文 | Tech商业
在AI大模型的狂欢里，算力一直被视作新时代的“石油”。
为了抢夺算力，硅谷的巨头们不惜砸下数百亿美元；而一众没有大厂背景的“新云”服务商（Neoclouds）更是通过疯狂借债，买下成千上万张英伟达GPU，试图在这场淘金热中扮演“卖铲人”的角色。
然而，这场击鼓传花的游戏，正在迎来它的终局。
近日，埃隆·马斯克（Elon Musk）在X上与知名投资人查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）进行了一场极其硬核的对话。马斯克一针见血地捅破了AI算力市场当下的虚假繁荣：过去几周，债务市场几乎完全切断了对那些没有独立于AI数据中心实质收入的公司提供资金的渠道。
随之而来的，是马斯克极其冷酷的预判：“这意味着，那200多家‘新云’服务商将经历孔雀鱼般的消亡与合并。有些能幸存下来，但数量不多。”
与此同时，曾精准做空次贷危机的投资人查马斯也补了一刀，指出2027年将出现大量“搁浅的算力（stranded compute）”和半途而废的数据中心。
这场对话，不仅撕下了AI算力狂热的遮羞布，更宣告了AI基建行业的大洗牌，已经提前拉开了帷幕。
债务市场的急刹车：200家新云的出局
要看懂这场危机，首先要明白什么是“新云（Neoclouds）”。
过去两年，像CoreWeave这样的公司异军突起。他们不像亚马逊AWS或微软Azure那样拥有庞大的通用云计算生态，他们只做一件事：从英伟达买来GPU，然后高价租给急需算力的大模型公司或AI创业团队。
这本质上是一场极其极端的“高杠杆+重资产”金融游戏。这些公司没有稳定的现金流支撑，只能依靠华尔街的债务融资来购买动辄几万美元一张的显卡。他们的商业模式极其脆弱：只要算力租赁的价格下跌，或者借新还旧的链条断裂，公司就会瞬间陷入绝境。
马斯克在对话中极其精准地点出了目前的死穴：债务市场的资金水龙头已经关上。
在经历了长期的狂热后，华尔街的金融机构终于开始审视风险。当他们发现，许多AI初创公司的收入根本覆盖不了高昂的算力账单，而“新云”公司不仅缺乏抵押物，且极度依赖高息债务时，资本的避险本能被彻底激活。
马斯克拿自己的公司举了例子：“SpaceX为GPU筹集了400亿美元的债务，但我们拥有投资级信用评级，并且有星链和发射收入。如果你没有实实在在的现金流，就别想从债务市场拿到一分钱。”
这就是最残酷的现实：那些没有“独立于AI数据中心实质收入”的200多家新云公司，即将面临资金链断裂。马斯克用“孔雀鱼般的大量消亡”来形容这场即将到来的倒闭潮。在资本退潮后，沙滩上只会剩下一地鸡毛和无法变现的二手显卡。
2027年的“搁浅算力”：基建狂飙后的废墟
在这场对话中，查马斯提出了一个更为致命的产业痛点：2027年的“搁浅算力”危机。
查马斯直言，他正在目睹一场因为官僚主义导致的灾难。很多科技巨头（Hyperscalers）在建设数据中心时，流程极度拖沓。“英伟达光是签个保密协议（NDA）就要三个月，委托期甚至要求18个月。”
这种迟缓的供应链和官僚主义，加上电网容量、建筑审批等物理限制，导致大量数据中心无法按时完工。与此同时，大厂们为了自己的大模型业务，会把所有自建的算力牢牢攥在手里，优先满足内部需求，其次才会分给第三方。
这就造成了一个极其荒诞的资本浪费：许多公司花了几十亿美元买来了最顶级的GPU，建了庞大的机房，但到了2027年，却因为没电、没网、没应用场景，导致这些算力变成了“半拉子工程”，堆积在仓库里吃灰，无法产生任何收益。
这些“搁浅的算力”，不仅是科技巨头的财务灾难，更是对那些靠租赁算力生存的中小AI公司的一记重击。钱已经花了，设备却成了毫无用处的废铁。
巨头的挤压：谁在偷笑，谁在流血？ 在这场算力寒冬中，并不是所有人都在瑟瑟发抖。
正如查马斯所指出的，这种错配只会让两类公司获益： 第一类，是能够快速建立起工业级规模的非官僚巨头。 比如马斯克的xAI，依托于强大的工程文化，能够极速建成拥有十万张H100的超级计算集群。 第二类，是资金极其充裕且足够敏捷的玩家。 比如微软和Meta，他们拥有庞大的现金流，可以无视短期债务市场的波动，继续囤积粮草。
而对于夹在中间的“新云”公司和大模型初创企业来说，他们面临的不仅是算力租不到，还有巨头们极其苛刻的“数字剥削”。
查马斯爆料称，大厂现在甚至不允许第三方模型公司直接接入，他们要求模型公司必须走大厂的“模型路由器（model router）”，并从中抽取极高的过路费。他甚至暗示，在Anthropic此前被曝光的S-1机密文件中，就提到了“大厂告知他们没有多余的算力”。这印证了算力短缺和巨头垄断，已经达到了极其危险的地步。
做空的阴云：马斯克与查马斯的“双簧”？
这则推文最有趣的一个细节，发生在评论区。
当马斯克和查马斯两人在对AI算力泡沫指点江山、表达悲观预期时，做空机构Hedgeye的创始人Jack Gee直接在下面开炮了：“你们俩配合得真棒。这看起来是个不错的分散风险、打击竞争对手的方法。”
Jack Gee的嘲讽可谓一针见血。查马斯在2021年的SPAC（特殊目的收购公司）泡沫中，曾靠着疯狂吹捧和虚假宣传，坑害了无数散户，套现离场。如今，他和马斯克两人一边重仓AI基础设施，一边在公开场合唱衰那些借债囤卡的“新云”公司。
他们的言论，客观上起到了打压竞争对手的作用。一旦市场恐慌加剧，债务利率飙升，那些依靠高杠杆买卡的竞争对手就会陷入死局，而像SpaceX和xAI这样有着稳定现金流和多元化融资渠道的巨头，反而能趁机抄底，完成行业的寡头垄断。
结语：AI大基建的“成人礼”
AI算力的狂奔，终于撞上了金融常识的南墙。
过去两年，人们以为只要买到英伟达的显卡，就能印钞。但这场由债务驱动的狂欢，掩盖了商业模式的巨大漏洞。AI基建，本质上是长周期、重资产、低容错的传统业务。它需要电网的配合、供应链的支撑、以及极其庞大的资本储备。
2027年，注定将是AI算力产业极其动荡的一年。大量不合格的“新云”服务商将被清洗，不合规的AI初创公司将被耗死，而缺乏电力和配套设施的“搁浅算力”，将成为这场泡沫最惨烈的遗迹。
在技术革命的浪潮中，那些只靠PPT和借债吹起来的泡沫，终将被重力无情地撕碎。最终能笑到最后的，永远是那些拥有最强现金流、最硬核工程能力和最冷静商业头脑的少数派。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Tech Business
