点击蓝字关注雨生
副标题： 别只盯着AWS和Azure了，看看Cursor背后的算力推手是谁
我是雨生。
最近华尔街那帮人又开始犯“躁郁症”了。一边看着甲骨文（Oracle）因为产能跟不上需求导致股价震荡，一边又疯狂鼓吹“AI泡沫破裂论”。
说实话，这帮分析师看Excel表格的时间太长，看真实世界的时间太短。
今天我们要聊的这家公司——Nebius，简直就是这波AI浪潮里的一朵奇葩（褒义）。
如果你做云计算，或者正在搞出海业务，这家公司你必须关注。
为什么？因为它刚从“前任”Yandex的废墟里爬出来，拍了拍身上的灰，转身就成了欧洲最硬核的AI基建之王。
这不仅仅是一个商业故事，这是给所有出海人的一堂生动的“反脆弱”课。
雨生视角：别被“泡沫论”忽悠瘸了
文章里提到一个很有意思的数据，关于“AI泡沫”的搜索量飙升了880%。
这让我想起那句老话：
当路边的擦鞋童都在谈论股票时，你就该离场了。但现在的逻辑反过来了：
当所有人都在喊泡沫的时候，真正的猎人正在进场。
Nebius是谁？简单说，它是前Yandex创始人Arkady Volozh带着核心技术团队和资产，在海外重组的“复仇者联盟”。
看看它的客户名单：
Microsoft给了它190亿美元的意向订单，Meta给了30亿。
但雨生觉得，这都不是重点。重点是Cursor、Mistral AI、Black Forest Labs都在用它。
各位，Cursor是什么？那是现在程序员手里的神兵利器，ARR（年度经常性收入）一年内从1亿干到10亿美金的怪物。
这些最懂技术的“AI原生”公司选择Nebius，不是因为它的销售请客吃饭厉害，而是因为它的**技术真香，价格真公道。
相比之下，那些只知道倒卖英伟达GPU、收个过路费的“云厂商”，才是真正的泡沫
（只有价格没有价值）
深度解读：为什么Nebius能成？
这篇深度分析文章（原文附后）扒开了Nebius的底裤，露出了它的核心竞争力：
1. 它不是二道贩子，它是全栈玩家：
Nebius不仅仅是买几万张H100插在机架上。他们自己设计服务器，自己搞数据中心架构，还有那个这一行里非常性感的“Token Factory”（代币工厂）。这玩意儿允许开发者按token付费，而不是按闲置的GPU时间付费。
商业逻辑：
这就像你打车，Uber是按里程收费（Nebius），而传统的租车公司是按天收费不管你开不开（AWS/Azure的部分模式）。
对于初创公司，哪个更省钱？显而易见。
2. 极高的转换成本（护城河）：
文章提到一个点很扎心：“AI工作负载具有极高的粘性。” 一旦你在Nebius的环境里跑通了模型，想搬家？那是脱层皮的事。
这就是为什么Nebius敢预测2027年搞到100-120亿美金的收入。
3. 毛利吓死人：
除去折旧，毛利率高达71%。
这哪是苦哈哈的IaaS（基础设施即服务），这简直就是印钞机。
出海实战指南：这对我们有什么用？
作为出海人，我们能从Nebius身上学到什么，或者怎么用它？
1. 算力成本优化（Cost Optimization）：
如果你正在做AI应用出海，不管是生图还是大模型套壳。别只盯着AWS、Google Cloud和Azure。
看看Nebius或者CoreWeave这种“Neo Cloud”（新云）。
Nebius的H100租赁价格（约$2.00/时）比CoreWeave（约$4.25/时）还要低。在烧钱的阶段，这省下来的都是纯利。
2. 技术选型要看“原生”：
跟着Cursor这种顶级技术公司选型大概率不会错。如果你的架构需要极致的推理性能，去试用一下Nebius的Token Factory。
3. 布局欧洲合规：
Nebius的数据中心主要在芬兰、法国、冰岛。对于那些担心GDPR合规的出海企业，使用欧洲本土的强力基建是一个很好的合规背书。
只有雨生敢说的大实话
当然，Nebius也不是没风险。它烧钱的速度和它扩张的速度一样快。
而且，它毕竟有“前世”的影子，虽然现在是荷兰公司，但在地缘政治复杂的当下，供应链和资本运作依然需要小心。
但无论如何，Nebius证明了一件事：
在AI时代，哪怕你是从废墟里走出来的，只要铲子造得好，就不怕没人来挖金矿。
---
互动话题：
你看好这种专注AI的“新云”厂商挑战AWS/Azure的霸权吗？还是觉得它们最终会被巨头收购？欢迎在评论区留言，雨生在评论区等你来杠！
**(文章结束)
*点击文末“阅读原文”或扫描下方二维码，加入雨生云计算&知识星球，获取更多出海避坑内幕。
---
### 3. 金句海报文案
海报1：
> “当所有人都在谈论AI泡沫时，
> 真正的猎人正在查看基础设施的毛利率。”
> —— 雨生云计算
海报2：
> “Cursor、Mistral都在用的云，
> 才是AI时代的真·风向标。
> 别让AWS的账单限制了你的想象力。”
> —— 雨生云计算 #出海必读
海报3：
> “Nebius不仅是从Yandex废墟中飞出的凤凰，
> 更是给高昂算力成本的一记响亮耳光。”
> —— 雨生云计算 #AI基建
### 5. 朋友圈文案模板
模板一（硬核风）：
Nebius这波分析太透了！雨生大佬一针见血：别整天担心AI泡沫，先看看Cursor这种独角兽都在用谁家的算力。H100两刀一小时，这价格还要啥自行车？出海做AI的朋友建议详读。👇
[文章链接]
模板二（焦虑制造风）：
你的AI应用还在给AWS交智商税吗？看看欧洲的Nebius是怎么靠技术和低价抢下微软百亿订单的。出海不降本，迟早两行泪。这篇深度解读值得收藏。强烈推荐关注雨生，干货真的多！
[文章链接]
模板三（吃瓜风）：
从Yandex分家出来的Nebius，现在要在欧美云市场大杀特杀了？这剧情比美剧还精彩。雨生这篇解读很有趣，不仅有技术分析，还有商业逻辑。这才是KOL该有的水平。进星球学习了！
[文章链接]
---
### 6. 新闻原文中英文对照版
(注意：以下内容为原文对照，未做删减，无星号)
Nebius: It's The Differentiation!
Dec. 13, 2025 8:56 AM ETNebius Group N.V. (NBIS) Stock
Bert Hochfeld
23.79K Followers
Nebius：这就是差异化竞争！
2025年12月13日 上午8:56 ET Nebius Group N.V. (NBIS) 股票
Bert Hochfeld
2.379万 关注者
Summary
I believe that Nebius (NBIS) is the technology leader in AI neo cloud, with a differentiated full-stack approach and strong momentum among AI-native customers.
NBIS's rapid ARR growth, marquee hyperscaler contracts, and expanding capacity underpin expectations for $10–12B revenue in 2027 and $20B by 2028.
Gross margins (ex-depreciation) are 71% and rising; high switching costs and sticky AI workloads support durable, recurring revenue.
I reaffirm a Buy rating on NBIS, seeing its valuation as modest given its visibility, execution, and leadership in the high-growth AI infrastructure space.
摘要
我相信 Nebius (NBIS) 是 AI 新云领域的技术领导者，拥有差异化的全栈方法，并在 AI 原生客户中保持强劲势头。
NBIS 的 ARR 快速增长、与顶级超大规模云服务商的合同以及不断扩大的产能，支撑了 2027 年 100-120 亿美元收入和 2028 年 200 亿美元收入的预期。
毛利率（不含折旧）为 71% 且在上升；高转换成本和具有粘性的 AI 工作负载支持持久的经常性收入。
鉴于其在高增长 AI 基础设施领域的可见性、执行力和领导地位，我重申对 NBIS 的买入评级，并认为其估值适中。
Nebius: No real signs of a bubble here
Nebius (NBIS) is one of the leaders in the AI neo cloud space. While not the revenue leader, I believe it is the technology leader, and has the overall strategy most likely to succeed in this burgeoning space. But does it matter? I find it essential to discuss bubble theories in some kind of rational fashion before trying to evaluate the position of Nebius.
