Microsoft’s True Value of $3.8T, Nvidia’s Cloud Revenue Cut, Can AI Be Conscious?

欢迎各位收看《The Information》的TITV节目。

Welcome everyone to the Information's TITV.

我是罗克特·德鲁。

My name is Rocket Drew.

今天是7月2日，星期四。

It is Thursday, July 2.

首先，《The Information》独家报道了英伟达同意为GPU采购提供资金担保，并参与收入分成的消息。

First up, the Information published exclusive reporting about Nvidia agreeing to financially backstop GPU purchases and taking cuts of the revenue.

我们的英伟达特派记者稍后会加入我们，分享她了解到的更多细节。

Our Nvidia reporter will join us to share more about what she learned.

我们的财经分析专栏作家认为，微软的估值可能高于它当前的股价。

And our financial analysis columnist says Microsoft may be valued at more than it's currently trading for.

我们会听听他这么判断的理由。

We'll hear the reasoning behind that.

我还会和几位嘉宾深入探讨一些技术话题，比如AI模型在视觉推理上为何仍显吃力，以及AI是否真的可能拥有意识。

I'll also dig into some technical topics with a couple of guests, like how AI models continue to struggle with visual reasoning and whether or not AI could ever be conscious.

最后，我会和同事劳拉·布拉顿一起，聊聊我们最新一期的《AI议程》通讯，主题是模型路由。

And to wrap things up, I'll be joined by my colleague Laura Bratton to talk through our latest AI Agenda newsletter about model routers.

这期节目内容很精彩，咱们马上开始。

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

《The Information》独家报道指出，英伟达正计划为那些租用其AI芯片的新兴云服务商提供财务担保，作为交换，他们将获得这些服务商的部分收入。

The Information published exclusive reporting that Nvidia is planning to financially backstop young cloud providers that rent out its AI chips in exchange for a share of their revenues.

我们的英伟达记者刘菲比率先披露了这条消息，现在她来到节目中，和我们分享她的发现。

Our Nvidia reporter Phoebe Liu broke that story, and she joins me now to share what she learned.

好的。

Alright.

欢迎来到节目，菲比。

Welcome on the show, Phoebe.

谢谢邀请我。

Thanks for having me.

不客气。

Of course.

所以你写到，这是英伟达内部一些人称为‘AI计算合作伙伴关系’计划的一部分。

So you wrote that this is part of a program that some people at Nvidia have dubbed the AI Compute Partnership.

能多讲讲你了解到的情况吗？

Tell me more about what you learned.

对。

Yeah.

没问题。

Totally.

所以在过去一个月左右，Furmis和Sharon AI这两家澳大利亚背景的新一代云服务商兼数据中心建设方，分别宣布了两笔交易。

So there were two deals that were kind of announced in the last month or so from Furmis and Sharon AI, which are both kind of Australian based Neo Cloud slash data center builders, somewhere along that spectrum.

这让我开始琢磨，既然已经有两家了，那肯定还有更多。

And it made me kind of think, okay.

于是我开始四处打听，了解这种情况有多普遍。

If there are two, there must be more.

结果发现，这其实是英伟达正在推行的一项相当成熟的计划：为那些较新、规模较小的云客户提供信用担保，帮他们以更低的利率购买英伟达那些价格不菲的GPU，作为交换，英伟达会分得一部分基于使用量的经常性收入。

So I started kind of calling around and asking people how widespread that was, and it turns out it is part of kind of a pretty codified program that Nvidia is starting to provide kind of younger, newer cloud customers a credit wrapper to help them buy Nvidia's, pretty expensive GPUs at lower interest rates in exchange for a cut of recurring usage based revenue.

这和以往的做法不太一样，以前英伟达是通过担保租赁、股权投资等方式来推动生态发展，但现在更像是在鼓励其他人去购买他们的产品。

And that's a departure from kind of previous strategy where Nvidia has backstopped leases, provided equity investments, and done other things to stimulate the ecosystem, but it's kind of part of that trend where Nvidia wants to get other people to buy its products.

很多公司都会这么做，但英伟达做得特别突出，毕竟它在生态系统里扮演的角色实在太关键了。

I think lots of companies do this, but I think Nvidia has done it at an extent that's very interesting given how big of a role it's playing in the ecosystem.

具体来说，英伟达会保证：如果算力卖不出去，他们会以每小时每张显卡固定金额的价格把它们租回来。

So the way it works is that Nvidia guarantees that if capacity is unsold, it'll lease it back at a flat rate of x dollars per GPU per hour.

我真想知道这个固定金额到底是多少。

I would love to know what x is.

然后收入分成会按超出部分的一定比例来算，而且随着合同期限逐渐递减，据我了解，这类合同一般持续五到六年。

And then the revenue share would be kind of a percentage of the upside and scaled down over the life of the contract, which tends to be five to six years as far as I understand.

这个保证让银行或其他贷款方更愿意给那些信用记录不好的新云服务商放贷。

And, basically, that guarantee makes banks or other lenders more willing to lend money to these newer clouds that, like, don't have good credit.

说实话，我也不太确定。

And I don't know.

他们可能能拿到的贷款利率大概在15%左右吧。

They would probably be offered, like, a 15% interest rate or something.

这对他们来说可能不是个可行的生意，或者他们没法那么快扩张。

And that may not be a viable business for them slash they wouldn't be able to scale as quickly.

但有了英伟达的支持，他们的贷款利率就能接近投资级水平。

But with Nvidia's backing, they can get their interest rates much closer to investment grade.