At the AZLLM carnival, calculus has been seen as the “oil” of the new era.
The giants of Silicon Valley were willing to destroy tens of billions of dollars in order to win the game; and Neoclouds, a group of “new cloud” servicemen without a big factory background, were trying to play the role of “thrower” in this gold rush by borrowing thousands of English-Wida GPUs in a crazy way.
However, this game of drumming flowers is coming to its end.
In recent days, Elon Musk has engaged on X in an extremely hard-core dialogue with Chamath Palihapitiya, a prominent investor. Mask has cut the current false prosperity of the AI calculus market: in the past few weeks, the debt market has almost completely cut off access to funding for companies that do not have a substantial income independent of the AI data centre.
And what follows is a very cold assumption by Mask: “This means that more than 200 new cloud service providers will experience the extinction and merger of peacockfish. Some survive, but not many.”
At the same time, Chamas, an investor who had been well qualified for the empty subprime mortgage crisis, added a knife to the point that in 2027 there would be a large number of “standed computes” and half-way data centres.
This dialogue, which not only tore down the veil of Al-Calculation fanaticism, but also proclaimed the great shuffle in the AI infrastructure industry, has opened up the curtain ahead of time.
Debt market brakes: 200 new clouds out
It is important to understand the crisis first and foremost what is “Neoclouds”.
In the past two years, companies such as CoreWeave have sprung up. Unlike the Amazon AWS or Microsoft Azure, they have a huge generic cloud computing ecology, and they do only one thing: buy GPUs from Inverda and rent them at high prices to LLM companies or AI start-up teams in need of critical credit.
This is essentially an extremely extreme “high leverage plus heavy assets” financial game. Without stable cash flows, these companies rely on Wall Street debt financing to purchase GPU, which is worth tens of thousands of dollars. Their business model is extremely fragile: companies are in a state of collapse as long as the price of a credit lease falls or the old chain breaks.
Mask was extremely precise about the current dead end in the dialogue: the capital tap of the debt market has closed.
After a long period of fanaticism, Wall Street financial institutions finally began to look at risks. When they find that the revenues of many AI start-ups simply do not cover high computing bills, and New Clouds not only lacks collateral but is heavily dependent on high-interest debt, the risk avoidance instinct of capital is fully activated.