Nebius：这里没有真正的泡沫迹象
Nebius (NBIS) 是 AI 新云领域的领导者之一。虽然不是收入的领导者，但我认为它是技术的领导者，并且拥有在这个新兴领域最有可能成功的总体战略。但这重要吗？我认为在试图评估 Nebius 的地位之前，有必要以某种理性的方式讨论泡沫理论。
In completing this article in the wake of Oracle’s (ORCL) revenue miss and share price implosion, I think it’s reasonable to comment that the read through ought not to be that Oracle missed revenue expectations by some de minimis amount, or that its RPO balance only grew by 15% sequentially, but that it raised its revenue guidance for its FY ’27 and laid out how it is engaging with customers who want to buy more of its AI cloud more rapidly than the company can actually provide it profitably. Some of the more sensational articles on SA seem to be based on the premise that orders, and the capex to fulfill them, are negatives and that cloud infrastructure will never be a profitable business. I believe that is exceptionally far from reality, but any readthrough of that unlikely premise to Nebius-and that is certainly possible, is an opportunity to acquire the shares at an attractive valuation. Today, apparently the issue is that of Broadcom's (AVGO) projected margins.
在甲骨文 (ORCL) 收入未达预期且股价暴跌之后完成这篇文章时，我认为合理的评论不应该是甲骨文的收入略微低于预期，或者其 RPO 余额环比仅增长了 15%，而应该是它提高了 2027 财年的收入指引，并阐述了它如何与那些希望购买更多 AI 云服务的客户接触，而这些客户的需求增长速度超过了公司在盈利前提下实际所能提供的速度。Seeking Alpha 上一些更耸人听闻的文章似乎基于这样一个前提：订单以及履行订单的资本支出是负面因素，云基础设施永远不会成为一项盈利的业务。我认为这极其脱离现实，但如果将这种不太可能的前提解读到 Nebius 身上——这当然是有可能的——那将是一个以极具吸引力的估值买入股票的机会。今天，问题似乎出在博通 (AVGO) 的预期利润率上。
I mean no disrespect to authors, observers and of course investors who believe that AI is, at the least, an investment bubble that is fated soon to pop. (From the lyrics of You’re the Top by Cole Porter). The bubble chorus has reached essentially deafening heights. A recent statistic was published on Monday, December 8th and before discussing Nebius, I think it is necessary to dispassionately look at some actual data about the AI revolution. The passage below comes from an article in Axios, a website owned by Cox Enterprises which often publishes on IT topics:
The big picture: Bubble talk is everywhere. Mentions of "AI bubble" rose 880% since last quarter's investor calls, according to AlphaSense.
What they're saying: Speakers at recent Axios events were buzzing about the risks of an overheated market and buzzing about the buzz itself.
我无意冒犯那些认为 AI 至少是一个注定很快会破裂的投资泡沫的作者、观察家，当然还有投资者。（出自 Cole Porter 的《You’re the Top》歌词）。泡沫的合唱已经达到了震耳欲聋的高度。最近的一项统计数据于 12 月 8 日星期一发布，在讨论 Nebius 之前，我认为有必要冷静地看一下有关 AI 革命的一些实际数据。下面的段落来自 Cox Enterprises 旗下的网站 Axios 的一篇文章，该网站经常发布有关 IT 主题的文章：
大局：泡沫言论无处不在。根据 AlphaSense 的数据，自上个季度的投资者电话会议以来，“AI 泡沫”的提及率上升了 880%。
他们在说什么：在最近 Axios 活动上的演讲者们都在热议市场过热的风险，并对这种热议本身议论纷纷。
While the article itself, to which I have linked here tells readers to beware of a possible AI investment bubble, it is this fixation on an AI bubble that can lead to some kind of fear of investing in the AI space. And that, I contend, would be a significant mistake.
虽然我链接的这篇文章本身告诉读者要提防可能的 AI 投资泡沫，但正是这种对 AI 泡沫的执念会导致对投资 AI 领域的某种恐惧。我认为，那将是一个重大的错误。
This is actually an article reaffirming my buy recommendations regarding the shares of Nebius from earlier this year. But any buy recommendation in the AI space has to go through what I can only label as the bubblemania barrier, in which investment merits are ignored in favor of some poorly drawn correlations to what are purported to be historical analogs. Is there a lot of enthusiasm and hype surrounding AI? Will there be AI casualties as investments? Yes, again. But there will be plenty of AI winners, and this article is about one such company, Nebius.
这实际上是一篇重申我今年早些时候关于买入 Nebius 股票建议的文章。但是在 AI 领域的任何买入建议都必须穿过我只能称之为“泡沫狂热障碍”的东西，在这种障碍中，投资价值被忽视，取而代之的是一些与所谓的历史类比之间拙劣的关联。围绕 AI 是否有很多热情和炒作？作为投资，会有 AI 受害者吗？是的，再说一次。但会有很多 AI 赢家，而本文就是关于这样一家公司，Nebius。
I imagine all readers and this writer will agree that bubbles are bad. The real estate bubble of the early years of this century almost brought down the global financial system. The dot.com bubble at the turn of the century destroyed lots of wealth.
我想所有读者和作者都会同意泡沫是坏事。本世纪初的房地产泡沫几乎摧毁了全球金融体系。世纪之交的互联网泡沫摧毁了大量财富。
That said, the other side of the argument is about productivity. AI is and will continue to have the strongest impact on productivity seen in many decades. It is in the process of becoming deployed broadly, and most investors are well aware that this deployment is dependent on infrastructure. I will go a bit further, and talk about specialized AI infrastructure as the requirement. That is just not a bubble but a response to an economic imperative. Almost everything stems from the fact that end-users believe they are or will achieve massive ROI from their AI investments.
话虽如此，争论的另一面是关于生产力。AI 正在并将继续对生产力产生几十年来最强烈的影响。它正在广泛部署的过程中，大多数投资者都清楚这种部署依赖于基础设施。我想更进一步，谈谈作为需求的专用 AI 基础设施。这不仅仅是一个泡沫，而是对经济当务之急的回应。几乎一切都源于这样一个事实：最终用户相信他们正在或将从他们的 AI 投资中获得巨大的投资回报率 (ROI)。
A well-known industry publication, The Information, recently carried a story about the ROI problem for Microsoft’s advanced AI. I have linked to that story here. I call out this story in particular because the headline was about Microsoft reducing quotas due to user pushback. What seems to be off is that Microsoft apparently told ”The Information” that its story about sales quotas was specifically inaccurate but the publication ran the story and reiterated the claim anyway. There are many similar claims about AI that have been disseminated.
著名的行业出版物《The Information》最近刊登了一篇关于微软高级 AI 的 ROI 问题的报道。我在此链接了该报道。我特别提到这篇报道，是因为标题是关于微软因用户抵制而减少配额。似乎不对劲的是，微软显然告诉《The Information》，其关于销售配额的报道特别不准确，但该出版物还是发布了报道并重申了这一说法。关于 AI 还有许多类似的说法在传播。
Of course AI is not a mature technology; if it were, its growth would have reached an apogee. And like any immature technology, it will suffer some speed bumps both on a customer level and in certain functional areas. The specific story in “The Information” related to Azure AI agent-building solution and their costs and inability to achieve desired productivity gains. While agent-building is an important function for continued AI growth, it is not something seminal nor does it mean that AI doesn’t work or that users have a hard time in achieving strong levels of ROI. I could easily call out numerous instances in which agent-building applications are working well and delivering the results that have been promised. Indeed, the link above points to 10,000 Microsoft customers and a fair number of successful references.
