加文·贝克 - 英伟达对决谷歌，规模法则与AI经济学 - [最佳投资之道，第451期]

Trainium二代，性能会稍好一些。

Trainium two, you get a little better.

Trainium三代，我认为是第一次达到还不错的水平。

Trainium three, it's, I think the first time it's like, okay.

然后，我觉得Trainium四代可能会很不错。

And then, you know, I think Trainium four will probably be good.

如果除了Trainium和TPU之外还有很多ASIC芯片，我会感到惊讶。

I will be surprised if there are a lot of ASICs other than Trainium and TPU.

顺便说一句，Trainium和TPU最终都会在客户自有的设备上运行。

And by the way, Trainium and TPU will both run on customer owned tooling at some point.

我们可以争论何时会发生，但刚才描述的成功经济学意味着这是不可避免的。

We can debate when that will happen, but the economics of success that I just described mean it's inevitable.

无论公司怎么说，纯粹的经济学原理和第一性原理都表明这绝对不可避免。

Like no matter what the companies say, just the economics make it and reasoning from first principles make it absolutely inevitable.

如果我

If I were

从宏观角度来看这一切，我发现这些细节令人难以置信地有趣，这就像是有史以来最宏大的游戏。

to zoom all the way out on this stuff, I find these details unbelievably interesting and it's like the grandest game that's ever been played.

这太疯狂了，追踪这些发展太有趣了。

It's so crazy and so fun to follow.

有时我会忘记退一步思考，然后问自己：那又怎样？

Sometimes I forget to zoom out and say, well, so what?

比如，好吧，那么预测三代之后的情况，超越鲁本或什么的。

Like, okay, so project this forward three generations past Ruben or whatever.

所有这些疯狂发展——我们不断降低预训练扩展模型的损失——给全人类带来的红利究竟是什么？

What is like the global human dividend of all this crazy development where we keep making the loss lower on these pre training scaling models?

谁在乎呢？

Like who cares?

我已经有段时间没向这个东西提问了，每次得到的答案都让我个人感到震撼。

Like it's been a while since I've asked this thing something that I wasn't kind of blown away by the answer for me personally.

这场疯狂的基础设施竞赛将让我们解锁接下来的哪些新事物，因为它们如此成功？

What are the next couple of things that all this crazy infrastructure war allows us to unlock because they're so successful?

如果要我预测发展路径，我认为Blackwell模型将会令人惊叹。

If I were to posit like an EventPath, I think the Blackwell models are going to be amazing.

GB300带来的每令牌成本大幅降低，可能更多是MI4.50而非MI3.55的功劳，这将使这些模型能够进行更长时间的思考，意味着它们将能完成新任务。

The dramatic reduction in per token cost enabled by the GB 300 and probably more the MI $4.50 than the MI $3.55 will lead to these models being allowed to think for much longer, which means they're going to be able to do new things.

这让我印象深刻。

I was very impressed.

GMI三代帮我订了餐厅。

GMI three made me a restaurant reservation.

这是它第一次为我做实事。

It's the first time it's done something for me.

我是说，除了检索信息和教我知识之外。

And I mean, other than like, go research something and teach me stuff.

如果能订餐厅，那离订酒店、机票和叫Uber也不远了。

If you can make a restaurant reservation, you're not that far from being able to make a hotel reservation and an airplane reservation and order me an Uber.

突然间

And all of a

突然间你就有了一个助手。

sudden you got an assistant.

是啊。

Yeah.

你可以想象大家都在讨论这个，但你可以想象它就在你的手机上。

And you could just imagine everybody talks about that, but you can just imagine it's on your phone.

我认为这在近期就能实现，但一些技术领先的大公司，超过50%的客户服务已经由AI完成，这是一个价值4000亿美元的行业。

I think that's pretty near term, but some big companies that are very tech forward, 50% plus of customer support is already done by AI and that's a $400,000,000,000 industry.

要知道，AI最擅长的是说服，这正是销售和客户支持的核心能力。

And then if, you know, what AI is great about is persuasion, that's sales and customer support.

因此从企业职能来看，无非就是生产产品、销售产品以及服务客户这几项。

And so of the functions of a company, if you think about them, they're to make stuff, sell stuff, and then support the customers.

所以到2026年底左右，AI很可能已经能出色胜任其中两项职能了。

So right now, maybe in late twenty six, you're gonna be pretty good at two of them.

我确实认为AI将对媒体行业产生重大影响。

I do think it's gonna have a big impact on media.

