Why Andy Jassy Sounded the Anthropic Alarm, Meta's ‘Tokenminimizing’, & Xbox Spin-Out Plans

欢迎各位来到《The Information》的TITV节目。

Welcome, everyone, to The Information's TITV.

我是阿卡什·帕什里卡。

My name is Akash Pasricha.

今天是6月15日，星期一。

It is Monday, June 15.

我们现在位于纽约的演播室，正在庆祝纽约尼克斯队夺冠。

We are here in our New York studio where we are celebrating the New York Knicks championship.

今天早上，我们还关注了几笔重大交易。

This morning, we are also watching a couple of big deals.

福克斯以两千二百亿美元收购了Roku。

Fox is acquiring Roku for $22,000,000,000.

这又是媒体行业的一次整合动作。

It is yet another media consolidation move.

每股160美元的股价，是Roku自疫情以来的最高价位。

The $160 share price is the highest that Roku would be trading since the pandemic era.

另外，Salesforce以36亿美元收购了AI客服公司Fin。

And in other deal news, Salesforce is buying AI customer service firm Fin for $3,600,000,000.

稍后我们会请我们的AI记者来聊聊这件事。

We'll talk about that shortly with our AI reporter.

今天节目中，亚马逊CEO安迪·贾西对Anthropic和Fable五提出了警告。

On today's show, Amazon CEO Andy Jassy has sounded the alarm over Anthropic and Fable five.

这一消息来自《The Information》的独家报道。

That is according to exclusive reporting from the information.

我们马上请到《AI议程》的作者斯蒂芬妮·帕拉佐洛，为大家详细解读。

We are going to bring on Stephanie Palazzolo, author of AI Agenda in just a second to break that down for us.

我们还会聊到Meta的最新动态——他们正在考虑限制员工使用AI。

We'll also talk about our Meta news that Meta is considering a move here to curb AI usage among employees.

我们的微软记者也马上上线，谈谈他挖到的独家消息：微软正在考虑把Xbox拆分出去。

We've also got our Microsoft reporter joining us shortly to discuss his scoop that the company has been thinking about spinning out Xbox.

最后，我会请两位创始人上节目，为本期节目收尾。

And I'm going to close out the show with two founders coming on.

我们接下来请到GPU基础设施公司Hydra Host，聊聊他们的融资进展，还会对话Abridge的联合创始人兼CEO拉奥博士。

We've got GPU infrastructure company Hydra Host coming on to talk about their funding round, and we will talk with Abridge cofounder and CEO, Doctor.

谈谈他和英伟达的新合作。

Rao, about his new play with Nvidia.

今天这期节目内容很精彩，咱们马上开始。

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

Fable Five在惊艳发布后，情况来了个大转弯。

Fable five had quite the U-turn after its impressive release.

特朗普政府出台了新的模型出口限制，因此Anthropic直接彻底关闭了该模型。

The Trump administration issued new export restrictions on the model, and Thropic, as a result, disabled the model entirely.

上周末，我的同事们打听到一些幕后讨论的细节，正是这些讨论促成了这个决定，我请来Stephanie Palazzolo，跟她聊聊她挖到的消息。

My colleagues over the weekend caught wind of the behind the scenes discussions that led to that decision, and I want to bring on Stephanie Palazzolo to share with us what she found out.

Steph，欢迎再次做客我们的节目。

Steph, welcome back to the show.

很高兴你今天能来。

It's great to have you here.

谢谢。

Thanks.

很高兴能来。

Great to be here.

好的。

Okay.

所以Fable Five发布了。

So Fable five comes out.

大家都夸它有多好，我们打听到安迪·贾西可能跟政府那边有过一些交流。

Everyone is trumpeting about how good it is, and we caught wind of a discussion that Andy Jassy may have had with folks in the government.

我们发现了什么？

What did we find?

是的。

Yeah.

正如你所说，这个周末对于Anthropic来说真是挺动荡的。

So as you mentioned, this weekend was a bit of a crazy one for Anthropic.

简单说，周五的时候，特朗普政府要求这家公司切断外国用户对Cloud Fable五和Mythos五的访问权限，这两款是Anthropic最先进的两个模型。

Essentially, you know, on Friday, the Trump administration said that the company would have to cut off access to Cloud Fable five and Mythos five, which are Anthropic's two most advanced models to foreign nationals.

结果Anthropic干脆直接对所有人关闭了访问权限，因为要区分谁是外国用户实在太难了，更别说公司里本身就有大量外国员工在参与这些模型的研发。

And then in response, Anthropic basically shut down access to the models for everyone because, you know, it's very hard for the company to tell who is and isn't a foreign national accessing the models, not to mention that there are a ton of those within the company itself actually working on the models.

所以上周末我和一些同事报道过，这件事最初其实源于亚马逊CEO安迪·贾西和特朗普政府高层的一次谈话，他向政府官员提了一些关于模型被破解和安全风险的担忧，结果事情越闹越大，到了周五，Anthropic不得不全面切断所有人的访问。

So as some of my colleagues and I reported over the weekend, a lot of this stemmed actually initially from a conversation that Amazon CEO Andy Jassy had with senior Trump officials, Trump administration officials, where he essentially kind of flagged some, you know, jailbreaks and other security related concerns that he had to the officials from the government, which then kind of snowballed into what we saw on Friday with, you know, Anthropic at the end the day having to cut off access to everyone.

那我们对安迪·贾西和特朗普政府之间的这次谈话，还有更多细节吗？

And so do we know any more details about this conversation between Andy Jassy and the Trump administration?

还是说，我们现在只知道这件事发生过，别的都不清楚？

Or do we is the extent of what we know just that it happened?

关于这次谈话，确实流传着各种不同的说法。

So it is you know, there are definitely different details floating around about the conversations.

综合我们和其他媒体的报道来看，安迪当时主要是向政府通报了一次亚马逊研究团队发现的模型破解事件。

Essentially, what we have gathered from ours and other outlet stories is it seems like Andy was basically flagging a jailbreak that Amazon researchers had found.