这看起来就像是英伟达帮助客户负担得起其GPU的又一种手段。

So it just seems like another tool in Nvidia's toolbox of ways that it's helping customers be able to afford its GPUs.

你觉得英伟达实际预期自己得租下部分算力吗？还是说他们押注这些新兴云服务商会非常成功，根本用不着自己出手？

Do you think Nvidia expects that in practice it will have to rent some of that capacity, or is the bet that it'll never come to that because the Neo-Clouds will be so successful?

是的。

Yeah.

我肯定他们内部讨论过这个问题。

I'm sure they've talked about it.

公开场合下，他们总说根本不会花一分钱，纯粹是靠自己雄厚的财务实力在做支撑。

I think publicly, they posture that they never expect to pay a cent, and basically, they're just leveraging their balance sheet because they have such a strong one.

我不清楚。

I don't know.

他们去年的净利润达到了一千六百亿左右，简直不可思议。

They I think they made a 160,000,000,000 in net income in the last year or something like that, which is crazy.

而且，没错，他们是目前全球市值最高的公司。

And, yeah, they're the world's largest company by market cap.

所以我认为，他们赌的是未来至少两三年内需求还会持续爆炸式增长。

So, yeah, I think they're betting that demand will continue to be astronomical over at least the next couple of years.

他们对外宣称的周期可能更长，但具体能撑多久，还得走着瞧。

I think they posture, like, longer than that, but, we'll see.

就拿这家公司来说，我跟他们的联合CEO蒂姆·罗森菲尔德聊过，他说只用了几个星期就签下了客户，合同全周期的订单都满了。

At least in the case of firmness, I talked to one of their co CEOs, Tim Rosenfield, and he said it took them just a couple weeks to get customer commitments for the full length of the contract.

对他们来说，这事儿一点都不难。

So for them, it wasn't hard.

我不确定其他公司会不会也这么顺利。

I don't know what the case will be for others.

但据我所知，他们到现在还没花什么钱，也不打算花。

But, yeah, as far as I know, they haven't had to pay much of anything yet and don't expect to.

不过，咱们走着瞧吧。

But, yeah, we'll we'll see.

网上很多人持不同看法，但这事还得时间来证明。

I think there are a lot of people on on the Internet that feel otherwise, but it's kind of a time will tell situation.

当然。

Sure.

网上总有人持不同看法，这很正常。

There usually are people on the Internet Yeah.

持不同看法。

That feel otherwise.

但如果真到了那一步，英伟达凭他们的资产负债表也负担得起。

But if it comes to that, Nvidia will be able to afford it because of their balance sheet.

他们拿那么多算力做什么用呢？

What would they do with that capacity?

如果真得把部分算力租回去，他们能用得上吗？

Would they have some use for it if they do end up having to rent some of that capacity back?

是的

Yeah.

完全正确

Totally.

我觉得已经有一笔这样的交易了，就是Lambda公司，Nvidia把一部分算力租回来，提供给自己的研究人员，这挺有意思的，因为你可能会觉得，既然他们为制造算力的公司工作，应该不缺计算资源。

So I think there is one such deal already, with Lambda where Nvidia rented some capacity back and gave it to its own researchers, which was interesting because you wouldn't think that Nvidia's own researchers would need more compute, given that they work for the company that makes it.

但之前报道的消息就是，Lambda那边确实发生了这种情况。

But, that's what the information previously reported, happened with Lambda.

我不清楚他们如果再从这些交易中租用更多算力，会怎么用。

I don't know what it's going to do with any additional capacity that it rents from these deals.

我不确定他们会不会考虑以更低的价格转租给别人，或者干脆自己用，因为看起来Nvidia正在越来越重注于训练自己的开源模型，这个押注很有趣，我真希望以后能再多聊聊。

I don't know if they would consider reselling it to other people at, I don't know, lower rates or something like that or just use it itself because it seems like Nvidia's placing a bigger and bigger bet on training its own open weight models, which is a fascinating bet that I would love to talk more about another time.

但我觉得正因如此，他们确实需要更多算力，也肯定用得上。

But I think because of that, that would require more capacity, and they would definitely have use for it.

明白了

Got it.

这些安排可能会带来哪些风险，对英伟达或者对这些新兴云服务商来说？

What are the kind of risks that these arrangements could create, both for Nvidia or for

对，完全没错。

the Neo-Clouds?

是的，没错。

Yeah, totally.

我觉得当Core Wave刚出现的时候，大家都吓坏了，说天哪。

So I guess I feel like when Core Wave first came onto the scene, everyone was like, oh my god.

这太冒险了。

This is so risky.

他们根本撑不下去。

Like, they are not going to survive.

但他们一直都在说，嘿。

And they kept going around saying, like, hey.

我们的债务虽然利率高，但有财务状况良好的客户合同作担保，还能用GPU的折旧价值来做抵押。

Like, all of our debt, even though it's high interest, is kind of guaranteed by people with good balance sheet by customer contracts and, like, take it out against the depreciating value of the GPUs.

我觉得现在还是有人担心，做Neo Cloud这种业务本身就风险很高，因为杠杆率太高了。

I think there's still some kind of fear that running a Neo Cloud business is inherently very risky because of how highly leveraged it is.

要筹集足够资金大规模采购GPU，本来就不容易。

Just like getting the capital to buy GPUs at a scale big enough to run such a business is pretty hard.

而且确实背负着大量债务。

And, yeah, there's a lot of debt.

这行特别烧钱。

It's very capital intensive.

而这就带来了很大风险，尤其是因为没人说得准。

And that inherently comes with a lot of risk, especially because nobody really knows.

