微软CEO推动Copilot，Cursor收购Graphite，Waymo与人形机器人面临的挑战 | 2025年12月22日

欢迎各位收看Information的TI TV。

Welcome everyone to the Information's TI TV.

我叫阿卡什·马斯里查。

My name is Akash Masricha.

今天是12月22日，星期一。

It is Monday, December 22.

距离圣诞节还有三天，本周我们为大家准备了三场精彩的节目。

We are three days away from Christmas, and we've got three exciting shows lined up for you this week.

今天，我们将为您带来新消息：微软CEO萨提亚·纳德拉已成为公司最具影响力的产品经理，并深度参与一线工作，与员工共同推动Copilot的改进。

Today, we've got new details for you on how Microsoft CEO Satya Nadella has become the company's most influential product manager and is very much in the trenches with employees trying to accelerate copilot improvements.

接下来，我将采访代码审查初创公司Graphite的CEO，该公司刚刚被Kershur收购。

I'll speak then with the CEO of code review startup Graphite, which was just acquired by Kershur.

随后，我们将邀请Madrona Ventures分享对2025年软件行业的看法，更重要的是，展望2026年的趋势。

We're then bringing on Madrona Ventures for a view of the software sector in 2025, and more importantly, what it looks like going into 2026.

最后，我们将以一段关于机器人监管的讨论结束本期节目，并单独分析福特决定削减部分电动汽车生产的决策。

And we will wrap the show with a segment on robotics regulation, and separately, some analysis on Ford's decision to do away with some of its electric vehicles production.

这将是一场精彩的节目，让我们马上进入正题。

It's going to be a fun show, so let's get right on into things.

几个月前，微软进行了一次结构重组，萨提亚·纳德拉卸下了一些管理职责，以便更专注于产品、研究以及数据中心建设等重大举措。

A few months ago, Microsoft underwent a structural change that saw Satya Nadella give up some management responsibilities so that he could focus more closely on product, research, and big initiatives like data center build outs.

我的同事亚伦·霍姆斯今天发表了一篇深入报道，揭示了纳德拉如何与一线团队紧密合作。

A new in-depth story from my colleague, Aaron Holmes, today reveals just how closely Nandela has been working with teams on the ground.

我想邀请亚伦来分享他在采访中了解到的内容。

I want to bring on Aaron to share what he's learned in his reporting.

亚伦，欢迎再次做客我们的节目。

Aaron, welcome back to the show.

很高兴你能来。

It's great to have you.

谢谢邀请我。

Thanks for having me.

萨提亚·纳德拉最近都是怎么安排时间的？

So how is Satya Nadella spending his time these days?

是的。

Yeah.

所以，正如我们今天报道的那样，萨提亚·纳德拉实际上已经深入参与到Copilot及其他AI产品的产品管理中。

So, I mean, as we reported in today's story, Satya Nadella has really been, I guess, diving into the weeds on product management of copilot and other AI products.

具体来说，这表现为他越来越积极地参与那些专注于提升Copilot产品的内部团队频道。

And, you know, specifically, that looks like him being increasingly active in internal teams channels that are focused on making copilot a better product.

我们听说，他一直在给负责Copilot的工程师发送详细邮件，指出当前产品中存在的具体问题。

You know, we've heard that he's been sending detailed emails to engineers who work on copilot pointing out specific flaws that he sees in the product right now.

此外，他还定期召开每周会议，邀请公司内所有参与AI开发的人员参加，有时会深入追问Copilot的进展情况。

And he's also been, you know, having these weekly meetings where he brings in people at all levels of the company involved in AI development, and will sometimes grill them on, know, what exactly is going on with copilot.

我认为，我们正看到这种极高程度的关注，因为他认为现在是Copilot以及微软整体发展的关键时期，并且他明确希望投入时间来塑造公司应对这一时刻的方式。

And I think we're seeing this really kind of high level of attention, because he feels that right now is a pivotal moment for copilot, and also for Microsoft more broadly, and has shown that he really wants to devote his time to shaping how the company will approach that moment.

那么，这与六个月前有什么不同呢？

So how is this different from, say, six months ago?

六个月前，他在Teams频道里也这么活跃吗？

Was he as active in the Teams channels, I guess, six months ago?

你感觉他因为过去几个月职责上的变化而变得更加积极了吗？

Your sense that he's become more active given this change in responsibility that we've seen over the past few months?

是的。

Yeah.

我的意思是，正如你提到的，他把一些职责交给了贾森·阿尔托夫，后者今年早些时候被任命为商业首席执行官，主要负责微软产品的销售，这给了萨提亚更多时间专注于高远的技术工作。

I mean, so like you mentioned, he handed off some of his responsibilities to Judson altoff, who earlier this year was named commercial CEO, and basically is in charge of selling Microsoft's products, which gives, you know, Satya a bit more time to focus on the high ambition technical work.

举个例子，过去十年里，萨提亚每年都是微软Ignite大会——也就是微软年度大会——的主旨演讲人。

And just as one example, I mean, we've seen Satya be the keynote speaker at Microsoft's Ignite conference, its annual conference every year for basically the last ten years.

但今年，他实际上并没有出席这次大会。

But this year, he actually wasn't present at the conference.

他让贾森·阿尔托夫来发表主旨演讲。

He let, you know, Johnson Altoff handle the the keynote speech.

我认为这只是一个例子，说明他现在更谨慎地分配自己的时间，把更多精力集中在打造和优化微软所销售的AI产品上。

And I think that's just kind of one example of how he's being more selective about how he spends his time and devoting more of his attention to specifically building and refining the AI products that Microsoft is selling.

从外部来看，这看起来是这样，我真希望有一种方式能量化这一点，也许我们真该量化一下。

Now, from the outside, what it looks like, and I wish there was a way to quantify this, maybe we should quantify it.

