Why SpaceX-Telsa Merger Makes Sense, Airtable’s $1.3B Sale, Claude Code Alternatives

欢迎大家收听Information的TITV节目。

Welcome everyone to the Information's TITV.

我是Stephanie Palazzolo，今天是8月4日，星期二。

My name is Stephanie Palazzolo and it's Tuesday, August 4.

首先，我们来聊聊今天早上的大新闻：曾经风光无限的效率工具初创公司Airtable被收购了，收购价还不到它上一次估值的一个零头。

First up, we'll discuss the big news of the morning, the acquisition of once high flying productivity startup Airtable for a fraction of its last valuation.

另外，今天早上，The Information在SpaceX发布财报前发表了一篇精彩的分析，主张SpaceX应该和马斯克旗下的另一家公司特斯拉合并。

Then the information published a great analysis this morning ahead of SpaceX earnings, making an argument in favor of a merger with Elon Musk's other company, Tesla.

我们的金融分析专栏作家仔细算了这笔账，接下来她会带我们看看她的分析。

Our financial analysis columnist crunched the numbers and she'll walk us through her story.

之后我们会讨论Coinbase和其他科技公司如何开发自己的人工智能编程助手，来替代Anthropic和OpenAI这些巨头提供的更昂贵的选择。

We'll then discuss how Coinbase and other tech firms are building their own AI coding agents as an alternative to pricier options from giants like Anthropic and OpenAI.

最后，我们会和AI可解释性初创公司Goodfire的CEO聊聊他们的新产品发布。

And we'll wrap the show with a discussion with the CEO of AI interpretability startup Goodfire about their new product launch.

我真的很期待这些讨论。

I'm really looking forward to all these discussions.

意大利企业集团Bending Spoons已同意以约13亿美元收购生产力初创公司Airtable，这比五年前的110亿美元估值大幅缩水，因为企业软件公司正面临越来越大的压力，需要在AI时代证明自己的价值。

Italian conglomerate Bending Spoons has agreed to acquire productivity startup Airtable for around $1,300,000,000 That's a steep drop from its $11,000,000,000 valuation it had just five years ago, as enterprise software companies face mounting pressure to prove their value in the AI era.

今天请到的嘉宾是Bain Capital Ventures的合伙人Aaref Hilaly，以及Marathon Management Partners的创始合伙人Chase Packard。

Here to comment is Aaref Hilaly, partner at Bain Capital Ventures, and Chase Packard, founding partner at Marathon Management Partners.

欢迎Aaref和Chase。

Welcome Aaref and Chase.

很高兴你们能来参加节目。

Great to have you both on the show.

谢谢。

Thank you.

很高兴来到这里。

It's great to be here.

感谢邀请。

Thanks for having us.

五年前，Airtable还是最热门的软件初创公司之一。

So five years ago, Airtable was one of the hottest software startups out there.

如今它却以远低于上次估值价格被收购了。

Today it's getting bought for a fraction of its last price.

到底发生了什么？你对这次收购怎么看？

What happened and what do you make of the acquisition?

我觉得现在有很多公司的估值只有2021年的10%。

Well, I think that there are a lot of companies that worth 10% of their 2021 valuations.

所以Airtable并不是个例。

And so Airtable is not uncommon.

发生这一切的原因就是AI，还有所谓的“SaaS末日”。传统SaaS公司在未来很可能被新的AI原生解决方案取代，所以它们的价值就大大缩水了。

And what happened is AI and, you know, the SaaSpocalypse, if you like, that legacy SaaS companies are just worth much less in a world where they're likely over time going to be displaced by new AI native solutions.

我觉得“AI导致的变化”这句话可以解释过去五年里发生的很多事情。

I think the phrase what happened as AI can apply to a lot of things that have occurred in the last five years.

Chase，你觉得人们对这笔交易有什么误解吗？

I mean, Chase, do you think people are misunderstanding anything about the deal?

是的。

Yeah.

有一件事大家很少提到，就是在交易前一周，Airtable把业务分成了核心业务和他们的新AI业务，叫HyperAgent。

One thing I haven't heard people talk about much is the week before the deal, Airtable split it into the core business and then their new AI business called HyperAgent.

目前还不清楚Bending Spoons到底收购的是整个公司，还是只收购核心业务。

It's unclear what exactly Bending Spoons required if it's the entire company or just the core business.

但如果真的把核心业务和AI业务分开，那这笔交易就完全不一样了。

But if it's splitting off the core business and the AI business, changes the deal completely.

这其实是在清理门户，适应AI优先的新现实。

It's cleaning house and adjusting to the new reality of an AI first world.

这听起来是个新消息。

So this seems like new information.

对，就像你提到的，他们好像把HyperAgent和核心业务分开了。

You know, as you mentioned, it sounds like they split off between HyperAgent and the core business.

那你知道这对投资者意味着什么吗？

I mean, do you know anything about what this means for investors?

他们会不会在分拆出来的HyperAgent里获得股权？

Are they going to get equity in the spun off entity, HyperAgent?

目前从我看到的来看，还不清楚。

It's unclear right now from what I've seen.

他们在两边都有股权。

They add equity in both.

拆分发生在合并宣布之前。

The split happened before the merger was announced.

所以如果接下来几周我们看到公告，说核心的Airtable团队专注于HyperAgent项目，而不是Bending Spoons旗下的Airtable，我也不会感到意外。

So I wouldn't be surprised if we've seen an announcement in the coming weeks of core Airtable team focusing on the HyperAgent project versus, Airtable within Benedict Spoons.

你对这件事的理解是，HyperAgent可能是那种AI原生或AI优先的产品，而且……

And your read on this is HyperAgent is the kind of maybe AI native or AI first product and

对。

Correct.

是的。

Yeah.

关于这一点。

On that.

对。

Correct.

是的。

Yeah.

他大概四个月前宣布的，而且很多工程团队最近都专注于这个项目。我觉得他认为这就是产品的未来。

How how he announced it, I believe, about four months ago, and it's where a lot of the engineering team's been focused recently, I believe what he views as the future of of the product.

斯蒂芬妮，我觉得从不同利益相关者的角度来看这笔交易很有必要，因为我认为这将是SaaS公司未来常见的现象。

I think it's worth breaking down this deal from the perspective of the different stakeholders, Stephanie, because I think this will be a common phenomena for SaaS companies as things go forward.

