AI的下一批赢家：来自私募市场的信号

欢迎收听《市场观点》。

Welcome to Thoughts on the Market.

我是摩根士丹利AI与科技主题股票策略师Sabrina Hao。

I'm Sabrina Hao, Morgan Stanley's AI and Tech Thematic Equity Strategist.

我是Adam Wood，美国软件研究部门的负责人。

And I'm Adam Wood, Head of U.

美国

S.

软件研究，摩根士丹利。

Software Research at Morgan Stanley.

我是Brian

And I'm Brian

Nowak，来自美国。

Nowak, of U.

美国

S.

互联网研究。

Internet Research.

今天，我们来聊聊私营公司对AI机遇的看法，以及这对公开市场意味着什么。

Today, what private companies are telling us about the AI opportunity and what it means for public markets.

今天是10月20日，星期二，纽约时间上午10点。

It's Tuesday, October 20 at 10AM in New York.

上周，我们在加利福尼亚参加了我们的Spark私营公司会议，汇集了超过100家初创公司，涵盖软件、机器人、AI基础设施等领域。

Last week, we were at our Spark Private Company conference in California, bringing together more than a 100 startups across software, robotics, AI infrastructure, and beyond.

甚至还有一家公司在太空中帮助制造药品。

There was even a company helping to manufacture pharmaceuticals in space.

考虑到市场上发生的一切，那周参加会议真的很有意思，但说实话，我们离开时对AI周期更加乐观了。

It was a really interesting week to be at the conference given everything that was happening in the market, but really we came away feeling even more positive about the AI cycle.

虽然公开市场的投资者还在等待回报来证明数万亿美元AI支出的合理性，但我们已经看到这些投入在私营市场得到了回报。

While public market investors are waiting for returns to justify trillions of dollars in AI spending, we're already seeing that pay off in private markets.

AI需求是真实的，公司正在为客户带来实实在在的回报，初创公司的发展速度也比以往任何时候都快。

AI demand is real, companies are delivering tangible returns for customers, and startups are scaling faster than ever.

亚当，我们先从增长说起吧。

Adam, let's start with growth.

我们听到一些初创公司以惊人的速度扩张。

We heard from startups scaling at extraordinary speeds.

有一家公司在两年内接近40亿美元的年度经常性收入，另一家不到一年就达到了10亿美元。

One was approaching 4,000,000,000 in annual recurring revenue in two years, another reaching 1,000,000,000 in under a year.

关于AI的采用速度，最让你印象深刻的是什么？

What really stood out to you about the pace of AI adoption?

为什么我们在成熟的软件公司那里没有看到同样的增长呢？

And why aren't we necessarily seeing that same growth at established software companies?

对，确实如此。

So yeah, absolutely.

这个速度确实很惊人。

The pace was remarkable.

而且很明显，这是这次会议的一大亮点。

And clearly, one of the standouts from the conference.

我们以前讨论的是，一家公司能多快达到所谓的十亿美元独角兽估值。

And we used to discuss how quickly a company could get to a $1,000,000,000 so called unicorn valuation.

现在我们在讨论的是，他们能多快达到十亿美元的年经常性收入，这对我来说相当不可思议。

And now we're discussing how quickly they can get to a unicorn revenue run rate of $1,000,000,000 which for me is quite incredible.

但我觉得，我不会把讨论简单地分成私营公司和上市公司。

But I think I wouldn't make the debate simply about private versus public companies.

我会区分两种不同类型的软件。

I'd make the distinction between two different types of software.

我觉得我们必须区分主要是编码和推理工具。

Think I we have to make a distinction between largely coding and inference tools.

会议上有很多编码和推理软件公司，还有应用软件公司。

There are a number of coding and inference software companies at the conference and application software.

我认为在把这些产品推向市场的时间和复杂性方面，存在很大的差异。

And I think there's a big difference in terms of time and complexity to bring these products to market.

我觉得编码和推理，也就是我所说的应用层下面的构建模块，它们可以更快地构建出来。

I I think the coding and inference, which tends to be the building blocks, as I would call them underneath the application layer, they can be constructed much more quickly.

而且我觉得，因为在我们还在定制开发的时候就可以采用，我觉得我们现在就处于这个阶段，还在做定制软件，而不是直接买现成的，所以这些工具也能更早地被采用。

And I think because we can adopt when we're still building bespoke, which I think is a lot of the phase we're in today, we're still building bespoke software, rather than buying off the shelf, They can be adopted earlier as well.

所以我们的构建速度更快，企业采用的速度也更快。

So we've got faster time to build and faster time to adopt by companies.

所以我觉得，在这些类型的业务中，看到它们快速增长是很自然的事情。

So I think it's natural where you can see very fast growth in those types of businesses very quickly.

