPUsontopofit.Thereisn’tapolicyoreconomicimperativeforafederalbailouthere,andonewillneverbeforthcoming.AftertheGlobalFinancialCrisis,thousandsofconstructionfirmswereallowedtocollapseentirely,totheextentthatthetotalnumberofconstructioncompaniesoperatinginAmericahalvedbetween2007and2012.

Youcouldarguethatfuturepresidentialadministrationswillhandoutsweethearttechsubsidiesortargetedrescuepackages,butthatspeculativetalkingpointisnotseriousfinancialanalysis.IfeveryrebuttaltostructuralAIbubbleriskreducestovaguefearmongeringaboutpresidentialcorruption,youarecatastrophizingratherthananalyzingmarketfundamentals.

And,mostcrucially,thevastmajorityofbigtechwillbefine,atleastintheshortterm,whenthebubblebursts.NVIDIAwilllikelyceasebeingthelargestcompanyonthestockmarket,andtheMagnificentSevenwillsufferadramaticdrawdowninmarketcap,butoutsideofunforeseenhorrendousfinancialdecisions,theworstoutcomeIcouldseewouldbemulti-billion-dollarassetwrite-downsforMicrosoft,Google,Meta,andAmazon,andpotentialSECregulatoryactionagainstNVIDIAifitisproventohaveillegallyroutedadvancedGPUstoChinainviolationofexportcontrols.

Thisdoesnotmeanthatretailinvestors,techworkers,semiconductorsupplychainemployees,pensionholders,andeverydaycitizenswon’tsuffermassivefinancialharmastheyalwaysdowhenassetbubblesburst.TensofthousandsofAIindustrylayoffs,trillionsinerasedpaperwealth,inflatedhardwarecostsforallconsumerelectronics,andcascadinglossesacrossprivatecreditfundswillcreatewidespreadeconomicpainforordinarypeoplewithnoconnectiontoSiliconValley’selitebubbleculture.

WhichiswhyIammakingafirm,clearstatementtoendthispiece.

WhenTheTimeComes,LetTheAIIndustryBurn

Irepeatmyself:Nobailouts,nohandouts,nospecialtreatment,notaxbreaks,noCHIPSactcarve-outs,andnosovereignwealthfunddedicatedtoproppingupgenerativeAI.ItistimetotelltheAIindustrytogofuckitself,becauseit’seffectivelydonetheexactsametotherestofsociety.Thesecompaniesmustbeforcedtostandontheirowntwofeetandcollapsewithdignityiftheirwretched,unprofitablebusinessmodelscannotsustainthemselveswithoutendlessexternalcapitalinfusions.

Theworld’sgovernmentshaverolledoverandshowntheirbelliestothetechindustryforfartoolong,andhavebeenaggressivelyconnedbysomeoftherichestpeoplealiveintobelievingthatSamAltmanandDarioAmodeiarebuildinganythingotherthantheworld’sleast-profitablemass-marketsoftwareproductline.

Wedonotneeda“sovereignAIstrategy,”nordoweneed“asovereignAIwealthfund,”nordoweneedto“makesureAmericaleadsinAI,”atleastnotwhenwe’retalkingaboutlargelanguagemodels—theunderlyingtechnologyofChatGPTandClaude,twoofthemostover-hypedanddeceptively-marketedpiecesofsoftwareinhumanhistory.

WhetherornotLLMsareaminornicheproductivitytoolisirrelevant,becausetheAIindustryhasdemandedtheworldhanditunlimitedland,unlimitedmoney,andunlimitedscarcenaturalresourcestocontinueproliferatingatechnologythathasonlyeverlostmoneyandhasnolong-termpathtosustainableprofitability.Theonlyreasonithasgainedanymainstreamtractionatallisbecausetheentiretechindustryunifiedarounditasadesperatedistractiontohidefromthehardtruththatithasnonextparadigm-shiftingconsumerorenterpriseproducttodrivegrowth,andnothinganLLMcanremotelydojustifiesthetrillionsincapitalpouredintoit.