就像我之前提到的机器人技术，终于要开始真正落地应用了。

Like I think robotics, you know, talked about the last time are gonna finally start to be real.

要知道，现在涌现了大量令人兴奋的机器人初创企业。

You know, there's an explosion and kind of exciting robotics startups.

我仍然认为主战场将在特斯拉的乐观主义者与中国企业之间，因为你知道，制造原型很容易。

I do still think that the main battle's gonna be between Tesla's optimist and the Chinese, because you know, it's easy to make prototypes.

但要大规模生产却很困难。

It's hard to mass produce them.

但这又回到了Andre Carpathi所说的观点：AI可以自动化任何可被验证的事情。

But then it goes back to that, what Andre Carpathi said about AI can automate anything that can be verified.

所以任何存在对错答案或结果的功能领域，你都可以应用强化学习让AI在这方面变得非常出色。

So any function where there's a right or wrong answer or a right or wrong outcome, you can apply reinforcement learning and make the AI really good at that.

有哪些

What are

你最喜欢的例子

your favorite examples of that

是迄今为止还是理论上的？

so far or theoretically?

我是说，模型能平衡吗？

I mean, just does the model balance?

它们会非常擅长建模。

They'll be really good at making models.

全球所有账本都能对得上吗？

Do all the books globally reconcile.

它们会非常擅长会计、复式记账，而且必须平衡。

They'll be really good at accounting, double entry bookkeeping, and has to balance.

这是可以验证对错的。

There's a verifiable you got it right or wrong.

支持还是销售。

Support or sale.

你完成销售了吗？

Did you make the sale or not?

这就像AlphaGo一样。

That's just like AlphaGo.

你赢了还是输了？

Did you win or you lose?

那个人转化了吗？

Did the guy convert or not?

客户在客服过程中要求升级了吗？

Did the customer ask for an escalation during customer support or not?

其最重要的功能之所以重要，是因为它们可以被验证。

It's most important functions are important because they can be verified.

所以我认为如果这一切开始发生并在2026年实现，Blackwell将会产生投资回报，然后这一切将继续推进。

So I think if all of this starts to happen and starts to happen in '26, there'll be an ROI on Blackwell, and then all this will continue.

然后我们会有鲁本，那将是另一个巨大的自旋量子——鲁本与MI450以及TPU V9。

And then we'll have Ruben and then that'll be another big quantum of spin Ruben and the MI four fifty and the TPU V nine.

然后我认为最有趣的问题是：人工超级智能的经济回报是什么？

And then I do think just the most interesting question is what are the economic returns to artificial super intelligence?

因为在这场伟大博弈中的所有公司，都陷入了囚徒困境。

Because all of these companies in this great game, they've been in a prisoner's dilemma.

他们害怕一旦减速就会永远失去机会。

They're terrified that if they slow down Gone forever.

而如果竞争对手没有减速，这就是生存危机。

And their competitors don't, it's an existential risk.

你知道，微软今年早些时候有大约六周时间动摇了。

And you know, Microsoft blinked for like six weeks earlier this year.

是啊。

Yeah.

我想他们会说对此感到后悔。

I think they would say they regret that.

但有了Blackwell，尤其是Rubin，从决策和投入角度看，经济规模将完全主导囚徒困境，因为数字实在太庞大了。

But with Blackwell, and for sure with Rubin, the economics are going to dominate the prisoner's dilemma from a decision making and spending perspective, just because the numbers are so big.

这又回到了关于AI投资回报率的问题。

And this goes to kind of the ROI on AI question.

人工智能的投资回报率在实证、事实和明确性上都是正向的。

And the ROI on AI has empirically, factually, unambiguously been positive.

我一直觉得奇怪，这有什么可争论的，因为购买GPU的最大开销方都是上市公司。

I just always find it strange that there's any debate about this because the largest spendages on GPUs are public companies.

他们会公布一种叫经审计的季度财报的东西。

They report something called audited quarterly financials.

你可以用这些数据计算所谓的资本回报率。

And you can use those things to calculate something called a return on invested capital.

如果进行这种计算，就会发现这些大公司在GPU上的资本回报率比增加投入前还要高。

And if you do that calculation, the ROIC of the big public spenders on GPUs is higher than it was before they ramped spending.

你可能会说，这其中部分是运营成本节省带来的。

Then you could say, well, part of that is, you know, OpEx savings.

某种程度上，这正是你期望从AI获得的投资回报。

Well, at some level that is part of what you expect the ROI to be from AI.