我一直在追踪萨提亚·纳德拉今年参加播客的次数，看起来他确实像在进行一场——我不想说‘新书巡演’，但他几乎在每个播客上都在宣传他的书，除了你这档节目，我跟你说，他非来不可。

I've been tracking the number of podcast appearances that Satya Nadella has been doing over the course of the year, And it seems like he's been on a bit of a, I don't want to say, you know, a book tour, but he's really been talking his book on every podcast, except for the he's got to come on this show, I'm telling you.

我的意思是，我们得让他上这个节目。

I mean, we've to get him on here.

但他的确到处跑，非常活跃。

But I mean, he's really out there.

这是否是他策略的一部分？你知道的，我想知道他到底想干什么，通过这些对话来了解人们的想法？

Is this part of his, you know, strategy here to sort of, you know, I don't know what it is, finding out what people are thinking about in these conversations?

他是在试图传播自己的观点吗？

Is he trying to get his own message out?

你怎么看这件事？

What do you make of this?

是的，我确实也注意到了同样的情况。

Yeah, I mean, I've definitely noticed the same.

我认为我们看到的是，萨提亚对整个软件行业面临的一些关键问题有着非常明确的看法。

I think that what we are seeing is that, you know, Satya has pretty strong opinions on some of the key questions that are facing the entire software industry.

比如，AI 什么时候、在哪些领域会对企业真正产生价值？

Things like, you know, when and where will AI prove really valuable for enterprises?

为了提升 AI 模型，数据中心的扩展将如何发展？AI 软件普及的受益者会是谁？而微软恰好处于所有这些问题的交汇点上。

What what's going to happen with the scaling up of data centers to improve AI models, who are going to be the beneficiaries of an uptake in AI software, and Microsoft is kind of uniquely situated at the intersection of all of those questions.

我认为，我们看到的是，萨提亚希望积极塑造科技界正在发生的讨论。

And I think that what we're seeing is Satya wants to be out there really kind of like shaping the the conversation that is happening across tech.

但同时，我认为这也反映了他坚定地致力于确保微软在这个关键时刻取得胜利。

But also, I think it reflects just, you know, how locked in he is on making sure that Microsoft can win in this moment.

那么，微软赢了吗？

So is Microsoft winning?

这是一个复杂的问题。

You know, it it's a complex question.

Copilot 本身已经有大约一亿月活跃用户，而 ChatGPT 的月活跃用户接近九亿，相比之下这只是很小的一部分。

I mean, Copilot itself has about a 100,000,000 monthly active users, which is a tiny fraction of the almost 900,000,000 monthly active users that chat GPT has.

谷歌的 Gemini 也有更多的用户。

Google also has, you know, many more users of Gemini.

同时，我们并不清楚微软从销售AI软件中究竟赚了多少收入。

And at the same time, you know, we we don't know exactly how much revenue Microsoft is making from selling AI software.

但它无疑通过像OpenAI这样的AI客户获得了大量收入，这些客户使用其Azure云服务来训练和运行模型。

But it certainly is making a lot of revenue from AI customers like OpenAI that are using its Azure cloud services to, you know, train and run their models.

由于AI的发展，云服务器的租赁量一直在激增。

Rentals of cloud servers have been spiking thanks to AI.

所以我认为，微软得益于其多元化的业务，能够在许多领域展现出胜利态势，即使其软件进展可能仍有些不明朗。

So I think Microsoft kind of has the benefit of having this diverse business that can show that it's winning in a lot of areas, even if maybe it's it's, you know, software progress is still a little bit unknown.

那编程呢？

And what about coding?

是的，这也是一个有趣的话题。

Yeah, that's another interesting one.

我的意思是，GitHub Copilot基本上是市场上第一个AI编程工具。

I mean, GitHub Copilot was basically the first AI coding tool on the market.

第三方估计显示，截至一年前，它几乎垄断了整个AI编程市场。

And third party estimates showed that as of a year ago, it kind of had the entire AI coding market unlock.

但自那以后，像Cursor和Anthropic的Claude Code这样的竞争对手一直在抢占市场份额，这一点实际上得到了萨提亚的承认，他表示他欢迎竞争，并且仍然认为GitHub Copilot整体上仍处于领先地位，尽管它正在失去一些市场份额。

But since then, you know, competitors like cursor, and anthropics Claude code have been eating up market share, which is actually something that, you know, Satya has basically acknowledged, saying that he loves the competition, and that he still thinks that GitHub Copilot overall is on top even if it is losing some market share.

但我认为，这又是另一个我们看到他希望塑造GitHub Copilot的领域，让他成为无论用户使用哪个模型生成AI代码都会想要使用的工具。

But I think that's another space where we're seeing that, you know, he wants to shape GitHub Copilot and make it this tool that people are going to want to use regardless of which model that they're using to generate AI code.

这恐怕是公司面临激烈竞争的另一个领域。

And that's, I guess, another area where the company is seeing really tough competition.

很好。

Great.

好了，亚伦，这是一个精彩的故事，让我们深入了解了他是如何在一线管理事务的。

Well, Aaron, it was a great story and a fascinating look at how he is managing things on the ground.

感谢你今天来做客。

I want to thank you for coming on.

这位是《信息》杂志的微软记者亚伦·霍姆斯。

That is Aaron Holmes, our Microsoft reporter here at The Information.

好的。

Okay.

Cursor 又完成了一项收购。

Cursor has made another acquisition.

该公司收购了代码审查公司 Graphite，其投资者包括 Excel、Menlo Ventures 和 Andreessen Horowitz。

The company has bought code review company Graphite, which counted Excel, Menlo Ventures, and Andreessen Horowitz among its investors.

Kerser 近期已进行多项收购，但我希望能邀请到 Graphite 的联合创始人兼首席执行官 Merrill Lutzky，谈谈他决定出售公司的原因。

Kerser has made a number of acquisitions as of late, but I want to bring on the co founder and CEO of Graphite to talk a bit about his decision to sell the company, Merrill Lutzky.