所谓‘破解’，就是用户设法绕过模型的安全限制，让它做出原本被禁止的行为。

And what a jailbreak is is essentially a way for a user of a model to kind of bypass the the guardrails on the model.

所以这次，Cloud Fable Five 的安全机制会阻止用户提问与网络安全相关的问题，因为这个模型在这一领域特别擅长。

So in this case, Cloud Fable five has guardrails that prevent, you know, from asking cybersecurity related questions since that's an area that the model is especially good in.

Anthropic 担心用户会利用这个模型去研究如何攻击网上各种网站和软件。

And Anthropic was worried that users would use this model to, you know, come up with different ways to hack, you know, websites and software floating out there.

看来亚马逊发现了一种能绕过这些安全限制的漏洞，并向特朗普政府的官员报告了这件事。

And so sounds like basically Amazon found a jailbreak that worked to get around those guardrails and kind of flagged this to Trump administration officials.

目前还不清楚 Andy Jassy 当时到底预期这件事会发展到什么程度。

Now it's not really clear kind of how much Andy Jassy expected this to escalate.

他很可能只是想提醒一下，让相关部门注意这个风险，根本没想到会闹到今天这一步——现在 Anthropic 索性直接对所有人关闭了这些模型的访问权限。

It's totally possible that he was just kind of flagging this as like something to look out for and had no idea that it was going to reach the point that it did where now you see Anthropic basically cutting off access to these models for everyone.

这里有个挺有意思的现象，因为亚马逊本身就是 Anthropic 的重要投资方。

And it is kind of an interesting dynamic here because Amazon, of course, is a big investor in Anthropic.

一方面，发现模型有漏洞并上报，AWS 确实有责任确保服务对客户来说足够安全。

I guess And on the one hand, flagging the fact that the model can be jailbroken, AWS certainly has obligations to its customers here to make sure that everything is as secure as possible.

但另一方面，我不禁好奇，Dario 和 Andy Jassy 当时的对话到底是怎样的？毕竟整个周末我都在用 Claude，屏幕上一直显示着‘Fable Five 目前不可用’，点‘了解更多’还能看到说明。

On the flip side, I doubt I wonder what that discussion between Dario and Andy Jassy was like, because, I mean, all weekend long, I was on Claude, and you had that bar saying Fable five is unavailable right now, and I guess you could click learn more to see what it is.

但我一直在想，你觉得这件事最后会怎么发展？

But I guess what I'm sort of thinking about here is how you think this plays out.

我们知道这次通话确实发生了。

I mean, we know that this call happened.

你觉得达里奥会不会说服安迪再打个电话给特朗普政府，说其实没什么大不了的？

Do you think maybe Dario convinces Andy to make another phone call to the Trump administration saying it's actually fine?

你觉得呢？

I mean, what do you think?

这确实让Anthropic陷入了非常棘手的境地。

Yeah, it's a really tough situation for Anthropic to be in.

退一步说，这也不是Anthropic第一次和特朗普政府产生矛盾了。

And just taking a step back, this isn't the first time that Anthropic has clashed with the Trump administration.

我们几个月来一直在报道这件事，其实Anthropic和政府之间一直有不少摩擦，早在今年早些时候，他们就因为Claude模型能否用于军方项目跟五角大楼争执过，核心问题是Claude在哪些情况下能用、哪些不能用。

Obviously, we've been reporting on this for months, but there's been a lot of kind of issues between Anthropic and the government, you know, sending all the way back to earlier this year when it was fighting against the Pentagon for the Claude models to be used in, you know, these cases where, you know, they were arguing basically about when Claude could and couldn't be used.

至于接下来会怎样，现在还不清楚，Anthropic已经派了几位高管去华盛顿，和特朗普政府方面商讨这件事。

And so, in terms of what happens next, it's not really clear, you know, right now, Anthropic has sent some of its leaders to to DC to discuss all of this with the Trump administration.

在我看来，我觉得Claude Fable Five很可能在下周左右重新上线。

And so in my opinion, I do feel like they're think it's semi pretty likely that we're gonna see Claude Fable five come back in the next week or so.

但显然，Anthropic得给特朗普政府一些让步，才能把这事谈妥。

But obviously, I think Anthropic is gonna have to make some concessions here to, you know, sweeten the deal for the Trump administration.

比如加强Fable Five的使用限制，甚至最极端的情况——暂时放缓公开发布，只保留一小部分测试用户。

And whether that is, you know, maybe upping the guardrails on Fable five, maybe, you know, even most drastic case scenario, like maybe rolling back the public rollout a bit and still keeping it to a pretty small group of testers.

最近还流传着一种说法，特朗普政府正考虑投资像OpenAI和Anthropic这样的顶级AI公司。

Another thing that's been floating around a lot is this idea that the Trump administration is looking to take stakes actually in, you know, the biggest AI labs like OpenAI and Anthropic.

这或许就是特朗普政府的一种谈判策略，想趁机拿下Anthropic的股份，毕竟它是目前估值最高的AI初创公司之一。

So maybe this could be kind of a negotiating tactic that the Trump administration is using to try to get a stake in Anthropic as one of the most, you know, highly valued AI startups out there today.

目前还不好说结果会怎样，但我认为，面对客户和投资者的双重压力，Anthropic必须在接下来几天或几周内找到出路。

So it really has yet to be seen, but I think just given the amount of pressure from customers, from investors on Anthropic, think they're going to have to find a way out of this in the next couple of days or weeks.

我想再回过头聊聊Anthropic和OpenAI之间的竞争。上周你提到，虽然Fable Five在客户中反响不错，但OpenAI也在加紧开发自己的最新模型。

I want to bring this back for a moment here to the race between Anthropic and OpenAI because last week, I believe you had reported that while Fable five was certainly making inroads with customers, and it had a pretty positive reception, you had reported that OpenAI has also been working on its own latest suite of models.

你能再跟我们说说，他们现在的模型进展到什么程度了？

Can you just remind us about where its model development efforts are at?

是的。

Yeah.