大家都知道GPU长期来看都有用，但它的实际使用寿命和经济寿命是两回事。

Like, people know that GPUs continue to be useful for a long time, but useful life and kind of economic life is a little bit of a different story.

所以很难预测未来GPU在财务上能值多少钱，这种不确定性已经深深埋在里面了。

So I think it's difficult to predict what a GPU will be worth in financial terms down the line, and that risk is kind of baked in.

不过话说回来，我觉得Nvidia的兜底措施已经很大程度上缓解了这种风险。

But, again, I think it's pretty much mitigated by Nvidia's backstops.

如果连英伟达都负担不起这笔支出，那整个市场，还有AI热潮——不管是泡沫还是崩盘，总之都会出大问题。

And then if it gets to the point where Nvidia is not afford not able to afford this, I think we have bigger problems with the market and kind of the AI the AI boom, bubble, bust, whatever you want to call it in general.

没错。

Sure.

对。

Yeah.

到时候我们的问题就更大了。

We'll we'll have bigger problems.

是的。

Yeah.

那这个协议里的收入分成部分呢？

And then what about the the revenue sharing part of this agreement?

为了获得英伟达的兜底支持，Neo云服务商必须承诺把一部分收入让出来。

So in order to get Nvidia's backstop, the Neo-Clouds have to promise a portion of their revenue.

对。

Yeah.

这个我还正在打听，挺想了解清楚收入分成具体是怎么算的。

So I this is something I'm still asking around on, and I would love to know more about what the revenue share exactly is.

我的理解是，Nvidia 会承诺以某个固定价格回购算力。

My understanding is that it's kind of so Nvidia promises that it'll buy back the compute at some kind of fixed rate.

我听说的是——虽然还没完全确认——如果云服务商能以远高于 Nvidia 回购价的价格卖出算力，那 Nvidia 的分成就是这两者之间的差价，你懂我的意思吧。

And what I've heard and I I'm not a 100% sure on this yet, but, the theory is that if a cloud provider is able to sell at kind of a far higher rate than what Nvidia has guaranteed it would buy it back for, Nvidia's revenue share would be kind of a difference of those two rates, if that makes sense.

这其实是在赌当前需求远远超过供应，从而推高价格。

So it's kind of a bet that demand is far, far higher than supply right now, which would impact pricing.

据我所知，随着合同推进，这个分成会逐渐降到零。

And that would kind of scale down to zero over the life of the contract as far as I understand it.

所以合同初期，Nvidia 能拿到更多现金。

So it kind of be more cash to Nvidia at the beginning of the contract.

到了后期，就相当于让 Neo-Cloud 自己飞起来，独立运营了，之类的。

And then by the end, it's kind of letting the Neo-Cloud go, like, fly its wings and be independent or whatever.

对。

Yeah.

至于收入分成协议本身，我觉得虽然英伟达一直强调它并没有强制任何人必须 exclusively 使用英伟达的芯片和全套产品，但这种紧密且财务上高度依赖的关系，实际上还是会促使客户保持忠诚，从而反过来壮大英伟达的生态系统。

And then, I guess, in terms of the revenue share agreement more generally, I feel like although Nvidia has been very strong on how it's not explicitly forcing anyone to exclusively use Nvidia chips and the full stack of Nvidia products, I think having such a close and financially dependent relationship on Nvidia would kind of implicitly encourage that loyalty, which would help the Nvidia ecosystem as well.

对。

Yeah.

所以你把这看作是英伟达迈向成为整个生态系统‘中央银行’的最新一步。

So you set this up as sort of the latest step, that Nvidia is taking towards being the central bank, playing central bank for this whole ecosystem.

接下来你最想了解这个故事里的哪一部分？

What are you interested in learning next in this story?

你觉得接下来还可能发生什么？

What do you think could could happen next?

是的。

Yeah.

完全同意。

Totally.

我特别想了解一些更具体的数据。

Definitely interested in kind of more specific numbers.

我理解这些协议大多是逐案处理的，很大程度上取决于每个客户的具体情况。

My understanding is that a lot of these deals are case by case and kind of depend on each customer's specific situation.

所以弄清楚背后有哪些因素在起作用，会非常有意思。

So trying to understand kind of what factors are at play there would be really interesting.

我知道Nvidia这个方案推广得挺广的，但具体有多少人真的接受了，我就不清楚了。

And just trying to see I think Nvidia has offered this pretty broadly, but I don't know how many people have actually taken up taken them up on it.

我知道有几个人拒绝了，因为他们根本用不上。

I know a couple people have declined it because they don't need it, for example.

还有，现在有哪些人暗地里在用，我们根本不知道？

And it's just like, who else is doing this right now that we don't know about?

这个倒是挺值得去挖一挖的。

That would be interesting to find out.

还有我们刚才聊的，这种需求到底会有多普遍？

And also kind of what we've been talking about, like, how how widespread will demand be?

Nvidia最后真得掏钱吗？

Will Nvidia ever have to pay a cent?

我觉得还有很多类似这样的兜底协议。

I think there are a lot of other kind of backstopping deals like this.

实际上，英伟达去年九月就跟CoreWeave签过一个。

Nvidia did one actually with CoreWeave last September.

我不清楚，也不觉得那个协议里包含收入分成条款。

I don't know or don't think there is a revenue share agreement as part of that.

但当英伟达说要买下CoreWeave所有未售出的算力时，投资者们都松了一大口气。

But I think when Nvidia said they would buy all of CoreWeave's unsold capacity, investors, like, let out a big exhale.

