优步考虑收购SpotHero，xAI新建超大规模数据中心，英伟达预测 | 2025年12月31日

欢迎各位收看《信息》的TI TV。

Welcome everyone to The Information's TI TV.

我是阿卡什·帕什拉查。

My name is Akash Pashracha.

今天是12月31日，星期三。

It is Wednesday, December 31.

我们距离新年仅剩几个小时，今晚将带来一场精彩的节目，为2025年画上句号。

We are just hours away from the New Year and we have a fun show to close out 2025.

首先，《信息》发布了关于优步正在考虑的一项潜在收购的新报道。

First up, The Information has new reporting around a potential acquisition that Uber is considering.

此外，我们还将讨论xAI计划建设的第三个大型数据中心。

And separately, we are talking about a third big data center that xAI is looking to build.

接着，我们将头脑风暴一下英伟达明年可能在并购方面关注哪些领域，还会为您带来另一个关于甲骨文2026年云业务的精彩预测。

We'll then have a bit of a brainstorm about where NVIDIA could look for more M and A next year, and we have another great prediction for you about Oracle's cloud business in 2026.

我们还独家报道了布鲁克菲尔德在云业务方面的雄心壮志，节目最后将播放一段精彩集锦，回顾TI TV在2025年的最高光时刻。

We also have some exclusive reporting about Brookfield's ambitions in the cloud business, and we're going to end the show with a fun highlight reel recapping TITV's biggest moments of 2025.

这将是一场精彩纷呈的年终节目，让我们马上进入正题。

It's going to be a great show to close out the year, so let's get right on into things.

《信息报》独家报道，优步正在考虑收购停车应用SpotHero，该服务允许司机预订停车位。

The Information has exclusive reporting that Uber is considering buying parking app SpotHero, a service that allows drivers to reserve parking spaces.

今天来和我们聊聊我们所知情况的是我们的财经部门副局长科里·温伯格。

Here to talk about what we know is Corey Weinberg, our Deputy Bureau Chief of Finance.

科里，欢迎再次做客我们的节目。

Corey, welcome back to the show.

很高兴你来到这里。

It's great to have you here.

谢谢，阿卡什。

Thanks Akash.

新年快乐。

Happy New Year.

新年快乐。

Happy New Year.

好吧，SpotHero 是一家公司。

Okay, so SpotHero is a company.

我之前不知道这个。

I didn't know this.

也许我曾经在那里停过车，但自己却不知道用过它。

Maybe I've parked there and I have used it unknowingly, perhaps.

嗯，阿卡什，我现在是洛杉矶人了，经常要找停车位，所以我对 SpotHero 非常了解。

Well Akash, I'm an Angelino now, and I have to find parking often, so let me tell you, I know SpotHero.

你了解 SpotHero，好的。

You know SpotHero, okay.

你手机上有这个应用吗？

You got the app on your phone?

有，我有。

I do, I do.

好的。

Okay.

所以你提前预订车位，这正是Uber非常想支持的业务吗？

And so you reserve the spot, and this is what, a business that Uber just really wants to get behind?

我们在这里知道些什么？

What do we know here?

是的。

Yeah.

实际上，我们今天早上刚报道了这件事。

So we just reported this this morning, actually.

Uber正在考虑收购SpotHero。

Uber is considering buying SpotHero.

我们还不知道具体的收购价格，但这件事对我们来说有几点值得注意。

We don't know the exact price yet of it, but it's notable to us for a couple reasons.

一是SpotHero是为数不多的能够长期存活下来的初创公司之一。

One is just SpotHero is one of these startups that has that has endured.

它已经存在很久了。

It's been around.

它自2011年就一直存在了。

It's been around since 2011.

它是一家罕见的消费者停车公司，挺过了科技界普遍认为停车行业已死的阶段——无论是网约车还是自动驾驶汽车，都没人需要停车了。

It's a and it's one of the rare sort of consumer parking companies that has outlasted sort of the stages of everyone in tech saying that parking is dead and that whether it's ride sharing or whether it's autonomous vehicles, no one will need to park anymore.

SpotHero成功证明了这种观点是错误的，并且我们认为，它通过帮助人们预订停车位，建立了一个相当不错的业务。

SpyHero has been able to prove that wrong and build a pretty nice business, we think, getting people to be able to reserve parking spaces.

所以我在想，Uber进入停车行业，这是否意味着，当我打开Uber应用时，这只是一个我可以点击的额外功能？

And so I'm just trying to think here, Uber getting into the parking business, what is this, a horizontal play here in terms of when I open up the Uber app, like another feature that I can click basically?

他们已经长期在开发这些功能了。

They've been building these features out for a while.

我的意思是，如果你现在打开Uber应用，显然叫车和配送是两大主要功能，但他们已经与一家名为Turo的汽车租赁公司建立了合作关系。

I mean, if you go in your Uber app now, obviously, ride hailing and delivery are the two main uses, but they have a partnership with a a car rental company called Turo.

他们在欧洲已经推出了预订火车票的功能。

They've rolled out in Europe being able to book train tickets.

你甚至可以预订滑雪缆车票，这有点奇怪，哇。

You can book ski lift tickets, weirdly enough Wow.

优步。

Uber.

你知道的。

You know?

是的。

Yeah.