所以我们分开看Bending Spoons，先看Bending Spoons。

So if we look separately at Bending Spoons let's look for Bending Spoons.

Bending Spoons捡了个大便宜。

Bending Spoons got a great deal.

这家公司有50万客户，财富500强中有80%在使用他们的产品，年经常性收入4.8亿美元，增长20%，而Bending Spoons以净现金12.85亿美元收购了它。

It's 1,285,000,000 net of cash for a company with 500,000 customers and 80% of the Fortune 500 using their product, $480,000,000 in ARR growing 20%.

所以这相当于4.7倍的收入倍数，或者净现金后2.7倍的股权倍数。

So that's a 4.7 revenue multiple multiple or 2.7 equity net of cash.

我觉得他们确实捡了个大便宜。

I think they objectively got a great deal.

对于投资者和创始人来说，创始人可能早就套现了一部分，现在还能分到一笔钱，投资者也能拿回本金，然后继续往前走。

For investors and for the founders, well, the founders probably took liquidity along the way and will still get a chunk of the proceeds and the investors get their money back and get to move on.

所以我觉得他们没什么问题。

So I think they're fine.

这其实就是创业和投资的一部分。

This is all kind of like part of the thing of starting a company and investing in startups.

我觉得最让人同情的是员工，有些人会觉得特别委屈。

I think the the the people you have to feel for are the employees, some of whom will feel really hard done by.

而且这会对他们造成冲击和震惊，因为就在不久前，Airtable在二级市场的估值还是40亿美元。

And it'll be jarring and a shock to them because in the secondary market, Airtable was trading at 4,000,000,000 pretty recently.

我觉得他们今天肯定很担心，自己的股份还能不能换到钱。

And they will be, I think today, very worried about will they get anything for their shares.

现在很难说，但算一下的话，这笔交易应该能给普通股东留下大约8.5亿美元。

I I think it's hard to tell off the bat, but it looks like if you do the math, there should be about 850,000,000 left over for common shareholders on this deal.

创始人大概持有公司20%的股份，所以他们会拿到那部分。

The founders probably own about 20% of the company, so they'll take that.

员工持股计划如果有15%的话，那就是900个人分1.27亿美元。

The ESOP, if it has 15%, that would mean $127,000,000 spread across 900 people.

所以普通股东应该能从这笔交易中得到一些回报，只是远没有他们期望的那么多。

So common shareholders should get something from this, just nowhere near what they were hoping for.

听起来员工确实有点吃亏了。

It does sound like maybe employees were left kind of holding the short end of the stick.

我是说，事后诸葛亮当然容易，但Aaref，我很好奇，回顾过去，你觉得他们有没有什么应该做得不一样的地方，或者应该做出什么转型？

I mean, obviously hindsight is twentytwenty, but Aaref, I'm curious, looking back, do you think there's anything that they should have done differently or a pivot that should have happened?

还是说这其实超出了他们的控制范围，他们押注了无代码，而现在AI基本上完全取代了它？

Or is it just kind of out of their control and they were betting on no code and now AI has kind of just totally subsumed that?

我觉得当管理层同时要处理这么多事情的时候，真的很难去质疑他们的决策。

I think it's just it's really hard to second guess management teams when they're dealing with so many different things at the same time.

当然，如果能回到过去，我相信Airtable团队在过去五年里肯定会做出不同的选择。

Of course, if you go back in time, I'm sure the Airtable team would do things differently over the last five years.

他们可能更早地押注AI，但事后诸葛亮容易，毕竟他们是在经营一家公司。

They might've leaned into AI more heavily, but it's easy to armchair quarterback that when they're the ones building a company.

别忘了，他们确实打造了一个非常成功的产品，拥有成千上万的客户和数百万满意的用户。

And let's not forget, built a really successful product here that has thousands and thousands of customers and millions of happy users.

我觉得他们都有很多值得骄傲的地方。

And I think they all have a lot to be proud of.

对，我觉得现在回头看，我们掌握了所有信息，但当时很难做出判断。

Again, yeah, I think it's tough to, you know, look back now with all the information that we do have in that moment.

真的很难知道下一步该怎么走。

I think it's super hard to know what to do next.

退一步看，我觉得有意思的是，这对更广泛的老牌企业软件公司意味着什么。

I mean, taking a step back, I think it's interesting to think about what this is going to mean for the kind of broader class of older enterprise software companies.

Chase，你觉得这是个例吗？

I mean, Chase, do you think this is an isolated case?

我们会不会看到更多这样的交易？

Are we going to see more of these types of deals?

如果是这样的话，你觉得哪些企业软件公司可能会面临这种情况呢？

And if so, you know, what types of enterprise software companies do you think are on the table here for this to happen too as well?

对，我觉得Airtable不是个例。

Yeah, I think Airtable isn't the exception.

这只是个开始。

It's the opening act.

有两种风险定价错误的情况：一种是买错了，另一种是建错了然后被删掉。

And there's two kinds of at risk pricing that was wrong that gets bought and built wrong that gets deleted.

我觉得第一类是在2021年融资时估值过高。

I think this bucket one, they raise ahead of their Cs in 2021.

他们会被私募股权公司、战略投资者或像Bending Spoons这样的运营者整合，Toma Bravo在2025年完成了超过420亿美元的交易量。

They get consolidated by PE strategics, operators like Bending Spoons, Toma Bravo did over 42,000,000,000 of transaction volume in 2025.

Airtable就属于这一类。

This is where Airtable lives.

他们找到了归宿，战斗结束了，之后唯一的争论就是优先权结构的问题。

They find homes, the fight is over and that's where the the only debate after that is where the preference stack is.

我觉得第二类是那些产品本身就有问题的公司。

I think bucket two is the companies that are built wrong.

这些公司面临AI带来的风险。

These have AI risks.

这些公司的产品本身被AI商品化了。

These are where the product itself is commoditized by AI.

2026年2月的SaaS大崩盘在一周内蒸发了近3000亿美元的市值，因为AI确实起作用了。

The February 2026 SaaS pocalypse wiped out almost 300,000,000,000 of market cap in a week because AI worked.

随着AI做更多的事情，代理接手更多的工作，客户正在削减席位。

Customers are cutting seats as AI does more and agents take more of this work.