股价一下子就涨了不少。

The stock went up a ton.

这确实让市场情绪缓和了不少。

And it kind of just, like, soothes the market.

所以我很好奇，这种模式会不会持续下去，或者能不能帮更多新兴云服务商上市。

So I'm curious if that's gonna continue to play out slash if that's gonna allow more Neo-Clouds to go public.

对。

Yeah.

说到底，最大的问题还是，这种疯狂的需求还能持续多久。

I think the biggest question at the end of the day is how long will demand continue to be as crazy it has been as it has been.

如果我知道答案，说不定早就转行了。

And, yeah, if I knew the answer to that one, I would maybe be in a different industry.

所以嘛。

So yeah.

说得有道理。

So fair.

确实有道理。

So fair.

或者，谁知道呢，也许我们还是会留在新闻行业。

Or I don't know, maybe we'd stay in journalism.

我觉得。

Think there's

对。

a Yeah.

嗯，从各个要素来看

Well, of elements to

这可是个重大进展。

it's a big development.

感谢你来参加节目，为我们详细解读，菲比。

So thanks for coming on and breaking it down for us, Phoebe.

谢谢，罗克特。

Thanks, Rocket.

微软目前的市值是2.9万亿美元。

Microsoft is currently trading at a $2,900,000,000,000 valuation.

但我的同事阿妮塔·拉马斯瓦米在她的《真实价值》专栏中发布了一项新分析，认为这家公司的实际价值最高可达3.8万亿美元。现在她加入我们，来解读她的评估。

But my colleague Anita Ramaswamy is out with a new analysis for her true value column that values the company at up to $3,800,000,000,000 Anita joins me now to break down her assessment.

阿妮塔，欢迎来到节目。

Anita, welcome to the show.

很高兴你能来。

It's great to have you.

很高兴能再回来，Rocket。

It's great to be back, Rocket.

我们得先把这个说清楚。

So we gotta get it out of the way first.

是因为微软持有大量OpenAI股份吗？

Is it just because Microsoft owns so much OpenAI?

这就是它被低估的原因吗？

Is that why they're undervalued?

是的，这是个好问题，尤其是现在大家都在讨论OpenAI何时上市。

Yeah, it's a good question, especially as all of these discussions swirl around OpenAI and when they're going public.

我觉得目前来看，这可能帮不上忙，毕竟大家对它的商业模式和长期前景还有很多疑问，而且从整体来看，微软持有的股份其实并不算大。

My thinking is that it probably doesn't help right now considering a lot of the questions around the business and its long term future, but also it's a pretty small stake in the grand scheme of things.

你得记住，微软是一家市值2.9万亿美元的公司。

What you have to remember is Microsoft is a $2,900,000,000,000 company.

规模非常庞大。

It's massive.

这也就是微软今天的企业估值。

That's what their enterprise value is today.

而它们在OpenAI的持股，最多也就值几百亿美元。

And their stake in OpenAI is probably worth a couple 100,000,000,000 at most.

所以归根结底，微软股价被低估的原因，并不在于这一点。

And so at the end of the day, that's not really what seems to be driving the undervaluation in Microsoft shares.

是的，这个说法很有说服力。

Yeah, that's pretty compelling.

我整天写AI相关的内容，都忘了除了OpenAI和Anthropic之外，还有其他公司。

I guess writing about AI all day long, I forget there are other companies besides OpenAI and Anthropic.

那你最初是怎么想到要仔细看看微软的估值的？

So what made you decide to take a closer look at Microsoft's valuation in the first place?

嗯，有几个原因。

Well, it's a couple of things.

首先，我觉得微软今年股价出现的这20%回调，不管你怎么称呼它，确实挺引人注目的。

I mean, first of all, I think this 20% correction that's happened, whatever you want to call it, the stock side that's happened this year for Microsoft is notable.

这比所有其他超大规模云服务商加起来还要多。

That's more than all of its other hyperscaler peers.

你看谷歌、亚马逊这些公司，虽然也有起起落落，经历了不少波折，甚至上个月表现很差，但今年到目前为止，它们的股价跌幅都没微软这么大。

If you look at the Googles of the world, the Amazons of the world, they've had some ups and downs, some bumps, and certainly a tough month, but they are not trading down as much as Microsoft is on a year to date basis.

我想深入看看，微软这家公司的资源极其丰富，资金雄厚，从生成式AI一兴起就开始大力投入了。

I wanted to examine, this is a company that has tons of resources, very deep pockets, and has been investing in AI from the early days really when it comes to Gen AI.

那为什么它的股价表现不佳？是不是已经失去了领先优势？

And so how is it that their stock is not doing well and have they lost their lead?

我之所以觉得现在研究微软的估值很有意思，还因为一个常见的现象：像这种旗下有多个业务板块的大型集团，通常整体估值都会打个折扣，远低于各个业务拆开来单独估值的总和。

The other reason I thought it was interesting to look at Microsoft's valuation right now is just the idea that, you know, a lot of times when you see these conglomerate sort of businesses that have different business units that do different things, they tend to trade at a discount relative to what each of those business units would trade if they were just standalone businesses.

所以我觉得这是一个挺有意思的思考角度。

So I thought it was an interesting thought exercise.

也许再过几年，我们真会看到微软、亚马逊或者谷歌这样的公司选择拆分，因为它们的各个业务板块估值差异太大，甚至吸引的投资者群体都不一样。

You know, maybe years from now, we could even see a scenario where companies like Microsoft or Amazon or Alphabet choose to break up because their businesses are valued so differently and sometimes even attract different investor bases.