然后你会发现，很大部分其实只是通过TPU应用，把支撑广告和推荐系统的大型推荐引擎从CPU迁移到了GPU上。

And then you say, well, a lot of it is actually just applying TPUs, moving the big recommender systems that power the advertising and the recommendation systems from CPUs to GPUs.

而且你们已经获得了巨大的效率提升。

And you've had massive efficiency gains.

这就是为什么这些公司的收入增长都加速了。

And that's why all the revenue growth at these companies has accelerated.

但这又怎样呢？

But like, so what?

投资回报率已经摆在那里了。

The ROI has been there.

这确实很有趣。

And it is interesting.

就像所有大型互联网公司一样，负责营收的人对分配给研究人员的GPU数量感到非常恼火。

Like every big internet company, the people who are responsible for the revenue are intensely annoyed at the amount of GPUs that are being given to the researchers.

这是个非常简单的等式。

It's a very linear equation.

如果你给我更多GPU，我就能创造更多收入。

If you give me more GPUs, I will drive more revenue.

是啊。

Yeah.

把那些GPU也给我，我们就能创造更多收入、更多毛利，然后赚到大钱。

Give me even those GPUs, we'll have more revenue, more gross profit, and then we get good money.

所以这是每家公司在持续面临的斗争。

So it's this constant fight at every company.

囚徒困境的一个因素是，每个人都怀着这种宗教信仰般的信念，认为我们终将实现人工超级智能。

One of the factors in the prisoner's dilemma is everybody has this like religious belief that we're going to get to ASI.

说到底，他们真正想要的是什么？

And at the end of the day, what do they all want?

几乎所有人都想永生不死。

Almost all of them want to live forever.

他们认为人工超级智能能帮他们实现这个目标。

And they think that ASI is going to help them with that.

没错。

Right.

不错的回报。

Good return.

这是个不错的回报，但我们并不确定。

That's a good return, But we don't know.

而且如果作为人类，我们已经突破了物理、生物和化学的界限——这些支配宇宙的自然法则——那么或许ASI带来的经济回报并没有那么高。

And if as humans, have pushed the boundaries of physics, biology, and chemistry, the natural laws that govern the universe, then maybe the economic returns to ASI aren't that high.

我很好奇你有时会想到的那些泼冷水的观点。

I'm very curious about your favorite sort of throw cold water on this stuff type takes that you think about sometimes.

其中一个可能是——我很好奇你的看法——哪些因素会导致对计算需求的变化，甚至改变其发展趋势。

One would be like the things that would cause, I'm curious what you think, the things that would cause this demand for compute to change or even the trajectory of it to change.

有一个非常明显的看跌因素，那就是

There's one really obvious bear case and it is just

边缘AI。

edge AI.

而这与ASI的经济回报息息相关。

And it's connected to the economic returns to ASI.

三年后，在更大更笨重的手机上，为了容纳所需容量的DRAM，电池续航可能也会缩短，你将能够运行精简版的Gemini五或Grok四、Grok 4.1或ChatGPT，每秒30到60个token。

In three years, on a bigger and bulkier phone to fit the amount of DRAM necessary, you know, and the battery won't probably last as long, you will be able to probably run like a pruned down version of something like Gemini five or Grok four, Grok 4.1 or ChatGPT at 30, 60 tokens per second.

而且这还是免费的。

And then that's free.

这显然是苹果的战略。

And this is clearly Apple's strategy.

我们就是要成为AI的分发商，保证隐私安全，让AI在手机上运行。

It's just we're gonna be a distributor of AI and we're gonna make it privacy safe and run on the phone.

然后你可以随时调用云端的大型模型，无论你有什么问题。

And then you can call one of the big models, you know, the the god models in the cloud, whatever you have a question.

如果这种情况发生，如果每秒30到60个token、115的智商水平就够用了，我认为那...

And if that happens, if like thirty, sixty tokens a second at a one fifteen IQ is good enough, I think that's

是个熊市案例。

a bear case.

除了规模定律失效的情况。

Other than just the scaling laws break.

但就假设缩放定律持续有效而言，我们现在知道它们至少还会延续一代预训练阶段，而在后训练、中期训练、RLVR（无论人们怎么称呼）以及测试时计算推理这两条新缩放定律方面，我们还处于非常早期的阶段。

But in terms of if we assume scaling laws continue and we now know they're going to continue for pre training for at least one more generation, and we're very early in the two new scaling laws for post training, mid training, RLVR, whatever people want call it, and then test time computed inference, we're so early in those and we're getting so much better at helping the models hold more and more context in their minds as they do this test time compute.

这其实非常强大，因为大家都会问：模型要怎么知道这些？

And then that's really powerful because everybody's like, well, how's the model going to know this?