“SpaceX raised $40 billion in debt for the GPU, but we have investment-grade credit ratings, as well as satellite chains and launch revenues. If you don't have a real cash flow, you won't get a penny from the debt market.”
This is the cruelest reality: more than 200 new cloud companies that do not have “substantial income from the AI data centre” are about to face a breakdown in the financial chain. Mask described this impending collapse as “a massive extinction of peacockfish”. Following the retreat of capital, only one chicken hair left on the beach and a second-hand GPU that could not be converted.
2027 "Standing the Count": the ruins of the capital boom
During this dialogue, Chamas proposed a more deadly industrial pain point: the crisis of 2027, the “sliding calculus”.
Chamas, to be blunt, is witnessing a catastrophe caused by bureaucracy. Many of the technology giants (hyperscalers) are extremely slow in building data centres. “It takes three months for Yvesda to sign a confidentiality agreement (NDA) and even 18 months for the commissioning period.”
This slow supply chain and bureaucracy, coupled with physical constraints such as grid capacity and building clearance, has resulted in a large number of data centres not being completed on time. At the same time, for their own LLM operation, the major plants will hold all self-built calculations firmly in their hands, giving priority to internal needs and then to third parties.
This has led to an extraordinary waste of capital: many companies have bought the top GPUs at billions of dollars, built huge aircraft rooms, but by 2027, these calculations were turned into “half-laze projects” and piled up in warehouses and could not produce any gains.
These “grounded calculations” are not only a financial disaster for the technology giants, but also a blow to small and medium-sized AI companies that survive on leasing. The money has been spent and the equipment has become useless scrap.
Giant squeeze: Who's laughing and who's bleeding? In this cold winter, not all people are shaking in the sand.
As Chamas has pointed out, this mismatch can only benefit two types of firms: the first is a non-bureaucratic giant capable of rapidly establishing an industrial-scale size. For example, in Mask, XAI, based on a strong engineering culture, has been able to build a supercalculator cluster with 100,000 H100s at an extreme speed. The second category is that of extremely well-funded and fast players. Microsoft and Meta, for example, have large cash flows and can continue to hoard grass despite the volatility of short-term debt markets.
For the middle company New Clouds and the LLM start-ups, they face not only a lack of credit, but also an extremely demanding “digital exploitation” by giants.
According to the Chamas explosion, large plants are now not even allowed direct access by third-party model companies, which are required to take a model router (moder router) from the large plants and draw very high transit charges from it. He even suggested that in the previously exposed S-1 confidential document of Anthropic, reference was made to “the large plant informing them that there was no surplus of calculations”. This attests to a shortage of computing power and a monopoly of giants, which has reached an extremely dangerous level.
An empty cloud: "The Spring of Mask and Chamas"?
One of the most interesting details of this tweet is in the comment area.
When Mask and Chamas were pointing at the AC foam and expressing their pessimistic expectations, Jack Gee, founder of the empty agency Hedgeye, fired directly down there: “You two worked well together. It looks like a good way to spread the risk and hit the competition.”
Jack Gee's sarcasm is blood. During the SPCC (Special Purpose Acquisition Corporation) bubble in 2021, Chamas used to blow up and falsify, plaguing numerous scattered houses and leaving. Today, he and Mask are slamming AI infrastructure while singing publicly about the New Clouds company that borrowed the bonds.
Their statements have objectively served as a disincentive to competitors. Once market panic increases, interest rates on debt soar, competitors who rely on high leverage to buy cards end up dead, and giants like SpaceX and XAI, who have stable cash flows and diversified financing channels, take advantage of the opportunity to achieve a oligopolistic monopoly in the industry.
Concluding remarks: "The bar mitzvah" in the capital of AI
The al-calculation rush finally hit the financial common sense southern wall.
In the past two years, people thought they could print money if they bought GPU from Inverda. But this debt-driven carnival masks a huge loophole in the business model. AI Foundation is essentially a traditional business with long cycles, heavy assets and low tolerance for error. It requires grid cooperation, supply chain support and extremely large capital reserves.
In 2027, it will be an extremely volatile year for the AI computing industry. A large number of unqualified “new cloud” service providers will be cleansed, non-compliant AI start-ups will be depleted, and the lack of electricity and the lack of “sliding capacity” to support them will be the worst legacy of the bubble.
In the wave of the technological revolution, the bubbles that are blown up by the PPT and the borrowing will eventually be torn to pieces by gravity. Ultimately, it will always be the few with the strongest cash flow, the hardest nuclear engineering capability and the coolest commercial mind.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.