为什么市场会给这种多元业务的组合打折扣呢？

What is the intuition for discounting that basket of different businesses?

会不会是其中某个业务线突然就被砍了，我们根本没法预知结果？

Is it that maybe one of those business lines might get cut on a whim and we don't know how it's going to go?

我一直觉得挺奇怪的，因为集团折价这个说法其实说不太通，但我猜大概是不同类型的投资者对各个业务板块的看法不一样吧。

You know, I've always wondered this because the conglomerate discount doesn't make a whole lot of sense to me, but I think it's something along the lines of, you know, just different types of investors looking at different businesses in different ways.

比如一些专注半导体的投资者，可能会给这一块很高的估值，但对软件业务就不那么看重。

And so some investors who are, let's say focused on semiconductors might value that a certain way versus they might not place as much emphasis on a software business.

我觉得这还跟公司的战略重心和资源投入方向有关。

I also think that part of it has to do with the focus of the company and where they're investing their resources.

明白了。

I see.

所以你是说，某种程度的折价是合理的，但微软的折价实在太大了。

So you say it's okay, it's reasonable to have some kind of discount, but Microsoft's is too steep.

那我们来一个个板块看看，你刚才提到的微软这些业务部分。

So let's go, sort of bucket by bucket through each portion of Microsoft's business that you consider here.

你主要看的是云服务、软件和个人电脑业务。

So you look at cloud, software, and personal computing.

我们一个个来分析。

Let's take them each in turn.

先从云业务开始。

Let's start with the cloud side.

你在这块发现了什么？

Tell us what you found there.

是的。

Yeah.

我在云业务这块发现，这是微软最难估值的部分。

So what I found on the cloud side, Rocket, is that this is the hardest part of Microsoft's business to value.

正如你提到的，他们有三个不同的业务部门，而云业务增长最快，在人工智能方面也最有前景。

As you mentioned, they have three different business units and cloud is the fastest growing, and it seems to be the most promising when it comes to AI.

他们每年的增长率大约在40%，云计算的需求推动了这种迅猛发展。

They are growing at a rate of about 40% year over year and it's been really sort of breakneck growth with the demand for cloud computing.

他们的Azure部门就属于这个云业务板块，而Azure的增长率正好在40%左右。

Their Azure unit is housed within this cloud division and Azure is the one that's growing around 40%.

如果你看整个云业务板块，里面还包括一些支持云基础设施运营的软件产品。

If you look at the overall cloud segment, there's a couple of other things in there like software that helps support some of their cloud infrastructure operations.

如果看整个云部门的整体增长率，会稍微低一点，但真正推动增长的核心引擎还是Azure，也就是他们的核心云业务。

That growth rate is a little bit lower when you look at the cloud unit overall, but the engine and the growth driver really is Azure, which is their core cloud business.

这个问题在于，相比微软其他核心的软件销售业务，云业务的利润率要低得多。

The problem with this business is that it's on the lower margin side compared to Microsoft's other core business of selling software.

他们目前的运营利润率大概在42%左右，而且增长速度非常快。

They're doing around 42 operating margins and you know, this is growing fast.

投资者们都在问：长期来看，这会不会拉低公司的整体利润率？

The question investors are asking themselves is, will this dilute the margin over time?

明白了。

Got it.

明白了。

Got it.

好的。

Okay.

那我们就说到软件这块了。

So that takes us into the software side.

这里的利润率确实更高，但你觉得软件业务还有什么被低估的地方吗？

The margins are better there, but do you think there's something else that's undervalued about the software side of the business?

是的。

Yeah.

在深入谈软件之前，我想先提一下，关于云业务的估值，尽管这很难判断，但我认为它的价值大概在1.7万亿美元左右。这个数字是参考了谷歌和亚马逊这些类似公司得出的，但说实话，我们找不到一个纯粹做云业务的公司来和微软直接对比。

So I guess before just getting into software, I did want to mention that the conclusion I came to in terms of actually valuing the cloud business, despite how challenging it is, is it's probably worth somewhere in the ballpark of $1,700,000,000,000 That's based on some similar companies like Google and Amazon, but we didn't really have a good comp in terms of a standalone cloud business that we could compare Microsoft's cloud business to.

这个估值可能有点偏高，但我个人觉得，它的价值应该在1.4万亿到1.7万亿美元之间。现在我们来聊聊软件，这才是微软的核心业务。

So it's possible that's a little too high, but my estimate is that that's worth the 1,400,000,000,000 to $1,700,000,000,000 Now let's talk about software because this is their core business.

这是他们最大的业务。

It's their biggest business.

微软之所以出名，很大程度上就是因为企业用户广泛使用他们的Office套件，这一点优势非常明显。

It's the thing that Microsoft is really known for, and they have a huge advantage in the fact that so many enterprises use Microsoft's Office Suite.

不过软件业务虽然利润率很高，差不多有58%，但增长速度远不如云业务那么快。

But you know, the hard part about software, while it is very high margin, around 58% operating margins, they are not growing nearly as fast as the cloud business.

他们的增长率大概在百分之十几的高位。

They're growing at around, you know, the high teens growth rate.

不过，他们也在一定程度上受到了SaaS行业担忧的影响。

Well, imagine they're also getting swept up in the SaaS pocalypse concerns to some extent here.

这算一个因素吗？

Is that a factor?

是的， definitely。

Yeah, definitely.