最终你会感觉能掌握足够的上下文信息。

Well, eventually you can feel you can hold enough context.

你可以把公司里每一条Slack消息、Outlook邮件和企业手册都纳入上下文。

You can just hold every Slack message and outlook message and company manual in a company in your context.

然后你就能计算新任务，并将其与你对世界的认知、你的想法、模型的判断等所有上下文进行比较。

And then you can compute the new task and compare it with your knowledge of the world, what you think, what the model thinks, all this context.

可能超长的上下文窗口就是解决当前诸多限制的关键方案。

And it may be that like just really, really long context windows are the solution to a lot of the current limitations.

这得益于KV缓存卸载之类的巧妙技术实现。

And that's enabled by all these cool tricks, like KV cash offload and stuff.

但我认为，除了缩放定律放缓或超级智能经济回报率低之外，边缘AI对我来说是目前最可信也最可怕的利空情景。

But I do think other than scaling laws slowing down, other than there being low economic returns to ASI, edge AI is to me by far the most plausible and scariest bear case.

我喜欢用不同的S曲线来可视化，就像你通过iPhone投资所经历的那样。

I like to visualize like different S curves you invested through the iPhone.

我特别喜欢看iPhone型号的演变过程，从最初笨重的砖块模样，逐步发展到如今的样子——每一代都只是稍作改进，显然在外形设计上我们已经达到了一个瓶颈。

And I love to like see the visual of the iPhone models as it sort of went from this clunky bricky thing up to the what we have now where like each one's like a little bit, you know, obviously we've sort of petered out on its form factor.

如果你为前沿模型设想类似的演进路径，是否感觉它们正处于技术范式发展的某个特定阶段？

If you picture something similar for the Frontier models themselves, does it feel like it's at a certain part of that natural technology paradigm progression?

如果你付费使用Gemini Ultra或Super Grok这些顶级AI服务，其实很难察觉到它们之间的差异。

If you're paying for Gemini Ultra or Super Grok and you're getting the good AI, it's hard to see differences.

除非我深入研究某些专业问题，比如：你认为PCI Express和以太网协议哪个更适合扩展网络规模？为什么？

Like I have to go really deep on something like, do you think PCI Express or Ethernet is a better protocol for scale up networking and why?

给我看看相关科学论文。

Show me the scientific papers.

当你在不同模型间切换，并提出这种你非常熟悉的深度问题时，才能真正看出差异。

And if you shift between models and you ask a question like that where you know it really deeply, then you see differences.

我确实玩梦幻橄榄球。

I do play fantasy football.

奖金都捐给了慈善机构，但你知道吗，这些新模型在帮助我决定该派谁上场方面表现得相当不错。

Winnings are donated to charity, but it is like, you know, these new models are quite a bit better at helping who should I play?

它们的思考方式要复杂得多。

They think in much more sophisticated ways.

如果你是个历史上表现不错的梦幻足球玩家，但这个赛季表现不佳，这就是原因，因为你没用它。

If you're a historically good fantasy football player and you're having a bad season, this is why, this is why, because you're not using it.

是啊。

Yeah.

你懂吗？

You know?

我认为我们会在越来越多的领域看到这种情况，但我确实认为它们已经达到了一定水平，除非你是真正的专家或拥有超乎想象的智力，否则很难看出进步。

And I think we'll see that in more and more domains, but I do think they're already at a level where unless you are a true expert or just have an intellect that is beyond mind, it's hard to see the progress.

这就是为什么我认为我们需要从追求更智能转向更有用。

And that's why I do think we need to shift from getting more intelligent to more useful.

除非更高的智能开始带来重大科学突破，比如我们在2026、2027年攻克癌症。

Unless more intelligence starts leading to these massive scientific breakthroughs and we're curing cancer in '26 and '27.

是啊。

Yeah.

我不认为我们即将攻克癌症，但从投资回报率曲线来看，我认为我们需要从追求智能转向追求实用性。

I don't know that we're going be curing cancer, but I do think from almost an ROI curve, we need to kind of hand off from intelligence to usefulness.

然后实用性将不得不接力给科学突破，从而创造全新的产业。

And then usefulness will then have to hand off to scientific breakthrough just that creates whole new industries.

在你看来，实用性的基础要素是什么？

What are the building blocks of usefulness in your mind?

就是能够持续可靠地完成任务。

Just being able to do things consistently and reliably.

这其中很大程度上是要保持所有上下文。

And a lot of that is keeping all the context.