欢迎来到 TI TV。

Welcome to TI TV.

很高兴你来到这里。

It's great to have you here.

太棒了。

Awesome.

谢谢你邀请我，Akash。

Thanks for having me, Akash.

恭喜你完成收购。

Well, congrats on the acquisition.

这算是为2025年画上一个重磅句号了。

Is this is a big news to end off 2025.

谢谢。

Thank you.

是的。

Yeah.

我们想在科技新闻周期结束时，再轰动一下。

We wanted to end the the tech news cycle with one last one last bang here.

是的。

Yeah.

很高兴你能来，尤其是我们六个月前刚推出这个节目时就邀请过你，当时我们聊过一点Graphite与编程平台的区别。

Well, it's exciting to have you on here, certainly because we had you on when we first launched this show about six months ago, and we talked a little bit about what makes Graphite different from the coding platforms.

但你能说说，是什么原因让你决定出售公司吗？

But tell me a little bit about the decision as to why you decided to sell the company.

是的。

Yeah.

嗯。

Yeah.

不。

No.

这是一段有趣的旅程，因为我们最初并不是为了追求业绩，但我们的指标却一直表现强劲。

It's it's it was an interesting journey because we we really came into it not with the intention to metrics have been as strong as ever.

上个季度我们增加的收入比2024年全年还要多。

We added more revenue last quarter than we did in all of 2024.

就像你说的，一切都在顺利进行，公司历史上从未有过这么好的状态。

Like, you know, everything everything was going, you know, about as as as well as it ever has in in the company's history.

所以这个决定其实源于一次国家级的合作——今年夏天，我们开始与Cursor团队讨论将他们的后台代理集成到Graphite中。

So it really came out of, what felt like a national partnership where back in the summer, we started talking with the Cursor team about integrating their background agents into Graphite.

让我们的用户可以直接从Graphite的代码审查界面启动Cursor的后台代理。

So our users could launch Cursor background agents, from Graphite's code review surface.

这样，你就可以在一个地方完成创建、审查和合并代码变更。

So, you know, you can kind of create review and merge code changes all from one place.

从那时起，我觉得我们越聊，就越发现我们对软件开发未来的愿景有着非常共同的看法。

And from there, I think the the more that we got to talking, the more we realized that we shared a really common vision for where the future of software development is going.

我们在构建公司的方式上有很多相似之处。

We think about building our company in a lot of the same ways.

你知道，我们俩都是每周五天线下办公。

You know, both of us are five days a week in person.

我们都非常重视高人才密度、高工艺水平，我们在很多理念和方法上都有很多共同点。

We both really value like high talent density, high craft, you know, a lot of a lot of the same, way that approaches that that we have, we we shared in common.

对。

Right.

这感觉像是一次非常自然的合作。

It felt like a really it felt like a really natural partnership.

我认为我们都看到了结合两个平台最佳优势的可能性，创造出一种非凡的、端到端的AI软件开发体验：从AI驱动的IDE，到拉取请求和代码审查界面，再到测试和合并队列，所有这些都连成一个完整、连贯的体验。

And I think we both saw the ability to, combine the best of of both platforms and create something really phenomenal, which is this, like, end to end experience, for building software with AI where you go from, you know, an AI powered IDE, into, you know, the pull request and the code review surface, into, like, testing and the merge queue, and that's just, like, all one connected cohesive experience.

现在，通过将Graphite和Cursor结合，我们将在未来几个月内实现这一目标。

And now with combining Graphite and Cursor, that's that's exactly what we'll be able to do in the coming months.

六个月前你上节目时说过一件事，我当时问过你——昨晚我重看了录像——我问的问题是：为什么这个代码审查功能不能被像Cursor这样的工具直接添加并跟你竞争？

One of the things that you had said when you came on the show about six months ago is I had asked, and I watched the tape last night actually, I asked the question, Why isn't this code review feature something that something like a cursor could just add on and compete with you?

你当时提到的观点是，让狐狸看守鸡舍并不总是好事，由一个第三方来评估和审查来自各种编码代理（比如Claude或Kerser）的代码反而更有益。

And the point that you made was you said, Well, it's not always great to have the fox watching the henhouse, and it's actually beneficial to have a third party that is assessing and reviewing the code from all these different coding agents, whether it's Claude or Kerser.

现在你已经成为Kershur的一部分了。

Now you're part of Kershur now.

那么，这种无偏见的审查会怎样呢？

So, I mean, what happens to that unbiased review?

你们还会继续审查来自所有这些不同代理的代码吗？

Are you still going be reviewing code from all these different agents?

还是说这个故事发生了变化？

Or how does the story change?

是的。

Yeah.

当然。

Absolutely.

我认为需要强调的关键一点是，我们将继续将 Graphite 作为独立产品运行。

I I think one of the key one of the key things to emphasize about this is that, we're we're running continuing to run Graphite, as an independent product through this.

我们仍在构建与 QuadCode、Codecs 等工具的后台代理集成。

We're still actually building background agent integrations with tools like QuadCode and Codecs and, others.

我们希望在开发者所在的地方与他们相遇，与团队正在使用的各种工具合作，以如今的 AI 方式进行开发。

Like, we wanna meet developers where they are and meet and and work with all the tools that the teams are using, in order to in order to build with AI now.

当然，有些功能会与 Cursor 的工具更紧密地集成，但我们非常希望保持 Graphite 的兼容性，让它能与所有现有工具协同工作。

So, we're not gonna be obviously, there's some things that will be, more tightly integrated with with Cursor's tools, but, we very much want to keep this as, you know, as something that works with every single tool out there.

比如，有很多用户在 Cursor 中使用 QuadCode。

Like, there are a lot of there are a lot of users that, that run quad code in Cursor, for instance.

我认为，我们确实会继续对支持的工具保持高度灵活。

Like, I think that there's, yeah, I think we're we will continue to be really flexible about the the tools that we support.