对。

Yeah.

我同意。

I agree.

我的意思是，我觉得对于一家初创软件公司来说，有两种策略。

I mean, I feel there are two strategies for a start up com a software company, I'll say.

一种是风投所设定的模式。

One is what venture capital is set up for.

你要增长一百倍，年增长率要达到三位数。

You're gonna grow a 100 you're gonna have triple digit ARR growing it by in triple digits.

你要成为时代的弄潮儿。

You're gonna be part of the zeitgeist.

你要征服世界。

You're gonna take over the world.

而另一种策略则完全不同。

And then the alternative strategy is kind of far from that.

那就是接受20%的增长。

It is we're gonna embrace 20% growth.

我们不会试图征服世界。

We're not gonna try and take over the world.

我们会削减成本。

We're gonna cut costs.

我们会停止投机性的研发，只做渐进式的改进。

We're gonna stop speculative r and d and make incremental improvements.

最重要的是，我们会提高价格，从现有客户身上变现，就像IBM大型机那种策略。

And most importantly, we're gonna raise price and monetize our existing base, sort of the IBM mainframe strategy.

我觉得，夹在这两种策略之间的公司，日子会很难过。

And that anyone stuck between those two, I think, has a hard time.

所以Airtable通过这次收购，正在从第一种策略转向第二种策略。

And so Airtable is moving from strategy one to strategy two with this acquisition.

你觉得哪种策略更好，还是说这取决于具体的公司？

Do you have a opinion on which strategy is better or is it just kind of depend on the specific firm?

这个嘛，其实没得选。

Well, I mean, don't get to pick.

如果你能采用征服世界的策略，实现百分之百的增长，谁会不这么做呢？

If you can do the takeover the world strategy and grow it a 100%, then who would not do that?

那当然。

Of course.

大家都想这样。

Everyone would love that.

但如果做不到，如果那不可能，那就得重新考虑B计划了。

But if you can't do that, if that is not possible, then you have to look again at plan B.

我觉得有个问题想问你们两位，这跟现在很多投资者看自己投资组合里的公司时面临的问题很像。你知道，虽然AI可能已经吃掉了SaaS，或者这只是个流行说法，但我觉得大家现在担心的是，AI实验室可能会吃掉AI应用，或者至少分走一大块蛋糕。

I mean, I think a question for both of you, I think there definitely are parallels here to a question that I think many investors are facing now looking at their own portfolio companies where we, you know, as much as AI maybe ate up SaaS or this is like a narrative that's going around, I think people are worried now that the AI labs might eat up AI applications or at least take a big chunk out of, their, out you know, kind of pie.

这怎么影响你们对自己投资的AI应用的看法，以及你们怎么看待那些也在开发竞争性应用的模型开发商带来的威胁？

How does this kind of influence how you're thinking about the AI apps that you've invested in and maybe the threat from model developers who are also building competing apps?

Aaref，要不你先来？

Maybe Aaref, we can start with you.

好的。

Yeah.

当然。

Sure.

我合作过不少这样的公司，比如Cognition、Decagon、Simile。

I mean, I I work with a number of these companies like Cognition, Decagon, Simile.

我看到的情况是，没有哪个单一模型能通吃所有。

And what I see is there is no one model that will take over them all.

而且我们大多数公司并不只是用那些实验室的基础模型产品。

And most of our companies are not just using foundation model products from the labs.

他们也在用开源权重模型，而且发现这些模型在很多方面确实很管用。

They're also using open weight models, and they're finding them really effective at a bunch of things.

所以我觉得实验室完全接管应用层的可能性是零。应用公司可以通过为合适的任务挑选合适的模型，以及帮助解决人类适应AI的问题——也就是让人们真正接受AI、使用AI，这本身就是一个大挑战——来创造巨大的价值。

And so I think the chances that the labs take over the application layer completely is zero And that the application companies can deliver a ton of value by selecting the right model for the right job and by helping with the human diffusion problem of actually getting people comfortable with AI and using AI, which itself is a big challenge.

对，我同意。

Yeah, I agree.

基础模型现在正面临前所未有的压力，这真是太好了。

The foundational models are under the most pressure they've been in quite some time, which is great to see.

开源模型越来越受欢迎。

Open source is getting more and more popular.

我觉得在应用层，最稳妥的永远是那些有下载量、有监管模式，或者是在特定垂直领域里，对终端用户或客户数据比较敏感的那些应用。

I think the safest at the application layer is always gonna be things with downloads, regulatory modes, or is it a specific vertical where there's sensitivities around end users or customer data or things like that.

这些领域总是会有自己的模式，不管是金融模型方面还是垂直应用方面。

Those are always going to have modes, whether on the financial model side or on the vertical app side.

好的，我们下次一定会再请你们两位来聊聊下一笔大型企业软件交易，我相信很快就会有的。

Well, we'll have to have both of you back on for the next, you know, big enterprise software deal, which I'm sure will happen very soon.

再次感谢两位的到来，Aaref，Bain Capital Ventures的合伙人，以及Chase Packard，Marathon Partners的创始合伙人。

Aaref Again, thank you both for coming on and Hilaly, partner at Bain Capital Ventures and Chase Packard, founding partner at Marathon Partners.

谢谢你邀请我们，Stephanie。

Thank you for having us, Stephanie.

谢谢你的邀请。

Thank you for having us.

SpaceX今天发布了有史以来第一份季度财报。

SpaceX is reporting its first ever quarterly results today.

在财报发布之前，The Information发表了一篇分析SpaceX和特斯拉合并的文章，探讨了这对股东可能带来的好处。

Ahead of that report, the information published a story analyzing a SpaceX Tesla merger and how it could benefit shareholders.

我们的金融分析专栏作家Anita Ramaswamy现在和我一起聊聊她的看法。

Our financial analysis columnist, Anita Ramaswamy, joins me now with her verdict.

欢迎你，Anita。

Welcome, Anita.

嗨，Steph。

Hi, Steph.

很高兴来到这里。

Happy to be here.

那么，SpaceX和特斯拉合并，对两家公司的股东来说，是好事还是坏事？

So a SpaceX Tesla merger, yay or nay for shareholders of both companies?