比如存在大量的上下文信息。

Like there's a lot of context.

如果有人要为我规划行程，你知道，我已经形成了一些特别的偏好。

If someone wants to plan a trip for me, like, you know, I've acquired these strange preferences.

比如我关注那个叫安德鲁·休伯曼的人。

Like I follow that guy, Andrew Huberman.

所以我喜欢有朝东的阳台，这样我就能享受晨光，明白吗？

So I like to have an east facing balcony so I can get morning sun, you know?

所以AI必须记住，这是我喜欢乘坐飞机的方式。

So the AI has to remember, here's how I like to fly.

这些是我在那方面的偏好。

Here are my preferences for that.

对我来说，乘坐配备星链网络的飞机很重要。

Being on a plane with Starlink is important to me.

好的。

Okay.

这些是我历来喜欢的度假村。

Here are the resorts I've historically liked.

这些是我喜欢的区域类型。

Here are the kinds of areas I've liked.

这些是我在每个地方真正想住的房间。

Here are the rooms that I would really like at each.

这需要很多上下文信息。

That's a lot of context.

要记住并权衡所有这些因素，这是个难题。

And to keep all of that and kind of weight those, it's a hard problem.

所以我认为上下文窗口是其中很重要的一部分。

So I think context windows are a big part of it.

你知道，有个叫meter task的评估体系。

You know, there's this meter task evaluation thing.

它能持续多久

How long it can

能持续工作多久。

work How long it can work for.

你可以认为这与上下文有一定关联，虽不精确，但任务长度需要不断扩展，因为预订餐厅虽然具有经济价值，但价值并不那么大。

And you could think of that as being related in some way to context, not precisely, but that just task length needs to keep expanding because booking a restaurant and booking is economically useful, but it's not that economically useful.

但为我安排整个假期，并了解我父母、姐姐、侄女和侄子的偏好，那是个困难得多的问题。

But booking me an entire vacation and knowing the preferences of my parents, my sister, my niece, and my nephew, that's a much harder problem.

这种事情人类可能要花三四个小时来优化。

And that's something that like a human might spend three or four hours on optimizing that.

如果你能做到这一点，那就太棒了。

And then if you can do that, that's amazing.

但话说回来，我认为它必须很快在销售和客户支持方面表现出色。

But then again, I just think it has to be good at sales and customer support relatively soon.

在那之后，它必须进入，我认为它已经在这里了。

And then after that, it has to be in, I think it is already here.

我确实认为我们将看到各类产品卓越性的加速提升。

I do think we're going to see an kind of an acceleration in the awesomeness of various products.

工程师们正在利用人工智能让产品变得更好、开发得更快。

Engineers are using AI to make products better and faster.

我们都投资了Fortell这家100%公司，它简直令人惊叹。

We both invested in Fortell, the 100% company, which is just absolutely remarkable.

就像我从未想过的事情。

Like something I never would've thought of.

我认为我们将在每个垂直领域都看到类似的情况，那就是AI被用于任何公司最核心的功能——产品设计。

And we're gonna see, I think something like that in every vertical and that's AI being used for the most core function of any company, which is designing the product.

然后，已经有很多例子显示AI被用于帮助制造产品和更高效地分销，无论是优化供应链，还是通过视觉系统监控生产线。

And then it will be, you know, there's already lots of examples of AI being used to help manufacture the product and distribute it more efficiently, whether it's optimizing a supply chain, having a vision system, watch a production line.

很多事情正在发生。

A lot of stuff is happening.

我认为在整个投资回报率方面真正有趣的是，财富500强企业总是最后采用新技术的。

The other thing I think is really interesting in this whole ROI part is Fortune 500 companies are always the last to adopt a new technology.

它们很保守。

They're conservative.

它们有很多法规限制，很多律师。

They have lots of regulations, lots of lawyers.

初创企业总是第一个吃螃蟹的。

Startups are always the first.

让我们想想云计算，这是企业领域上一次真正具有变革性的新技术。

So let's think about the cloud, which was the last truly transformative new technology for enterprises.

能够将所有计算资源和云服务结合使用SaaS。

Being able to have all of your compute and the cloud and use SaaS.

所以它总是保持最新状态。

So it's always upgraded.

它总是很棒，诸如此类。

It's always great, etcetera, etcetera.

你可以在任何设备上使用它。

You can get it on every device.

我记得第一届AWS re:Invent大会应该是在2013年。

I think the first AWS re:Invent, I think it was in 2013.

而到了2014年，地球上所有初创公司都在云上运行了。

And by 2014, every startup on planet earth ran on the cloud.