Andithasonlygottenthisfarbecauseofacapturedbusinessandtechmediaecosystemthatsystematicallyoverstatesitscapabilitiesandhand-wavesawayitsobviousaccuracyflawsandexistentialeconomicinstability.Therearefartoomanyjournalists,analysts,andcommentatorseasilywooedbycharismaticwealthyfounderswhopromisethey’rebuildingsentientmachineintelligence,andwhenthemarketsbleedredbythetrillions,thesepeoplemustaccepttheirshareofresponsibilityforstokingmassspeculativemania.Somuchoftheso-calledAIjournalismpublishedsincelate2022hasbeenweaponizedtofurtherenrichthealreadyultra-wealthyandinflateacatastrophicassetbubblethatwillinflictlastingfinancialharmonhundredsofmillionsofregularpeopleglobally—allwhileSamAltmanandDarioAmodeiremainbillionaireseveniftheircorecompaniescollapseentirely.

Whenthetimecomes,theAIindustrymustburn.Itmustbeallowedtofailwholesale.GenerativeAIhasalreadybeengivenfartoomuchmoney,mediaoxygen,politicaldeference,andscarceplanetaryresources,andifitcannotsurvivewithoutcontinualventurecapitalinfusionsandunceasingmediacoddling,itisentirelyunworthyofsocietalprotectionandmustfacethecold,hardmarketrealitythateveryregularpersonfaceswhentheirbusinessorcareercollapses.

Thereisnomagic“bailingout”thesebrokenfirms.Handing$42billioninpubliccapitaltoOpenAIorAnthropicwillnotfixtheirstructurallybrokenuniteconomics,norwillitconjureupthe$400billionormoreinannualorganiccommercialrevenuerequiredtosubstantiatejustNVIDIA’sprojectedAIGPUsalesthroughtheendof2027.

Thesepeoplearenotbuildingatransformativesharedfuture—they’reinventingnewmechanismstore-entrenchexistingwealthconcentration,togiveMicrosoft,Google,AmazonandMetanewexcusestoinflatetheirrecurringrevenuestreamsandcentralizeglobalcomputinginfrastructureunderthemarketingauspicesof“innovation.”

Ifanypolicymakersreadthis,knowthatyou’vebeensystematicallyconnedbytheAIindustrylobby.Theywantyoutobelievethey’reeconomicallyessentialsolelysoyou’llbailthemandtheirwealthyventurecapitalbackersoutwhenthebubblepops,orfunnelunlimitedtaxpayerfundsintobuildingthemtax-subsidizeddatacentercampusesacrossyourdistricts.Theyarenotbuildingautonomousgeneralartificialintelligence,norwilltheyeverdeliveronthatdecades-oldpromisewithcurrentLLMarchitectures.

Ithinkit’snaivetoimaginetherewouldeverbemeaningfulpersonallegalconsequencesforthearchitectsofthisbubble,butifsystemicregulatoryandpolicyreformsemergeafterthecrash,thepeopletoholdfullyaccountableareSamAltman,DarioAmodei,SatyaNadella,SundarPichai,AndyJassy,JensenHuang,MarkZuckerberg,andeveryotherC-suiteexecutivewhoforcefullymanufacturedmasspublicconsentforatechnologicaldeadendandlaidthefinancialgroundworktoservetheworlditsnextgreatglobalfinancialcrisis.

Untilfundamentalstructuralreformhitsthetechsector,SiliconValleywillneverbecapableofbuildinganythingotherthanconsensus-drivenhypecyclesandtoolsthatreinforcetheexistingwealthstatusquo.