表面上看，考虑到埃隆·马斯克过往的记录和一些交易，这似乎不是个最好的主意，但在我和投资者聊过，并深入分析了两家公司的财务状况后，我的结论是，这是好事。

So it sounds on the surface like it might not be the greatest idea considering some of Elon Musk's track record and past deals, but what I ultimately concluded after talking to investors and really analyzing the financials of both companies is that it's a yes.

那太好了。

That's awesome.

那我们来听听你的理由。

So let's get into your reasoning a little bit here.

你提到的一个原因就是消除埃隆·马斯克公司之间的利益冲突。

You know, one of the reasons you talk about is eliminating conflicts of interest between Elon Musk's companies.

你知道，众所周知，他经营着SpaceX，经营着特斯拉，还经营过xAI，现在xAI已经并入SpaceX了。

You know, as we all know, he runs SpaceX, he runs Tesla, he ran XAI, which is now part of SpaceX.

这对他来说是不是就是顺理成章的下一步？

Is this just kind of the next natural step for him?

从某些方面来说，是的，Steph。

In some ways, yes it is, Steph.

问题是，我接触过的很多马斯克的股东，不管是特斯拉的、SpaceX的还是两家的，他们都非常相信马斯克的长期愿景。

So the thing is that a lot of Elon Musk shareholders who I talked to, who are shareholders in either Tesla or SpaceX or both, they're really believers in Elon Musk's long term vision.

他们愿意在这些前沿领域下注，比如相信乐观主义和人形机器人，相信火箭技术的未来，或者相信他能把xAI和SpaceX打造成一个长期的AI业务，不管是靠出租算力，还是靠开发自己的AI模型。

They're taking a lot of these frontier edge bets where they might believe in optimists and humanoid robotics, they might believe in the future of rocketry, or they might believe in his ability to turn xAiSpaceX into a long term AI business, whether that's through renting out compute or whether that is through building their own AI models.

所有这些可能性都还没有实现，而且可能还需要好几年才能真正落地。

And all of these possibilities have not materialized and probably will take several years to actually get there.

但与此同时，很多我采访过的投资者都在说，如果我是特斯拉的股东，我希望马斯克的利益和我的利益是一致的，这样他才能为我创造价值。

But in the meantime, a lot of the investors who I spoke to were saying that, look, if I'm an investor in Tesla, I want Elon Musk's incentives to be aligned to create value for me.

所以，如果他把所有时间都花在SpaceX上，那我就不一定能从马斯克运营这些合并公司中获益了。

And so if he's spending all his time on SpaceX, then I'm not necessarily going to capture the value of Musk running those combined companies.

我还想提一个历史相似案例，因为很多人跟我聊到的都是特斯拉在2010年代收购太阳城的那笔交易。

And I do want to draw a historical parallel because I think the deal that a lot of people were talking to me about was the deal for the acquisition of Solar City that Tesla did back in the 2010s.

那是一个非常不同的交易，当时人们觉得是埃隆·马斯克在拯救他表弟运营的太阳能公司，然后把它合并进特斯拉。

That was a very different transaction where it was perceived as Elon Musk sort of rescuing this company that was run by his cousin's solar business and folding that into Tesla.

我这篇文章想说的重点是，这次的情况其实不一样。

And the point that I wanted to make with this column is that this time around is actually different.

这不是对一家濒临倒闭公司的救助。

It's not a bailout of a failing company.

这次是合并两家正在进行高风险押注的企业，这些赌注结果如何还不好说。

It is combining these two businesses that are making some really risky bets that may or may not play out.

但说到底，如果你相信埃隆·马斯克的长期愿景，你会希望看到他整合自己的商业利益，消除潜在的利益冲突，然后专注于运营合并后的公司。

But ultimately, if you're a believer in the long term vision of Elon Musk, you're going to want to see him sort of align and consolidate his business interests, eliminate any potential conflicts of interest, and focus on running the combined company.

我注意到你的文章里还提到了，这两家公司其实已经在共享资源了，所以合并的话，共享起来会更方便。

And I think something that you also mentioned too in your column is this idea that these two companies are already sharing resources, it sounds like, and so combining would just make it easier to do so.

你多讲讲这一点吧。

Talk a bit about that.

是的，他们在很多方面都在共享资源。

Yeah, so they are sharing resources in a lot of ways.

比如说，埃隆·马斯克就经常分享。

I mean, example is that Elon Musk shares.

他喜欢在自己的各个公司之间自由调动高管和员工。

He tends to move executives and employees around between his various companies pretty freely.

而且你知道，如果想想实际的产品和它们现在的运作方式，XAI的Grok模型正在被用在特斯拉汽车里，帮助指导一些自动驾驶的功能和技术。

And you know, if you think about the actual products and how they're being operationalized right now, Grok models through XAI are being used in Tesla cars to help guide some of the self driving capabilities and technologies.

所以现在，比如说特斯拉实际上是从XAI购买计算资源，他们可能是在按零售价购买。

And so right now, let's say Tesla is actually buying its compute from XAI, they're doing so probably at retail prices.

如果这些公司真的合并了，他们就能从中节省一些成本。

And if these companies were actually to consolidate, they can find some cost savings through that.

也许他们不用再为计算资源付全价，也许他们的激励会更一致，也许他们能找到一些节省基础设施和更广泛AI建设成本的方法。

Maybe they don't have to pay full price for compute, maybe their incentives are more aligned, maybe they can find some ways to save on infrastructure and the broader AI build out.

这些协同效应很多都不是最明显的。

A lot of these synergies are not the most distinct.

现在就是投资者在说，哦，他可能会这么做，或者那么做。

Like right now it's investors sort of saying, oh, he could do this or he could do that.

最终，我们还是得看马斯克自己打算怎么提议，如果他决定推进这笔交易的话。

And ultimately we will have to see what Elon Musk himself decides to pitch if he decides to go forward with this deal.

而且，你也谈到了这笔潜在交易的数字问题。

And you know, you talk a bit about the math of this potential deal as well.

我觉得有意思的一点是，SpaceX股票最近的表现，虽然可能让SpaceX的投资者不太开心，但实际上可能让这种合并变得更容易。

You know, I think an interesting point here is that the recent performance of SpaceX's stock, while maybe not, something that makes SpaceX investors super happy, it could actually make this sort of combination easier.

你在数字方面发现了什么？

What did you kind of find there with that, with that math?