当然，AI 不是一项成熟的技术；如果是的话，它的增长早就达到顶峰了。像任何不成熟的技术一样，它在客户层面和某些功能领域都会遇到一些减速带。《The Information》中的具体报道涉及 Azure AI 代理构建解决方案及其成本和无法实现预期的生产力收益。虽然代理构建是 AI 持续增长的一个重要功能，但它并不是什么开创性的东西，也不意味着 AI 不起作用或者用户难以获得高水平的 ROI。我可以很容易地举出许多实例，说明代理构建应用程序运行良好并交付了所承诺的结果。事实上，上面的链接指向了 10,000 名微软客户和相当数量的成功案例。
It happens that I have some background in what is called cliometrics. That is really a quantitative approach to economic history. And with that background, and with all respect to many writers on both SA and elsewhere, the fact that there has been a dramatic and massive investment in AI is no bubble but a rational response to the ROI that has been and is likely to continue to result from substantial productivity gains.
碰巧我有一些计量历史学 (cliometrics) 的背景。这实际上是一种经济史的定量方法。有了这个背景，恕我冒昧地对 Seeking Alpha 和其他地方的许多作者说，对 AI 进行的戏剧性和大规模投资并不是泡沫，而是对已经产生并且可能继续由巨大的生产力收益产生的 ROI 的理性反应。
There are almost always segments of the economy that are rising and others that are falling on a relative basis. At this point, with the expected ROI of so many AI applications at very high levels, anything other than very rapid growth of AI investment would really be an outlier in terms of economic history. OpenAI recently published a report detailing where enterprise AI is gaining momentum and why that is happening. I have linked above to that report.
经济中几乎总是有一些部门在上升，而另一些部门在相对下降。在这一点上，随着如此多 AI 应用程序的预期 ROI 处于非常高的水平，除了 AI 投资的非常快速增长之外，任何其他情况在经济史上都将是一个异常值。OpenAI 最近发布了一份报告，详细说明了企业 AI 在哪里获得动力以及为什么会发生这种情况。我在上面链接了该报告。
Eventually, like almost all other economic segments, there will be overinvestment. In this country, we wound up with 5 transcontinental railways when 3 would have been more than sufficient for ultimate demand. Back in the 19th century, no one knew just how much demand there would be for rail freight capacity because never before had it been possible to move freight so quickly and at such a low price. What happened then, and is happening now with AI, is the ability to do something new, at a dramatically different cost is creating its own demand.
最终，像几乎所有其他经济部门一样，将会出现过度投资。在这个国家，我们最终建成了 5 条横贯大陆的铁路，而 3 条就足以满足最终需求。早在 19 世纪，没有人知道铁路货运能力会有多大的需求，因为以前从未可能以如此快的速度和如此低的价格运输货物。当时发生的事情，以及现在 AI 正在发生的事情是，以截然不同的成本做新事情的能力正在创造其自身的需求。
This is no new phenomenon, and no one really knows the limits of AI demand, because no one really has specific and dispositive ideas on the limits of what the technology can do. It is my belief that when a company such as Nebius creates infrastructure that is cost effective and easy to use, it is creating its own demand, and while it is the case that most of the headlines in the debate about AI bubbles are about giant contracts for capacity that have been awarded by the hyperscalers, in the case of Nebius, much of their growth is and will be coming from AI natives who are inventing new solutions that can only be undertaken through the use of AI. I will provide a few examples later in this article.
这不是什么新现象，也没有人真正知道 AI 需求的极限，因为没有人真正对这项技术能做什么的极限有具体和决定性的想法。我相信，当像 Nebius 这样的公司创建具有成本效益且易于使用的基础设施时，它就在创造自己的需求，虽然关于 AI 泡沫辩论的大多数头条新闻都是关于超大规模云服务商授予的巨大容量合同，但在 Nebius 的案例中，他们的大部分增长正在并将来自 AI 原生企业，这些企业正在发明只能通过使用 AI 才能实现的新解决方案。我将在本文稍后提供几个例子。
I recently wrote about Oracle and how its future was a function of its OCI, the specialized infrastructure that facilitates the deployment of AI applications at an attractive price and with differentiated performance. And, I believe, that there will be multiple winners and many losers in the race to get AI infrastructure into the hands of users. Essentially, the investment case for Nebius is similar to that of Oracle-similar but not the same as Nebius is not offering a hybrid cloud and has no legacy customers for whom to cater. Nebius has a better mousetrap, and that is already showing up in terms of orders from the most successful and sophisticated AI users.
我最近写了一篇关于甲骨文的文章，谈到它的未来取决于其 OCI，这是一种专用的基础设施，能够以有吸引力的价格和差异化的性能促进 AI 应用程序的部署。而且，我相信，在将 AI 基础设施交付给用户的竞赛中，将会有多个赢家和许多输家。本质上，Nebius 的投资案例与甲骨文相似——相似但不相同，因为 Nebius 不提供混合云，也没有需要照顾的传统客户。Nebius 拥有更好的捕鼠器（更好的产品），这已经体现在来自最成功和最成熟的 AI 用户的订单上。
While Nebius shares have doubled over the past 6 months, they have also fallen 11% over the past month and by about 13% over the past week. They are down by about 33% since there recent peak despite an enormous increase in their growth forecast. To me, Nebius is the anti-bubble whose shares simply do not reflect their current and future opportunity.
虽然 Nebius 的股价在过去 6 个月翻了一番，但在过去一个月也下跌了 11%，过去一周下跌了约 13%。尽管其增长预测大幅增加，但自近期高点以来已下跌约 33%。对我来说，Nebius 是反泡沫的代表，其股价根本没有反映出其当前和未来的机会。
What does Nebius Offer?
Nebius is one of several neo-clouds that burst upon the investment scene about 1 year ago. The company was part of what was left after the deconstruction of Yandex. Its founder, Arkady Volozh, left Russia, sold the component of Yandex that was located in Russia, and was left with Nebius, some cash and other smaller businesses of which the most significant are Avride, an autonomous driving vendor and ClickHouse which is a leading vector data base.
Nebius 提供什么？
Nebius 是大约一年前闯入投资界视野的几家新云公司之一。该公司是 Yandex 解体后剩下的部分。其创始人 Arkady Volozh 离开了俄罗斯，出售了位于俄罗斯的 Yandex 部分，留下了 Nebius、一些现金和其他较小的业务，其中最重要的是自动驾驶供应商 Avride 和领先的向量数据库 ClickHouse。
At the moment, the company operates 7 data centers in Finland, France, Iceland, the UK, 2 in the US and one in Israel. It is spending $5 billion this year to increase capacity severalfold, and it expects that it will increase its connected capacity by about 4 fold (from 220MW to 900+MW) over the coming year. The company now expects to reach a run rate of between $7-$9 billion of ARR by the end of next year, probably leading to actual revenues of $10-$12 billion by 2027.
目前，该公司在芬兰、法国、冰岛、英国、美国（2 个）和以色列（1 个）运营着 7 个数据中心。该公司今年将花费 50 亿美元将产能提高数倍，并预计在来年将其连接容量增加约 4 倍（从 220MW 增加到 900MW 以上）。该公司目前预计到明年年底 ARR 运行率将达到 70-90 亿美元之间，这可能导致 2027 年的实际收入达到 100-120 亿美元。
As many readers know, part of this stunning increase in ARR and revenue is a function of an order Nebius received from Microsoft (MSFT) in September for as much as $19 billion for the rental of its AI infrastructure. More recently, Nebius received an order from Meta (META) for $3 billion; this order was limited to $3 billion because of capacity constraints.
正如许多读者所知，ARR 和收入这种惊人的增长部分源于 Nebius 在 9 月份收到的来自微软 (MSFT) 的一份订单，该订单价值高达 190 亿美元，用于租赁其 AI 基础设施。最近，Nebius 收到了一份来自 Meta (META) 的 30 亿美元订单；由于产能限制，该订单被限制在 30 亿美元。
I think a brief review of Nebius products will be a foundation for understanding the differentiation that Nebius offers. The company’s core product currently is the recently released Nebius AI Cloud 3.0 “Aether.” (Aether comes from Greek mythology and is usually referred to as ether in English but regardless it means an invisible, weightless medium that permeates all of space.) This new release adds several security features along with advanced governance and control capabilities. One of the things that makes an AI Cloud, AI, so to speak, is performance. This new release has set performance records for inference processing. It also offers substantial storage performance advantages. The company’s service provides users guarantees for reliability and, flexibility and continued operations. In the world in which Nebius competes, enhancing developer productivity is a key area of evaluation. The new release offers a new UI and has better support for other components that are part of the application creation process.