我们想打造一个涵盖所有出行和配送服务的平台。

It's sort of a we wanna be the everything mobility app, the everything delivery app sort of sort of play.

但我认为这也是对另一个现实的承认，即人们乘坐网约车或租车的次数，在整体出行次数中占比极小，尤其是在美国。

But I think it also is an acknowledgement of another reality, which is the number of ride hailing trips or car rental trips that people take is a very tiny percentage of the overall trips that are taken, particularly in The US.

绝大多数出行都是人们自己开车从A点到B点，而这在两端都需要停车位。

Vast, vast, vast, vast, vast majority are people driving themselves from point A to point B, and that requires a parking space on either end.

很好。

Great.

科里，你的报道非常出色。

Well, Corey, it was some great reporting.

我期待看到最终会形成什么样子。

I look forward to seeing what actually does shape up.

这是科里·温伯格，我们在《The Information》的副首席财务官。

That is Corey Weinberg, our Deputy Raradero Chief of Finance here at The Information.

好的。

Okay.

XAI 已经拥有两个大型数据中心来支持其人工智能运营，我的同事西奥·韦特的一篇独家报道揭示，该公司正计划建设第三个大型设施，以继续采取与其它顶级人工智能模型公司截然不同的战略。

XAI already has two giant data centers powering its AI operations, and a new exclusive story from my colleague, Theo Wait, reveals that the company is planning a third giant facility as it continues to pursue a markedly different strategy than other top AI model companies.

我想请西奥来跟我们详细说说他了解到的情况。

I want to bring on Theo to tell us more about what he's learned.

西奥，欢迎再次做客我们的节目。

Theo, welcome back to the show.

很高兴你来这里。

It's great to have you here.

很高兴能回来。

Good to be back.

那么，关于这个第三个数据中心，我们了解些什么？

So what do we know about this third data center?

它现在是一个仓库，前面停着一堆卡车，

It's currently a warehouse with a bunch of trucks in front of it,

但那不就是每个数据中心的样子吗？老兄。

but That's like every data center, though, man.

得了吧。

Come on.

嗯，其实不是的。

Well, no, actually.

因为有些数据中心必须从零开始建造，

Because some of them they have to build from the ground up,

而这个

and the

下一个

next

有人正在进去购买那些原本为其他用途建造的设施，并将它们改造过来。

guy is going in and and buying these other, you know, facilities that are built for other things and converting them.

但没错，这规模真大。

But, yeah, it's huge.

它有81万平方英尺，紧邻着另一个仍在建设中的大型XAI数据中心。

It's 810,000 square feet, and it's right next to another huge XAI data center that is still under construction as well.

这个第三设施叫什么名字？

And what's the name of this third facility?

在我们的报道发布后，马斯克发帖确认了这个消息，并称它叫MacroHarder。

Well, after our story came out, Elon tweeted about it and confirmed the story and said it was called MacroHarder.

XAI附近那个第二数据中心叫MacroHard，这个词最初是马斯克对微软名字开的一个玩笑，你大概也能猜到，是个孩子气的玩笑。

XAI's second data center that it's close to is is called MacroHard, which is a term that kind of started out as Elon making a joke about Microsoft's name, you know, a childish joke as as you can probably tell.

但现在这已经变成一件正经事了，他用这个词来指代XAI在AI软件开发方面的努力。

But now it's actually a serious thing, and it's what he's calling XAI's attempts to build software with AI.

我问你个问题。

Let me ask you something.

作为一名记者，当埃隆·马斯克在推特上确认你的报道时，是什么感觉？

What does it feel like as a reporter when Elon Musk confirms your story on Twitter?

我的意思是，你看到了那条推文。

I mean, you see that tweet.

你心里在想什么？

What what what's going through your mind?

我的意思是，这感觉很棒。

I mean, it was cool.

你知道，这种事并不常发生，因为我联系这家公司求证时，他们会给每个记者发自动回复，说传统媒体在撒谎。

Like, you know, it doesn't happen that much because, you know, I reach out to this company for a comment, and they have an autoresponder for every journalist that says legacy media lies.

以前是这样，我不知道我们算不算传统媒体。

It used to be a I I don't know if we count as legacy media, though.

大概不算，因为我们没有

I guess not because we didn't

撒谎在

lie in

这种情况。

this case.

但在此之前，它曾经是一个便便表情符号。

But it used to be a poop emoji before that.

所以，通常来说，他们对记者抱有一种非常对抗的态度。

So, you know, generally, there's this very confrontational attitude toward reporters and that they have.

因此，当我看到他们承认我们报道的内容是正确的时，这非常罕见，我也很感激。

And so when I see them acknowledge that something we reported is correct, it's it's rare and and, you know, I appreciate it.

我应该说，这件事发生得也很快。

And I should say, it happened pretty quickly too.

我的意思是，就在你的报道发布几分钟后。

Mean, was minutes after your story was released.

我的意思是，很明显他在阅读，这正是我想表达的重点。

I mean, clearly he's reading, is the point that I'm making.

希望他真的读了我们的报道，而不仅仅是看到了关于它的推文。

Hope he read our story and didn't just see a tweet about it.

我希望如此。

I hope.

不，我认为你读过这篇报道。

No, think you've read the story.

时间点太完美了。

The timing is too perfect.