就像我上周提到的，Fable Five发布的时候，OpenAI可没闲着。

So as I had reported last week around the time that, you know, Fable five came out, OpenAI is obviously not just sitting around on its hands.

他们一直在开发下一代模型。

It is working on developing its next generation of models.

一些早期测试者告诉我，他们对OpenAI正在研发的下一代模型印象非常深刻。

And some early testers told me that they were actually very impressed by the next models that OpenAI is working on.

下一个版本很可能是GPT-5.6。

The the next one will be GBT 5.6 most likely.

所以从整体竞争来看，这短期内确实会给Anthropic带来压力。

And so, in terms of what this means for the broader race, this is obviously going to handicap Anthropic for a little bit.

但我认为OpenAI自己也可能有点担心，对吧？

But I think OpenAI is probably a bit worried about this too, right?

因为如果他们想推出一个能和Fable或Mythos比肩的模型——这可是他们跟上AI竞赛的必要条件——那就有很大风险，特朗普政府也会对OpenAI如法炮制，要求他们限制最先进模型的访问。

Because if they want to come out with a fable or mythos level model, which of course they're going to need to do to, you know, stay up in the AI race, there is very much this risk that the Trump administration is going to do the same thing to OpenAI and also tell it to, you know, cut off access to its most advanced models.

所以，我觉得OpenAI肯定在密切关注他们和Anthropic的谈判进展。

And so, you know, I think OpenAI is definitely keeping an eye out on this to see how the talks go with Anthropic.

不过说实话，历史上OpenAI和特朗普政府的关系一直比Anthropic要好。

Although to be fair, I think historically, OpenAI has had a better relationship with the Trump administration than Anthropic has.

没错。

Right.

我现在在想，我们都知道Anthropic在OpenAI里也有很大股份。

And I'm just thinking now, we of course know that Anthropic has a big stake in OpenAI as well.

我觉得，如果Dario和Sam Altman对Andy Jassy频繁联系特朗普政府感到不满，那他们最好把Andy的日程摸得清清楚楚，因为……

I think, you know, people just need to Dario If and Sam Altman are annoyed with the calls that Andy Jassy is making to the Trump administration, then I think they better have a good handle on his calendar because

是啊，我真想当个墙上的苍蝇，等咱们这篇报道一出来，听听Dario和Andy之间的对话，我敢打赌Dario肯定气得不轻。

Yeah, I would love to be a fly on the wall after our story broke and, you know, overhear the conversation between Dario and Andy because I'm sure Dario is not very happy.

我的意思是，如果OpenAI推出更强大的模型，同样引发类似担忧，也许也会被以类似方式突破限制，对吧？

Well, all I mean to say is, you know, if OpenAI comes up with a more powerful model, which poses similar concerns, maybe it can be jailbroken in similar ways, right?

我的意思是，我猜Andy Jassy这次和政府沟通时，态度可能会更谨慎一些，但对话内容大概还是差不多。

I mean, I can imagine Andy Jassy would have a perhaps maybe a more cautious conversation this time, but a similar conversation with the government.

总之，这件事发展得太快了，非常感谢你来参与，帮我们理清了头绪。

Anyway, it is all a fast moving story, and I want to thank you for coming on and helping us make sense of it.

这位是斯蒂芬妮·帕拉佐洛，《信息》杂志AI议程通讯的作者。

That is Stephanie Palazzolo, author of our AI Agenda newsletter, here at The Information.

曾经风靡一时的令牌最大化策略，如今已被Meta彻底扭转——我们独家获得了一份由记者乔蒂·曼撰写的报告，揭示了公司对员工AI使用政策的重大调整。

Token maxing had its moment, but a new exclusive report from the information's Meta reporter, Jyoti Mann, reveals the company has essentially gone the opposite direction now with its AI guidance for employees.

我想请乔蒂来跟我们详细说说目前的情况。

I want to bring on Jyoti to share with us more about what we know.

乔蒂，欢迎再次做客我们的节目。

Jyoti, welcome back to the show.

很高兴你能来。

It's great to have you here.

谢谢邀请我。

Thanks for having me.

能跟我们说说你拿到的那份内部备忘录吗？

Tell me about this internal memo that you got ahold of that ran at Meta.

当然。

Sure.

这份我们看到的内部备忘录显示，Meta在内部使用AI上已经投入了数十亿美元，现在打算通过推出一项效率计划来控制成本。

So this internal Meta memo, which we reviewed shows that it's spending billions on internal use of AI, and now it wants to rein in the costs by setting up an efficiency program.

过去一年里，公司一直鼓励员工大量使用AI，比如写代码、做研究，或者处理日常工作任务。

Now over the past year, the company's been encouraging employees to adopt AI heavily by using it for coding, research or everyday work tasks.

但这份备忘录表明，现在公司正转向一套更系统化的方式来追踪和管理这些开支。

But the memo shows that it's now moving towards a more structured system for tracking and managing those costs.

这实际上是Meta在看到AI使用量暴增之后，开始对内部AI工具施加财务管控。

So this is really Meta trying to place financial controls around internal AI tools after seeing just how quickly usage has exploded.

那这个系统化的计划具体是怎样的？

And so what is that structured plan here?

他们打算怎么限制token的使用？

How do they plan to clamp down on token usage?

他们正在开发一个叫AI Gateway的内部平台，可以实时追踪AI的使用情况和相关支出。

Well, it's building an internal platform called AI Gateway, which will track AI usage and spending in real time.

不同部门或团队将能查看自己消耗了多少令牌、花了多少钱，未来还会实行预算和配额制度，公司因此可以限制员工使用AI令牌的数量或支出额度。

So different orgs or teams will be able to see how many tokens they're consuming, what they're spending, and eventually they'll operate under budgets and allocation systems so that could see the company place limits on how AI tokens employees can use or how much they're able to spend.

这个计划的一部分还包括设置异常支出预警，以及引入更精准的预测工具。

And part of that plan actually involves trying to introduce alerts for unusual spending spikes and other better forecasting tools.

