文 | 字母AI
黄仁勋还在努力扶住Neocloud这堵墙，马斯克却已经预言了它的倒塌。
AI泡沫的第一张多米诺骨牌，可能就藏在这批公司之中。
10月25日，马斯克在X上发出警告：最近几周，债务市场几乎已经不再向那些缺乏AI数据中心之外实质收入的企业提供资金。
他预言，200多家Neocloud将迎来“孔雀鱼式”（guppy-level）的死亡与整合，最终可能只有少数公司能够存活。
但就在三天前，英伟达CEO黄仁勋给出了几乎相反的判断。
10月22日，黄仁勋在接受CNBC采访时坚称，Neocloud的算力需求依然极其旺盛，真正限制它们发展的只是营运资金不足，而融资困难终将得到解决。
两人都看好AI算力需求，却对经营算力租赁业务的公司能否活下去，给出了截然不同的判断。
而越来越多的迹象表明，这门曾经被AI热潮捧上天的生意，正在出现裂缝。
黄仁勋扶，马斯克推
Neocloud，简单来说，就是专门出租AI算力的新型云计算公司。与亚马逊AWS、微软Azure等业务广泛的传统云厂商不同，这些公司主要围绕AI计算需求提供服务。它们采购大量GPU，建设数据中心，再把算力租给需要训练、运行AI模型的企业。
其中最有代表性的公司，就是CoreWeave、Nebius和Nscale。
从进入AI算力租赁市场开始，它们的生意就高度依赖英伟达的GPU。对于英伟达而言，Neocloud自然是非常重要的客户。
黄仁勋对它们的支持并不只存在于口头上，他拿出了真金白银，甚至亲自为Neocloud的算力收入托底。
2025年9月，英伟达与CoreWeave签订了一份初始价值63亿美元的算力采购协议。根据约定，如果CoreWeave的数据中心算力无法全部卖给其他客户，英伟达将按协议购买剩余算力，相关承诺持续至2032年。
今年1月，英伟达又进一步向CoreWeave投资20亿美元，支持其加快数据中心建设，双方计划到2030年建成超过5吉瓦的AI计算基础设施。
3月，英伟达又宣布向另一家Neocloud企业Nebius投资20亿美元，帮助其扩大AI云计算业务。
除此之外，英伟达还设计了一套更直接的扶持机制。在今年8月26日的财报电话会上，英伟达管理层披露，公司开始为部分Neocloud提供最低收入保障，承诺购买项目中的一部分算力，以帮助它们获得银行融资。
作为交换，当这些算力产生的收入超过保底水平时，英伟达会参与分成。
“按照这种模式，我们可以赚两次钱：一次来自硬件销售，另一次来自租赁收入分成。”英伟达首席财务官科莱特·克雷斯直接表示。
甚至有媒体评论干脆把英伟达比喻为“AI行业的中央银行”，形容它一边向Neocloud出售GPU，一边通过股权投资、算力采购承诺等方式为这些公司提供资金支持，帮助它们继续获得贷款、扩建数据中心，维持整个AI算力市场的运转。
黄仁勋如此积极地扶持Neocloud，当然有非常现实的商业利益。这些公司需要不断购买最新的GPU、扩大数据中心规模，才能满足客户需求。Neocloud持续扩张，意味着英伟达能够不断获得新的芯片订单。
只不过，英伟达愿意为Neocloud提供融资支持，不代表资本市场也愿意继续借钱给它们。
这也是马斯克发出警告的原因。
值得注意的是，马斯克自己也在大规模举债购买GPU。10月25日，他刚刚宣布，SpaceX用于购买英伟达GPU的400亿美元债务融资已经“关闭”，主要对应2027年第一、第二季度交付的GPU。
但与主要依靠算力租赁收入的Neocloud不同，SpaceX还有Starlink和火箭发射等成熟业务作为支撑。在马斯克看来，缺乏其他收入来源的Neocloud，很可能撑不过接下来的行业洗牌。
当有网友猜测，马斯克是在故意唱衰Neocloud，以便趁股价下跌低价收购时，马斯克的回答是：
“SpaceX不会购买任何Neocloud。这毫无意义。”
Neocloud的好日子到头了吗？
2023年以来，随着生成式AI爆发，英伟达GPU供不应求。微软、Google、亚马逊等传统云厂商无法及时满足市场需求，Neocloud抓住机会迅速崛起。
它们不需要建设完整的云服务体系，只要能买到GPU、部署好集群，就有机会拿到大额订单。
CoreWeave是其中的典型代表：2023年这家公司的全年营收只有2.29亿美元，到2025年就已经达到51.3亿美元，两年增长约22倍。
Nebius的AI云业务增长同样惊人，2023年营收仅960万美元，2025年已经达到4.8亿美元。
但营收暴涨，并不意味着能够长期赚钱。
尽管2025年CoreWeave营收达到51.31亿美元，但其净亏损为11.67亿美元，全年净利息支出更是高达12.29亿美元。
Neocloud需要提前投入巨额资金购买GPU、建设数据中心，再依靠未来的租赁收入收回成本。但随着GPU的更新换代，租赁价格也在下降。
麦肯锡2025年11月的一份研究显示，2021年至2025年，英伟达A100的公开租赁报价从每小时3.10美元降至1.50美元，跌幅超过50%。
麦肯锡还以B200 GPU为例测算，在完全使用股权资金、不考虑债务融资的情况下，GPU租赁业务扣除折旧等成本后的营业利润率仅约8%。如果设备利用率降至80%以下，或者租赁价格小幅下降，投资回报就可能被抹平。
与此同时，微软等Neocloud的大客户，也在不断建设自己的AI基础设施。谷歌、亚马逊等传统云巨头，同样在扩张AI算力业务。
Neocloud既要承担巨额资本开支，又要面对GPU贬值和大客户的竞争。一旦供需关系发生变化，过去的优势就可能迅速消失。