我觉得在应用软件领域，情况有点不同，因为这些应用必须为客户提供完整的业务解决方案。

I think in the application landscape, it's a little bit different because those applications have to deliver complete business solutions for customers.

在金融和客户服务领域，这意味着要整合数据、理解工作流程并建立控制措施。

In finance and customer service, that means integrating data, understanding workflows and putting controls in place.

需要构建和部署的东西要多得多。

There's much more to build and much more to implement.

而且一般来说，这些工具的采用会稍晚一些，是在我们经历了企业自己构建的定制阶段之后。

And generally, the adoption of those tools takes place a little bit later after we've gone from the bespoke phase where companies are building themselves.

然后进入一个阶段，企业发现市面上有现成的软件包，是其他公司已经开发好的。

We go to a phase where companies see that there are off the shelf packages that are available that companies have built out.

所以我觉得这既关乎构建的复杂性，也关乎我们在软件采用周期中所处的阶段。

So I think it's both a complexity of build debate, and also a time in terms of where we are in the software adoption cycle.

第二点我想说的是，在那些已有大型现有业务的上市软件公司里，有时会存在一个增长缓慢的大型现有业务，同时还有一个增长非常快的AI业务。

The second thing I'd say is, you know, in public software companies that have a big incumbent existing business, sometimes there can be that big incumbent business that's not growing very quickly, and a very fast growing AI business in there as well.

而那个增长非常快的AI业务，可能会被这家大型上市软件公司的整体形象所掩盖。

And that very fast AI growth business can be hidden a little bit by that large public software company that exists.

所以我觉得不能简单地说私营公司增长快、上市公司增长慢，我认为这里面有几层复杂性，取决于你处理的是哪种类型的软件。

I So don't think it's as straightforward as just saying, you know, private is growing very quickly public not, I think it's there's a couple of layers of complexity in there in terms of what type of software you're dealing with.

然后还要看软件公司现有业务的规模有多大，以及现有软件的增长是否稍显缓慢。

And then how big the existing kind of business is in a software company and what is kind of existing software growing a little bit less quickly.

而AI部分则可能增长非常快。

And then the AI bit that could potentially be growing very fast.

完全同意。

Absolutely.

这非常有道理。

It makes a ton of sense.

布莱恩，说到全栈解决方案这个话题，我们听说了从AI助手到能实际执行工作的智能体的转变。

Maybe on this theme of full stack solutions, you know, Brian, we heard about the shift from AI assistants to agents that can actually execute work.

你在消费者和企业领域都做了很多这方面的研究。

You've done a lot of work on this both across consumer and enterprise.

你了解到智能体的发展有多快？这又如何影响你覆盖的互联网公司的前景？

You know, what did you learn about how quickly agents are advancing and how does it shape the outlook for the Internet names that you cover?

现在是智能体应用非常激动人心的时期，无论是在消费者端还是企业端。

This is a very exciting time for agentic adoption on both the consumer side and the enterprise side.

我们估算，消费者端和企业端各有30到40万亿美元的潜在支出市场，我们认为这些市场可以通过智能体工具来覆盖。

You know, we've sized potential addressable markets of $30 to $40,000,000,000,000 of spend on both the consumer side and the enterprise side that we think have been addressed by agentic tools.

在消费者端，人们在购物、规划旅行、安排假日购物。

On the consumer side, people are shopping, they are planning their travel, they're planning their holiday shopping.

在企业端，我们看到垂直领域的专用智能体开始出现。

On the enterprise side, we're seeing both vertical specific agents start to emerge.

比如法律领域的智能体，或者医疗保健领域的智能体。

So agents within the legal area or agents within the health care area.

同时，我们也看到一些大型私营公司推出的通用型代理开始崭露头角。

And then we're also seeing broad based agents from some of the larger private companies also start to bud through.

所以，在消费者端，各种代理正在不断涌现。

So there is a smattering of agents that are popping through on the consumer side.

Muse 是早期的领先者，我们预计到2027年会有更多类似的产品出现。

Muse is your early leader, and we do expect more to come in 2027.

然后，在企业端，无论是垂直领域的专用代理，还是更通用的代理，预计都会有所增长。

And then I think on the enterprise side, they're both vertical specific as well as broader based agents expected to grow.

代理确实引发了一个关于颠覆的重要问题。

Agents really raise an important question around disruption.

也许对亚当来说，你写过关于软件领域的护城河和演进的文章。

Maybe for Adam, you've written about the moat and the journey in software.

这不仅仅是关于一家公司是否拥有竞争优势，而是关于它如何应对向AI驱动世界的转型。

It's just it's not just about if a company has a competitive advantage, but how it navigates that transition to an AI driven world.