我对你这个报道中另一个觉得有意思的地方是，听起来他们其实是在鼓励员工多用公司自研的AI工具，而不是外部工具，这样能更好地掌控和监控使用情况。

And one of the other things I found interesting about your story is that it sounds like they're actually encouraging people to use their own AI tools more than external AI tools so that they can have more control or visibility, I guess, over the usage.

我理解得对吗？

Am I understanding that correctly?

没错。

Absolutely.

这个计划的一部分还包括逐步转向公司内部的工具，比如Meta Code这样的编程工具。

So part of this plan involves migrating and transitioning towards some of its in house tools like Meta Code, for example, which is a coding tool.

他们希望减少对外部服务商的依赖，转向自研工具，这样有望降低相关成本。

So they want to kind of reduce their reliance on outside providers and being able to transition towards in house tools will hopefully help them bring down some of those costs.

这个决定挺有意思的，因为如果Meta Code真那么好，我猜Meta早就把它当成产品对外销售了。

That's kind of an interesting decision, though, because if Meta Code was so good, then I imagine it would be a product that Meta would try to sell externally.

所以，假设一下吧——我本人从来没用过，也没写过一行代码，不敢说Meta Code就一定比Claude Code差，但要是真那么好，他们早该拿去卖了。

So I guess if we operate on the assumption, and again, I've never used it, I've never coded a thing in my life, so I don't know if we could say that Meta code is not as good as Claude code, but if it was that good, I imagine they would try to sell it.

这就像是在告诉员工：我们明白用这个可能影响效率，但咱们实在负担不起让你们用最顶尖的工具。

And so it's sort of like telling your employees, Hey, we understand this might make you less productive, but we can't really afford for you to use the latest and greatest.

这中间的权衡，确实有点尴尬。

That's sort of the awkward calculus here.

确实如此，也正因如此，新成立的应用AI工程团队才被赋予任务，去优化Meta Code，好让公司内部能更依赖它。

Absolutely, and that's why we're seeing the newly formed Applied AI Engineering org has been tasked with trying to improve Meta code so that they can rely on it more internally.

但我觉得这份备忘录最有趣的地方在于，它揭示了一个不只是Meta面临的难题——很多公司现在都遇到类似情况：AI在内部被用得太猛，账单高得吓人，现在只能想办法压降成本。

But I think what makes this memo so interesting is that it highlights a problem that not just Meta's dealing with, a lot of companies are starting to face where AI's been adopted so heavily internally that the bill's becoming enormous, and now it comes down to trying to reduce those costs.

不过，Jyoti，Meta的钱袋子可不瘪，虽然我们知道他们的资本支出确实惊人。

Now, Jyoti, Meta is not exactly hurting for cash, although we do know that its CapEx bills are extraordinary.

我觉得在这种时候讲求效率，是个明智的选择。

I think efficiency is a smart play here.

但咱们有没有想过，为什么Meta的AI代币使用量会这么快从‘尽量多用’转向了你故事里说的‘尽量少用’？

But do we have any idea why token maxing at Meta turned so quickly into what you call in a story as this notion of token minimizing?

我的意思是，真的是高层看到账单后才决定改变的吗？为什么事情会这么快就变了？

I mean, was it really just senior leaders seeing the bills, or why did things change so quickly?

我觉得部分原因可能是AI令牌的浪费使用。

I think in part it could be the wasteful use of AI tokens.

当初大家疯狂使用令牌的时候，公司高层其实一直在鼓励大家多用AI。

So back when token maxing was happening, there was a lot of encouragement from senior leaders at the company to embrace AI more.

有些员工就把这当成内部竞赛，争着上排行榜，好证明自己是AI重度用户。

And some employees were kind of turning that into the internal competition to climb a leaderboard that was around at the time to show that they were a power user.

这自然就推高了AI令牌的使用量。

And that inevitably drove up usage of AI tokens.

但现在他们的重点变了，讨论的核心不再是多用AI，而是怎么更有效地用，真正提升工作成果。

But now they're kind of shifting their focus and saying the conversation's more about not use more AI, but use it more meaningfully so that it improves outcomes.

这种转变，其实整个行业也开始出现了。

So that's a transition that we're beginning to see across the broader industry as well.

员工们对这种新的‘减少令牌使用’的模式反应如何？

How have employees been reacting to this new paradigm with token minimizing?

是的，说实话，我跟一些员工聊过，他们并不觉得意外。

Yeah, mean, well, it's not come as a surprise to some of the employees I've spoken to.

他们觉得这是必然的，毕竟有人确实浪费了大量令牌。

They saw this as sort of inevitable given that some staffers were using tokens in a wasteful way.

他们认为这项措施很合理，毕竟在这样的规模下，无限使用是不可能持续的。

So they see this move as quite sensible as they recognize that unlimited consumption's never going to be sustainable at this scale.

但对那些深度使用AI、并且用得很有意义的员工来说， tighter controls 可能会让他们感到不满。

But for some employees who have embraced AI heavily and are doing so in a meaningful way, there could be some frustration around tighter controls.

你觉得这会不会影响Meta在自家模型和产品上的进展？

And do you think that this could curb the progress at all that Meta has been making on its own models, its own products?

这其实又回到了我们之前讨论的那个问题。

This kind of goes back to the earlier question we were talking about.

Meta最终得推出更好的模型，推出更好的产品。

Meta ultimately needs to I mean, needs to get a better model out there, it needs to get better products out there.

这会拖慢这个进程吗？

Could this hamper that?

如果Meta只是在员工日常使用AI工具方面变得更加有纪律，这大概不会对开发造成太大影响。

Well, if Meta was simply becoming more disciplined about how employees use AI tools day to day, this probably isn't gonna hurt development very much.

反而可能提升效率。

It's probably just gonna improve efficiency.

更大的风险在于，如果这些限制扩展到从事AI模型研发的研究人员和工程师身上，他们可能无法自由地做实验，或者失去所需的计算资源。

The bigger risk would be if these cross controls extend to researchers and engineers who are using it for AI model development, where they might not be able to experiment as freely or lose access to the compute resources they need.