而最近，资本市场已经开始重新审视这门生意。
10月以来，Neocloud代表企业CoreWeave股债双双承压。截至10月26日，其股价较10月1日收盘价下跌近38%，收于55.19美元。10月中旬，其部分债券收益率一度升至15%至16%；到10月26日，CoreWeave部分2033年到期债券的价格已跌至历史新低。
10月22日，据外媒报道，做空CoreWeave、Nebius等五家AI基建公司的投资者，当月账面获利约60亿美元。其中，CoreWeave股价当月跌幅接近30%。
有意思的是，10月26日，CoreWeave股价还经历了一次反弹。当天，其股价盘前一度因马斯克的悲观言论下跌约5%，随后受到Anthropic公开IPO招股书的消息提振，最终收涨1.76%。
不过，这次反弹并没有改变CoreWeave近一个月来的整体跌势。
资本市场之外，数据中心建设也遇到了麻烦。
10月19日，微软退出Neocloud企业Nscale位于英国的一个数据中心项目。该项目原计划2027年初投入运营，部署超过2.3万块英伟达GB300 GPU，但电力供应可能直到2030年代才能落实。
对于依赖债务融资建设数据中心的Neocloud而言，项目延期意味着收入暂时无法兑现，前期投入和融资成本却还在不断累积。拖得越久，财务压力就越大。
研究机构Gartner预计，到2030年，超过80%的Neocloud将因成本压力和缺乏差异化而停止独立经营，包括被收购、合并或退出市场。
这个依靠GPU短缺迅速崛起的行业，如今已经开始为自身脆弱的商业模式付出代价。
敢问路在何方
Neocloud眼下的麻烦，可能还只是开始。
10月25日，硅谷投资人、Social Capital创始人查马斯·帕利哈皮提亚 （Chamath Palihapitiya）在回应马斯克时，给出了一个大胆而悲观的预测：2027年，AI行业可能出现大量闲置算力。
查马斯本人持有一批已经接入电力的数据中心用地资产，因此，他对AI基础设施的建设情况有着直接的观察。在他看来，Neocloud普遍资金不足、组织混乱，虽然部分公司表现较好，但整个行业的建设能力仍然令人担忧。
不过，传统云厂商也好不到哪里去——查马斯抱怨，这些大公司的官僚主义严重，仅签署一份保密协议就可能花上三个月，还会要求长达18个月的资金托管期。
更大的问题在于，已经规划的数据中心项目正在出现延误。
查马斯认为，这些问题将在2027年集中暴露。届时，大型云厂商可能优先满足自己的算力需求，而大量已经购买的GPU，却可能因为数据中心未能完工、需求与供给错配等原因，无法充分投入使用。
他因此预计，2027年将成为行业动荡的一年，大量错误的投资决策会被清算，一些企业则将在整合中消失。
况且，即使这些建设和交付问题最终得到解决，Neocloud的处境也未必会好转。
研究机构Gartner预计，AI优化云基础设施市场的支出将从2027年的660亿美元增长至2030年的1740亿美元。即便如此，市场增长仍可能不足以让大多数AI算力供应商在现有一代芯片被替换前收回投资。
随着微软、Google、亚马逊等科技巨头持续扩建自己的AI基础设施，GPU供应增加，Neocloud赖以崛起的市场环境也可能逐渐发生变化。
麦肯锡早在2025年11月就指出，Neocloud最初是为填补GPU供应缺口而诞生的。如果不能摆脱单纯出租算力的商业模式，建立长期竞争优势，它们可能重演早期云计算公司的命运，最终被大型云厂商收购、挤出市场，或者退守细分领域。
英伟达可以提供融资支持、保底采购，帮助这些企业渡过眼下的困难，却无法保证它们永远能够依靠GPU短缺赚钱。
黄仁勋可以扶一时，但不能扶一世。
尤其，在整个AI产业链中，Neocloud可能是最容易被绕过、抗风险能力也最弱的一环。
对于真正需要算力的科技巨头而言，租用Neocloud的GPU只是满足需求的一种方式。它们可以自己建设数据中心，可以收购现成的算力资产，也可以向大型云厂商租赁算力。
但是，甲骨文、微软、亚马逊乃至SpaceX等大型企业，尽管同样承担巨额资本开支，但至少还有其他成熟业务提供资金支持，能够承受更长时间的市场波动。
Neocloud却没有什么退路，只要算力租赁的收入增长稍慢一点，原本建立在高速增长预期上的商业模式，就可能难以为继。
同样的市场变化，对大型科技公司可能只是一阵动荡，却可能成为Neocloud的生存危机。
当GPU短缺逐渐缓解，客户有了更多选择，而Neocloud又无法证明自己不可替代，这批企业就可能成为AI产业扩张过程中的第一批牺牲品。
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。
Letter AI
Wong In-hoon is still trying to hold the wall of Neocloud, but Mask has predicted its collapse.
The first dominoes of AI foam may be hidden in this company.
On 25 October, Mask warned on X that in recent weeks the debt market had virtually stopped funding businesses that lacked real income outside the AI data centre.