我认为简要回顾 Nebius 的产品将是理解 Nebius 提供的差异化的基础。该公司目前的核心产品是最近发布的 Nebius AI Cloud 3.0 “Aether”。（Aether 来自希腊神话，在英语中通常被称为以太，但无论如何，它意味着一种渗透整个空间的无形、无重量的介质。）这个新版本增加了一些安全功能以及先进的治理和控制能力。可以说，让 AI 云成为 AI 的要素之一就是性能。这个新版本创造了推理处理的性能记录。它还提供了巨大的存储性能优势。该公司的服务为用户提供了可靠性、灵活性和持续运营的保证。在 Nebius 竞争的世界中，提高开发人员的生产力是一个关键的评估领域。新版本提供了一个新的用户界面，并对作为应用程序创建过程一部分的其他组件提供了更好的支持。
The company’s service offering is based on various incarnations of Nvidia technology. That said, Nebius is a full stack vendor. That means that the service is based on a complete range of proprietary hardware that includes servers, racks, and data center designs in addition to software and tools that are specifically designed for AI development. I don’t want to suggest that there aren’t other vendors who make similar assertions and I also want to confess that I am investment analyst and not someone who evaluates competitive technologies.
该公司的服务产品基于各种版本的英伟达技术。也就是说，Nebius 是一家全栈供应商。这意味着该服务基于全套专有硬件，包括服务器、机架和数据中心设计，以及专为 AI 开发设计的软件和工具。我不想暗示没有其他供应商做出类似的断言，我也想承认我是一名投资分析师，而不是评估竞争技术的人。
Companies like Nebius offer specific services tailored for the requirements of the users to which they want to sell. For example, a company such as Oracle optimizes OCI for its target user base which includes legacy ERP, CRM and finapp customers. I have linked here to one comparison between Oracle Cloud and Nebius which may highlight the functional differences for readers. When it comes to AI cloud infrastructure, I believe that there is room for several approaches.
像 Nebius 这样的公司提供针对其想要销售的用户需求量身定制的特定服务。例如，像甲骨文这样的公司为其目标用户群（包括传统 ERP、CRM 和财务应用程序客户）优化 OCI。我在此链接了甲骨文云和 Nebius 之间的一项比较，这可能会向读者强调功能上的差异。谈到 AI 云基础设施，我相信有多种方法的空间。
Last month, Nebius launched what it has described as the Token Factory. This is a unified platform that facilitates tuning, large-scale inference processing and model optimization. It is another type of service appealing to developers and digital enterprises to deploy their LLMs at scale. The service is based on pay-per-token pricing which is a major advantage for many users who get hit with hidden costs for things like infrastructure or idle GPU charges. The platform can leverage open source AI models from many vendors. I have linked here to one analysis that best describes how this new service enlarges the target market for Nebius and allows for it to compete more effectively against hyper scalers.
上个月，Nebius 推出了它所描述的 Token Factory（代币工厂）。这是一个统一的平台，便于调整、大规模推理处理和模型优化。这是另一种吸引开发人员和数字企业大规模部署其 LLM 的服务。该服务基于按代币付费的定价，这对于许多因基础设施或闲置 GPU 费用等隐性成本而受到打击的用户来说是一个主要优势。该平台可以利用来自许多供应商的开源 AI 模型。我在此链接了一项分析，该分析最好地描述了这项新服务如何扩大 Nebius 的目标市场，并使其能够更有效地与超大规模云服务商竞争。
The Nebius Growth Story: Its products enable it to capture the fastest growing AI Natives as customer
Planning for beyond the current market environment is, I believe, a key factor in investment differentiation between Nebius and its rivals. It was built by individuals with backgrounds in the technology space rather than individuals whose background has been in creating datacenters used in bitcoin mining or in hedging natural gas futures. It has enabled Nebius to choose a different approach focused beyond that of creating capacity in a market that currently is imbalanced between available supply and demand.
Nebius 的增长故事：其产品使其能够捕获增长最快的 AI 原生企业作为客户
我认为，针对当前市场环境之外的规划是 Nebius 与其竞争对手之间投资差异化的关键因素。它是由具有技术领域背景的个人建立的，而不是那些背景是创建用于比特币挖矿的数据中心或对冲天然气期货的个人。这使得 Nebius 能够选择一种不同的方法，专注于在当前供需失衡的市场中创造容量之外的事情。
That doesn’t mean that Nebius doesn’t compete for and win some capacity deals from hyper-scaler customers. It has two announced and as its capacity ramps next year, I expect it will be able to sell additional deals to other hyperscalers. Because of the high level of performance and reliability shown above, it can offer a very attractive TCO for the largest users and doing so provides the company with a base of demand that keeps the utilization of its data centers at optimum levels to maximize profitability.
这并不意味着 Nebius 不竞争并赢得超大规模云服务商客户的一些容量交易。它已经宣布了两项交易，随着明年产能的增加，我预计它将能够向其他超大规模云服务商出售更多交易。由于上面显示的高水平性能和可靠性，它可以为最大的用户提供非常有吸引力的 TCO（总拥有成本），这样做为公司提供了需求基础，使其数据中心的利用率保持在最佳水平以最大化盈利能力。
But in terms of looking for growth beyond the invariably limited period in which demand exceeds supply, an infrastructure vendor needs a strategy in which the growth of its customers is driving the usage of its platform. And, I believe, that is what Nebius has most recently accomplished.
但是，就寻求在需求超过供应的必然有限时期之外的增长而言，基础设施供应商需要一种策略，即其客户的增长推动其平台的使用。而且，我相信，这就是 Nebius 最近所取得的成就。
As busy as we are with these mega deals, our main focus is still to build our own core AI cloud business. We made great progress here with AI native start-ups like Cursor, Black Forest Labs and others. The economics and the cash flow of mega deals are attractive in their own right, but they also enable us to build our core AI cloud business faster. This is our real future opportunity.
尽管我们忙于这些大型交易，但我们的主要重点仍然是建立我们自己的核心 AI 云业务。我们在 AI 原生初创公司（如 Cursor、Black Forest Labs 等）方面取得了巨大进展。大型交易的经济效益和现金流本身就很有吸引力，但它们也使我们能够更快地建立我们的核心 AI 云业务。这是我们真正的未来机会。
To illustrate the point, I am presenting a list of some of the company’s users-not Meta (META) or Microsoft, but users that have adopted the platform because of its performance, cost, reliability and the tools that make it more productive and user friendly for AI natives:
为了说明这一点，我提供了一份公司部分用户的名单——不是 Meta (META) 或微软，而是因为其性能、成本、可靠性以及使其对 AI 原生企业更具生产力和用户友好的工具而采用该平台的用户：
Cursor is perhaps not the best known company in the world, but I expect that it soon will become far better known amongst investors. It is an incredibly useful AI coding tool that is already in use at companies such as BP, Stripe, Figma (FGMA), Nvidia (NVDA), Lyft (LYFT), Datadog (DDOG) and Fox. Cursor has grown from $100 million to $1 billion of ARR in less than a year. It offers, perhaps the fastest growing IT product of all-time. And it uses Nebius for its infrastructure.
Cursor 可能不是世界上最著名的公司，但我预计它很快就会在投资者中变得更加知名。它是一个非常有用的 AI 编码工具，已经在 BP、Stripe、Figma (FGMA)、Nvidia (NVDA)、Lyft (LYFT)、Datadog (DDOG) 和 Fox 等公司使用。Cursor 的 ARR 在不到一年的时间里从 1 亿美元增长到 10 亿美元。它可能提供了有史以来增长最快的 IT 产品。而且它使用 Nebius 作为其基础设施。
Black Forest Labs is also probably not the best known company in the world, but it has seen rather incredible growth.It is a company, not a cake or a German region, or a furniture type. It is called a frontier AI research lab that is known for Flux which is a suite of text-to-image research models. Its models are already used by Adobe (ADBE), Meta and Microsoft. In August, its ARR was $96 million and it has projected ARR of $300 million in 2026 because it already has closed $300 million of signed deals. And it uses Nebius infrastructure.