但我猜测，这些预算不会用在这些关键领域，用于AI模型开发等优先事项的计算资源应该还是会有的。

But I suspect that these budgets won't be placed on those critical areas and there will still be enough compute for priority areas like AI model development.

很好。

Great.

Jyoti，非常感谢你今天来做客。

Well, Jyoti, I want to thank you for coming on.

这位是Jyoti Mann，我们《The Information》的Meta记者。

That is Jyoti Mann, our Meta reporter, here at The Information.

微软正在考虑对Xbox部门进行一些重大调整。

Microsoft is considering some major changes to its Xbox division.

我们的微软记者亚伦·霍姆斯上周晚些时候独家挖到了一些内部讨论的细节。

Our Microsoft reporter Aaron Holmes scooped some of its behind the scenes discussions late last week.

我想请亚伦来跟我们分享他了解到的更多信息。

I want to bring on Aaron to share with us more what he knows.

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 here.

很高兴能来。

Happy to be here.

关于微软和Xbox，你打听到些什么？

What did you find out with respect to Microsoft and Xbox?

Xbox首席执行官阿莎·夏尔马正计划对整个游戏部门进行一次重大调整。

So Xbox CEO, Asha Sharma, is planning a pretty major overhaul of the entire gaming unit.

具体来说，我们知道裁员即将来临，她打算削减成本。

Specifically, we know that layoffs are coming and that she's planning to cut costs.

但我还了解到，夏玛计划在未来一年增加对游戏开发的投入，甚至可能比上一年 Xbox 的投入还要多。

But I also learned that Sharma wants to increase spending on game development in the coming year, potentially spending even more next year than Xbox spent in the past year.

特别是，她希望加快《光环》、《辐射》和《上古卷轴》等系列新作的开发进度。

And specifically, she wants to speed up some of the the new games that might be coming from franchises like Halo and Fallout and Elder Scrolls.

这些系列中有些已经十多年没出过新作了。

Some of those haven't had a new game in more than a decade.

与此同时，我听说，就在夏玛筹划 Xbox 的重大改革时，微软也在考虑重新调整 Xbox 与公司的关系，比如把它变成子公司，或者彻底分拆出去，以便未来更容易把 Xbox 卖给其他公司。

At the same time, you know, I've heard that as Sharma sort of plots this major overhaul of Xbox, Microsoft has considered, you know, potentially restructuring Xbox relationship to Microsoft to either, you know, make it a subsidiary or potentially spin it out entirely, you know, and possibly make it easier for Xbox to to be sold to another company.

我觉得短期内不太可能有行动，但我听说这些选项全都在讨论范围内。

I don't think there's any imminent plans for that, but I'm told that all of those options are on the table.

你觉得这是微软主导的决定，还是 Xbox 自己推动的？

Do you think this this is more of a of a Microsoft driven idea or an Xbox driven idea?

我挺好奇的。

I wonder.

我觉得两者都有吧。

You know, I think it's a mix of both.

我们注意到，Xbox的利润率远低于微软其他业务。

One thing that we've seen is that Xbox has much lower margins than the rest of Microsoft's business.

部分原因是他们整个硬件业务的利润率极低。

Part of that is because they have this entire, you know, hardware business that's extremely low margin.

近几年他们卖主机也挺困难，销量一直在下滑。

They've also had a hard time selling consoles in recent years, and those sales have been declining.

成本在大幅上升，不仅硬件方面，连员工的人力成本也在涨。

And costs have been rising significantly both in hardware, but also, you know, just in the the cost of all of the people that they employ.

这在一些微软股东眼里，已经成了个问题。

And it's kind of becoming a a problem in the eyes of some Microsoft shareholders.

所以确实有人认为，把Xbox拆分出去、卖掉，或者至少从财报里剔除，能让微软的股票对投资者更有吸引力。

So there is definitely an argument to be made that spinning out Xbox or selling it, or, you know, just taking it off the balance sheet would make Microsoft a more kind of attractive stock for investors and shareholders.

但同时我也听说，CEO萨提亚·纳德拉非常支持阿卡夏玛的愿景，真心希望Xbox能成功，甚至觉得这种重组可能是让Xbox比现在作为微软子公司更成功的途径。

But at the same time, I'm told that, you know, CEO Satya Nadella very much believes in Akasharma's vision and really wants Xbox to succeed and, you know, potentially sees this restructuring as one way to to make Xbox even more successful than it is as part of Microsoft.

你之前提到过利润率的问题。

Now you talked about the margins earlier.

现在的Xbox本身是盈利的吗？

Is Xbox right now as is, is it a profitable enterprise?

微软有没有给Xbox额外拨款？或者Asha Sharma的团队有没有独立的开发预算可用？

Does Microsoft provide any additional funding, or does it have a development budget at all that that Asha Sharma and company can can make use of?

Xbox本质上只是微软的一个部门，所以我们并不清楚它的具体盈亏情况，公司财报里也没明确披露。

So Xbox is technically just a division of Microsoft, meaning we don't actually know exactly what it's, you know, p and l looks like that the company doesn't clarify that in their earnings.

有可能Xbox的某些部分是靠微软其他业务在补贴着。

And it's possible that, you know, some parts of the business are being subsidized by other parts of Microsoft.

我们只知道，Sharma上周给员工发过一封邮件，后来她把这封邮件公开了。

What we do know is that Sharma actually said to staff in an email last week, which she then made public.

她把那封邮件发布了出来。

She published the email.

她说，他们的责任利润率是3%，这是一个内部指标，大概相当于EBITDA。

She said that they have accountability margins of 3%, and that's an internal metric that is, you know, roughly similar to EBITDA.

也就是说，他们的利润率大概是3%。

So that, you know, suggests that 3%.

就这些了。

That's it.

对。

Yeah.

这简直微乎其微。

That that's that's like nothing.

我的意思是，

I mean, like

没错。

Right.

薄得像刀刃一样。

That's razor thin.

是啊。

Yeah.

就勉强保本而已。

Just like a break even business.

有收入，但几乎微不足道。

Money, but barely any.