甚至从模型角度来看，你也会看到，用于生成我们审查评论的模型和流程，与用于在 Cursor IDE 中生成代码的模型和流程并不相同。

And even from a a model standpoint, I think you'll see the, the models and the pipelines that go into, generating our review comments, are not gonna be the same as as the ones that go into, generating code in in the cursor IDE for instance.

所以我认为，只要继续将这两者视为独立实体，并确保我们构建的审查体验具有灵活性，能够支持工程师如今使用 AI 编码的各种方式即可。

So I think as long as as long as you continue to approach, those those two as separate entities and, and make sure that the the review experience that we build, is is flexible and supports all the different ways that that engineers are are writing code with AI now.

我认为，今后我们会继续避免这个问题。

I think that you'll you know, we'll continue to avoid that problem going forward.

所以，你仍然会全职投入 Graphite，把它作为一个独立的产品和某种程度上独立的业务来运营，与……

So so you're you're gonna be still working full time on Graphite as sort of an independent product and an independent business in in in some ways from

没错。

Exactly.

Cursor？

Cursor?

是的。

Yeah.

没错。

Exactly.

我们清楚地强调过，这笔交易的一个关键点是：Graphite 将作为独立产品继续存在。

We we see you know, we one of the things that that we've we've emphasized with with this deal is that graphite's gonna continue independently as a product.

整个 Graphite 团队将继续致力于 Graphite 的开发。

The entire graphite team is is continuing to work on graphite as part of this.

我们真的很想确保保留Graphite最好的那些方面。

We really wanna make sure that we we keep, you know, the best aspects of graphite.

另外需要注意的是，有些人一直有点担心，比如Graphite会不会被关闭，或者会不会完全转向AI方向？

And another another, another thing to to note there is that there's, you know, some folks have have been worried a little bit about, like, oh, is Graphite going to get shut down or is Graphite gonna focus entirely on on the AI side of things?

Cursor团队和我们讨论过，Graphite的价值其实体现在许多接口、基础设施和功能上，比如我们的合并队列和堆栈TR，这些功能并不完全是AI原生的，但却能极大加速代码审查流程。

And one thing that that the Cursor team and and we've talked about is that really the value of Graphite is in a lot of the the interfaces and infrastructure and, features like our merge queue and stack TRs that aren't necessarily as AI native, but help to accelerate the code review process dramatically.

由于这次收购合作，我们只会进一步加强这些功能。

And we're we're only going to double down on those as a result of of this acquisition partnership.

我认为他们也真正认识到Graphite的价值，它是一个战略资产，像Shopify、Snowflake、Robinhood、Datadog这样的团队每天都依赖它进行代码审查。

And, I think they really see the value of of Graphite as, you know, as like a strategic asset that, teams like Shopify and Snowflake, Robinhood, Datadog, all depend on every single day for for code reviews.

没错。

So Right.

这些都不会消失。

That's not going anywhere.

它将继续作为独立产品存在。

It's it's very much going to continue as an independent product.

这显然并不意味着，整合这两个界面存在一些绝佳的机会。

That's not obviously, there's some amazing opportunities for integrating the two surfaces.

比如，我们曾讨论过将 Cursor IDE 中的聊天历史记录作为额外上下文显示在拉取请求中，让你能无缝接续之前的工作。事实上，我们一直觉得，你在本地用于创建代码更改的代理，却无法与你在审查及其他流程中使用的代理进行沟通，这简直太荒谬了。

Like, we've talked about having the the chat history from the cursor IDE showing up in the pull request as additional context and why don't you just like pick up where you left off and, it really should be like, we we've always thought it was it was silly that the agents that you're working with locally to create a code change can't really talk to the ones that you're then working with in review and the rest of the process.

如果我们能将所有这些服务连接起来，实现无缝衔接，并拥有一个贯穿审查、测试直至上线全过程的一致性代理来指导和协助这一变更，那将是我们能为开发者创造的最理想的体验。

And if we can just connect all of those those services and make that that super seamless and have, you know, a consistent agent that's guiding that's helping to guide that change, you know, through, through review, through testing, out to production, that's really the the best possible developer experience that we can create here.

让我问你一个问题。

Let let me ask you a question here.

Kerser 是以多少钱收购这家公司的？

How much did Kerser buy the company for?

我们不会透露这笔交易的经济细节，但我可以这么说，我们的所有投资者和团队都对这一结果感到非常兴奋。

So we're not sharing details of of the economics, of the deal, but, I mean, I will say that all of our investors, and our team are thrilled about the outcome.

这是一次极具战略意义的收购。

It's a it's a very strategic acquisition.

对我们来说，这是一个绝佳的结果。

It's a fantastic outcome for, for us.

我认为这里最重要的是，我们所有人都对这里的巨大潜力感到非常兴奋。

And, I think the main thing to emphasize here is that, there's so much I think we're all we're all really, really excited about the upside potential here.

一旦你将这两家公司结合起来，我们能共同创造出什么？

And like, you know, once you combine these two, you know, these two companies, what can we create together?

那么，这个未来的代码审查平台会是什么样子？我们将本地开发、后台引擎、拉取请求全都连接起来。

Like, what does that that future code review platform look like, where we've connected, everything from local development, background engines, pull requests.

我认为，这种尚未释放的价值是极其非凡的，这里每个人都非常期待在未来几年内将其实现。

Like, I think the the the value of that that is left to be unlocked, is is really, you know, phenomenal and something that everybody here is is really, really excited about unlocking in the coming years.

太好了。

Great.

恭喜你们完成这项收购。

Well, congrats on the acquisition.

正如我所说，这对你们团队来说是结束2025年的绝佳消息，我也期待看到明年产品如何继续演进。

Like I said, it's some great news for your team to end 2025, and I'm looking forward to seeing how the product continues to evolve next year.