So,spitinthefaceofanypolitician,lobbyistorexecutivewhoevenhintsatatargetedAIbailout,refusetoentertaintheirdemandsfornewtaxbreaksandstatesubsidies,anddemandthattheydothecomplex,ugly,unglamorousworkofweighingtheactuallong-termeconomicconsequencesiftheentireindustry’scollectiveassumptionsaboutgenerativeAIturnouttobecompletelywrong.WhenthisAIeraends,wewillneedtothoroughlyexcavateeverylayerofthecollapsetomakesurethisexactspeculativemanianeverhappensagain,identifyingeverymediaoutlet,venturefund,corporateexecutiveandpersonalitythatwasusedtomanufacturemassconsentandspreadunsubstantiatedmythologyaboutLLMsasworld-changingtechnology.

Everymajorassetbubblethathaseveroccurredhasmostlyleftthearchitectsofriskunexaminedandunpunishedafterthedustsettles.TheeconomiccarnageIfearwillfollowthisera’scollapsewillbewidespreadandbrutalforordinaryhouseholds,andwemustdoeverythinginourpowertoboththoroughlyunderstandexactlyhowwearrivedatthismomentandimplementpermanentguardrailstoensureitcannotrepeatitself—whichwillinvolvemanypainfulnationalconversationsaboutourunregulatedprivatefinancialsystem,capturedmainstreammediaecosystem,andhowspeculativetechnologicalinnovationisfunded,valued,acquiredandsubsidized.

ThesamereckoningappliestotheonlinecultacolytesofthisAIera.Therearemillionsofpeopleonlinewhohavedevelopedagenuinehostileresentmenttowardanyonewhodoesnotimmediatelyacceptafor-profittechcorporation’smarketingmessagingasobjectivescientifictruth.Thiscult-likefanaticismwithinthebroadertechaudienceisadeepculturalsicknessthatmustbedismantledonceandforall.

Muchofthiscomingmarketresetwillbeunavoidable,becauseIthinkwhatdirectlyfollowstheAIbubblewillbeamassiveindustry-widerevaluationoftheentiretechsector,anecessaryrealitycheckforaSiliconValleyculturethat’sfarmorebeholdentoWallStreetcapitalthanitistohumanprogressorbroadsocietalbenefit.Thecultsofpersonalitythatdominatethisindustrydonotcareaboutyou,orme,oranyworkingpersonotherthanthebillionairestheyworshipandtheirimaginedplaceinastratifiedsocietydominatedbytheultra-richandtheirunelectedcronies.

Irefusetoaccepttheirnarrow,self-servingvisionofthefutureasaninevitablefateforhumanity.

AsIsaidafewweeksago:TheAIbubbleissoldtothepublicashumanity’sinevitabletechnologicalfuture,butitactuallyresemblesthedrawn-out,expensivedeathofmid-eraSiliconValley.Onlyatechindustryentirelydominatedbysymbolicpaperwealthandhollowvaluecreationwouldevertolerateatrilliondollarsoftotalwastechasingastill-theoretical,unprovenlong-termoutcome,andonlyanintellectuallyhollow,capital-addictedValleyculturewouldbesoeasilygriftedbycharismatichucksterslikeDarioAmodeiandSamAltman.

Thiseramustend,andallfailedAIfirmsmustbefullyallowedtofailwithoutpublicrescue.

LetAIburn.

Abouttheauthor

EdZitronViewallarticles手机号码：15222026333相关词条:铁皮保温    塑料挤出机     钢绞线    玻璃卷毡厂家    保温护角专用胶

1.本网站以及本平台支持关于《新广告法》实施的“极限词“用语属“违词”的规定，并在网站的各个栏目、产品主图、详情页等描述中规避“违禁词”。
2.本店欢迎所有用户指出有“违禁词”“广告法”出现的地方，并积极配合修改。
3.凡用户访问本网页，均表示默认详情页的描述，不支持任何以极限化“违禁词”“广告法”为借口理由投诉违反《新广告法》，以此来变相勒索商家索要赔偿的违法恶意行为。