He predicted that more than 200 Neoclouds would receive the death and integration of the “peacockfish” and that, ultimately, only a few companies might survive.
But just three days ago, the CEO Wong gave almost the opposite judgment.
On 22 October, Huang In-hoon, in an interview with the CNBC, asserted that the computing needs of Neocloud remained extremely high and that the only real constraints on their development were insufficient operational resources, and that financing difficulties would eventually be addressed.
Both of them looked to AI ' s computing needs, but gave a very different judgement as to whether the company operating the computing leasing business would survive.
And there are increasing signs that the door is cracking in a business that was once dominated by Ai.
Wong In-joon. Mask push.
Neocloud, in short, is a new type of cloud computing company that rents AI computing. Unlike traditional cloud manufacturers with a wide range of operations, such as the Amazon AWS and Microsoft Azure, these companies provide services mainly around AI computing needs. They procure large quantities of GPUs, build data centres and lease computing power to enterprises requiring training to run AI models.
The most representative of these companies are CoreWeave, Nebius and Nscale.
Their business has been highly dependent on GPUs in England since they entered the ACLC market. Neocloud is naturally a very important client for Inverda.
Wong In-hoon's support for them does not exist only in words, he has come up with real silver and silver, even personally to build up Neocloud's income.
In September 2025, Young Weida entered into an initial $6.3 billion credit purchase agreement with Core Weave. It was agreed that, if the data centre of CoreWeave could not be fully sold to other customers, the remaining capacity would be purchased by Weida as agreed and the commitment would continue until 2032.
In January of this year, Weida further invested $2 billion in CoreWeave to support its accelerated data centre, with the parties planning to complete an AI computing infrastructure of more than 5 Gwa by 2030.
In March, Weida announced a further $2 billion investment in another Neobius enterprise to help expand its AI cloud computing operations.