Spark 是否改变了你对哪些软件现有企业最有能力捍卫自己的业务或抓住新增长机会的看法？

Did Spark change your thinking about which software incumbents are best positioned to defend their businesses or capture new growth?

所以我觉得Spark确实印证了我们之前提出的那个关于护城河和转型路径的框架。

So I think Spark really confirmed a lot of our views on that moats and journey framework that we've put forward.

护城河保护你现有的业务，而转型路径则是把AI变成新的增长机会。

The moat protecting the existing business that you have, and the journey turning AI into new growth opportunities.

而且在某种程度上，企业的护城河越深，转型的路径可能就越长。

And in some ways, the deeper the moat for businesses, the longer that journey can be.

那些已经嵌入到复杂关键任务流程中的现有企业，因为数据集成和客户信任，很难被替换。

And incumbent embedded in complex mission critical workflows is hard to replace because of the data integrations and customer trusts it has by definition.

但同样的要求也可能让AI引入到其平台变得更加困难。

But those same requirements can make introducing AI into its platform more demanding.

你不能牺牲那些工作的可靠性。

You can't compromise the reliability of that work.

所以这就是为什么我们把业务的持久性和AI转型的速度分开来看。

So that's why we've separated the durability of the business from the speed of its AI transition.

对于我们看到的一些编码推理企业，这些增长极快、收入增速极高的企业，变现显然已经非常迅速了。

For some of the coding inference businesses we heard from, again, these incredibly fast run rate, fast growth businesses, the monetization has obviously moved very, very quickly.

我觉得那里的投资者争论可能会反过来。

And I think there the investor debate is going to be a little bit the opposite way around.

这些公司建立在快速增长的收入基础上，它们的竞争优势有多持久、多大。

How durable and how big a competitive advantage of these companies built on that very fast growing revenue base.

这并不是说这些企业不能有护城河，它们可以通过性能、模型定制和集成来创造真正的差异化。

And that's not to say these businesses can't have moats, they can, you know, performance, model customization, and integration can create real differentiation.

但我认为快速增长和可防御性之间有很大区别。

But I think there is a big difference to be made between rapid growth and defensibility.

举几个例子，关于一些AI原生的垂直和商业软件公司正在利用的护城河，有一家法律AI公司正在雇佣前律师成为前部署律师，我们之前有前部署工程师，现在有了前部署律师。

And to give a couple of examples here, in terms of the moat that some of these AI native vertical and business software companies are leveraging, there's a legal AI company that's hiring former lawyers to become forward deployed lawyers, we had forward deployed engineers, now we've got forward deployed lawyers.

但原因是他们带来了对法律工作流程的理解，应用到产品和实施中。

But the reason there is they bring that understanding of legal workflows to the product and its implementation.

所以价值不仅仅是访问模型，而是让模型和技术在客户的业务中发挥作用，并深入了解那些业务流程。

So the value isn't just access to the model, it's making the model and the technology work in the customer's business and understanding in-depth those business workflows.

我们之前讨论过AI采用者这个主题，要让这个交易成功，我们需要的是AI带来的可量化的好处。

And we've talked about this theme of AI adopters and really what we need to get this trade to work are those quantifiable benefits from AI.

所以开放权重是其中的关键部分。

So open weights are a key part of that.

也许再深入一个话题，你刚才也提到了，模型本身越来越成为基础条件，而不是护城河，价值正在转移到模型周围的一切，无论是你的数据、工作流程，还是让AI变得有用的工具。

Maybe to pull on another thread and you kind of alluded to this, the models themselves are increasingly becoming table stakes rather than moats and the value is shifting to everything around the model, whether that's your data, your workflows, the tools that make AI useful.

那你在Spark哪里最清楚地看到了这一点？

So where did you see this most clearly at Spark?

这是否改变了你对软件领域持久价值将如何积累的看法？

And does it change your view on where durable value will accrue in software?

所以，我想再次回到开放权重模型日益重要的话题上。

So again, I mean, I think I'll come back to the growing importance of open weight models.

这是会议上一个反复出现的主题，几乎每家公司都跟我们谈到了实施。

I mean, it was a recurring theme at the conference, almost every company spoke to us and spoke to, you know, the implementation.

这再次强化了多模型世界的观点。

And again, it reinforces that view of a multi model world.

推理公司都在谈论降低使用模型的成本。

You've got the inference companies talking about bringing down the costs of using the models.

还是我们一直在说的那个扩散概念。

Again, that diffusion concept we talk about.

在应用层，有一家法律软件公司说，他们用200个标准来评估18个模型，以决定哪个模型来执行任务。

The application layer, a legal software company said it benchmarks 18 models across 200 criteria to decide which model should run the task.