如果你拿微软和阿尔法贝特比较，我说过，今年微软的股价跌幅比阿尔法贝特更大，但要是看Salesforce、ServiceNow、SAP这些公司，微软今年的股价走势其实更接近这些纯软件公司的表现。

So when you look at Microsoft versus Alphabet, I said that, you know, Microsoft has traded down more than Alphabet year to but if you look at Salesforce, ServiceNow, SAP, Microsoft is actually their stock performance this year has more closely mirrored some of the pure play software names.

这说明投资者在看微软时，觉得它可能会受到像Claude这样的AI工具的冲击。

So that suggests that investors are looking at Microsoft and thinking that they are going to be vulnerable to disruption from things like Claude Co.

还有Anthropic推出的其他AI工具。

Work out of Anthropic and other AI tools.

没错。

Sure, yeah.

有人坐下来用Claude Co，

Someone can sit down and use Claude Co.

就能写出自己版本的Microsoft Word，那他还需要真正的Microsoft Word干嘛？

To vibe code their own version of Microsoft Word, then what do they need the actual Microsoft Word for?

但另一方面，这也说明他们的云业务前景不错，毕竟大家会这么频繁地使用AI工具。

On the other hand, that's a good sign for their cloud business, presumably that people would use AI so much.

没错，确实如此。

Yeah, it totally is.

我觉得软件业务里还是有一些亮点的。

And there are some bright spots in the software business I would say.

比如Copilot，这是微软自己推出的AI助手，一直在大力推广。

Mean, have Copilot, is Microsoft's own sort of AI assistant agent that they've been promoting.

但问题是，到目前为止，我接触过的投资者和客户都没说他们更喜欢用Copilot而不是其他服务。

But the challenge is that so far, at least the investors that I spoke to and the customers, have not really said that they prefer using Copilot over other services.

他们用Copilot，更多是因为它已经集成在现有的软件系统里了。

It's more that they're using it just because it already comes in their software stack.

你知道吧，我们得看看这项技术接下来会怎么发展。

You know, we have to see how the technology evolves.

但关键是，罗杰，如果你看看软件行业里类似的公司，这个业务估值大概在1.7万亿美元左右，不过这个数字可能高得离谱。

But the bottom line here, Rocket, is that this business, if you look at comps that are similar in the software business, is probably worth around $1,700,000,000,000 It's possible that that is way too high of an estimate.

这只是我个人的一点看法。

You know, that's just my 2¢ on it.

不过你要是去查一下，有些投行比如高盛，早就给出过类似的估值。

But if you look, there are some investment banks that have floated estimates like Goldman Sachs.

他们觉得软件业务的估值更接近5000亿美元。

They value software business closer to 500,000,000,000.

所以归根结底，这取决于你对SaaS未来会变成什么样怎么看。

So it just depends on what you think the future of the SaaS apocalypse is going to look like.

明白了。

Got it.

好的，那这就是云业务和软件业务的情况，最后一个板块是个人计算。

Okay, so that's the cloud side, the software business, and the last bucket is personal computing.

你在那里发现了什么？

What did you find there?

这是微软规模最小的业务板块。

So this is Microsoft's smallest line of business.

投资者对它的关注也最少，主要是因为它的收入占比最小。

It's the least sort of scrutinized by investors, and that's because it's smallest in terms of revenue.

它占了公司整体营收的大约百分之二十。

It makes up around 20% of their overall top line.

利润率非常低。

The margins are really low.

大概在20%左右，远低于另外两个板块的40%到50%。

They're closer to the 2020s range versus the other two in the 40s and 50s.

这个部门负责游戏业务，目前正经历一场重大调整。

This is the unit that houses the gaming division and they're going through a big shakeup in gaming right now.

他们已经换了新的管理层。

They have new leadership.

我们之前对此做过深入报道，但归根结底，这只是一个规模较小的部门。

We've reported on that quite in-depth, but at the end of the day, this is just a small unit.

还有Surface电脑，如果你想想这个个人计算部门所投入的硬件成本，它和微软另外两大业务线完全不同。

And the Surface PCs, and if you think about the hardware investment involved in this personal computing unit, it is very different from the other two lines of business Microsoft has.

所以我为了评估这个部门的价值，参考了一些可比公司，比如苹果、惠普、索尼和艺电。

So what I did to try and value this unit is that I took some comparable companies such as Apple, HP, Sony, and Electronic Arts.

索尼和艺电都是做视频游戏的，所以它们能在一定程度上反映这个业务的价值。

Sony and EA are video game makers, so that kind of reflects some of the value in this business.

这个业务的收入增长其实不是在萎缩，而是正在放缓。

This business is actually shrinking in terms of, or not shrinking, sorry, it's decelerating in terms of its revenue growth.

这在未来会是一个值得关注的问题。

And so that's an area of concern for the future.

因此，我最终根据同行的估值，给这个部门定出了3500亿美元的价值。

And so the value I ultimately ascribed to it was $350,000,000,000 based on some of its industry peers.

明白了。

Got it.

好的。

Okay.

所以这就是为什么整个业务的估值能达到3800亿美元。

So that's how you get to this total of up to $3,800,000,000,000 for the whole business.

但就像你说的，过去一年股价跌了20%。

And yet, like you said, the stock is down 20% over the past year.

这是为什么呢？

So why is that?

微软目前面临的主要担忧有哪些？

What are the big concerns that Microsoft is facing?

我采访的一位投资者把这个问题讲得很透彻，他把微软现在的处境形容为‘双重打击’。

One of the investors I talked to for this story put this really nicely, and he sort of framed this as a bit of a double whammy that Microsoft is going through right now.

一方面，大家担心SaaS行业会迎来‘末日’，而且很多人依然把微软看作传统的核心软件公司。

On one hand, have the SaaSpocalypse fears and people see Microsoft as a core software business.