SpaceX最近上市了，上市时估值比现在高很多。

So SpaceX went public pretty recently and when it went public, it was valued a lot more highly than it is today.

股价最高达到每股211美元。

It reached a high share price of $211 per share.

那对应的市值大约是2.8万亿美元。

And so that was like a $2,800,000,000,000 equity valuation for the market cap.

现在SpaceX的市值接近1.5万亿美元。所以当你把SpaceX和特斯拉作为两家独立公司来比较时，它们现在的远期市盈率其实非常接近。

And now the market cap of SpaceX is closer to $1,500,000,000,000 So when you look and compare SpaceX and Tesla as two standalone companies, they are actually now trading at a forward price to earnings ratio that is very similar.

我觉得这让这笔交易对股东来说更有吸引力了。

And I think that makes the deal look a lot more palatable to shareholders.

而且你知道，作为SpaceX的投资者，你可能会高兴，当然，你的股票值211美元一股。

And you know, as a SpaceX investor, maybe you were happy, yes, of course, that your stock was worth $211 per share.

但同时，那个估值从根本上来说并不现实。

At the same time, that valuation was just not rooted in reality fundamentally.

所以我认为市场回调是合理的，SpaceX的股价估值已经降下来了。

And so I think the market correction makes sense that the SpaceX valuation of the stock has come down.

虽然这可能会让事情变得有点棘手，乍一看你可能会想，好吧，这会变得更难，因为SpaceX没有那么多股票可以用来收购，或者他们的股票价值没那么高了。

Even though it might make it a little bit harder, at first glance you might think, Okay, it's going to make it harder because SpaceX doesn't have as much stock that they can, or their stock isn't worth as much to be able to do the acquisition.

我认为估值回落到一个现实的水平是健康的。

I think that it's healthy that the valuation has come down to a level where it is realistic.

这两家公司估值相近，更容易让人看到它们如何互相受益。

And these two companies trading at a similar valuation makes it an easier sell to see how they can benefit one another.

所以这可不是那种一边倒的交易，比如SpaceX在救特斯拉，或者特斯拉在救SpaceX。

So it's not just a deal that's lopsided like, Oh, you know, SpaceX is bailing out Tesla or Tesla is bailing out SpaceX.

其实不是这样，它们的股价真的非常接近。

It's like, No, they actually trade at really similar levels.

所以把两家公司合并起来，就更有道理了。

And so combining the two makes a lot more sense.

而且显然今天是它们的第一次财报发布。

And obviously today is their first earnings.

特别激动人心。

Very exciting.

我觉得很多人都在关注这件事。

I think a lot of people are looking to it.

你有哪些比较重要的问题，想在今天得到答案，好让这个合并的事，

What are the kind of big questions that you're looking to perhaps get answered today that might make this question of

变得更明朗一些？

a merger clearer?

这个合并的想法其实是在特斯拉的财报电话会上提出来的。

Well, this whole idea of a merger really came up in the Tesla earnings call.

之前也提过很多次，大家也猜测过，也许这就是埃隆·马斯克的宏伟计划，是他想做的事。

I mean, had come up many times before and people had speculated, okay, maybe this is Elon Musk's grand plan and this is what he wants to do.

几周前的特斯拉财报电话会上，他含糊其辞地回答说不会回答关于这笔交易的问题。

During the Tesla earnings call a couple of weeks ago, he sort of coyly answered that he was not going to answer questions about the deal.

他说如果真有这样的事，也得走正规流程。

He said if any such thing happened, it would have to go through an appropriate process.

所以我非常怀疑今天电话会上会有分析师问他会不会推进这件事。

So I highly suspect that he's going to get some questions from analysts on the call today about whether or not he would proceed with such a thing.

看他怎么回应会很有意思。

It's going to be interesting to see how he chooses to respond.

他可能会正式拒绝评论，但以埃隆的风格，我们总能从他的肢体语言或其他暗示中看出他下一步的动向。

He will probably officially decline to comment, but with Elon, we always tend to get some signal and some, you know, sort of other body language or indicators of what his next move is.

我觉得投资者们认为这件事非常有可能发生。

I think this is just something that investors are seeing as very, very much on the table.

所以今天要留意一下，看马斯克会怎么回答这个问题。

So keep a lookout for maybe the way Elon Musk will answer that question today.

从另一个角度看，这两家公司都在烧很多钱，因为他们在投资一些尚未被验证的技术。

To kind of take the other side of this argument, you know, both companies are burning a lot of cash because they're investing in kind of, you know, in some cases unproven technology.

合并似乎并不能解决这个问题。

It doesn't seem like a merger would help with that.

实际上，合并可能会让情况看起来更糟，因为把这么多亏损的业务整合在一起。

In fact, it might just make everything look worse because you're consolidating all these, you know, hugely, unprofitable businesses together.

这对合并意味着什么？

What does that mean for the merger?

这是否是人们

And is that something that people

所担心的？

are worried about?

是的，绝对是这样。

Yeah, absolutely.

你说得完全对，这点确实值得指出来。

And you're totally right to point it out.

就像你说的，SpaceX和特斯拉那些大胆的押注，基本上都还没被验证过。

Pretty much all of the big ambitious bets that SpaceX and Tesla are making are unproven, to your point.

关于这个，我想说两点。

I'll say two things to this.

首先，SpaceX和特斯拉确实在烧钱。

Firstly, SpaceX and Tesla are burning cash.

这是事实。

That is true.

但至少特斯拉在电动车销售上，有持续产生健康自由现金流的记录。

But Tesla at least has a track record of generating healthy free cash flow through their electric vehicle sales.

至于Optimus，我们还得看他们怎么走。他们现在选择在这个周期投资，是为了给长期的人形机器人业务和AI基础设施打基础。

We'll have to see where they go with Optimus, they are choosing to invest in this cycle to sort of lay down the foundation for the long term humanoid robotics play and for their AI infrastructure.

但至少对我来说，看到他们有一条能产生现金流的路径，而且单从卖车来看，每卖一辆都是赚钱的，能带来现金，这还是挺让人鼓舞的。

But it, at least to me, was sort of encouraging to see that there is a path to generating cash flow and at least on a unit level when they sell pars, those are profitable and generating cash.

所以我觉得这能让人稍微安心一点，但你说得完全对，这两家公司确实都在烧钱。

So I think that gives a little bit of comfort, but your point is still totally correct that these are two cash burning businesses.