好的，你在报道中提到的一件事是，xAI 与 OpenAI 和 Anthropic 等其他快速发展的模型公司所采取的策略有所不同。

Okay, so one of the things you talk about in your story is the differing strategy here that xAI is taking compared to other fast growing model companies like OpenAI and Anthropic.

你能稍微谈谈，为什么他们选择自己建设数据中心，几乎是从零开始，而其他人工智能公司——我不是说那些原本就有数据中心的超大规模云服务商，也不是说 Meta，而是 OpenAI 和 Anthropic——似乎大量依赖合作伙伴来完成这件事，对吧？

Talk a little bit about why they are moving to build their own data centers and basically from the ground up, whereas other AI companies, and I'm not talking about the hyperscalers, the cloud companies that obviously have data centers for other purposes, or Meta for that matter, but OpenAI and Anthropic, they seem to be working with a lot of partners to do this, right?

这是一种略有不同的做法。

It's a bit of a different approach.

是的。

Yeah.

我们有这么多名字，比如 Crusoe、CoreWeave、Oracle，这些公司都与 OpenAI 有关联，各自承担不同任务，但最终都是为了帮助他们实现算力上线的目标。

We we, you know, have all these names, Crusoe, CoreWeave, like Oracle, you know, all these companies that we associate with with OpenAI that, you know, do different things, but ultimately are kind of part of their mission of of getting his compute online.

而且，你知道，亚马逊正在为Anthropic建造数据中心。

And then, you know, there's also Amazon building a data center for Anthropic.

你知道，这类交易到处都是。

You know, there there's all kinds of these deals going on.

XAI并不是完全独立的，但相比这些其他公司，它的独立性要强得多。

And, you know, XAI is not entirely independent, but is definitely more independent than these these other companies.

我的意思是，他们直接去购买房产并自行开发，或者租下房产然后自己开发。

I mean, they are going in and, you know, buying properties and developing them or, you know, renting and buying properties and developing them themselves.

这意味着他们可以采取一些激进甚至有些鲁莽的行动，而其他公司可能不会觉得舒服这样做。

And, you know, that means they can be kind of aggressive and and, you know, like brash sometimes in a way that other companies wouldn't be comfortable doing.

比如，把数据中心建在孟菲斯，这本身就是一个大胆的举动。

I mean, like, putting the data centers where they are in in Memphis, you know, is a bold move in itself.

通常情况下，公司不会在人口百万的城市里建造数据中心。

Like, companies don't usually build data centers in cities that have a million people in them.

你知道，Meta最大的数据中心项目位于路易斯安那州一个极其偏远的地区，那里甚至连给建筑工人住的酒店都没有。

You know, Meta's biggest data center project is in a part of Louisiana that's extremely rural, and there aren't, like, even hotels for construction workers building it to stay in.

所以在城市里建设，让XAI获得了大量基础设施资源。

So building, you know, in a city has given XAI access to a lot of infrastructure.

这让他们能够找到现有的建筑，比如仓库和工厂，将其改造为数据中心，从而更快上线；但同时也使他们面临大量审查和反对。

It's let them find, you know, existing buildings like warehouses and factories that they can transform into data centers and, you know, get online faster, but it's also kind of exposed them to a lot of scrutiny and opposition.

这似乎在某种程度上体现了埃隆·马斯克的行事风格——他希望对更多事物拥有更多控制权。

This seems to be sort of part of, you know, Elon Musk's way of operating, is that he wants to have more control over more things, essentially.

他不想把任何事交给别人，以免自己对时间表或其他事情的掌控力减弱。

He doesn't want to leave anything to somebody else, you know, where he might not have as much say over the timeline or or something like that.

没错。

Right.

我是说，前几天我正好在工地外面，看到圣诞节到新年期间，那里有大量工人在忙碌。

I mean, know, I was I was at I was, like, you know, outside the site the other day, and you could see there were tons of people working between Christmas and New Year's.

我只是觉得，埃隆不相信，如果雇别人来做，他们真的会24小时不间断地按这种节奏工作。

And, you know, I just I don't think Elon trusts that if you hire someone else to do that, they're really gonna be out there twenty four seven on on that kind of schedule.

你知道的？

You know?

我觉得他更喜欢自己掌控。

I I think he feels better having it in house.

现在，我再问你一个问题。

Now, let me ask you one more question here.

XAI的这些数据中心，现在有两个，马上要第三个了。

These data centers that XAI, there's two of them and now there's a third.

我们是否知道这些数据中心只供XAI使用？

Do we know are these only going to be for XAI's operations?

埃隆有没有更宏大的计划，把这些数据中心也用于他的其他公司，甚至其他公司？

Does Elon have sort of a broader agenda here in using these for his other companies or maybe companies outside of his empire too?

你知道，理论上，它们当然可以用于特斯拉或SpaceX，而特斯拉和SpaceX也确实在将一些XAI服务整合到他们的产品中。

You know, I I think theoretically, yeah, they they could be used for Tesla or SpaceX, and and, you know, both Tesla and SpaceX are, you know, integrating some XAI services into their products.

但据我所知，目前这些数据中心主要还是用于开发最新的Grok模型。

But, you know, for now, it's it's pretty much all to develop the the latest Grok models as as far as I know.

很好。

Great.

好了，西奥，感谢你前来做客。

Well, Theo, I wanna thank you for coming on.