这里是Graphite的联合创始人兼首席执行官梅瑞尔·卢茨基，为您带来TI TV的报道。

That is Merrill Lutzky, the cofounder and CEO of Graphite here on TI TV.

好的。

Okay.

今年，软件行业经历了巨大的变革，每个人都试图将产品套件转向人工智能。

The software sector has undergone tremendous transformation this year, with everyone trying to gear their product suites towards AI.

玛德罗纳风险投资公司凭借其涵盖各种规模、阶段和子领域的公司组合，亲历了这一切。

Madrona Ventures had a front seat to it all with a portfolio of companies of every size, stage, and subsector.

该公司还在生物技术领域进行了大量投资，我希望能深入探讨这一领域。

The company also does a ton of investing in the biotech sector, which I'm hoping to get into.

我想邀请玛德罗纳的董事总经理马特·麦克莱恩，来聊聊这些话题。

I want to bring on Matt McElwain, Managing Director at Madrona, for a conversation about all that.

马特，欢迎来到TI TV。

Matt, welcome to TI TV.

很高兴你来到这里。

It's great to have you here.

谢谢邀请我参加节目，阿卡什。

Thanks for having me on the show, Akash.

所以我想聊聊当前软件投资的大致情况。

So I want to talk a little bit about software investing broadly right now.

我们这里抱着预测的心态，我知道我们正在预测2026年的情况。

And we're in the spirit of predictions here, and I know we're making predictions for 2026.

我一直在思考未来几年软件市场的结构会发生怎样的变化。

I've sort of been thinking about how the structure of the software market is going to change over the next couple years.

关于上市公司，我有一个想法，想跟你探讨一下，听听你的看法。

And, you know, as it relates to the public companies, the working theory I have, and I want to run this by you, you tell me what you think of it.

我们现在大约有八九十家顶级的SaaS上市公司。

You know, we've got these, I don't know, eighty, ninety top public SaaS companies right now.

它们已经在市场上存在一段时间了。

They've been on the market for a while.

我认为在未来三年里，我们可能会看到这前90家公司的数量略有减少，因为一些公司会私有化。

I think over the next three years, you know, what we might see is the number of those top 90 companies, it might shrink a bit because some companies will go private.

还会出现一些整合，因为许多公司开始互相抢占市场。

There'll be some consolidation because a lot of these companies are starting to sort of step on each other's toes.

然后你会看到，比如说，数量从90减少到60左右，对吧？

And then you'll kind of, you know, let's say it goes from 90 to, I don't know, 60, right?

另外30家将是那些随着成长而上市的顶级AI初创公司。

The other 30 will be these top AI startups that start to go public as they grow.

我认为最终留下的会是一些庞大而多元化的软件企业。

And I think what you might be left with is some of these big, chunky software businesses that do a lot of different things.

它们收购了一些自己原本不知道需要的东西。

They buy some things they didn't know they needed.

然后你还会有一批增长更快的公司群体。

And then you have the faster growing cohort.

从你的角度来看，这是软件投资者普遍如何看待这个趋势的吗？

Is that kind of the way software investors see it broadly from your perspective?

不是。

No.

我认为，我们正处在一个颠覆的时代，同时也在质疑任何企业的持久性。

Look, I think that we are in both a time of disruption and we're in a time of questioning the durability of any business.

这一点对于这些大型上市SaaS公司以及那些看似快速增长的私营公司都成立。

And so that's both true about these larger public SaaS companies, as well as it's true about the seemingly fast growing private companies.

为什么会这样呢？

And why is that?

这是因为整个技术栈的每一层都在发生变化。

Well, it's because every layer of the stack is changing.

用户界面正在变化。

The user interfaces are changing.

数据模型正在变化。

The data models are changing.

工作流程正在变化。

The workflows are changing.

我们正从记录系统转向智能系统。

We're going from systems of record to systems of intelligence.

这意味著什么？

What does that mean?

这意味着我们既要处理结构化数据，也要处理非结构化数据。

Well, it means that we're dealing with both structured data and unstructured data.

结构化的工作流程和非结构化的工作流程。

Structured workflows and unstructured workflows.

因此，如果你是Salesforce、ServiceNow或Atlassian，你就必须非常灵活。

So you've got to be pretty nimble if you are Salesforce or ServiceNow or Atlassian.

我认为，其中一些公司将在这一轮中表现卓越，而另一些则不会那么突出。

I think that some of those will excel in this cycle and some of those will not as much.

私营公司也是如此。

And the same thing can be true of private companies.

我们发布了这份智能应用40强榜单。

We do this intelligent application 40 list.

我们已经连续五年做这个榜单了。

We've been doing it for five years.

这是40家顶尖的私营人工智能公司，重点在于应用型人工智能。

It's the 40 top private AI companies with an emphasis on applied AI.

有趣的是，除了名列这十大赋能者之一的Databricks之外，没有一家公司能连续五年出现在这份榜单上。

And what's interesting is that no company except Databricks, which is one of the 10 enablers in that list, has been on the list all five years.

这表明，即使在私营公司内部，持久性也是一个挑战。

So that tells you that even within private companies, there's a challenge in durability.

最后我想说的是，我认为我们将会看到一个2026年的预测：我们将看到更多私营公司之间的并购。

The last thing I would say is that, well, I think we are going to see, here's a prediction for 2026, we're going to see more private to private mergers.

今年下半年，我们已经看到了这种趋势的上升。

We already saw that uptick in the back half of this year.

比如，我们投资组合中的两家公司——Clari并购了SalesLoft，Statsig被OpenAI收购，都是私营公司之间的交易。

Two of the companies in our portfolio, for example, Clari merging with SalesLoft and then Statsig being bought by OpenAI, two private to private companies.

我认为，基于你的观点，不是所有公司都能存活下来，而界定赢家的一种方式就是通过并购，包括私营公司之间的并购，以及上市公司收购私营公司的交易。

I think you're going to see, because of the thesis you have, that not all are going to survive, that one of the ways you're going to define the winners is by mergers, both private to private mergers and then public to private mergers.