那种自己购买服务器、存储设备和路由器的想法变得很荒谬。

The idea that you would buy your own server and storage box and router was ridiculous.

这种情况可能发生得更早，甚至在首届Reinvent大会之前就已经出现了。

And that probably happened like even earlier that that had probably already happened before the first Reinvent.

第一批大型财富500强企业开始将其标准化，大约是在五年之后。

The first big fortune 500 companies started to standardize on it, like maybe five years later.

在AI领域你也能看到这种趋势。

You see that with AI.

我相信你在创业公司中已经见识过这一点了。

I'm sure you've seen this in your startups.

我认为风投比公开市场投资者更普遍看好AI的一个原因是，风投看到了实实在在的生产力提升。

And I think one reason VCs are more broadly bullish on AI than public market investors is VCs see very real productivity gains.

有大量图表显示，在同等收入水平下，如今的公司员工数量比两年前的公司显著减少。

There's all these charts that for a given level of revenue, a company today has significantly lower employees than a company of two years ago.

原因是AI正在承担大量销售、支持工作，并帮助完善产品。

And the reason is AI is doing a lot of the sales, the support and helping to make the product.

我是说，你知道Iconic有些图表，a16z的David George是我的好友，他是个很棒的人。

I mean, there's, you know, Iconic has some charts, a16z, by the way, David George is a good friend, great guy.

你知道，他有这个模型破坏者的理论。

You know, he has this model busters thing.

所以有非常明确的数据表明这种情况正在发生。

So there's very clear data that this is happening.

因此那些了解风险投资领域的人都能看到这一点。

So people who have a lens into the world of venture see this.

我确实认为这在第三季度非常重要。

And I do think it was very important in the third quarter.

这是首个季度我们看到科技行业外的财富500强企业给出AI驱动增长的具体量化案例。

This is the first quarter where we had fortune 500 companies outside of the tech industry give specific quantitative examples of AI driven uplift.

比如C。

So C.

H。

H.

罗宾逊公司的盈利增长了约20%。

Robinson went up something like 20% on earnings.

我该告诉大家C.

Should I tell people what C.

H.

H.

Robinson是做什么的吗？

Robinson does?

是的。

Yeah.

比如说，有辆卡车从芝加哥开往丹佛。

Like, let's just say a truck goes from Chicago to Denver.

而卡车司机住在芝加哥。

And then the trucker lives in Chicago.

所以它要从丹佛返回芝加哥。

So it's gonna go back from Denver to Chicago.

C.那里有个空载。

There's an empty load at C.

H.

H.

罗宾逊公司与这些卡车司机和运输公司建立了所有关系，他们通过匹配货主需求与空载供应来提高运输效率。

Robinson has all these relationships with these truckers and trucking companies, and they match shippers demand with that empty load supply to make the trucking more efficient.

你知道，他们是一家货运代理公司，实际上有很多类似的公司，但他们是规模最大、最具主导地位的。

You know, they're a freight forwarder, you know, there's actually lots of companies like this, but they're the biggest and most dominant.

所以他们做的最重要的事情之一就是报价和确认可用性。

So one of the most important things they do is they quote price and availability.

比如有客户打电话给他们说：'嘿，我急需三辆十八轮大卡车从芝加哥运到丹佛。'

So somebody, a customer calls them up and says, Hey, I urgently need three eighteen wheelers from Chicago to Denver.

但在过去，他们需要花费大约十五到四十五分钟来处理。

But in the past, they said it would take them, you know, fifteen to forty five minutes.

而且他们只对60%的进站询价请求进行了报价。

And they only quoted 60% of inbound requests.

借助人工智能，他们现在能对100%的请求进行报价，并且只需几秒钟就能完成。

With AI, they're quoting a 100% and doing it in seconds.

因此他们公布了一个亮眼的季度报告，股价上涨了20%，这完全得益于AI驱动的生产力提升。

And so they printed a great quarter and the stock went up 20% and it was because of AI driven productivity.

这正在影响收入线、成本线等方方面面。

That's impacting the revenue line, the cost line, everything.

我其实非常担心Blackwell芯片的投资回报率会出现空档期，因为我们投入了巨额资金。

I was actually very worried about the idea that we might have this Blackwell ROI air gap because we're spending so much money on Blackwell.

那些Blackwell芯片仅用于训练模型，而训练本身无法产生直接收益。

Those Blackwells are being used for training and there's no ROI on training.

训练就是在构建模型。

Training is you're making the model.

真正的投资回报来自模型推理阶段。

The ROI comes from inference.