这位是西奥·韦特，我们《信息》杂志上报道埃隆·马斯克的记者。

That is Theo Waite, our Elon Musk reporter here at The Information.

新年快乐。

Happy New Year.

新年快乐。

Happy New Year.

好的。

Okay.

上周英伟达与Grok的大笔交易并不是今年该公司做的第一起人才收购。

NVIDIA's big Grok deal last week wasn't the first acquihire that the company did this year.

它还收购了另一家名为Enfabrica的初创公司，而在几个月前，它收购了芯片软件公司CentML。

It also did that with another startup called Enfabrica, and a few months before that, it bought chip software company CentML.

这些交易规模小得多，但所有这些反思让我产生了一个疑问。

Those were much smaller deals, but all of this reflection has prompted a question in my mind.

NVIDIA 在2026年可能会收购什么？或者至少，它可能会瞄准哪家公司进行另一次收购兼团队并购？

What could NVIDIA buy in 2026, or at least who might it look to for another acquihire?

为了帮助我们头脑风暴，我想邀请创意策略公司的高级分析师奥斯汀·莱昂斯。

To help us brainstorm, I want to bring on Austin Lyons, a senior analyst at Creative Strategies.

奥斯汀，很高兴见到你。

Austin, it's great to see you.

欢迎再次做客我们的节目。

Welcome back to the show.

嘿，卡拉什。

Hey, Kalash.

谢谢邀请我。

Thanks for having me.

上一段我没好好向西奥道新年快乐，所以我想先向你道一声新年快乐。

I didn't wish Theo Happy New Year properly in the last segment, so I'm going to start by saying Happy New Year to you.

呃，谢谢你，先生。

Well, thank you, sir.

好吧，让我们来头脑风暴一下。

Okay, so let's brainstorm here.

你认为英伟达明年可能会收购什么？

What do you think that NVIDIA could buy next year?

这是个很好的问题。

So that's a great question.

我认为，如果你看看英伟达的历史，我们可以推测一下，他们一直有收购技术公司以获取人才和知识产权的倾向。

And I think if you look at NVIDIA and if we just speculate here, they have a history of buying technology companies for their talent and for their intellectual property.

当然，从Mellanox时代就开始了，那次收购让他们获得了横向扩展网络和纵向扩展网络技术。

Obviously, going all the way back to the Mellanox days, which gave them scale out networking and scale up networking.

一个有趣的方向是，如果你在思考网络的未来，那么网络下一步会往哪里发展？

One interesting so if you ask, thinking about networking, if you say, where is networking going next?

我们已经看到，在2025年，英伟达提到他们将开发用于横向扩展网络的共封装光学交换机，而横向扩展网络已经通过光纤传输光信号进行通信。

We've already seen in in 2025, NVIDIA mentioned that they are going to build co package optics switches for scale out networking, which scale out networking already communicates using photons over optical signals over fibers.

但是

But

给我们降低一个层次解释一下。

Bring bring it it down one one level for us.

如果我们向一个不懂技术的人解释的话。

So if we were explaining it to somebody who who wasn't in in tech.

是的。

Yeah.

完全正确。

Totally.

完全正确。

Totally.

当GPU需要在长距离通信时，使用光比使用电子更合适。

So when GPUs have to talk over long distances, it's better to use light instead of electrons.

这背后有根本性的效率原因。

There's fundamental reasons why that's more efficient.

当GPU彼此非常接近时，历史上它们是通过铜线传输电信号进行通信的。

When GPUs are really close together and historically, they've talked using electricity over copper traces.

但随着行业开始以越来越快的速度传输越来越多的数据用于训练和推理，通过电力进行通信实际上将面临一些物理限制。

But now as the industry starts to push more and more data around at faster and faster speeds for training and for inference, there actually will be some physical limits to communicating over using electricity essentially.

还有芯片之间如何相互通信，基本上就是这样。

And how how the chips talk to each other, basically.

没错。

Exactly.

没错。

Exactly.

因此，在扩展规模时，这是指当这些GPU彼此非常接近时，它们仍然会相互通信。

And so for scale up this is scale up when these GPs are really close, they still talk to each other.

实际上，行业开始转向使用光子技术进行扩展，是有技术原因的。

There's actually technical reasons why the industry is starting to look to using, optics for scale up.

例如，Marvell刚刚收购了一家公司，叫Celestial AI，他们制造了一种叫做光子互连的技术。

So for example, Marvell just bought a company, Celestial AI, and they make something called a photonic fabric.

因此，这使得即使芯片彼此非常接近，也能通过光来相互通信。

And so this is just letting chips, even if they're really close together, talk to each other using light.

看到英伟达在这里进行收购并不令人意外。

It would not be surprising to see NVIDIA make an acquisition here.

想到的一家公司是IR Labs，这实际上是一家英伟达已经投资过的公司。

A company that comes to mind is IR Labs, which is a company actually that NVIDIA has invested in already.

所以他们对这家公司非常熟悉。

So they're intimately familiar with them.

IR Labs生产光学IO芯片和这些光学互连器件。

IR Labs makes optical IO chiplets in these optical interconnects.

我知道这听起来非常技术性和复杂，但这绝对会是一次收购。

So that is I know it's a it's very sort of meaty and technical, but that would definitely be an acquisition.