Added to this is the design of a more direct support mechanism. At a press conference on 26 August this year, the management of Ingweida disclosed that the company had begun to provide a minimum income guarantee for part of the Neocloud, committing to buy a portion of the project to help them access bank financing.
In exchange, when income from these calculations exceeds the floor level, Yvette will participate in the split.
“In this pattern, we can make two payments: one from the sale of hardware and the other from the share of rental income.” Colette Criss, Chief Finance Officer of Inverda, said so directly.
There are even media comments that simply refer to British Weeda as “the central bank of the industry” and describe it as selling GPUs to Neocloud, and providing financial support to these companies through equity investments, calculus purchase commitments, etc., to help them continue to access loans, expand data centres, and maintain the functioning of the entire ACMC.
Hoang In-hoon has been so active in supporting Neocloud and certainly has very real commercial interests. These companies need to continuously purchase up-to-date GPUs and expand data centres to meet client needs. The continued expansion of Neocloud means that New Chip Orders are continuously available to Inverda.
However, the willingness of Inverda to provide financing support for Neocloud does not mean that capital markets are willing to continue lending to them.
This is why Mask warned.
It is noteworthy that Mask himself is also making large-scale debt to purchase GPU. On 25 October, he had just announced that SpaceX's $40 billion in debt financing for the purchase of GPU in Weida had been “closed”, mainly for the delivery of GPU in the first and second quarters of 2027.
However, unlike Neocloud, which relies mainly on revenue from computing leases, SpaceX is supported by mature operations such as Starlink and rocket launch. In the view of Mask, Neocloud, lacking other sources of income, is likely to fail in the next trade to shuffle cards.
When netizens speculated that Mask had deliberately sung Neocloud in order to take advantage of the downfall in stock prices, he replied:
“SpaceX will not buy any Neocloud. It's pointless.”
Is neoclaud's good day over?
Since 2023, with the generation of AI, the supply of GPUs in Weida has fallen short of demand. Traditional cloud producers, such as Microsoft, Google and Amazon, were unable to meet market demands in a timely manner, and Neocloud seized the opportunity to emerge rapidly.
They do not need to build a complete cloud service system and have access to large orders as long as GPUs are bought and clusters are deployed.
CorreWeave is a typical representative of this: it received only $229 million a year in 2023, reaching $5.13 billion by 2025, an increase of about 22 times over two years.
The growth of the Nebius AI cloud operation was equally alarming, with only $9.6 million collected in 2023 and $480 million in 2025.
But the surge does not mean that money can be earned in the long run.
Although the CoreWeave battalion received $5,131 million in 2025, its net loss was $1,167 million, with net interest expenditure up to $1,229 million throughout the year.
Neocloud would need to invest large sums of money in advance to purchase GPUs and build data centres, before cost recovery would depend on future rental income. However, as the GPU was upgraded, the lease price was falling.
A study by McKinsey in November 2025 revealed that between 2021 and 2025, the open lease offer for A100 in Weidar declined from $3.10 to $1.50 per hour, a decline of over 50 per cent.
McKinsey has also taken B200 GPU as an example, where the turnover of GPU leasing operations after deduction of costs such as depreciation is only about 8 per cent, with full use of equity funds and without taking debt financing into account. If the utilization rate of equipment fell below 80 per cent, or if rental prices declined slightly, the return on investment could be wiped out.
Meanwhile, large clients of Neocloud, such as Microsoft, are constantly building their own AI infrastructure. Traditional cloud giants such as Google and Amazon are also expanding the AI computing business.
Neocloud had to cover both large capital expenditures and the devaluation of GPU and competition from large clients. Once supply-demand relations change, past advantages may quickly disappear.
Recently, the capital market has begun to revisit the business.
Since October, Neocloud has represented the company CoreWeave's double-collateral debt. As at 26 October, its stock price was nearly 38 per cent lower than the closing price on 1 October, receiving $55.19. In mid-October, the rate of return on some of its bonds rose to 15 to 16 per cent; by 26 October, the price of some of the CoreWeave bonds that matured in 2033 had fallen to record lows.
On 22 October, foreign media reported that five AI capital companies, such as CoreWeave and Nebius, had made approximately $6 billion in their books during the month. Of these, the CoreWeave stock price fell by almost 30 per cent that month.