还有一种看法是，Copilot在技术上没能跟上像Cloud Cowork这样增长更快的替代方案。

There is this perception that Copilot has not kept up technologically to some of the faster growing alternatives like Cloud Cowork.

所以现在大家担心的是SaaSpocalypse。

So there's SaaSpocalypse fears.

另一方面的担忧是，投资者都在问所有超大规模云服务商一个问题：你们计划投入、也已经投入的数十亿美元资本开支，真能换来AI产品带来的收入吗？

And then in the second bucket of concerns, is the question that investors are asking of all the hyperscalers about are the billions of dollars that you guys plan to spend and have been spending on capital expenditures to build out your cloud businesses, is that going to be justified by the amount of revenue that your AI products actually bring in?

这一点目前还不明朗，最近所有在AI上大举投资的大型科技公司都因此受到了市场惩罚。

And so that remains to be seen, and that's something that all of these bigger tech companies that have been investing heavily in AI have gotten punished for lately.

而微软现在正好处在这两个负面预期的交汇点上。

And Microsoft is at the perfect nexus and intersection of those two sort of negative outlooks right now.

明白了。

Got it.

这确实让微软在行业中处于一个很特别的位置。

Well that certainly puts Microsoft in an interesting position in the industry right now.

非常感谢你来分享你对这个问题的分析。

Thanks so much for coming on and walking us through your thinking about it.

这家公司在未来几个月肯定会非常值得关注。

Definitely a company that will be very interesting to watch in the coming months.

谢谢您邀请我。

Thanks for having me.

目前最先进的AI模型，仍然解决不了某些连孩子都能搞定的视觉谜题。

The most advanced AI models still cannot solve some visual puzzles that even kids can solve.

安德鲁·戴在谷歌工作了十四年，现在他创办了新公司Elorian，专门研究AI的视觉推理能力。

Andrew Dai worked at Google for fourteen years, but now he's focusing on the visual reasoning capabilities of AI models at his new startup, Elorian.

现在我们有请安德鲁来跟我们详细聊聊。

Andrew joins us now to tell us all about it.

欢迎来到节目，安德鲁。

Welcome on the show, Andrew.

谢谢，罗伯特。

Thanks, Robert.

谢谢您邀请我。

Thanks for having me.

是的。

Yeah.

当然。

Absolutely.

我特别期待这次对话，因为视觉推理现在是人工智能领域一个非常有趣的方向。

I'm excited for this conversation because visual reasoning is a really fascinating area of AI right now.

我想先了解一下当前模型的局限在哪里。

I wanna start by understanding the limitations of current models.

有些人听到说有局限，可能会觉得奇怪。

I think some people will hear the idea that there are limitations.

他们会问：你说的局限到底是什么意思？

They'll say, well, what do you mean there are limitations?

ChatGPT 不是经常给我生成特别棒的图片吗？

ChatGPT makes me great pictures all the time.

那不就是视觉推理吗？

Isn't that visual reasoning?

那现在你看到的差距到底在哪？

So where's the gap that you see right now?

是的。

Yeah.

ChatGPT 制作图片确实很棒，我经常用 Google Lens 来识别植物和花卉。

ChatGPT is great at making pictures, and I use Google Lens all the time to identify, like, plants and flowers.

我们现在用的这些模型和工具，在物体识别方面非常出色，也就是计算机视觉里说的分割和定位。

These current models we have and these tools are really good at object recognition, what in the computer vision industry we call segmentation and localization.

这些功能每天都被各种工具广泛使用。

These are used every day by all sorts of tools.

但问题是，尽管它们在模式识别上很厉害，面对更复杂的视觉推理任务——比如白领工作中常见的那些——像设计，或者你是房产中介，需要画出和你卖的房子一模一样的平面图，又或者你是建筑师，要设计一个恰好容纳 50 个人的办公室，不多不少——这些模型的视觉推理和理解能力还远达不到人类水平。

But the fact is that even though they're very good at pattern recognition for complex visual reasoning tasks, so the more the more common type of tasks you see in kind of like, white collar visual work, so these might be things like design or, say, you're a realtor and that you need to place a floor plan that looks exactly, you know, like the house that you're selling, or you're an architect that needs to, design an office for, like, exactly, 50 people, no more, no less.

目前这些模型的视觉推理和理解能力，还赶不上人类。

These models, their visual reasoning, their visual understanding isn't at the level of humans yet.

如果画面里有个餐桌，你问能坐几个人，它们能答得很好。

So they will be great if there are, say, dining table, and you're asking how many people sit around the dining table.

但要是画面里人挤人，你问有多少人，它们就真搞不定了。

But if you are in a crowded room and you ask how many people are there, then they really struggle.

这是因为对于简单的任务，它们能识别模式。

And this is because for simple tests, can do pattern recognition.

你大概能凭印象判断出某些东西，比如特定类型的客户，或者桌边坐了三四个人。

You kind of find an impression of something, whether that's a certain kind of client or three or four people around the table.

但一旦问题需要花一秒以上的时间仔细看图才能回答，这些模型就做不到这种推理了。

But as soon as you have a question that requires more than one second of looking at an image to figure out the answer, then these models aren't able to reason like that.

明白了。

Got it.

好的。

Okay.

这些例子真的很具体、很有帮助。

Those are some really helpful concrete examples.

说实话，视觉推理居然落后了，这有点让人意外。

I guess it's a little surprising that visual reasoning has gotten left behind.