Black Forest Labs 可能也不是世界上最著名的公司，但它经历了相当惊人的增长。它是一家公司，不是蛋糕、德国地区或家具类型。它被称为前沿 AI 研究实验室，以 Flux 闻名，Flux 是一套文本到图像的研究模型。其模型已被 Adobe (ADBE)、Meta 和微软使用。8 月份，其 ARR 为 9600 万美元，预计 2026 年 ARR 将达到 3 亿美元，因为它已经完成了 3 亿美元的签约交易。而且它使用 Nebius 基础设施。
Mistral AI is a fairly well known builder of LLM models. It competes with OpenAI, Google (GOOG) and Anthropic. Since the release of “Le Chat” its revenues trebled in 100 days. ARR is projected to treble again to E300 million in 2026. It is another user of Nebius infrastructure.
Mistral AI 是相当知名的 LLM 模型构建者。它与 OpenAI、Google (GOOG) 和 Anthropic 竞争。自发布“Le Chat”以来，其收入在 100 天内翻了三倍。预计到 2026 年 ARR 将再次翻三倍至 3 亿欧元。它是 Nebius 基础设施的另一个用户。
Another hyper growth Nebius user is a company called Higgsfield AI. It is perhaps the fastest growing GenAI start-up of all time. ARR actually went from $11 million in February to a recently reported $100 million. This company facilitates the creation of short videos by creators, marketers and others using text prompts, images and sketches. Its claim to uniqueness is its ease of use and rapid iteration capabilities
另一个高速增长的 Nebius 用户是一家名为 Higgsfield AI 的公司。它可能是有史以来增长最快的 GenAI 初创公司。ARR 实际上从 2 月份的 1100 万美元增长到最近报道的 1 亿美元。这家公司通过使用文本提示、图像和草图，帮助创作者、营销人员和其他人创建短视频。其独特的卖点在于易用性和快速迭代能力。
I could have included additional companies on this list, some tiny and others much more substantial. When investors think about why Nebius has an exceptional growth ramp, one part of the answer is that it has customers whose usage of Nebius infrastructure is being driven by their own hyper growth.
我本可以在这个名单中加入更多的公司，有些很小，有些则大得多。当投资者思考为什么 Nebius 拥有非凡的增长速度时，答案的一部分是它拥有的客户对其基础设施的使用是由他们自身的高速增长驱动的。
Of course not all Nebius customers are growing at these kinds of rates and these kinds of rates are not sustainable as customers such as these scale. Nebius has many different classes of customers and some of them are tiny and not all of them are going to be successful. And Nebius will not capture every hot AI native start-up.
当然，并非所有 Nebius 客户都以这种速度增长，而且随着这些客户的规模扩大，这种速度也不可持续。Nebius 有许多不同类别的客户，其中一些很小，并不是所有的都会成功。Nebius 也不会捕获每一个热门的 AI 原生初创公司。
One advantage Nebius has is that it has a global footprint and has become the “go-to” AI infrastructure vendor for AI native vendors domiciled in Europe. 2 of the 4 companies called out above are headquartered in Europe, and Higgsfield, while headquartered in San Francisco, also has location in the UK and Kazakhstan.
Nebius 的一个优势是它拥有全球足迹，并已成为注册在欧洲的 AI 原生供应商的“首选” AI 基础设施供应商。上面提到的 4 家公司中有 2 家总部位于欧洲，而 Higgsfield 虽然总部位于旧金山，但在英国和哈萨克斯坦也有分支机构。
My thesis is that Nebius has the best technology and the best geographic coverage for AI Native start-ups. And it has a competitive offering to fulfill capacity requirements for Hyper-scalers. While the reality is that it is too soon to try to project a multi-year CAGR for Nebius and really the other neo-clouds as well, I think that the combination of technology, customer momentum, geographic reach and a management team likely to react quickly and successfully to new trends in this space are likely to help Nebius to grow at faster rates for longer than other companies in the space.
我的论点是，Nebius 拥有最适合 AI 原生初创公司的技术和地理覆盖范围。而且它拥有具有竞争力的产品来满足超大规模云服务商的容量需求。虽然现实情况是，现在试图预测 Nebius 以及其他新云公司的多年复合年增长率还为时过早，但我认为技术、客户势头、地理覆盖范围以及可能对该领域新趋势做出快速成功反应的管理团队的结合，可能会帮助 Nebius 比该领域的其他公司以更快的速度增长更长时间。
The issue of depreciation
Nebius sells infrastructure based on Nvidia chips and the cost/performance of Nvidia chips is seeing substantial year on year improvement, and that is likely to continue for years to come. It has not gone unnoticed that Nebius and its peers have assets whose value is notionally depreciating at a significant rate. That said, it needs to be understood that Nebius sells a differentiated service far beyond access to Nvidia chips. It is really the price of the service that matters and not the price/performance of the latest Nvidia chips.
折旧问题
Nebius 销售基于英伟达芯片的基础设施，而英伟达芯片的性价比正在逐年大幅提高，这可能会在未来几年持续下去。人们已经注意到，Nebius 及其同行的资产价值在名义上正在以显著的速度贬值。也就是说，需要理解的是，Nebius 销售的是一种差异化的服务，远不止是访问英伟达芯片。真正重要的是服务的价格，而不是最新英伟达芯片的性价比。
Nebius depreciates its hardware assets at over a 4 year span. As reported, depreciation soared by more than 4X year over year and was $99 million or 68% of reported revenues last quarter. Depreciation is rising by about 30% a quarter but because revenues are rising so rapidly, as a percentage of revenues, depreciation charges are falling.
Nebius 将其硬件资产的折旧期定为 4 年以上。据报道，上个季度折旧同比飙升 4 倍以上，达到 9900 万美元，占报告收入的 68%。折旧每季度增长约 30%，但由于收入增长如此迅速，作为收入的百分比，折旧费用正在下降。
The two large infrastructure deals that Nebius has signed are multi-year commitments that go beyond the 4 year depreciation period. While data centers do have substantial operating expenses, last quarter these had come down to 29% of revenues for Nebius. So, the margins on fully depreciated data centers are going to be high.
Nebius 签署的两笔大型基础设施交易是多年承诺，超过了 4 年的折旧期。虽然数据中心确实有大量的运营费用，但上个季度 Nebius 的这些费用已降至收入的 29%。因此，完全折旧的数据中心的利润率将会很高。
I assume that when the deals with the hyperscalers reach the end of their current terms, Nebius will renegotiate these contracts with its customers at a considerable discount, but without depreciation charges on the income statement, pricing flexibility is going to be substantial. These very large deployments are custom designed and optimized for the hyper-scaler users so their stickiness is substantially greater than might be imagined.
我假设当与超大规模云服务商的交易达到当前期限结束时，Nebius 将以相当大的折扣与客户重新谈判这些合同，但由于损益表上没有折旧费用，定价灵活性将非常大。这些非常大的部署是为超大规模云服务商用户定制设计和优化的，因此它们的粘性比想象的要大得多。
Nebius’ smaller customers pay for their service based on usage charges calculated on minutes of access. No doubt, over time, these per minute charges will decrease-that has been the case for cloud pricing almost since there has been cloud pricing. Needless to say, that hasn’t prevented cloud revenues from growing. That relates to demand elasticity which is almost certainly very high.