这非常非常像是一个长期投资，适合那些相信愿景、相信他们最终会实现盈利的投资者。

This is a very, very much like a long term play for investors who believe in the vision and that they will eventually get to profitability.

我觉得这里还有一点很重要，就是合并会让SpaceX更容易为这些雄心壮志提供资金，SpaceX和特斯拉，实际上为这些雄心壮志提供资金。

The other point that I think is important to consider here is that the merger would make it easier for SpaceX to actually fund some of these ambitions, SpaceX and Tesla, actually fund some of these ambitions.

所以从长远来看，如果他们要去资本市场，他们需要以某种方式筹集大量股权或债务，才能真正为他们的AI建设、Robotaxi和人形机器人建设提供资金。

So in the long term, if they're going to go to the capital markets, they're going to need to raise a lot of equity or a lot of debt one way or the other to be able to actually fund their AI build out and to fund robotaxis and to fund the build out of humanoid robotics.

我和一位信贷高管聊过，他告诉我，他们需要这么多资金，当两家公司合并后，他们去债务市场融资会更容易。

And I talked to one credit executive who was kind of telling me that the idea that they need all this capital, that'll actually become easier for them to go to the debt markets and raise when the two companies are combined.

从信贷角度来看，这两家公司的评级都相当高。

They're both pretty highly rated from a credit perspective.

现在对投资级信贷的需求很大。

There's a lot of demand for investment grade credit right now.

通过合并，通过扩大公司规模，他们可能更容易为一些长期目标融资。

And by bringing them together, by increasing the scale of the company, it might make it easier for them to finance some of their long term ambitions.

我觉得挺有意思的，因为我们显然在讨论所有这些不同的战略理由。

I mean, is interesting because I think we are obviously talking about all these different kinds of strategic rationales.

但你知道，就像你在文章里提到的，这很大程度上归结为投资者就是押注和追随埃隆·马斯克，而且就像你之前说的，他们不希望他分散注意力，他们非常相信他，觉得如果他能把全部精力都放在合并后的业务上，那对股东会更好。

But you know, as you mentioned in the piece, a lot of this boils down to investors just betting and just following Elon Musk and this idea that, you know, I think as you mentioned earlier, they don't want him to be dividing his attention and they're such big believers in him that they feel like, you know, if he could focus all of his attention on the combined businesses, that would be better for shareholders.

但确实很令人兴奋，听起来今天的财报电话会议会有很多值得关注的内容。

But very exciting and sounds like we'll have a lot to look out for in the earnings call today.

再次非常感谢你，安妮塔，来参加节目。

Again, thank you so much, Anita, for coming on.

刚才那是安妮塔·拉马斯瓦米，我们《信息》杂志的金融分析专栏作家。

And that was Anita Ramaswamy, our financial analysis columnist here at The Information.

随着AI成本上升，公司都在寻找节省开支的方法。

As AI costs rise, companies are looking for ways to save.

对有些公司来说，这意味着开发自己的智能体。

And for some, that means developing their own agents.

劳拉·布拉顿，我们《应用AI》通讯的作者，写了关于这些情况发生的地方。

Laura Bratton, author of our Applied AI newsletter, wrote about where this is happening.

劳拉现在加入我们，分享她的发现。

Laura joins me now to share what she learned.

欢迎，劳拉。

Welcome, Laura.

嘿，斯蒂芬。

Hey, Steph.

最近怎么样？

How's it going?

挺好的。

Great.

那么，你在今天的专栏里写到，Coinbase 在四月份开发了自己的 AI 编程代理，叫 Forge，这个代理还能集成到 Slack 里。

So, you know, you wrote in today's column that Coinbase built its own AI coding agent in April called Forge, and this is an agent that's also able to integrate into Slack.

这个时间点意味着它甚至比 Anthropic 的 Claude Tag Agent 还要早，后者也集成到 Slack。

So this timing means that it came even before Anthropic's Claude Tag Agent, which also integrates into Slack.

你在写这篇专栏的过程中学到了什么？

What did you learn through the process of writing this column?

我了解到，Coinbase和其他那些投入开发自己内部编码助手的公司，现在开始看到这些投资有了回报。

I learned that Coinbase and other companies who have invested in developing their own internal coding agents are now starting to see that investment pay off.

这是因为现在公司们开始想，哎呀，也许我们不应该只依赖某一家前沿AI实验室，因为政府可能会出台新政策，限制我们怎么用这些模型。

That's because just now companies are starting to think about, Oh man, maybe we shouldn't be locked into one frontier AI lab as, you know, the government might introduce new policies that restrict how we use these models.

你知道，从我们自己的报道和Fable five的情况来看，也是这样。

You know, as we've seen from our own reporting and with Fable five.

而且，随着公司可能会调整编码工具的成本结构，拥有自己的内部工具或者多样化你的工具组合，就显得很有道理。

And, you know, as companies might change their cost structures with coding tools, it just makes sense to have your own internal tool or to diversify the set of tools that you have.

所以Coinbase就是一个例子，他们很早就投资开发了自己的内部编码助手。

And so Coinbase is an example of a company that early on invested in creating its own internal coding agent.

有趣的是，它比Anthropic的Claude Tag Agent还要早推出。

And, you know, interestingly, it came out before Anthropic Squad Tag.

Coinbase和Shopify都有编码助手，通过Slack提供给所有员工，你可以在频道里@它，然后通过Slack和这个编码助手一起工作。

Both Coinbase and Shopify have coding agents that are available through Slack to all their employees, and you can sort of tag it in channels and work with the coding agent through Slack.

不过，你在专栏里确实提到，尽管Coinbase有内部选项，但他们还是说Claude Code是他们最常用的AI编码助手。

I mean, do mention in your column that in the case of Coinbase, they still do say that Claude Code is their most used AI coding agent despite, you know, also having this in house option as well.

那Forge在那里到底起了多大作用？

So how much is Forge really making a difference there?

我觉得这只是一个例子，说明公司在多样化他们的工具集。

I think that it's just an example of companies diversifying their set of tools.

所以即使Claude Code在很多公司仍然是最受欢迎的工具，他们也开始探索其他选择，不管是自己的内部编码代理还是其他的。

So even though Claude Code remains the most popular tool at many companies, they're beginning to explore other options, whether that's their own internal coding agent or others.