跟我说说，这整个游戏行业都是这样吗？

And tell me, is this like the gay like, the gaming sector at large?

我的意思是，Xbox 是不是个特例？还是说整个游戏行业在业绩上都经历了大起大落？

I mean, is is is Xbox kind of an anomaly here, or has gaming, broadly speaking, kind of being been been through the ringer as well in terms of business performance?

是的。

Yeah.

分析师普遍认为，游戏行业的毛利基准大概在15%左右，这比普通软件行业的利润率低得多。

I mean, the the general sort of benchmark according to analysts for, you know, gross margins in the gaming industry is, like, roughly 15%, which is still, you know, much lower than, you know, typical software margins.

但与此同时，看起来Xbox目前的利润率甚至可能比这个数字还低。

But at the same time, it it looks like potentially Xbox is even lower than that right now.

我觉得夏玛想解决的问题有两个：一是砍掉那些不太可能带来收入的开支，二是把资源集中在那些已有知名IP的领域，多投入开发粉丝们早已喜爱的系列新作，这些作品几乎肯定能带来可观的销量。

And and I think that's part of what Sharma wants to fix is, you know, a, cutting costs on things that don't seem likely to drive a lot of revenue and, b, spending in areas where, you know, they have these game franchises that are already known and beloved by fans and and trying to invest more in putting out new games from those franchises that are almost guaranteed to drive a lot of sales.

对。

Right.

艾伦，我得提醒大家，几个月前你写过一篇关于阿莎·夏尔马的精彩长文，我们会在节目笔记里附上链接，强烈推荐大家去读一读。

And Aaron, I should remind folks that you had a great weekend big read a couple months ago on Asha Sharma that we will link in the show notes and that I encourage everyone to read.

不过，能不能结合你当时报道的内容，再跟我们说说，上周关于游戏开发的那些变动，是怎么契合阿莎·夏尔马对Xbox的整体战略的？

But just remind us from that story and from the reporting you did, changes that you reported on with respect to the game development last week, how does that fit into Asha Sharma's broader strategy for Xbox coming into this role?

说实话，我觉得她现在走的是一条很微妙的路。今年年初刚上任时，她特别想向Xbox的玩家和自家员工证明，她不是来单纯裁员、压缩成本、只盯着利润率的。

You know, in a lot of ways, I think that she is kind of walking a fine line where on the one hand, you know, when she first came into the role earlier this year, she really wanted to prove to, you know, Xbox fans and also to her own employees at Xbox that she is not just there to, you know, wind this organization down or to cut costs and improve margins.

当然，这些事她也得做，但那不是她的唯一目标，她特别希望传达出这个意思。

Obviously, that is something that she does need to do, but it's it's not her sole focus or she really wanted to get that message across.

现在，她努力在维持当初的承诺——继续投资游戏开发，推出更多玩家真正想玩的新作。

And I think that now what she is trying to do is kind of maintain that promise that she wants to keep investing in game development and actually making more games that fans want to see.

但与此同时，她也得让这门生意真正可持续下去。

But at the same time, she does need to to make it a sustainable business.

从我上周做这组报道了解到的情况看，她可能打算减少对微软旗下那几十家小型游戏工作室的投入。

And she's been signaling, and from what I learned reporting this story this past week, I think that that might look like potentially spending less money on the dozens of smaller game studios that Microsoft owns.

我一点都不意外，未来几周或几个月，微软可能会关闭一些规模较小的游戏工作室，把资源更集中地投入到那些已经成功的大型IP上——顺便说一句，这些IP很多都是微软过去几年花数十亿美元收购来的，目标就是把它们打造成Xbox未来内容的核心引擎。

I wouldn't be surprised if we see Microsoft start to shutter some of its smaller game studios in the coming weeks or months while basically, you know, consolidating more money into the larger already successful franchises, you know, including many of the franchises that Microsoft has spent billions of dollars to acquire over the past few years, by the way, and basically making those more powerhouses of of new content for Xbox.

对。

Right.

我觉得你提到的一点特别关键，至少在短期内可能决定走向——正如你所说，萨提亚·纳德拉对阿沙·夏尔马很有信心，至少暂时会支持她把业务做起来。

And I think one of the most important things you mentioned that I think could sort of determine this, at least in the short term, is, as you said, Satya Nadella seems to have a lot of faith in Akasharma in terms of building the business, at least for a little while here.

这或许能为她争取一些时间，不至于立刻拆分业务，也能更快推动那些重大的调整。

Maybe that will sort of help buy some runway here in terms of not spinning it out immediately and helping to get some of those bigger changes underway quickly.

Aaron，非常感谢你今天来做客。

Aaron, I want to thank you for coming on.

这位是Aaron Holmes，我们《The Information》的微软记者。

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

GPU托管公司Hydra Host完成了1亿美元融资，估值达8亿美元。

GPU hosting company Hydra Host raised $100,000,000 at a $800,000,000 valuation.

本轮融资由Kindred Ventures领投。

The deal was led by Kindred Ventures.

现在邀请到Hydra Host的联合创始人兼首席执行官Ginn，来聊聊这一重要里程碑。

Joining me now to talk about the milestone is Hydra Host cofounder and CEO, Ginn.

亚伦，欢迎再次做客节目。

Aaron, welcome back to the show.

很高兴你来。

It's great to have you here.

是的。

Yeah.

谢谢你提到他们刚刚完成了史上最大一轮融资，800亿美元，对吧？

Thanks for announcing, I guess, the largest series they ever, right, with 800,000,000,000.

是吧？

Right?

所以我刚才在想

So I was

我本来想说，8000亿美元？

gonna say, I was like, 800,000,000,000.

如果我们真有这个数字，你绝对能排在节目头条

I think we would've put you at the top of the show if

如果这算的话。

if that's something.

对。

Yeah.

是的。

Yeah.

我们有埃隆。

We have Elon.

对吧？

Right?

没错。

Yes.

是的。

Yeah.

没错。

Yeah.

好的。

Okay.

八亿美元。

Dollars 800,000,000.