Nebius 的较小客户根据按分钟计算的使用费支付服务费用。毫无疑问，随着时间的推移，这些每分钟的费用将会下降——自从有云定价以来，云定价几乎一直是这种情况。不用说，这并没有阻止云收入的增长。这与需求弹性有关，这种弹性几乎肯定非常高。
But in addition, the applications that are adopting Nebius for their infrastructure requirements are also exceptionally sticky. Switching costs are high, the effort to switch is considerable and time consuming and the benefits are unlikely to be significant. I expect that most of the AI natives that have become and will become customers are likely to remain Nebius users for many years into the future. Here is a quotation about switching from a 3rd party blog that is pretty straightforward and dispositive:
但除此之外，采用 Nebius 满足其基础设施需求的应用程序也具有极高的粘性。转换成本很高，转换工作量巨大且耗时，而且收益可能并不显著。我预计大多数已经成为和将要成为客户的 AI 原生企业可能会在未来许多年继续作为 Nebius 用户。以下引用自第三方博客关于转换的一段话，非常直截了当且具有决定性：
This advantage is fortified by high switching costs. AI workloads are “notoriously sticky,” as migrating them once optimized on a platform is prohibitively expensive. This dynamic locks in clients, secures long-term recurring revenue, and embeds Nebius deeply in their AI operations.
这种优势因高昂的转换成本而得到加强。AI 工作负载“极其具有粘性”，因为一旦在平台上进行了优化，迁移它们的成本高得令人望而却步。这种动态锁定了客户，确保了长期的经常性收入，并将 Nebius 深度嵌入到他们的 AI 运营中。
Switching costs include data egress fees, lots of engineering effort, ecosystem and tooling migration and redoing an app in order to optimize it for new hardware. It is almost unheard of for users to switch infrastructure vendors and given the specifics of Nebius infrastructure services, that chance is close to a vanishing level.
转换成本包括数据出口费用、大量的工程工作、生态系统和工具迁移以及重做应用程序以针对新硬件进行优化。用户更换基础设施供应商几乎闻所未闻，考虑到 Nebius 基础设施服务的具体情况，这种机会几乎微乎其微。
I think that there is very little chance that the chips being used by Nebius currently to deploy infrastructure for its customers won’t continue to be deployed and functioning for most customers, both hyperscalers and AI natives long beyond the 4 year depreciation period. There is a risk that some of the start-up AI natives won’t make it-but even in those cases the likelihood is that the customer will be bought out or it will sell its technology to a 3rd party so that Nebius data centers will continue to achieve high levels of utilization.
我认为，Nebius 目前用于为客户部署基础设施的芯片在超过 4 年折旧期后，很有可能继续为大多数客户（包括超大规模云服务商和 AI 原生企业）部署和运行。存在一些 AI 原生初创公司无法成功的风险——但即使在这些情况下，可能性也是客户将被收购或者将其技术出售给第三方，以便 Nebius 数据中心将继续实现高水平的利用率。
Nebius Competition
Nebius has loads of competitors of various kinds. Of course the most formidable competitors are the well known hyperscalers including AWS, Google’s Cloud, Azure and Oracle. It also faces more specialized vendors such as CoreWeave, Lambda, and IREN (IREN).
Nebius 的竞争
Nebius 拥有各种各样的竞争对手。当然，最强大的竞争对手是众所周知的超大规模云服务商，包括 AWS、Google Cloud、Azure 和 Oracle。它还面临更专业的供应商，如 CoreWeave、Lambda 和 IREN (IREN)。
At the moment, given that there is more demand for AI infrastructure services than supply, Nebius has one competitor, Microsoft, who has contracted for up to $19 billion of service over a 5 year term. Other deals with hyperscalers seem likely depending on how much capacity Nebius might be able to offer.
目前，鉴于 AI 基础设施服务的需求大于供应，Nebius 有一个竞争对手，即微软，它已签订了为期 5 年、价值高达 190 亿美元的服务合同。与其他超大规模云服务商的交易似乎也有可能，这取决于 Nebius 能提供多少容量。
Actually, regardless of the capacity issue, Nebius does have some noticeable advantages compared to the hyperscalers when it comes to AI workloads. The single biggest of these advantages is performance that I alluded to earlier in this report. Then there is the custom hardware that Nebius uses to optimize GPU utilization and minimize power and cooling requirements. Because of the specialization Nebius has, it can and does provide customers with both enhanced support and a superior user experience. And, as mentioned, because of the location of Nebius data centers, it can provide customers with data sovereignty and green compliance that are mandatory for some users at this point.
实际上，无论容量问题如何，在 AI 工作负载方面，与超大规模云服务商相比，Nebius 确实具有一些明显的优势。其中最大的单一优势是我在本报告前面提到的性能。其次是 Nebius 用于优化 GPU 利用率并最小化电源和冷却要求的定制硬件。由于 Nebius 拥有的专业化，它可以并且确实为客户提供了增强的支持和卓越的用户体验。而且，如前所述，由于 Nebius 数据中心的位置，它可以为客户提供数据主权和绿色合规性，这对于某些用户来说是目前的强制性要求。
Its major weaknesses vs. the hyper scalers relate to its size, its cash burn, execution risk and brand recognition. Thus far, Nebius has been able to overcome those weaknesses and exploit its advantages.
与超大规模云服务商相比，其主要弱点与其规模、烧钱速度、执行风险和品牌知名度有关。到目前为止，Nebius 已经能够克服这些弱点并利用其优势。
Nebius, as will be seen in the next section but one, has a much less leveraged balance sheet than its neo-cloud competitors. Nebius is, to be sure, going to need to raise capital, but on a relative basis, it has a more conservative capital model, and most of its borrowings are essentially secured by the contracts it has with its hyperscaler customers.
Nebius，如下一节所见，其资产负债表的杠杆率远低于其新云竞争对手。毫无疑问，Nebius 将需要筹集资金，但相对而言，它拥有更保守的资本模型，而且其大部分借款基本上由其与超大规模云服务商客户的合同担保。
Nebius offers a full stack. It designs its own data centers including servers, racks and cooling systems. That is a significant cost advantage since it doesn’t have to pay OEM mark-ups for its hardware and its hardware is specifically optimized for typical AI use cases. Part of the reason why Nebius has gotten so many of the high growth AI native startups is its very flexible terms for those customers. As mentioned, its pricing is pay-as-you-go with no lock-ins.
Nebius 提供全栈服务。它设计自己的数据中心，包括服务器、机架和冷却系统。这是一个显着的成本优势，因为它不必为其硬件支付 OEM 加价，而且其硬件针对典型的 AI 用例进行了专门优化。Nebius 能够获得如此多高增长 AI 原生初创公司的部分原因是其为这些客户提供的非常灵活的条款。如前所述，其定价是按需付费，没有锁定。
Here is a current price comparison chart for two flavors of Nvidia GPUs. The chart speaks for itself when it comes to pricing:
以下是两种英伟达 GPU 的当前价格比较表。在定价方面，该表不言自明：
Provider: Nebius
GPU Model: H100
Estimated Price per GPU per Hour (On-Demand): ~$2.00
Notes: Requires minimum monthly commitments; final cost is higher with storage/networking.
供应商：Nebius
GPU 型号：H100
预估每 GPU 每小时价格（按需）：约 2.00 美元
备注：需要最低月度承诺；加上存储/网络后的最终成本会更高。
Provider: CoreWeave
GPU Model: H100 (PCIe)
Estimated Price per GPU per Hour (On-Demand): ~$4.25
Notes: Pricing is à la carte (GPU only); full instance can reach ~$6.15/hr per GPU.
供应商：CoreWeave
GPU 型号：H100 (PCIe)
预估每 GPU 每小时价格（按需）：约 4.25 美元
备注：定价是单点的（仅 GPU）；完整实例每个 GPU 可能达到约 6.15 美元/小时。
Provider: IREN
GPU Model: H100
Estimated Price per GPU per Hour (On-Demand): Not publicly listed
Notes: IREN focuses more on the physical data center infrastructure; software layer is emerging.
供应商：IREN
GPU 型号：H100
预估每 GPU 每小时价格（按需）：未公开列出
备注：IREN 更专注于物理数据中心基础设施；软件层正在兴起。
Provider: Nebius
GPU Model: A100 (80GB)
Estimated Price per GPU per Hour (On-Demand): ~$1.80+
Notes: Offers highly competitive A100 pricing, often undercutting competitors.