特别是Coinbase告诉我，他们看到Ford的使用率增长得和Claude Code一样快，甚至更快。

So Coinbase in particular told me that they see usage of Ford rising as fast, if not faster than Claude Code.

所以我觉得这很重要，他们认为这长期来看会带来显著的成本节省，因为通过他们的网关或路由器，你可以连接到你想要使用的任何AI模型。

And so I think that that is significant and they think that it will result in significant cost savings over time because it lets you sort of connect to any AI model that you want to use through this gateway or router that they have.

而使用Claude Code时，你主要用的是Anthropic的模型。

Whereas with Claude Code, you're primarily using Anthropic models.

虽然技术上你可以设置Claude Code使用不同的AI模型，一些公司也在找有趣的方法这么做，但这肯定不是Anthropic鼓励的，也不是他们推广使用的方式。

And while technically you could set up Claude Code to use with different AI models and some companies are finding interesting ways to do that, that's certainly not encouraged by Anthropic and not the way they're promoting its usage.

你还提到了Shopify，这是另一家有自己的代理的公司，也可以通过Slack使用。

You also mentioned Shopify, which is another company that has its own agent as well that's accessible through Slack.

那这个采用情况跟Coinbase那边相比怎么样？

How does that adoption look like, compared to, you know, what you're seeing with Coinbase?

两者是不是差不多，增长也差不多？

Is it similar, you know, similar growth with both?

对，我确实觉得两者增长差不多。

Yeah, I would definitely say similar growth with both.

在5月的一次演示中，Shopify提到公司大约75%的人都在用River，这是他们的编程代理。

In a May presentation, Shopify said that about 75% of the company is using River, which is their coding agent.

他们说，我们全力投入River。

And they said that, you know, we're all in on River.

River就是我们用来构建Shopify的工具。

River is what we use to build Shopify.

如果要说的话，我觉得他们对内部编程代理的投入程度甚至比Coinbase还要高，我跟Coinbase聊过。

If anything, I feel like they were even more all in on their internal coding agent than, you know, Coinbase when I talked to Coinbase.

不过有意思的是，Shopify还在用Cloud Code和Codex，但我没听说Coinbase用Codex。

Although interestingly, Shopify still also uses Cloud Code and Codex, whereas I didn't hear Coinbase using Codex.

我觉得这里贯穿的一个主线，你也提到了，就是公司们都在想方设法在AI工具上省钱，而不是直接默认用那些大供应商。

I mean, obviously kind of the running through line here, which you did touch on, is this idea that companies are coming up with creative ways to save money on AI tools rather than just, you know, defaulting automatically to the big providers.

这现在是不是成了企业特别关心和关注的问题？

I mean, is this becoming such a big concern and focus for businesses?

你觉得这种思路管用吗？

And do you think this sort of approach is working?

对，我觉得现在说结果还太早，很难说会怎么发展。

Yeah, I think that the approach, it's too early to really know how this will play out.

有些情况下，我预见到未来公司自己开发内部编码工具，然后这可能会变成一种技术债务，工程师得花时间和精力去维护这个工具。

In some cases, I could see a future in which companies have developed their own internal coding agents and then it becomes a form of tech debt where, you know, they're stuck with this tool that it takes time and energy for engineers to keep up with.

但另一方面，如果公司技术实力强，工程师有远见，愿意投入时间和精力，那这么做就很有意义。

In other cases, I I think that it makes a ton of sense if you have a tech savvy company, and really forward thinking engineers who are willing to spend the time and energy on these things.

以Coinbase为例，两个工程师花了大概七个月才建好这个工具。

In Coinbase's case, took about seven months for two engineers to build this tool.

所以，我觉得越来越多的公司都在采取这种策略，而且考虑到他们想避免被供应商锁定，以及前沿实验室的定价越来越按使用量来算，这么做是合理的。

So yeah, I think that the strategy we're seeing among more and more companies, and you know, it makes sense as they look to avoid vendor lock in and as the pricing shifts more and more to usage among the frontier labs.

七个月可不是小数目。

I mean, seven months is not nothing.

工程师花这么多时间来开发这样一个工具，时间成本确实挺高的。

That's a pretty significant amount of time for engineers to be spending their time building on- building such a tool.

挺有意思的。

It's interesting.

这让我想起ChatGPT刚出来那会儿，大家也在讨论类似的问题，就是企业要不要自己训练AI模型。

It does remind me a lot of kind of the early days post ChatGPT where it felt like this same discussion was happening but around, oh, should businesses be training their own AI models?

后来慢慢地，这个讨论就变成了：我作为企业，为什么要花几百万美元去训练自己的模型，直接用开源或者闭源的模型不就行了？

Then it kind of felt like over time that discussion settled around, you know, why should I, a business, be spending millions of dollars to train my own model when I can just use, you know, open or closed source ones?

现在感觉又回到了类似的讨论，就是我要不要自己开发智能体，来减少对AI供应商的依赖。

Now it seems like it's a similar discussion around, oh, should I be building my own agents, to try to, you know, lower my, lower the amount of dependence I have on, on AI providers.

不过感觉很多人其实是在走中间路线，比如用其他公司的AI工具，但用你提到的那个模型路由器来调度，这样就不会太依赖某一家模型供应商。

It does kind of feel like there is just like this middle ground though, that a lot of people are doing, which is like, maybe I use AI tools from other companies, but I have this model router that you talk about where I can route and that way I'm not so dependent on any one model provider.

完全同意。

Totally.

我觉得这点说得很好。

I think that's a great point.

而且我觉得，随着开源模型越来越受欢迎，就像我之前在TITV上提到的，就在几个月前，我还没听到任何CIO说他们对使用Anthropic以外的模型做编程任务感兴趣。

And I think that as we see open source models become more and more popular, I think, you know, as I've mentioned previously on TITV, just a couple of months ago, I wasn't hearing from any CIOs that they had any interest in using models other than Anthropics for coding tasks.

但现在我们越来越多地听到公司在尝试用中国的开源模型来做编程工作。

And now more and more, we're hearing companies experiment with Chinese open source models to, you know, do coding work.