我不觉得2026年这项收购会不在考虑范围内。

I wouldn't be surprised if it's on the table in 2026.

嗯，这个光子学的概念我们在这个节目中已经讨论过了。

Well, look, this idea of photonics is an area that we've talked about on the show.

维诺德·霍斯拉曾指出，这是他看好的最具前景的领域之一。

Vinod Khosla pointed to it as one of the most promising areas that he was looking.

因此，我认为英伟达考虑在该领域扩展业务的想法对我来说确实很有道理。

And so I think the fact that NVIDIA is looking to that NVIDIA might look to expand its operations there certainly makes sense to me.

除此之外，您还在关注这个领域的其他公司吗？

Outside of that, are there other names that you're looking at in

这个领域？

this space?

在这个领域，还有其他参与者。

So in this space, there are other players.

例如，Lightmatter这家公司做的事情基本非常相似，他们实际上专注于三维光子学。

For example, Lightmatter is a company that does essentially very similar things, and they're actually focused on three d photonics.

那么，如果能将光学引擎和GPU堆叠在一起，让它们实现垂直通信会怎样？

So what if you could actually stack the optical engine and the GPU, and they could actually communicate vertically?

所以这更有趣一些，可能稍微更远期一些。

So a little bit more interesting, maybe a little bit further out.

这有点像他们在做的一个架构上的赌注。

Kind of that would be an architectural bet that they're making.

但我会说，IR Labs 是许多关注这一领域的人首先想到的公司。

But I would say IR Labs is one that definitely comes to mind for a lot of people tracking this space.

那让我问问你关于云业务的事。

So let me ask you about the business of the cloud.

本月早些时候，我们有一些报道称，NVIDIA 已基本放弃了与 AWS 竞争、建设云业务的雄心。

We had some reporting earlier this month that NVIDIA has basically put aside its ambitions to compete with AWS as far as building out a cloud business.

我们也报道过，他们已经重新分配了负责这项工作的部分高管。

We also reported that they had basically reassigned some of the executives that were overseeing this work.

我原本认为，对于 NVIDIA 来说，这会是一个合乎逻辑的步骤，尤其是当它

I kind of thought that this would have been a logical step for NVIDIA, especially as it

看起来

looks

试图垂直整合时——也许不是在芯片生产的上游，而是在下游，即如何使用芯片及其相关产品方面。

to vertically integrate itself, maybe not upstream with the chip production, but downstream insofar as what you do with the chips and stuff like that.

我原本觉得，收购一家云公司或许是个不错的选择。

I kind of thought maybe a cloud acquisition would have made sense.

你认为2026年或2027年可能会发生类似的事情吗？

Do you think anything like that could be in the cards for 2026 or 2027?

这又是一个有趣的问题。

So that's again, interesting question.

这在战略上非常有意思。

This is very strategically interesting.

NVIDIA应该超越硅芯片，真正运营一个云业务吗？

Should NVIDIA go beyond silicon and actually run a cloud business?

早期，他们曾试探性地涉足这一领域，许多分析师和行业观察者都感到有些惊讶，因为人们担心NVIDIA会与自己最大的客户——比如微软——产生竞争。

Early on, you know, they dipped their toes into this, and a lot of analysts and industry followers were a little surprised because it felt like the danger there is that NVIDIA would be competing with their biggest customers, you know, Microsoft.

就连谷歌也会购买GPU并将其出租。

Even Google buys GPUs and rents them out.

对吧？

Right?

因此，人们不禁质疑：如果NVIDIA开始运营一个竞争性的云业务，会不会惹恼他们的客户？

And so there was a question of, oh, is NVIDIA gonna make their customers mad if they start a competing cloud business?

现在，如果你向前看一点，思考一下，NVIDIA 要重新进入这个领域需要什么条件？

Now if you fast forward a little bit and ask, well, what would need to come true for NVIDIA to get back into that game?

如果这真的关乎商业模式和利润捕获的话。

If it's really about business model and and capturing margin.

而在当前的价值链中，这就像卖铲子和镐子的生意。

And in the value chain, right now, it's it's like a picks and shovels play.

制造硬件的人捕获了所有的利润。

The people that make the hardware are capturing all the margins.

在未来，如果你相信由于竞争压力，这些利润会受到压缩——无论是来自其他 GPU 厂商如 AMD，还是来自超大规模云服务商自研 XPU，比如亚马逊，未来可能不需要购买那么多 GPU，因为他们正在自研 Tranium。

In the future, if you believe in a world where those margins compress because of competitive pressures, whether it's from other GPU vendors like AMD or from hyperscalers making their own XPUs, so Amazon, maybe they don't need to buy as many GPUs in the future because they're building their own Tranium.

可能会出现一种情况，利润捕获会向上游转移，转向应用层或推理与 API 层。

There could be a world where the margin capture actually shifts higher in the stack to the application layer or to the inferences and API layer.

如果 NVIDIA 感觉到自己的利润会被压缩，而他们仍有大量现金，我能看到他们考虑重新入场，并思考：我们能否在这里赚钱？

And if NVIDIA felt that their margins would be compressed and they compressed and they still had lots of cash, I could see them considering getting back in the game and asking, could we make money here?

而且这不会是像裸金属 GPU 租赁那样的模式。

And it wouldn't be on, like, bare metal GPU rentals.