现在不是有个趋势吗？模型在推理能力上进步很快，尤其是在代码和数学方面。

I mean, there's this this trend right now of reasoning in models that has uplifted code and math reasoning.

视觉推理被落下，确实有点让人意外。

It's a little surprising that visual reasoning has gotten left behind.

是因为视觉推理任务很难验证，没法自动评估模型的表现吗？

Is that because visual reasoning tasks are hard to verify, it's hard to automatically grade a model's performance on them?

还是说，视觉推理被忽视的其他原因是什么？

Or what are the reasons that visual reasoning is left behind?

是的，原因有几个。

Yeah, there's several reasons.

我觉得最主要的是，大家平时用AI的方式都比较受限。作为普通用户，你用图片或视频，多半就是想认出公园里看到的那颗行星。

I'd say the main one is that the way people are used to using AI is in these restricted settings, As an average user, most of your cases for using images and video, like I said, it's just to identify this planet you saw in your park.

对吧？

Right?

但像理解CAD图纸这类更复杂的用途，人们根本不会觉得这是日常需求，而且大家平时也不用AI来做这些事。

But for more complex use cases, like I mentioned, understanding like CAD drawing, People don't think of that as an everyday use case, and the thing is that people don't use AI for these other types of work.

这就导致整个领域更关注那些简单的识别问题。

And what that has led to is that that's led to the community focusing a lot on these recognition problems.

然后你会看到一些像MMU这样的基准测试，它的全称是大规模多任务多模态基准。

And then you do see some of the benchmarks like MMU, which is also called massively multitask multimodal benchmark.

在很多情况下，这些基准测试其实不需要看图片就能解答。

Those benchmarks actually can be solved without looking at the image in a lot of cases.

另外还有一个挺有名的基准测试，叫AI2D。

And then the other benchmark that is quite well known is also called AI Agei.

这个通常被当作多模态推理的基准测试来讨论。

This is commonly brought up as a multimodal reasoning benchmark.

但即便在这个基准测试里，展示给模型的图像分辨率也只有32x32或64x64像素。

But even in this benchmark, the images that are being shown to the models are only 32 by 32 or 64 by 64 pixels.

要把现实世界的问题，比如户型图，压缩到这么低的像素级别，难度非常大。

And it's very hard to take a real world problem, like a floor plan and shrink it down into that many pixels.

这根本不可能实现。

It's just not possible.

不过，现在也有一些新的视觉推理基准测试正在推出。

But there are some visual reasoning benchmarks that are coming up.

让我印象最深的是今年早些时候发布的那个叫 Baby Vision 的基准测试。

And the the one that has really struck for me was this one called Baby Vision released earlier this year.

在这个测试里，他们发现，即使是目前最先进的模型，其推理能力也只相当于学龄前儿童的水平，而任何小学生在这些视觉任务上都能击败所有前沿模型。

And in this benchmark, they do show that the frontier models, the the best of them still reason around the age of a preschooler and any elementary school kid in their benchmarks were able to beat all the frontier models on these visual tasks.

这个基准测试里包含哪些类型的任务？

What are the kinds of tasks in that benchmark?

Baby Vision？

Baby vision?

是的。

Yeah.

这些任务不只是识别物体，更侧重于空间推理，比如对几何关系的理解。

So these are tasks that are not just object recognition, it's about spatial reasoning, more like geometry understanding understanding differences.

比如，让模型穿过迷宫、沿着一根绳子追踪另一端是什么、数清楚桌面上分散的物品、找出两张图片的差异、识别哪个是与众不同的那个。

So they have tasks like navigating through remains, following one string along to see what's at the other end of a string, counting things that are spread across, like, a table or a desk, finding differences between two images, finding, like, things that are the odd one out.

很多任务其实都是多步骤的。

So a lot a lot of this is kinda like more multistep.

有时候人们说的视觉思维，就是得在脑子里想象各种场景来解决问题，比如走迷宫。

What what we what sometimes people call visual thinking, where you have to, think, mentally, and mentally imagine different kinds of scenarios to solve the the problems, like navigating a maze.

我明白了。

I see.

好的。

Okay.

所以业界在视觉推理上做得不够，是因为现有的评测标准有局限。

So the industry is dropping the ball on visual reasoning because of the limitations of benchmarks.

这些评测标准并没有真正衡量视觉推理能力。

The benchmarks don't reward visual reasoning.

它们没有以自然的方式去评估视觉推理。

They're not sort of measuring visual reasoning in a naturalistic way.

你提到的 Arc AGI，就是想说明这一点。

That's your point with Arc AGI.

现在的测试环境太简单了。

The settings are kind of too toy.

但你的意思是，人们根本没用这些模型来做视觉推理。

But your point that people aren't using the models for visual reasoning.

你担心的是，我们没法从实际使用中收集到数据，了解人们到底怎么用这些模型，然后用这些数据来训练吗？

Is the concern there that we're not able to get data from production about how people are actually using the models and then use that for training?

还是说，实验室觉得根本没这个需求，所以不愿意投入这些能力？

Or that the labs don't think that the demand is there, so they're not investing in these capabilities?

是的。

Yeah.

我觉得两方面都有。

So I think it's a bit of both.

实验室确实有投入多模态理解，包括我之前担任数据负责人时的Gemini。

So the the labs definitely, including Gemini, which is where I was a data area lead.

像Project Video、Nano Banana这些项目，都能看到他们在多模态上的投入，但主要集中在生成方向。

There is investment in multimodal understanding, and you see that through, like, Project Video, Nano Banana, but it's much more on the generative side.

对。

Mhmm.