供应商：Nebius
GPU 型号：A100 (80GB)
预估每 GPU 每小时价格（按需）：约 1.80 美元以上
备注：提供极具竞争力的 A100 定价，通常低于竞争对手。
Provider: CoreWeave
GPU Model: A100 (80GB)
Estimated Price per GPU per Hour (On-Demand): ~$2.21
Notes: GPU component only; full instance price is generally over $3.00/hr.
供应商：CoreWeave
GPU 型号：A100 (80GB)
预估每 GPU 每小时价格（按需）：约 2.21 美元
备注：仅 GPU 组件；完整实例价格通常超过 3.00 美元/小时。
As mentioned earlier, Nebius gross margins are not being constrained by this pricing. To repeat, its gross margins, excluding depreciation are 71% and are on a rising trajectory so far this year.
如前所述，Nebius 的毛利率并未受到这种定价的限制。重申一下，今年迄今为止，其不含折旧的毛利率为 71%，且处于上升轨道。
Nebius Other Bets
Nebius does have some other bets, and their valuation is a significant component of the company’s current valuation. Probably Avride, the autonomous driving company that Nebius controls is its most valuable non AI infrastructure assets. At this point, the company provides food delivery through autonomous vehicles in Austin, Dallas and Jersey City as well as on the campus at Ohio State. Its first robotaxi services was recently launched downtown Dallas. Avride recently raised money in a financing with a valuation of $3.75-$4 billion. Avride and Uber have a partnership and but Avride is considered a wholly owned subsidiary. Employees of the venture own 17%, while Uber has a convertible note that apparently can be converted into 10% of the equity in the venture. Avride is far behind Waymo but of course this is a nascent space.
Nebius 的其他赌注
Nebius 确实有一些其他赌注，其估值是公司当前估值的重要组成部分。Nebius 控制的自动驾驶公司 Avride 可能是其最有价值的非 AI 基础设施资产。目前，该公司在奥斯汀、达拉斯和泽西市以及俄亥俄州立大学校园通过自动驾驶汽车提供食品配送服务。其首个机器人出租车服务最近在达拉斯市中心推出。Avride 最近以 37.5-40 亿美元的估值筹集了资金。Avride 和 Uber 建立了合作伙伴关系，但 Avride 被视为全资子公司。该企业的员工拥有 17% 的股份，而 Uber 拥有一张可转换票据，显然可以转换为该企业 10% 的股权。Avride 远远落后于 Waymo，但这当然是一个新兴领域。
Tokola is considered to be an independent company. Its current estimated revenues are about $60 million+ more than double last year’s revenue, Nebius owns more than 50% of Tokola and the balance is owned by Bezos Ventures and by the CTO of Shopify. Tokola helps companies train and fine-tune AI models by offering what are called annotated datasets. Customers currently include Microsoft, Amazon (AMZN), Shopify (SHOP) and Anthropic.
Tokola 被认为是一家独立公司。其目前的预估收入约为 6000 万美元以上，是去年收入的两倍多，Nebius 拥有 Tokola 50% 以上的股份，余额由 Bezos Ventures 和 Shopify 的首席技术官拥有。Tokola 通过提供所谓的带注释数据集，帮助公司训练和微调 AI 模型。目前的客户包括微软、亚马逊 (AMZN)、Shopify (SHOP) 和 Anthropic。
Nebius has a 28% interest in ClickHouse. ClickHouse has a very high performing database that is geared toward both web analytics and AI applications. It uses Vector technology and has been quoted as having a valuation of greater than $6 billion. ClickHouse is another company seeing exceptional growth with ARR having doubled to $96 million between May of this year and the end of 2024. Large customers include Meta, Tesla (TSLA) and Anthropic. It is possible that ClickHouse will go public in the next several months and will provide required capital for Nebius when that happens.
Nebius 拥有 ClickHouse 28% 的权益。ClickHouse 拥有一个非常高性能的数据库，面向网络分析和 AI 应用程序。它使用向量技术，据称估值超过 60 亿美元。ClickHouse 是另一家经历非凡增长的公司，ARR 在今年 5 月至 2024 年底之间翻了一番，达到 9600 万美元。大客户包括 Meta、特斯拉 (TSLA) 和 Anthropic。ClickHouse 有可能在未来几个月内上市，届时将为 Nebius 提供所需的资金。
Nebius Valuation
Nebius is an asset that is not easy to value using typical valuation metrics. I would like to use some kind of DCF analysis at least to set upper and lower bounds, but it really isn’t possible to estimate all of the variables necessary for a credible DCF analysis at this point. One of the basic issues is that the faster Nebius' growth, the greater the capital requirements, and in turn the more capital, some of which will be equity, will be required.
Nebius 估值
Nebius 是一项很难使用典型估值指标进行估值的资产。我想使用某种 DCF 分析至少设定上限和下限，但在目前确实无法估计可靠 DCF 分析所需的所有变量。基本问题之一是 Nebius 增长越快，资本需求就越大，反过来也就需要更多的资本，其中一些将是股权。
About a week ago, an “analysis” so called was published by Zack’s and I have linked to it here. The methodology is deeply flawed, I believe, because there is an essential disconnect between capex and revenue in the analysis. Of course that is not the case. The author doesn't seem to believe that capex for a company like this is the foundation for revenue growth. But I believe that some of the current valuation angst relates to this article.
大约一周前，Zack’s 发布了一篇所谓的“分析”，我已在此链接。我认为该方法存在严重缺陷，因为分析中资本支出与收入之间存在本质上的脱节。当然情况并非如此。作者似乎不相信像这样的一家公司的资本支出是收入增长的基础。但我认为目前的一些估值焦虑与这篇文章有关。
Before trying my own hand at valuation, I at least need to clear away some of the inaccuracies in this article. The company’s capex will, by definition, not influence EBITDA margins which were 19% last quarter and which are projected to rise somewhat more in this current quarter. EBITDA means earnings excluding depreciation and amortization-it seems unfortunate that the Zack’s analyst is unaware of that.
在尝试自己估值之前，我至少需要清除这篇文章中的一些不准确之处。根据定义，公司的资本支出不会影响 EBITDA 利润率，上个季度为 19%，预计本季度将有所上升。EBITDA 意味着不包括折旧和摊销的收益——Zack’s 的分析师似乎不知道这一点，这很遗憾。
The company does intend to spend $3 billion on capex this quarter per the conference call script.
I think the basic question to address is if it is reasonable to make a $5 billion/year of capital commitments in order to increase ARR by $8 billion in a single year with a gross margin in the low 70% range. This ARR is going to persist for years, without requiring additional capex. The answer should be fairly self-evident.
根据电话会议记录，该公司确实打算本季度在资本支出上花费 30 亿美元。
我认为要解决的基本问题是，为了在单一年份增加 80 亿美元的 ARR（毛利率在 70% 低区间），每年做出 50 亿美元的资本承诺是否合理。这个 ARR 将持续多年，而不需要额外的资本支出。答案应该是不言而喻的。
I find it difficult to believe that any profit-maximizing business would not make that kind of investment, particularly if they have available sources of capital to fund the expenditure. The AI space is new, and using traditional financial analysis methodology is simply not appropriate. But this article has apparently had an impact on the shares and that has apparently continued up to this writing.
我觉得很难相信任何追求利润最大化的企业不会进行这种投资，特别是如果他们有可用的资金来源来资助这项支出的话。AI 领域是新的，使用传统的财务分析方法根本不合适。但这篇文章显然对股价产生了影响，而且这种影响显然持续到了本文撰写之时。
How do I look at valuation? Nebius hasn’t presented a specific revenue forecast for 2026. It has projected that its ARR will be $7-$9 billion by the end of next year. It apparently has substantial visibility into this forecast: more than half of this ARR has already been committed more than a year before the forecast period.
我如何看待估值？Nebius 尚未提出 2026 年的具体收入预测。它预计到明年年底其 ARR 将达到 70-90 亿美元。它对这一预测显然有很高的可见度：超过一半的 ARR 已经在预测期前一年多得到了承诺。
I have laid out the rapid growth of some of Nebius's new AI native customers. Those customers don’t add to ARR but they certainly will add to revenue as their usage grows along with their own business. There will, in addition, be additional customers-probably a couple of hyperscalers but scores of AI native customers in many verticals.