所以我真的很担心，我们可能会面临长达三个季度的资本支出高到难以想象的时期。

So I was really worried that, you know, we're gonna have maybe this three quarter period where the CapEx is unimaginably high.

那些Blackwell芯片目前仅用于训练用途。

Those Blackwells are only being used for training.

对。

Right.

我们的保持不变，他们的在上升。

Ours staying flat, eyes going up.

是的，没错。

Yeah, exactly.

所以投资回报率下降，你可以看看Meta的情况，Meta公布的财报显示，因为他们一直没能研发出前沿模型。

So ROIC goes down and you could see like Meta, Meta they printed, you know, because Meta has not been able to make a frontier model.

Meta公布的季度财报显示投资回报率下降，这对股价很不利。

Meta printed a quarter where ROIC declined and that was not good for the stocks.

我当时非常担心这一点。

I was really worried about that.

我认为这些数据点很重要，它们表明我们或许能度过这段投资回报率的潜在空窗期。

I do think that those data points are important in terms of suggesting that maybe we'll be able to navigate this potential air gap in ROIC.

是啊。

Yeah.

这让我思考，在这个市场中，与其他所有公司不同，市值最高的那10家公司吸引了所有注意力。

It makes me wonder about in this market, unlike everybody else, it's the 10 companies at the top that are all the market cap, more than all of the attention.

标普500指数中还有490家公司。

There's four ninety other companies, the S and P 500.

你也研究过那些公司。

You studied those too.

比如，你对那群公司有什么看法？

Like, what do you think about that group?

现在似乎没人谈论、也没人真正关心的那群公司中，有什么让你感兴趣的点？因为它们既没有带来回报，在整体指数中的占比也越来越小。

Like what is interesting to you about the group that now nobody seems to talk about and no one really seems to care about because they haven't driven returns and they're a smaller percent of the overall index?

我认为如果越来越多的公司公布这类财报，人们就会开始关注。

I think that people are going to start to care if you have more and more companies print these C.

C。

H.

H。罗宾逊式的季度业绩。

Robinson like quarters.

我认为那些历史上经营良好的公司，之所以能保持长期的成功记录，是因为不充分利用技术就无法取得成功。

I think the companies that have historically been really well run, the reason they have a long track record of success, you cannot succeed without using technology well.

因此，如果企业内部具有实验和创新的文化氛围，我认为它们会在AI领域表现出色。

And so if you have a kind of internal culture of experimentation and innovation, I think you will do well with AI.

我敢打赌，最优秀的投资银行会比那些落后的银行更早、更好地采用AI技术，历史往往就是未来的序幕。

I would bet on the best investment banks to be earlier and better adopters of AI than maybe some of the trailing banks, just sometimes past his prologue.

在我看来，这种情况很可能会发生——我有一个强烈的观点：所有这些风险投资公司都在建立控股公司，声称要利用AI来提升传统企业的业务表现。

And I think it's likely to be in this case, one strong opinion I have, all these VCs are setting up these holding companies and, know, we're gonna use AI to make traditional businesses better.

他们是非常精明的风险投资家，有着出色的业绩记录，但这正是私募五十年来一直在做的事。

And they're really smart VCs and they're great track records, but that's what private equity has been doing for fifty years.

你不可能在私募的游戏中胜过他们。

You're just not gonna be private equity at their game.

这就是Vista早期所采取的策略。

This is what Vista did in the early days.

对吧？

Right?

是的。

Yeah.

我确实认为私募股权最近经历了一段艰难时期。

And I do think this is actually private equities, maybe had a little bit of a tough run.

估值倍数已经上升了。

Just multiples have gone up.

现在私募资产变得更昂贵了。

Now private assets are more expensive.

融资成本也增加了。

The cost of financing has gone up.

让公司上市变得很困难，因为公开市场的估值比私募估值低了30%。

It's tough to take a company public because the public valuation is 30% lower than the private valuation.

所以私募确实遇到了困境。

So PE's had a tough run.

实际上我认为这些私募股权公司在系统化应用AI方面会做得相当不错。

I actually think these private equity firms are gonna be pretty good at systematically applying AI.

我们还没怎么讨论过Meta、Anthropic或OpenAI。

We haven't spent much time talking about meta, anthropic, or open AI.

我很想听听你对基础设施领域所有进展的看法，就像我们之前讨论的那样。

And I'd love your impression on everything that's going on in this infrastructure side that we talked about.

在这场宏大棋局中，这三家都是举足轻重的参与者。

These are three really important players in this grand game.