Interestingly, on 26 October, CoreWeave also experienced a rebound. On the same day, its stock fair was once set down by about 5 per cent because of the masqueeic pessimism, which was subsequently boosted by Anthropic ' s announcement of an IPO book, culminating in an increase of 1.76 per cent.
However, this rebound did not change the overall decline of CoreWeave over the past month.
Outside capital markets, data centre construction has also experienced difficulties.
On 19 October, Microsoft withdrew from the Neocloud Enterprise Nscale data centre project in the United Kingdom. The project, which was to be operational by early 2027, deployed more than 23,000 British Weeda GB300 GPUs, but the power supply could not be implemented until the 1930s.
For Neocloud, which relies on debt financing for the construction of the data centre, the extension of the project means that income cannot be realized for the time being, while the cost of prior-period inputs and financing continues to accumulate. The longer it takes, the greater the financial pressure.
The research institute Gartner expects that by 2030, more than 80 per cent of Neocloud will have ceased to operate independently, including through acquisitions, mergers or exits from the market, owing to cost pressures and lack of differentiation.
This rapidly rising industry, which relies on GPU shortages, has now begun to pay for its fragile business model.
How dare you ask where the road is.
Neocloud's present trouble is probably just the beginning.
On October 25th, Chamath Palihapitiya, a Silicon Valley investor and founder of Social Capital, gave a bold and pessimistic prediction in response to Mask: In 2027, the AI industry may have significant idle capacity.
Chamas himself holds a number of data centre land assets that are already connected to electricity, so he has direct observation of the construction of the AI infrastructure. In his view, Neocloud was generally underfunded and organized, and while some companies performed well, capacity-building across the industry remained a concern.
But traditional cloud manufacturers are not much better. — Chamas complains that these large companies are highly bureaucratic, and that it may take three months to sign a confidentiality agreement alone and require a period of up to 18 months to hold funds.
The larger problem is that delays are occurring in the planned data centre project.
According to Chamas, these problems will be concentrated in 2027. At that time, large cloud manufacturers may give priority to meeting their own computing needs, while a large number of purchased GPUs may not be fully operational because of the failure to complete the data centre and the mismatch between demand and supply.
He therefore anticipated that 2027 would be a year of industrial turmoil, with a large number of erroneous investment decisions being liquidated and some enterprises disappearing in integration.
Moreover, even if these construction and delivery problems were finally resolved, the situation of Neocloud would not necessarily improve.
The research institute Gartner predicts that AI spending on optimizing the cloud infrastructure market will increase from $66 billion in 2027 to $174 billion in 2030. Even so, market growth may not be sufficient to allow most AI computing suppliers to recover their investments before the current generation of chips is replaced.
As technological giants such as Microsoft, Google and Amazon continue to expand their AI infrastructure, the supply of GPUs has increased, and the market environment in which Neocloud is built may gradually change.
McKinsey noted as early as November 2025 that Neocloud had originally been born to fill the GPU supply gap. If they do not move away from a purely rental-calculative business model and create long-term competitive advantages, they may repeat the fate of early cloud computing companies, eventually bought by large cloud manufacturers, squeezed out of the market, or retreated from the subdivisions.
Young Weidar can provide financing support, secure purchases, and help these enterprises to overcome their current difficulties, but cannot guarantee that they will always be able to make money from the GPU shortfall.
Wong In-hoon can help for a while, but he cannot help the First.
In particular, within the AI industry chain as a whole, Neocloud is probably the most vulnerable to circumvention and the weakest risk resistance.
For a technology giant that really needs credit, renting Neocloud's GPU is only one way of meeting demand. They can build their own data centres, acquire off-the-shelf computing assets or lease power from large cloud manufacturers.
However, large businesses such as Oracle, Microsoft, Amazon and even SpaceX, while also bearing large capital expenditures, have at least other mature operations that provide financial support that can withstand longer market fluctuations.
Neocloud has little to lose, and as long as the growth of rent-based income is slower and built on a business model based on high growth expectations, it may be difficult to sustain it.
The same market changes, which may be a wave of turbulence for large technology companies, could become a existential crisis for Neocloud.
When the GPU shortage gradually subsided and clients became more selective, and Neocloud could not prove itself irreplaceable, the enterprises could be the first victims of the expansion of AI.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.