这可能会是类似推理即服务这样的模式。

It would probably be on the, like, inference as a service type play.

对。

Right.

但你所强调的是，制造芯片、销售芯片以及围绕芯片的软件，这些环节的经济模式必须大幅压缩，NVIDIA 才会考虑向下游拓展他们可能提供的服务。

But but but the point you're making is that the the economics of making the chip and selling the chip and the software around the chip, I mean, that would have to compress considerably for NVIDIA to look downstream in terms of what they might seek to offer.

完全正确。

Totally.

而且，我们目前并没有看到这样的迹象。

And, you know, we don't necessarily see a sign of that.

NVIDIA 的利润率高达 70%，但这可能是一个长期趋势。

NVIDIA has wonderful margins in the 70%, but this could be a longer term thing.

所以 2026 年还不好说，但未来会不会发生呢？

So 2026, not so sure, but could it happen in the future?

有可能。

Possibly.

对。

Right.

不过，我要说一下。

Although, I will say look.

这一点和云的问题不同，但软件对这些芯片至关重要。

To to and this is now a different point than the cloud point, but software is so key to these chips.

为了保持软件护城河的强大，某种收购可能是他们考虑的另一个方向，这一点我们在这档节目中也讨论过。

Some sort of acquisition to sort of keep that software moat strong, I mean, that could be another area that they look to, because and we've talked about this on the show.

他们实际上是在向客户推销包含芯片的端到端完整系统。

It's really it's the they're trying to sell people the full system end to end with with their chips.

所以这确实值得深入思考。

So definitely interesting to think about.

奥斯汀，感谢你做客节目，新年快乐！

Austin, I wanna thank you for coming on, and Happy New Year to you.

这位是来自Creative Strategies的高级分析师奥斯汀·莱昂斯，正在TI TV上做客。

That is Austin Lyons, senior analyst at Creative Strategies here on TI TV.

好的。

Okay.

今天，我们发布了《2026年新闻室预测系列》的最新文章。

Today, we are publishing our latest piece in our newsrooms 2026 prediction series.

昨天，我们的苹果记者谈到了他为何认为苹果将扭转其人工智能的低迷局面。

Yesterday, our Apple reporter talked about why he thinks Apple will reverse its AI slump.

今天，我们的财经记者分享了他对甲骨文如何创新其云基础设施融资方式的看法。

Today, our finance reporter has some thoughts on how Oracle could get creative with the way that it finances its cloud build out.

我想邀请迈尔斯·克鲁帕来帮助我们理解他的想法。

I want to bring on Miles Krupa to help us understand what he's thinking.

迈尔斯，欢迎再次做客我们的节目。

Miles, welcome back to the show.

很高兴你能来。

It's great to have you here.

谢谢，阿卡什。

Thanks, Akash.

你的2026年预测是什么？

What is your 2026 prediction?

嗯。

Yeah.

我的2026年预测是，甲骨文将推出全球首例芯片支持、资产支持的证券。

My 2026 prediction is that Oracle is going to do the first of its kind, chipped backed, asset backed security.

我知道这听起来有点拗口。

I know that's a mouthful.

芯片支持、资产支持的证券。

Chip backed, backed security.

好的。

Okay.

是的。

Yes.

很好。

Great.

我认为这是个很好的结束点。

I think that's a great place to end it.

所以

So

不。

No.

所以，告诉我

So Tell

这代表什么。

us what that means.

对。

Right.

所以，资产支持证券市场非常庞大。

So, the asset backed securities market is huge.

今年到目前为止，市场规模已经超过8万亿美元，并且已经被用于资助数据中心等项目。

It's a bit over 800,000,000,000 so far this year, and it's already been used to finance things like data centers.

因此，我提出的建议是，甲骨文将利用这个市场，以现有的芯片作为抵押来筹集资金，以便购买更多芯片，用于为OpenAI建设数据中心，并跟上其大力推动的大型云扩张计划。

And so what I'm proposing is basically that Oracle will look to this market to be able to raise money against its existing chips so that it can buy more chips to do things like build data centers for OpenAI and, you know, stay on track for this big cloud expansion that they're pushing.

我只是想确认一下我是否理解清楚了。

And I just want to make sure I'm clear on this.

我的意思是，看看甲骨文拥有芯片，然后以这些芯片作为抵押品，这部分我明白。

So, I mean, look, the idea of Oracle owning chips and then taking it alone that is collateralized by those chips, I get that part.

至于他们将如何使用这笔资金，是用于购买更多芯片，还是真正用于自己建设数据中心？

In terms of what they would spend that money on, would they be spending that money on buying more chips or would it be really building the data centers themselves?

是的。

Yeah.

这正是我们的想法。

That would be the idea.

芯片是任何数据中心——尤其是云中的AI数据中心——中最昂贵的部分。

The chips are really the most expensive part of any data center well, any AI data center, in the cloud.

到目前为止，甲骨文在让其他公司为其建设数据中心的物理外壳方面做得相当不错，并且已经通过几种不同的方式分摊了这些前期成本。

And so Oracle's done a pretty good job of getting others to build the physical shells of the data centers for them so far, and they've been able to defray the upfront costs of that in a few different ways.

但我认为，他们明年需要开始认真考虑如何为这些数据中心中的芯片融资，以及如何降低这些成本。

But what I think they're going to need to start doing next year is really thinking about how they're financing the actual chips that go in those data centers and how to bring down those costs.