我已经列出了 Nebius 一些新的 AI 原生客户的快速增长。这些客户不增加 ARR，但随着其使用量随自身业务增长，肯定会增加收入。此外，还将有更多客户——可能有几个超大规模云服务商，但在许多垂直领域会有数十个 AI 原生客户。
We are also winning new customers and expanding use cases in healthcare and life sciences, where partners such as Basecamp Research, Prima Mente, and Sword Health are building models to help solve some of the world’s biggest medical challenges.
我们还在赢得新客户并扩展医疗保健和生命科学领域的用例，Basecamp Research、Prima Mente 和 Sword Health 等合作伙伴正在构建模型，以帮助解决世界上一些最大的医疗挑战。
The above is just a brief listing of some new Nebius customers in a single vertical. The first of these companies is small and the second is tiny. That said, Sword Health has already reached $240 million in ARR and is seeing revenues double year-on-year. Sword is a company that provides virtual physical therapy and pain management using both AI and human intervention.
以上仅仅是 Nebius 在单一垂直领域的一些新客户的简要列表。这其中的第一家公司很小，第二家微不足道。也就是说，Sword Health 的 ARR 已经达到 2.4 亿美元，收入同比翻了一番。Sword 是一家使用 AI 和人工干预提供虚拟物理治疗和疼痛管理的公司。
It is looking at just those kinds of statistics that lead me to believe that Nebius is likely to achieve $10-$12 billion in revenue in 2027 and as much as $20 billion of revenue in 2028-in other words a consistent level of dollar revenue growth. I expect that gross margins will likely rise to around 75% with opex falling to 35% of revenue, down from current the latest reported level of 92%. That would lead to an EBITDA margin of 40% about double current levels. The adjusted EBITDA margin in my model reaches 45%.
正是看着这些统计数据，让我相信 Nebius 很可能在 2027 年实现 100-120 亿美元的收入，并在 2028 年实现高达 200 亿美元的收入——换句话说，美元收入增长保持一致的水平。我预计毛利率可能会上升到 75% 左右，运营支出将下降到收入的 35%，低于目前最新报告的 92%。这将导致 EBITDA 利润率为 40%，约为当前水平的两倍。我模型中的调整后 EBITDA 利润率达到 45%。
I really don’t think it is feasible at this point to project either depreciation or free cash flow. They depend to a great deal on capex and on the trend of capex and its level 2-3 years from now.
我认为目前预测折旧或自由现金流是不可行的。它们在很大程度上取决于资本支出以及资本支出的趋势及其在 2-3 年后的水平。
Nebius had 252 million shares outstanding at the end of last quarter and it is selling another 25 million shares through an “at-the market” process. It had a cash balance of $4.8 billion as of Sept. 30, 2025. The share sale will raise another $2.3 billion. My guess is that capex through 2028 will aggregate about $18 billion, averaging $6 billion/year with operating income through that period aggregating $8-$9 billion. I think that will mean that the company will need another equity raise during the period-for the sake of prudence I am using another 25 million shares. Nebius will likely be able to raise debt funding based on the firm orders it will be getting from hyperscalers.
Nebius 上季度末已发行股票为 2.52 亿股，并通过“市价”流程再出售 2500 万股。截至 2025 年 9 月 30 日，其现金余额为 48 亿美元。股票出售将再筹集 23 亿美元。我猜测到 2028 年资本支出总计约为 180 亿美元，平均每年 60 亿美元，该期间的营业收入总计 80-90 亿美元。我认为这意味着该公司将需要在此期间再进行一次股权融资——为谨慎起见，我使用的是再增加 2500 万股。Nebius 很可能能够根据其从超大规模云服务商获得的确认订单筹集债务资金。
The way I look at it, at the end of the forecast period, Nebius will have an outstanding share count of 305 million and a revenue run rate of at or greater than $20 billion with a strong EBITDA margin, and likely a strong level of free cash flow. Those estimates create a potential share price several multiples of the current share price level, i.e. $88/share which equates to a market cap of $27 billion.
依我看，在预测期结束时，Nebius 将拥有 3.05 亿股流通股，收入运行率达到或超过 200 亿美元，拥有强劲的 EBITDA 利润率，并且可能拥有强劲的自由现金流水平。这些估计产生的潜在股价是当前股价水平的数倍，即 88 美元/股，相当于 270 亿美元的市值。
I make it a point to try to avoid publishing price targets when the assumptions relate to expectations several years in the future. That said, I would find an EV/S of 5X unusually compressed and an EV/S of 10X within an expected range if AI infrastructure growth continue on at levels I anticipate.That produces a return of several time on the current share price and that is good enough for me.
我特别注意尽量避免在假设涉及未来几年的预期时发布价格目标。也就是说，如果 AI 基础设施增长继续保持我预期的水平，我会发现 5 倍的 EV/S 异常压缩，而 10 倍的 EV/S 在预期范围内。这会在当前股价基础上产生数倍的回报，这对我来说已经足够好了。
Risks to the Investment thesis
Of course the biggest risk to the investment thesis relates to the progress of AI applications and their deployment. I have looked at just a few of many companies whose applications have been designed to run on Nebius infrastructure. Their success, or lack thereof, will be a major determinant in Nebius’s growth over time.
投资论点的风险
当然，投资论点的最大风险与 AI 应用程序及其部署的进展有关。我只查看了众多将其应用程序设计为在 Nebius 基础设施上运行的公司中的几家。随着时间的推移，它们的成功与否将是 Nebius 增长的主要决定因素。
Nebius will face material execution challenges as it scales. Building out capacity at the rate it has forecast is going to be a challenge. It needs to acquire land, power, and design and build data centers at a rapid pace. That will be a material undertaking requiring the best execution skills.
Nebius 在扩张时将面临重大的执行挑战。以其预测的速度建设产能将是一个挑战。它需要快速获取土地、电力，并设计和建造数据中心。这将是一项需要最佳执行技能的重大任务。
Nebius faces competition from both hyperscalers and from other neoclouds. Challenges will come in terms of functionality, performance and the maintenance of cost advantages.
Nebius 面临来自超大规模云服务商和其他新云公司的竞争。挑战将来自功能、性能和成本优势的维持方面。
The shares will be vulnerable to risk-off sentiment and to read-throughs from the performance of other vendors.
股价将容易受到避险情绪和其他供应商表现解读的影响。
Finally, the overall macro environment could upend projections for a relatively benign economy and the ability of customers to spend on AI solutions.
最后，整体宏观环境可能会推翻对相对良性经济以及客户在 AI 解决方案上支出能力的预测。
Wrapping Up-The case to buy Nebius shares now
Nebius, I believe, is the leading neocloud from an investment perspective because of its advantages in technology, cost and its AI Cloud Native customer base. It has an experienced and tech savvy leadership team who have designed and developed AI infrastructure offerings that appeal to the fastest growing AI native startups. And it can support capacity requirements of hyperscalers efficiently and at competitive costs. It has strong visibility with this current quarter probably sold out and more than 50% of its projected 8X increase in ARR next year already committed.
总结——现在买入 Nebius 股票的理由
我相信，从投资角度来看，Nebius 是领先的新云公司，因为它在技术、成本和 AI 云原生客户群方面具有优势。它拥有一支经验丰富且精通技术的领导团队，他们设计和开发了吸引增长最快的 AI 原生初创公司的 AI 基础设施产品。而且它可以高效且以具有竞争力的成本支持超大规模云服务商的容量需求。它具有很强的可见度，本季度可能已售罄，明年预计增加 8 倍的 ARR 中超过 50% 已经得到承诺。
It has been able to do this while managing costs effectively so that it has already reached 19% EBITDA margins for core Nebius operations. I believe its valuation is quite modest for its highly visible opportunities and that is why I have made it my largest single holding in my high growth portfolio.
它能够在有效管理成本的同时做到这一点，因此其核心 Nebius 业务的 EBITDA 利润率已达到 19%。我相信鉴于其高度可见的机会，其估值相当适中，这就是为什么我将其作为我高增长投资组合中最大的单一持仓。