这可不是小数目。

Look, it's not chump change.

这是一笔大额融资。

It's a big round.

恭喜你。

Congratulations.

我们之前请过你几次做客节目，但从来没专门聊过你的公司。

We've had you on the show a couple times before, but never actually to talk about your company.

所以再跟我们说说，公司现在发展到什么阶段了？

So remind us, where is the company at right now?

你们卖的是什么？

What do you sell?

我们稍后会谈到发展方向。

And we'll talk about the direction shortly.

是的。

Yeah.

没错。

That's true.

对。

Yeah.

我们一直都在讨论政策，以及全球范围内与GPU相关的各种变化。

We've always been talking about policy and different things happening across the world with GPUs.

不过，我们五年前创立这家公司时，目标是为数据中心提供一套软件，帮助它们更好地服务终端客户。

But so, yeah, we started the company over five years ago with a goal of providing a software for data centers so that they could serve end customers.

我们当时以为，租赁类型的编排层会逐渐整合，而公有云作为底层基础设施的概念会慢慢衰退。

And we thought that the rental kinda orchestration layer was gonna compress and that public cloud was going to start dying as a concept of the underlying infrastructure change.

但这个观点还处于非常早期的阶段，因为我们看到很多AI公司都在重复加密货币矿工的做法——离开云平台，以获得对硬件的更大控制权，同时也更好地掌控网络和编排层面。

And and we're very early in that thesis because we saw a lot of AI companies doing the same thing that crypto miners were doing, which was that they're leaving the cloud to get more control of the hardware, also have more control of the basically networking orchestration kind of layer.

他们需要更高的性能吗？

Do they need more performance?

但当时的数据中心根本没做好准备。

But data centers were not prepared for that.

他们过去只是公有云体系里的一个普通零件。

They were previously just a cog in the wheel of public cloud.

在交易中，他们根本无法获得大部分收益。

They were not in a position to make most of the upside in the transaction.

他们只是个成本中心。

They were just a cost center.

所以今天，我们已经是全球最大的新型云服务商之一。

And so we today are one of the largest Neo Clouds in the world.

我们真正提供的，其实是一个基础设施即服务层，而我们是一家软件公司。

And what we provide purely actually is a an I an IS layer and infrastructure as a service layer where we are a software company.

我们是目前最大的纯软件型新型云服务商，而硬件资产则由数据中心运营和拥有。

We are actually the largest Neo Cloud that actually is a software company, and the iron relying assets are run and owned by the data center.

他们现在有能力获得大部分收益，因为他们打算建设更多的数据中心，而这些都是我们所需要的。

And they are in the position to make the majority upside because what they're gonna do, they're gonna build more data centers, which we all need.

我们提供全部的技术支持、软件服务和客户交付，让他们不必自己从头搭建一套系统。

So we provide all the expertise and software support and customer delivery for the data center so they don't have to go build their own stack.

他们只需要购买 Hydra，把它部署在自己的基础设施之上，就能直接进入市场了。

And they just basically buy Hydra, plug it on top of their own infrastructure, and then now they can go to market.

没错。

Right.

我有个问题想问你。

Can I ask you?

你们和英伟达是紧密的合作伙伴。

You you guys are a a close partner of Nvidia.

你们只用他们的芯片吗？

Are you only using their chips?

你们也会用其他公司的芯片吗？

Are you using other companies' chips as well?

所以我们做的，是帮助数据中心完成采购。

So, yeah, what we do is we help the data center procure.

你可以把我们想象成一家特许经营公司。

So you can think that we're a little bit like a franchise company.

我常跟数据中心说，我们打造的是赛百味那样的品牌。

So I like to I like to tell data centers we build Chick fil A.

我们不搞塔可钟那样的。

We don't build Taco Bell's.

虽然塔可钟确实很老牌。

Even though Taco Bell is super OG.

塔可钟确实很老牌，但我现在再也不去吃了。

Even though Taco Bell is super OG, I don't eat there anymore.

不过上大学那会儿，我每周肯定要去好几次。

But in in college, I definitely did multiple times a week.

但现在有

But There's

今天吃塔可钟也没什么不对。

nothing wrong with eating Taco Bell today.

好的。

Okay.

我们继续说下去吧。

Let's keep going there.

没错。

Yeah.

如果你从食谱的角度来看，它可能是世界上最有意思的公司，因为他们就那六样东西，反反复复地重新组合。

It it it is if you think about, like, recipe wise, it is probably the most amazing company in the world because they have, like, the same six things that they just kinda remake over and over again.

所以，是的，没错。

But the so so yeah.

就像餐饮连锁或者酒店连锁一样，我们负责所有的采购。

So just like how your could be a food franchise or, like, a hotel franchise, we control all the procurement.

我们的采购方式完全跟市场趋势走。

And the way we do procurement is, like, it reflects the trend of the market.

我们确实拥有全球最顶尖的数据，因为每个数据中心都是独立运营的，各自有独立的运营成本结构。

So we do have the some of the best data in the world because every one of our data center locations is a independent operating data center with its own OpEx cost structure.

因此，我们能真实反映全球GPU价格的变动趋势。

And so we get to reflect a lot of the true pricing trends around GPUs.

没错。

Right.

你最近看到的那些供不应求的趋势，其实我们从去年年底就开始大力扩建了，因为那时我们就注意到利用率明显上升了。

So a lot of the the the current trends you were seeing about, like, basically, the sold out kind of orientation, we basically started increasing builds really aggressively at the end of last end of last year because we started seeing the utilization increase.

我们网络中GPU的分配方式，完全跟随市场动态，也就是杰森所描述的那样。

But our our GPU allocation across the network in terms of what we advised is reflects the market, which is at Jensen So

那现在TPU有没有成为新的讨论焦点？

is is there more is there more TPUs coming into the conversation then?

没有。

Or more No.

没有。

There's not.

再说一遍？

Say it again?

没有。

They're they're not.

对，没有。

No.

所以，没有。

So No.

他们确实没有。

So they yeah.

没有。

No.

没有。

There's not.