因此，这就是我看到他们转向资产支持证券市场的原因。

So that's that's where I see them turning to the ABS market.

你一开始提到这是前所未有的。

Now you started off by saying first of its kind.

所以，之前没有任何公司做过类似的事情吗？

So so no company has done anything like this before?

没错。

Right.

在公开交易的资产支持证券市场上，还没有任何公司做过这样的事情。

No company has done anything like this in the sort of publicly traded asset backed securities markets.

CoreWeave 公司率先推广了这种 AI 云模式，已通过芯片担保的债务融资约 130 亿美元，但他们通过向私人投资者在私人信贷市场进行定制化交易来实现的。

CoreWeave, which, you know, pioneered this kind of AI cloud model broadly, has raised about $13,000,000,000 of debt backed by chips, but it did that by going to private investors in the kind of private credit market in a more bespoke deal.

我所提出的方案则是一种更为标准化的资产支持证券发行方式，大型保险公司、养老基金等风险厌恶型投资者可以参与，并根据需要进行买卖。

What I'm proposing here would be a much more sort of standardized asset backed securities issuance that big insurance funds, big pension funds, you know, risk averse investors could get behind and then trade in and out of if they wanted to.

本质上是一种公开证券吗？

As a public security, essentially?

是的，本质上是的。

Essentially, yes.

好的。

Okay.

你还提到了你所描述的这个系统中的一个特殊目的载体。

Now, you also talked about a special purpose vehicle in this system that described.

这里的SPV除了是‘特殊目的’之外（这正是特殊目的载体的定义），它的作用是什么？

What would be the purpose of an SPV here other than being special, which is the point of a special purpose vehicle?

嗯。

Yeah.

这是ABS市场中一种常见的工具。

Well, it's a common tool in the ABS market.

基本上，甲骨文会将其用作抵押品的芯片放入这个特殊目的载体中。

And basically, Oracle would put the chips that it's using as collateral in that special purpose vehicle.

所以这些芯片基本上被锁定在里面，贷款方可以放心，如果情况恶化，他们仍然可以对SPV中的这些芯片采取行动。

So they're basically locked in there, and lenders could have comfort that if everything went south, they would be able to then do something with those chips in the SPV.

而且这还能额外带来一个好处，就是将部分债务移出甲骨文的资产负债表，而甲骨文的资产负债表已经显得有些吃紧。

And it could also have the added benefit of keeping some of this off of Oracle's balance sheet, which is already looking a bit stretched.

因此，它在一定程度上实现了双重目的。

So it sort of serves that dual purpose.

那么，为什么公司能够这样做呢？

And why why is it that companies can do this?

我的意思是，把债务移出资产负债表这个想法，对我来说似乎很有道理。

I mean, you know, the idea of keeping debt off the balance sheet, that seems to make a lot of sense to me.

但我就是难以理解，为什么一家公司可以把所有风险都转嫁给一个第三方实体，而这个实体本身却不必承担任何风险？

But I just I'm struggling to understand why a company could just basically allocate all that risk to this third party entity or something that doesn't beget any risk upon itself?

这是一种常见手段，实际上，自金融危机以来，银行一直大量使用这种做法。

Well, it's a common tactic, and it's actually something that the banks have done a lot more since the financial crisis.

所以，这是把这一概念应用到科技领域，而坦率地说，直到现在，科技行业还从未需要做这样的事情。

So it's sort of taking that concept and applying it to the tech world, which frankly, up until this point, hasn't had to do anything like this.

你知道，谷歌的云业务发展得相当迅速，但一开始几乎没什么利润。

You know, Google has grown a fairly big cloud business pretty quickly, and it was pretty unprofitable to begin with.

谷歌用其他业务的利润来补贴这一块。

And Google subsidized that with profits from the other part of its business.

现在不同的是，甲骨文没有这种优势。

What's different now is basically that Oracle doesn't have that luxury.

它正在烧钱，并且必须花大量资金来履行对OpenAI的承诺。

It's burning cash, and it has to spend a lot of money to fulfill this commitment that it's made to OpenAI.

所以，现在看到像甲骨文这样曾经非常盈利的科技公司，如今却不得不考虑采取这样的策略，确实是个有趣的时刻。

So it is kind of it's an interesting moment, to see a tech company like Oracle that used to be very profitable now in a position where I think it should be considering something like this.

那么，采取这种策略会有什么风险呢？

And what would be the risks with with pursuing a strategy like this?

嗯。

Yeah.

风险主要在于这些芯片——这一直是个热门话题，即芯片的实际使用寿命可能没有甲骨文和其他云服务商希望的那么长，原本预期能用五六年，结果可能两三年后，它们的性能就下降了，或者被其他芯片取代了。

Well, the risks are basically that the chips this has been a hot topic of debate, that the chips maybe don't last as long as Oracle and other cloud providers want them to, that instead of five or six years of useful life, it turns out maybe after two or three years, they're not as, powerful or they're usurped by other chips.

所以这是一个主要风险。

So that's one big risk.

而且，最终这些都会由云客户合同来支持。

And then, you know, ultimately, these will be backed by cloud customer contracts.

这可以说是关键的保障机制。

That's, kind of the key backstop here.