而且我觉得，当西方开源模型也进步到一定程度时，公司可能更倾向于使用像OpenCode这样越来越受欢迎的开源工具来做编程，而不是像你说的，投入大量时间和精力去构建一个内部编程工具，毕竟这不是他们业务的主要目标。

And I think that, as at some point, we see open source models from the West improve as well, It might make more sense for companies to use open source harnesses like OpenCode, which is increasingly popular to do coding work rather than invest all this time and energy, as you said, in building an internal coding tool that takes time and energy and resources when that's not the primary objective of their business in the first place.

显然，对于像Coinbase和Shopify这样的科技公司来说，这样做是有意义的，但也许对于硅谷以外的其他行业来说，就不一定了。

Obviously, it makes sense for companies like Coinbase and Shopify that are tech firms, but maybe not others more, you know, in other industries outside of Silicon Valley.

完全同意。

Totally.

而且我敢肯定，OpenAI和Anthropic不会坐视不管。

And I mean, I'm sure OpenAI and Anthropic aren't taking this without a fight.

他们采取了什么措施来应对这种趋势，让他们的模型更具成本效益呢？

What are they doing to kind of counteract this trend and make their models a more cost effective option?

对，我觉得OpenAI就是个很好的例子。

Yeah, I think OpenAI is a great example.

特别是他们真的在大力推广Codex。

In particular, they've really tried to market Codex.

首先，Codex本身是个开源的框架。

First of all, Codex is an open source harness.

他们给Codex的定位很明显是优化了OpenAI的模型，但你也可以用其他模型来跑Codex。

And they've positioned Codex, you know, obviously, it's optimized for OpenAI's models, but you can use different models with Codex.

他们想把它推成一个比Claude Code更灵活、更省钱的选择。

And they're sort of trying to push this as a more flexible option and more cost effective than Claude Code.

而且他们还会搞促销和打折。

And they also will have promotions and discounts.

另外，Cursor也值得一提，因为他们最近推出了一个模型路由器，还发了篇博客说，如果你用多个不同的模型，包括他们自己的composer模型，就能在编程上省下不少成本。

And then Cursor also, I think is worth noting here because they recently launched a model router and they published a blog showing that, you know, if you use multiple different models, including their own composer model, you can save on costs with coding.

所以我觉得，相比之下，Anthropic这边好像没怎么发力。

So I think that I haven't seen as much of an effort from Anthropic in particular.

而且我们看到，Claude Code至少在短期内，在企业领域还是有一定竞争优势的。

And as we're seeing Claude Code at least temporarily has has a competitive moat in enterprises.

但在其他领先的大公司中，特别是OpenAI和Cursor，我们看到他们在这方面取得了一些进展。

But with the other big firms that lead this space, particularly OpenAI and Cursor, we are seeing some headway made there.

太好了。

Great.

好的，非常感谢你，Laura。

Well, thanks so much, Laura.

再次感谢，刚才和我们交流的是Laura Bratton，她是The Information的Applied AI通讯作者。

Again, that was Laura Bratton, author of our Applied AI newsletter here at The Information.

AI可解释性实验室Goodfire今天宣布，向公众推出其AI研究平台Silico。

AI Interpretability Lab, Goodfire, announced today it's launching its AI research platform called Silico to the general public.

接下来，我们请到了Goodfire的联合创始人兼CEO Eric Ho来为我们介绍更多。

Here to tell us more about it is Goodfire's co founder and CEO, Eric Ho.

欢迎，

Welcome,

谢谢邀请我。

Thanks for having me.

最近怎么样？

How's it going?

当然。

Of course.

很高兴你能来。

Excited to have you on.

在我们聊你最新的产品发布之前，先退一步，简单介绍一下Goodfire是做什么的吧。

I mean, we get into your latest product announcement, let's take a step back and just tell me a bit about what Goodfire does.

好的。

Sure.

Goodfire是一家AI可解释性研究公司，也就是说，我们整天都在思考AI模型内部到底发生了什么。

Goodfire is an AI interpretability research company, which means that we spend all day thinking about what's going on inside the mind of an AI model.

最终，我们的目标和努力方向是迈向一个有意设计的未来。

Ultimately, our goal and what we're building towards is a future of intentional design.

我们说的意思是，能够像编辑和调试书面软件那样去理解和修改模型，而不是像现在这样靠反复试验来训练它们。

So what we mean by that is actually being able to edit and understand and debug models much more like written software rather than training them by trial and error like we do today.

所以我们想消除塑造和训练这些模型时的猜测成分，真正深入地理解它们，最终能够完全信任它们。

So we wanna remove the guesswork from actually shaping and training these models and really be able to deeply understand them and be able to ultimately be able to deeply trust them.

我知道你过去提到过这个想法，就是窥探AI模型的内心。

And I know you've talked about in the past this idea of kind of getting a look into the minds of AI models.

这个领域也叫机制可解释性，名字挺拗口的。

This is an area otherwise known as mechanistic interpretability, quite a mouthful.

确实拗口。

It's a mouthful.

快说三遍试试。

Say that three times fast.

对。

Yeah.

没错。

Exactly.

那么，如果我们能做到这一点，这对我们未来开发和运用AI模型的方式意味着什么呢？

So you know, if we are able to achieve that, what does that mean about the way that we're going to develop and use AI models moving forward?

我觉得首先，一切从理解开始。

So I think to start, it starts with understanding.

也就是说，我们如何真正深入地理解一个模型为什么会做出某个预测？

So how do you actually deeply understand exactly why a model is making a prediction?

在数万亿的参数和神经元中，它们是如何组合在一起的？

So out of maybe trillions and trillions of parameters and neurons, like how do they all compose?

它们是如何被激活的，以至于我们能理解产生那个预测的每一步计算？

How do they all activate such that we understand every bit of computation that goes into that prediction?

这真的非常重要，因为这为在训练过程中塑造模型、真正调试问题奠定了基础。

And that really, really matters because this kind of sets the foundation for being able to shape them during training, to be able to truly debug their issues.

现在，假设你的模型喜欢谈论哥布林，或者喜欢去攻击Hugging Face之类的。

So right now, let's say that your model likes to talk about goblins or likes to go and hack hugging face or something like that.

我们真的没有办法准确指出具体的机制和计算，来解释它为什么会输出这样的结果。

There's really no way to actually be able to pinpoint the specific mechanisms, the computations, why it's actually kind of outputting what it's outputting.