那种芯片的使用方向非常有限，就只有两个客户在用。

So so they they are like I would say there I would say the the orientation around that kinda chip is very isolated to, like, two customers.

而且 Beyond that，客户群体到底是谁，跟GPU或者OEM厂商卖的配套GPU相比，根本说不清楚，反正我们是支持他们的。

And and and it's really unclear the customer base beyond that versus the customer base for GPUs or, like, let's say, like, associated GPUs you can buy from an OEM, know, like like, like, we support them.

AMD，我们在我们的网络里用他们。

AMD, we have them in our network.

但那些专注超大规模数据中心的，通常都是采取防御性策略。

But the kind of hyperscaler focused ones, they're generally a defensive mode strategy.

所以对那些以盈利为目的的数据中心来说，这么做根本说不通。

So so it doesn't really make sense for data centers who are in the business to make returns.

他们干嘛要保护那个之前抢走他们所有生意的人呢？

Why would they defend the person that took all of their lunch before?

所以这确实没道理。

Like, so so it doesn't yeah.

所以我觉得，这种产品的市场其实很小。

So so I think the market for that product actually is quite small.

尽管那两家客户——Anthropic和Google——规模可能真的很大。

Even though those two customers Anthropic and Google could be actually quite huge.

你觉得Vera Rubin会对你的业务带来什么影响？客户对Vera和Blackwell的接受度会怎么分布？

How do you see Vera Rubin changing your business and also the makeup of the workloads in terms of how many customers will use Vera, how many are okay with Blackwell?

你怎么看这个问题？

How do you think about that?

确实如此。

Yeah.

我们跟数据中心说，把GPU当成买车来考虑。

We what we tell tell data centers is to think about GPUs like buying a car.

我个人是奥迪的粉丝，所以也是大众集团的拥护者。

If and so I'm personally an Audi fan, so I'm a Volkswagen Group fan.

我买保时捷或者奥迪的时候，期待的体验跟买大众是不一样的。

So when when I go get a Porsche and an Audi, I expect a different type of technology experience than when I get a Volkswagen.

但这并不意味着大众就没什么价值。

But it doesn't mean that the Volkswagen itself is not worthwhile.

我觉得这个市场很多人没搞明白，是因为他们没看清H100的使用率和回报率。

And I I I think for this market, I think not people understand it because of the utilization and the return profile of h one hundreds.

如果你在最初那波热潮时买了H100，现在你肯定是市场上回报率最高的GPU持有者之一。

If you had bought h one hundreds back during the original boom, you're literally making the best return profile probably to any GPU right now in the market.

这其实也跟买车的道理差不多。

And and so that's kind of the way cars work too.

高端车型就像大众旗下的旗舰产品，简单来说，这个逻辑适用于所有品牌——雷克萨斯、丰田、宝马、福特、雪佛兰，它们都有各自的高端系列。

Is that the the the halo product, which is introduced in the Volkswagen, just an easy way to understand this, but it applies to all cars, like Lexus and Toyota and and BMW and all of its excellent, you know, x and the number lineup, you know, Ford, Chevy.

它们本质上做的事都差不多：高端车型带动整个公司的技术升级，然后技术再慢慢下放到底层产品线。

It's like they all kinda do the same thing, which is that the HALO product actually drives the the, you know, the overall technology of the company, and then it diffuses down the line.

大众车才是销量主力，但保时捷率先推出了新技术。

And so the the bulk and the meat potatoes resides in Volkswagen, but Porsche now introduced the new technology.

对客户群体来说，这两者完全不同，应用场景也不一样，也就是工作负载不同。

So for the customer base, it's completely different, and the use cases are different, aka workloads.

当Vera Rubin发布时，它会吸引那些现在用A100或者大型P300集群的人吗？

So when VeriRuba comes out, it's is this gonna draw in the same people who are using GBs or a very large p 300 clusters.

他们会升级，因为他们的使用场景跟那些跑普通任务的人根本不一样。

They're gonna upgrade because their use cases are are different than the guy that's running.

他们做的批量差异计算任务，更像大众汽车那种。

They do us a batch differencing workloads, which is more like the Volkswagen.

所以这项技术还是会往下渗透，但升级周期其实正是客户真正想要的。

So so the the technology will continue to downstream through the stack, But the upgrade cycle is basically part of actually what customers really want.

当然，有人会抱怨，比如数据中心的人或者其他相关方。

Know, they people can complain about it and people, like, you know, data centers or whoever.

但说到底，这跟你们汽车经销商的情况是一样的。

But it's like the same thing, I think, like yourself, like your car dealership.

那种客户属于完全不同的类型，比如前沿实验室的客户。

Like, that customer is just a different type of customer, like, the Frontier Lab customer.

他们总是会需要，或者像那些大型开源公司一样。

Like, like, they're always gonna want it or or a very large scale open source company.

但市场上真正的主流和核心，还是这些推理平台。

But the, you know, like, the bulk and the meat and potatoes in market is gonna be a lot of these inferencing platforms.

他们会用，但主要用的还是中高端的显卡组合，规模很大的集群。

They're gonna use it, but they're mostly gonna use the, you know, advanced midlife kind of card sets, very large scale clusters there.

而让客户能够自由选择，反而让中端市场也能接触到最顶尖的技术。

And and they they having the ability for customers to pick and choose actually creates the ability for the midlife market to have access to the the best stuff.

但这需要时间，因为实际情况是……

It's just gonna take like Because it actually

确实，关键问题在于算力资源有多紧张。

it actually because, I mean, capacity is the big issue here.

你甚至能搞到芯片吗？

Can you even get your hands on the chips?

我明白你的意思。

And so I see your point.

这实际上为市场增加了流动性，没错。

It sort of adds liquidity to that market where Yeah.

你或许拿不到最顶级的芯片，但那些性能足够强的型号依然值得重点关注。

You can get your your your hands on those chips that, I mean, they might not be best in class, but they're powerful enough, so they are certainly one to focus on.

亚伦，非常感谢你今天来做客。

Aaron, I want to thank you for coming on.