对。

Right.

我想问你一个问题，如果我们暂时跳出初创公司这个范畴，你是否可以把一些这些数十亿美元的公司称为初创公司？

I want to ask you if we sort of take a step out of startup land, you know, to the extent that you could call some of these multibillion dollar companies startups.

让我们转向超大规模云服务商的世界。

You know, let's go over to the land of hyperscalers.

你住在西雅图，那里有亚马逊和微软，我相信你一直密切关注着它们的发展。

You are based in Seattle, and we've got Amazon and Microsoft there, and it's a story I'm sure you've been watching closely.

我想知道，你对亚马逊和微软在人工智能竞赛中的当前地位有何看法？

You know, I wonder what your sense is of Amazon and Microsoft's current place in the AI race.

我们一直听到很多关于谷歌的消息，对吧？

We have all this news about Google, right?

今天早上刚发布了一篇关于微软的文章，提到萨提亚·纳德拉正采取深入一线的方式，帮助加速其人工智能的采用。

And we just had this story published this morning about Microsoft, how Satya Nadella is really taking a really boots on the ground approach to helping them accelerate their AI adoption.

我的意思是，你认为这两个超大规模云服务商的故事在2026年会如何变化？

I mean, how do you think the story around those two hyperscurs particularly is going to change in 2026?

我们确实了解这两家公司，而且有趣的是，回想一年前，人们对谷歌还充满怀疑。

Well, we do know both those companies, and it is also interesting to think of a year ago, there was a lot of skepticism about Google.

完全正确。

Totally.

现在对亚马逊和AWS，尤其是微软，也有差不多这么多的怀疑。

There's nearly that much skepticism right now about Amazon and AWS in particular, or Microsoft.

但我认为它们目前都不太受欢迎。

But I think they're both a little out of favor right now.

但让我们把两件事分开来看。

But let's disaggregate two things.

我非常确信的是，它们将继续在云服务提供商层面、基础设施即服务层面表现出色。

One thing I'm highly confident they're going to continue to excel in is in the cloud service provider layer, the infrastructure as a service layer.

而这一点正是AI带来的巨大顺风。

And this is where there's a massive tailwind that AI is pulling.

如果你和马特·加曼聊聊，或者和萨提亚聊聊，他们很可能会告诉你，由于AI，越来越多的工作负载正涌入云端。

If you talk to Matt Garman, if you talk to Satya, what they're going to likely tell you is because of AI, we're getting way more workloads coming into the cloud.

因此，你会看到AWS的增速重新加快，例如Azure继续保持30%以上的强劲增长。

And that's why you see reaccelerating growth at AWS, for example, and continued strong 30 plus percent growth at Azure.

但战略问题在于，作为微软或AWS，我能否不仅将基础设施，还将AI模型融入工作流程、代理式工作流程和上层的智能应用中？

The strategic question though is, can I infuse, if I'm Microsoft or for that matter AWS, can I infuse not just the infrastructure, but the AI models into workflows, into agentic workflows and intelligent applications up the stack?

微软在所有Copilot 365方面拥有明显的优势地位。

Microsoft has a clear advantage position here with all the Copilot three sixty five.

如果你仔细听了马特在re:Invent大会上的主题演讲，他会重新聚焦，将讨论从前沿模型转向前沿代理。

AWS, if you listened carefully to Matt's speech at the re:Invent, his keynote, he was re shifting, moving the debate between frontier models to frontier agents.

逻辑上，他们将前沿代理应用在安全、DevOps和编程等领域，但还能在堆栈上走多远，以在应用层和代理层创造并捕获价值？

Now logically they took frontier agents that were in areas like security and DevOps and coding, but how much further can they go up the stack to create and capture value at the applied level and in the agentic level?

我认为微软也面临同样的问题。

And I think the same question exists for Microsoft.

我认为他们俩都会做得不错，但做得更好的那一家将捕获更多市场份额。

I think they both will do a good job personally, but the one that does the better job is going to capture more of that pie.

上周我们请来了我们的AI记者。

And we had our AI reporter on last week.

我们在做预测。

Were doing predictions.

她的一项预测是，AWS可能会收购一家主要的AI实验室，以帮助启动这一增长。

One of the predictions she had is that AWS might seek to acquire a major AI lab to sort of help kickstart that growth.

正如你所说，他们已经落入了一个被人认为目前没有最佳模型的类别。

Like you said, they've sort of fallen into this bucket where people don't think of them as having the best models right now.

你觉得2026年会发生这种情况吗？

Could you see that happening in 2026?

我认为这不太可能。

I think that's unlikely.

不太可能。

Unlikely.

不太可能。

Unlikely.

你知道，他们正在强调AWS一贯重视的方面。

You know, they are emphasizing the classic things that AWS emphasizes.

价格、便利性和选择性。

Price, convenience, selection.

因此，这里的‘选择性’体现在Bedrock上，也就是你可以使用任何模型，包括我们的。

So selection in this case comes in the form of Bedrock, which is their, you know, use any model you want, including ours.

他们有自己的Nova模型，包括与Anthropic非常紧密的合作关系。

They have their Nova models, including the very close relationship they have with Anthropic.

现在看来，他们将与OpenAI建立越来越好的合作关系。

And now it would appear that they're going to have better and better treatment, you know, in partnership with OpenAI.

因此，我认为他们在选择性方面已经做得相当不错了。

So that's an area that I think they're doing quite well already on selection.

他们始终致力于降低成本，这就是为什么他们会推出Sutranium芯片和Nova模型之类的产品。

They're always going to be driving for being the lowest cost, and that's why they come out with things like Sutranium chips, things like the Nova models.

所以我认为他们在这一点上已经有了一个相当不错的策略。

So I think they've got a pretty good playbook there.