所以你在效率上获得了巨大的提升。

So you're getting incredible gains in efficiency.

我觉得这对公共软件供应商有两个影响。

I think there are two implications for public software vendors.

首先，软件公司有机会帮助客户进行编排。

Think, you know, first of all, there's an opportunity here for software companies to help their customers with orchestration.

所以会有一个层面，就是如何把正确的任务分配给正确的模型，以合适的价格。

So there's going to be a layer around how do we get the right task to the right model at the right price.

而且我认为很多公共软件公司，在超大规模云服务商这个层面，他们可以在更广泛的企业采用层面提供帮助。

And I think a lot of the public software companies, you know, at the hyperscaler level, they can help at a broader enterprise adoption level.

然后在应用层，不管是CRM、ERP还是HR，软件公司也有机会在那个层面帮助进行编排。

And then down at the application layer, whether we're talking about CRM, ERP, HR, there's an opportunity for software companies to help with orchestration at that layer as well.

第二点，可能讨论得比较少的是，这里或许存在一个毛利率的机会。

The second thing, which is probably a little bit less talked about is maybe there is a gross margin opportunity here.

我觉得在软件行业，我们到目前为止讨论的都是毛利率面临的压力，因为必须把运行token的成本纳入软件商业模式，并作为套餐的一部分提供给客户。

I think in software, all we've discussed so far is the pressure that's going to come on gross margins from having to include the cost of running tokens in a software business model and offer that to customers as part of the package.

如果我们能大幅降低推理成本，并且作为软件公司能维持一些定价，那么中期来看，可能有一个少被提及的看涨理由，那就是软件毛利率可能不会像人们想的那么受压，我们或许能看到毛利率比人们预期的要好得多。

If we can bring down inference costs dramatically, and you can hold on to some of the pricing as a software company, Maybe there could be a little talked about bull case in the medium term that maybe software margins are not going to be as pressured as people think, and we can maybe see gross margins hold in a lot better than people think.

Brian，我想以一个可能看似反直觉的观点来结束。

Brian, I want to end with something that may seem counterintuitive.

在一个软件越来越容易构建的世界里，一些最强的竞争优势可能实际上在物理世界中。

In a world where software is becoming easier to build, some of the strongest competitive advantages may actually be in the physical world.

这次大会有很多公司，无论是物流、制造业还是现实世界的运营。

The conference had many companies, whether that was logistics, manufacturing, real world operations.

为什么数字世界和物理世界的交汇如此具有防御性？这如何影响你对在智能体世界中最终谁胜出的看法？

You know, why is that intersection of the digital and physical world so defensible and how does it influence your view of who ultimately wins in an agentic world?

我觉得人们需要记住，AI不仅仅是软件。

I think it's important for people to remember that AI is not just software.

人工智能将通过新型机器人走向实体世界，影响全球。

AI is going to get physical through new types of robots that impact the entire globe.

我们在多个行业已经看到了这一点。

We're seeing in multiple industries already.

在国防领域，我们看到无人机和其他类型的机制正在改变我们构建整体国防工具的方式。

We're seeing it on the defense side with drones and other types of mechanisms that are to change the way that we build out our overall defense tools.

在工业领域，我们看到仓库和配送方面的变化。

We're seeing on the industrial side, in warehouses, on delivery.

在农业领域，我们看到未来农业的发展方式。

We're seeing on the agriculture side, the way that we're going to the future of farming.

所以从长远来看，虽然我认为会有仿人机器人和家用机器人，但重要的是人们要记住，在未来两三年、五年内，会有很多实体和机器人形式的人工智能应用，真正改变我们做很多业务的方式。

And so over the long term, while I think we will have humanoids and we will have robots in our homes, I think it's important for people to keep in mind there are going to be a lot of physical and robotic implementations of AI over the next two, three, five years that are going to really change the way we do a lot of business.

完全同意。

Absolutely.

有很多令人兴奋的事情正在发生。

Lots of exciting things happening.

亚当、布莱恩，谢谢你们来参加，也谢谢大家的收听。

Adam, Brian, thanks for joining me and thank you for listening.

如果你喜欢关于市场的见解，请在你收听的地方给我们留个评论，并今天就把这个播客分享给朋友或同事。

If you enjoy thoughts on the market, please leave us a review wherever you listen and share the podcast with a friend or colleague today.

以下内容仅供参考，基于创建时可获得的信息。

The proceeding content is informational only and based on information available when created.

这不是要约或招揽，也不是税务或法律建议。

It is not an offer or solicitation, nor is it tax or legal advice.

它没有考虑你的财务状况和目标，可能不适合你。

It does not consider your financial circumstances and objectives and may not be suitable for you.