因此，如果突然间甲骨文的大量客户取消了他们的云合同，这对资产支持证券的发行也会造成问题。

And so, you know, if all of a sudden Oracle got these huge customers canceling their cloud contracts, that would also be problematic for the ABS issuance.

我只是想试着把这种情况推演一下。

And I I'm just trying to sort of play this out.

所以，假设这些芯片的折旧速度比人们想象的要快，那会发生什么呢？

So if we let's say if the chips do depreciate faster than people might think, then I guess what would happen here?

抵押品的价值基本上会缩水，对吧？

The value of the collateral would basically it would shrink, right?

它可能只剩下原估值的70%，那么他们就得想办法寻找新的抵押方式。

It might be 70% of what they thought it were, and then they'd basically have to find new ways to collateralize it.

这就是这个想法吗？

Is that is that the idea?

可能不是，但在下行情景下，这可能会带来潜在危害。

Probably not, but it it would just be potentially harmful in the downside scenario.

如果他们不得不没收这些芯片，也可能对通过该SPV流向最终偿还ABS投资者的收入造成损害。

If, You know, if they were to have to sort of seize the chips and also could be harmful for the revenues flowing through that SPV that ultimately go to repay the investors in the ABS.

你知道，如果英伟达——这种情况不太可能发生。

You know, if NVIDIA well, that's less likely.

但假设OpenAI突然不想再使用这些旧款英伟达芯片，这可能会影响流向投资者的现金流。

But let's say if OpenAI all of a sudden doesn't wanna use these older NVIDIA chips, you know, maybe that harms the cash flows going to the investors.

但有一些方法可以解决这个问题。

But there are ways to sort of solve for this.

你可以通过超额抵押，还有其他一些技术手段来确保这些现金流持续不断。

You can overcollateralize it, and there are other sort of techniques you can use to make sure that those cash flows keep keep flowing.

很好。

Great.

迈尔斯，我想转到你今天发表的关于布鲁克菲尔德的另一个故事。

Miles, I wanna pivot to another story that you published today about Brookfield.

布鲁克菲尔德是一家庞大的资产管理公司。

Brookfield is a giant asset management company.

我们在这里《信息》上很少写过他们。

We haven't written too much about them here at The Information.

你对他们进军云计算市场的雄心有什么发现？

What did you find about their ambitions in the cloud market?

是的。

Yeah.

这很有趣。

It's interesting.

他们在11月宣布了一项重大举措，计划收购价值1000亿美元的资产以支持人工智能。

They made this big announcement in November that they were going to acquire $100,000,000,000 of assets to power AI.

从土地、能源到数据中心，方方面面都包括在内。

So everything from land to energy to data centers.

在那份公告的小小备注部分，他们实际上提到正在打造一家名为Radiant的云公司，并且英伟达正作为该交易的一部分投资于布鲁克菲尔德的这个基金。

And there was a little notice part of that announcement where they actually said they were building a cloud company called Radiant, and that NVIDIA was investing in this Brookfield fund as part of that.

在这次对布鲁克菲尔德人工智能基础设施负责人进行的采访中，我了解到，他们将Radiant视为其在全球各地（如瑞典、法国、卡塔尔）建设的所有数据中心的首选云平台。

And so in this interview with, Brookfield's head of AI infrastructure, I was able to basically learn that they're see they're seeing Radiant as the sort of first cloud of choice for all of these data centers they're building around the globe, Sweden, France, Qatar.

简而言之，布鲁克菲尔德认为，他们可以亲自部署芯片，并通过向企业出租这些芯片来开展业务——这一点是我迄今为止尚未看到任何其他大型私募资本公司做过的事情。

Basically, Brookfield thinks that they can go in, install chips themselves, and build a business renting them out to companies, which is something that I haven't seen any of the other major private capital firms do so far.

所以，布鲁克菲尔德实际上是在暗示，它有可能成为另一个Oracle，作为一个云服务提供商，或者另一个AWS之类的企业。

So Brookfield basically is suggesting it could be another Oracle in terms of being a cloud provider or another AWS or something like that.

我的意思是，他们所竞争的正是这个市场。

I mean, that's the market they're competing in.

对吧？

Right?

是的。

Yeah.

没错。

Exactly.

他们通过这种主权人工智能细分市场找到了切入点，比如为大型政府服务，为欧洲需要本地存储数据的客户提供建设。

And it sees a way in through this kind of sovereign AI niche, you know, serving big governments, serving customers in Europe that need to store their data locally.

这一点上，他们著名的政府关系起到了作用，你知道，他们在加拿大和欧洲有着紧密的政府联系。

And that is something where their sort of government connections you know, famously, they have these tight government connections in Canada and Europe.

他们正利用这些关系来推动这项业务的发展。

You know, they're kind of leaning on that to try to build this business.

因此，我认为在未来几年里，这种策略很可能会带来回报。

And so I think you'll probably see that paying dividends over the next few years.

你认为他们的能源业务如何能为他们带来优势？

How much do you think about how their energy work could give them an advantage here?

因为他们大量投资能源和基础设施。

Because they do a lot of energy investing, they do a lot of infrastructure investing.

我只是想到，这些数据中心所需的电力。

I mean, I just think about the power that is needed for these data centers.

人人都在谈论这个问题。

Everyone can't stop talking about it.