我不认为他们需要这样做，而且我认为他们收购一家大型模型公司的可能性极低。

I don't think they need to, and I think it would be highly unlikely that they would buy one of the big model companies.

另一方面，他们一直面临困境。

Now, on the other hand, they have struggled consistently.

虽然像Connect服务这样的例子是个例外，但他们在应用层一直表现不佳。

There are outlier examples like their Connect service, but they have struggled up at the application layer.

我可以想象他们会收购更多类似Notion这样的横向应用型公司，随便举个例子。

And I could see them buying more of a horizontal application style of company, like let's say a Notion, just to pick an example.

对。

Right.

或者某种横向的工具，能帮你连接历史上作为记录系统的各个应用之间的流程和工作流。

Or something that's kind of a horizontal, helps you connect the dots of workflows and processes across applications that are historically systems of record.

这基本上就是微软所拥有的应用。

Which is basically what Microsoft I mean, has the applications.

你是在说，AWS并不一定拥有面向消费者或企业的这种应用层吗？

AWS doesn't necessarily have that consumer facing or enterprise facing sort of application layer, you're saying?

没错。

That's correct.

我认为关键在于，应用程序的价值是随着时间积累的。

And I think that's where they Keep in mind that the important thing about the applications is that they accrete in value over time.

所以，如果我有一个创新的界面，比如语音、文本，或者AI原生公司的现代界面，再加上我所使用的模型。

So if I have a innovative interface, some voice, some text, some sort of the modern interfaces of AI native companies, and then I have models I'm using.

我的意思是，在人工智能领域，没有任何一家应用公司只使用单一模型。

I mean, no application company in the AI world uses one model alone.

它们使用来自多个模型提供商的多种模型组合。

They use a mix of models from a mix of model providers.

然后你还会拥有数据，这些数据既包括最初使用的，也包括用户使用你的应用后产生的输出。

And then you've got the data that is both initially used and then the outputs of that data from people using your application.

这形成了一种推理飞轮，你需要应用层来驱动整个推理飞轮。

This becomes a reasoning flywheel, and you need the application layer to drive the total reasoning flywheel.

这就是为什么AWS最初从那些已经具备推理飞轮的领域入手，比如安全和DevOps。

That's why AWS started with things that they have a reasoning flywheel in, like security, like DevOps.

因此，你必须不断向上游移动这个技术栈。

And so you've got to keep moving up that stack.

所以，我知道我们只是随便举了Notion这个例子，但如果你要思考一下，你觉得AWS可能会在这一领域瞄准哪些公司？

So, I know Notion is just a name that we throw out there, but if you were to think of names, area, who could you see AWS going after in that bucket?

我提到Notion的原因是，我曾是Smartsheet的董事会成员，这家公司今年年初被黑石和Vista私有化收购了，我觉得亚马逊收购它会是个不错的选择。

The reason I mentioned Notion was because I was on the board of Smartsheet that got taken private at the beginning of this year by Blackstone and Vista, I thought that that would have been a good one for Amazon to buy.

他们在私有股权方面的这次收购表现得非常出色，但还有一些跨应用程序的横向工作流，我认为这些是非常有潜力的候选对象。

They did just fine with that acquisition on the private equity side, But there's these horizontal workflows that exist across applications that I think are very interesting candidates.

我不认为他们会收购Atlassian。

Now I don't think that they would buy an Atlassian.

我认为那会是难以消化的，ServiceNow也是如此。

I think that's too big of a bite or a ServiceNow for that matter.

但像Airtable、Notion这样的公司，在我看来更有可能属于他们感兴趣的范畴。

But something, Airtable, Notion, these kinds of companies seem to me to be more likely to be in the category of things they'd be interested in.

另一个领域是与身份验证相关的产品。

The one other area would be things that enable identity.

你刚刚看到了ServiceNow被Visa这样的身份公司收购。

You just saw ServiceNow by an identity company in Visa.

因此，像Okta这样的身份层可能是一个值得布局的方向，再次实现跨流程、跨数据系统的 workflows，这些流程可以轻松地依托所有底层的AWS服务运行。

So some of the identity layers possibly in Okta, that would be an interesting place to enable, again, cross process, cross system of record workflows that could be living very easily and leveraging from all of the underlying AWS services.

回到我之前的观点，他们的增长已经重新加速。

Back to my earlier point, they've reaccelerated growth.

我的意思是，他们正在推动业务增长。

I mean, they are growing the business.

每年超过200亿美元。

Well over $20,000,000,000 a year.

很少有软件公司能达到每年200亿美元的收入，而AWS去年的收入也超过了200亿美元。

There's not many software companies that have gotten to $20,000,000,000 in revenue a year, and AWS grew last year over $20,000,000,000

对。

Right.

马特，感谢你前来做客。

Matt, I want to thank you for coming on.

我们还没谈到另一类话题。

There was another category of things that we didn't get to talk.

我想聊聊你在麦德罗纳25年来的观察。

I wanted to talk about your observations from twenty five years now at Madrona.

今天早上我读了你的简历，你是2000年加入的。

I was reading your bio this morning, 2000 is when you started.

所以我会这么说，2026年再回来做节目吧，我们聊聊你在风险投资领域二十五年的一些感悟，因为这绝对是我想听听你看法的话题。

So I'll tell you what, come back on the show in 2026, and we'll talk about some of your reflections from twenty five years in venture, because that is definitely a topic I want to get your taste on.

节日快乐。

Happy holidays.

我们很快再聊。

We'll talk to you again very soon.

节日快乐。

Happy holidays.

非常感谢你，阿卡什。

Thanks so much, Akash.

很高兴能有这次交流。

Enjoyed my time.

我们再聊

Talk to

很快。

you soon.

人形机器人在2025年成为最受关注的技术之一。

Humanoid robots became one of the most closely watched technologies in 2025.

但在过去几个月里，这个故事中一个越来越受关注的角度是，这些机器人可能四处走动所带来的安全问题。

But over the past few months, one angle of this story that has started getting a lot more attention is the safety element of these robots possibly walking around all over the place.

今天有新报道披露，监管机构正在考虑制定相关标准，而开发这些机器人的初创公司对此有何反应。

The information is out with a new story today on how regulatory groups are thinking about setting these standards and how startups developing the robots have reacted to things.

结果发现，并非所有人都持相同意见。

Turns out not everybody is necessarily on the same page.

我想邀请我们的机器人记者火箭·德鲁，来告诉我们他发现了什么。

I want to bring on our robotics reporter, Rocket Drew, to tell us more about what he's found.

火箭，欢迎再次做客节目。

Rocket, welcome back to the show.

很高兴你能来到这里。

It's great to have you here.

嗨，阿卡什，很高兴能来这里。

Hi Akash, great to be here.

所以让我理解一下。

So let me understand this.

机器人四处走动可能很危险？

Robots walking around could be dangerous?

我的理解对吗？

Is that what I understand?

没错，没错。

That's right, that's right.

我认为要理解这一点，需要稍微转变一下思维。

I think to understand it helps to have a little bit of a mindset shift.

当我们通常购买家用小工具时，我们习惯于它们非常安全，对吧？

So when we think about typically buying gadgets to put in our own homes, we're used to them being very safe, right?

扫地机器人。

The Roomba.

扫地机器人。

The Roomba.

是的。

Yeah.

很轻。

Real light.

没错。

Exactly.

没错。

Exactly.

我买

I buy

一个灯泡。

a light bulb.

买一个吸尘器。

Buy a vacuum cleaner.

我买一个扫地机器人。

I buy a Roomba.

我知道，我不指望它会短路、着火或者在我面前爆炸。

Know, I don't expect it to short circuit and catch on fire or blow up on me.

我们习惯了这些设备非常安全。

We're used to these things being very safe.

但想象一下，你是一名工厂工人，每天上班时都被各种机器包围，比如巨大的冲压设备，它们会冲压钢板，还有快速摆动的大型工业机械臂，以及末端装有焊枪的焊接机器人。

But imagine instead that you're a factory worker and when you go to work every day, you're surrounded by machines like giant stamping equipment that punches out sheets of steel and big industrial arms that swing around at rapid paces and welding robots that have welding torches on the end of them.

这是一个完全不同的环境，在这种环境中，多年来已经有很多人受伤。

It's a very different environment and it's an environment where, you know, a lot of people have gotten hurt over the years.

因此，工业界建立了一套非常严格的安全部署标准，以确保工人不会在工作中被这些机器人伤害。

And so the industrial world has built up a set of very robust safety standards to ensure that workers don't get hurt by these robots on the job.

从这种思维出发，现在你想象一下，引入一种新的机器人，它能用两条腿行走，可以搬运各种物品，在你的设施内移动，并执行不同类型的工作，而且很可能由人工智能驱动——你就能理解为什么这会引起一些警觉。

From that mindset, now you imagine introducing a new robot that walks around on two legs, it can carry all sorts of things, it can move around your facilities, and it can perform different kinds of work, and it's likely powered by AI, you can start to imagine why it would raise some alarms.

人们会质疑：我们如何保证这些机器人能够安全运行？

People would have questions about how do we guarantee that these things are going to work safely.

那么，这些庞大的人形机器人有没有已经发生过撞倒人的事故？

So have there been issues yet with these giant honking humanoids knocking people over?

你知道，我

You know, I

我认为目前还没有出现大的问题，因为我们还处于早期阶段。

think there haven't been any big issues yet because we're so early on.

实际上，这些机器人还没有大量进入现实世界，而一些现有的安全标准大致适用。

Like, not a lot of these have actually made it out into the world yet, and some of the existing safety standards do roughly apply.

这些标准提供了一些指导，帮助我们实际应对这些机器人。

They apply, you know, give a little bit of a guideline for how to handle these things in practice.

但我们已经看到了一些所谓的早期预警信号或担忧的迹象。

But we've seen what I'd maybe call early warning shots or signs of concern.

在测试过程中，曾有机器人出现失控，在实验室里发疯的情况。

There have been instances where robots in testing have gone a little haywire and sort of gone berserk in the lab.

几周前还有一起备受关注的事件，特斯拉的Optimus机器人突然失控，把手重重砸在桌子上，无意中压扁了路上的几个水瓶。

Or there was one high profile incident a couple weeks ago where a Tesla Optimus robot kind of freaked out and threw its hand down on a table and inadvertently crushed a number of water bottles on its way down.

人们对于机器人竟然有足够力量压扁这些水瓶感到有些震惊。

I think people were a little astonished to see that the robot was strong enough to crush those water bottles.

因为这些again都是大型机器。

Because again, these are large machines.

只是它们通常由人操作，看起来有点脆弱。

It's just that they're usually operated with a they look sort of delicate.

当你在演示视频中看到它们时，它们看起来就像人类在你身边活动。

They look like a human acting near you when you see them in these demonstration videos.

所以你的故事中很大一部分是关于谁将监管这些机器人，以及谁来制定安全指南。

So a big part of your story was talking about who is going to regulate these robots, and who's going to set the safety guidelines.

这里的相关方是谁？他们为制定这些标准做了多少工作？

Who are the relevant players here and how much work have they done to actually set these standards?

是的，这是个很好的问题。

Yeah, that's a great question.

事实证明，工业安全组织多得像一堆字母缩写，我就不一一列举了，但有几个关键组织要知道：在国际层面，这些安全标准通常是ISO标准。

It turns out there's a huge alphabet soup of industrial safety organizations, and so I'll spare you all of them, but a couple of the key ones to know are that at the international level, these safety standards are typically ISO standards.

因此，ISO是国际管理机构，不同国家都有代表参与ISO流程，以确保自身利益得到体现。

So that's the international governing body, and different countries have delegates to the ISO process so that they're represented along the way.