我们目前讨论的这些发展将如何具体影响这几家公司？

How does all of this development that we've discussed so far impact those players specifically?

首先，让我从整体上谈谈前沿模型。

First thing, let me just say about frontier models broadly.

是的。

Yeah.

在2023和24年，我很喜欢引用埃里克·维克里亚的话，我们的朋友埃里克·维克里亚是个才华横溢的人。

In 2023 and '24, I was fond of quoting Eric Vichria and Eric Vichria's statement, our friend, brilliant man.

埃里克常说，基础模型是历史上增值最快的资产。

And Eric would always say, Foundation models are the fastest appreciating assets in history.

我认为他有90%是正确的。

And I would say he was 90% right.

我对这个说法做了修改。

I modified the statement.

我说过：'没有独特数据和互联网规模分发的基础模型，是历史上增值最快的资产'。

I said, Foundation models without unique data and internet scale distribution are the fastest appreciating assets in history.

而推理从根本上深刻改变了这一局面。

And reasoning fundamentally changed that in a really profound way.

用杰夫·贝索斯的话说，每个伟大的互联网公司核心都存在这样一个循环飞轮。

So there was a loop, flywheel to quote Jeff Bezos, that it was at the heart of every great internet company.

这个逻辑就是：你做出了好产品，

And it was, you made a good product.

获得了用户，用户使用产品产生的数据可以反馈到产品中使其变得更好。

You got users, those users using the product generated data that could be fed back into the product to make it better.

这个飞轮已经在Netflix、亚马逊、Meta和谷歌运转了十多年。

And that flywheel has been spinning at Netflix, at Amazon, at Meta, at Google for over a decade.

这是一个极其强大的飞轮效应。

And that's an incredibly powerful flywheel.

这就是为什么这些互联网企业如此难以被竞争超越。

And it's why those internet businesses were so tough to compete with.

这就是为什么它们具有规模报酬递增的特性。

It's why they're increasing returns to scale.

所有人都在谈论网络效应。

Everybody talks about network effects.

这对社交网络来说曾经很重要。

They were important for social networks.

我不确定Meta现在在多大程度上还能算作社交网络。

I don't know to what extent Meta is a social network anymore.

它更像是个内容分发平台，但正是由于这个飞轮效应，它们实现了规模报酬递增。

It's more like a content distribution, but they just had increasing returns to scale because of that flywheel.

而在AI具备推理能力之前的时代，这种动态机制并不存在。

And that dynamic was not present in the pre reasoning world of AI.

你预训练了一个模型，把它放到现实世界中，它就那样了。

You pre trained a model, you let it out in the world and it was what it was.

这实际上相当困难。

And it was actually pretty hard.

他们会通过RLHF（人类反馈强化学习）来尝试改进机器人模型。

They would do RLHF reinforcement learning with human feedback, and you try and make the bot model better.

或许你能从推特氛围中感觉到人们不喜欢这样。

And maybe you'd get a sense from Twitter vibes that people didn't like this.

于是你进行微调。

And so you tweak it.

虽然有上下箭头的小按钮，但要把这些反馈真正融入模型其实相当困难。

There are the little up and down arrows, but it was actually pretty hard to feed that back into the model.

虽然推理能力还处于早期阶段，但那个飞轮已经开始转动了。

With reasoning, it's early, but that flywheel started to spin.

这对这些前沿实验室来说意义深远。

And that is really profound for these Frontier Labs.

首先，推理能力从根本上改变了前沿实验室的行业格局。

So one, reasoning fundamentally changed the industry dynamics of Frontier Labs.

具体解释一下为什么是这样。

Just explain why specifically that is.

到底发生了什么？

Like what is going on?

因为如果很多人都在问类似的问题，他们要么一直喜欢这个答案，要么一直不喜欢。

Because if a lot of people are asking a similar question, they're consistently either liking or not liking the answer.

那你就可以利用这种可验证的反馈作为奖励信号。

Then you can kind of use that like that has a verifiable reward.

这是个好结果。

That's a good outcome.

然后你就可以把这些优质回答反馈给模型。

And then you can kind of feed those good answers back into the model.

目前这个飞轮才刚刚开始转动。

And we're very early at this flywheel spinning.

是的。

Yeah.

明白

Got

了。

it.

虽然现在操作起来很困难，但你能看到它开始运转了。

Like it's hard to do now, but you can see it beginning to spin.

因此，这是影响所有这些动态的首要重要事实。

So this is important fact number one for all of those dynamics.

其次，我认为非常关键的是，Meta的马克·扎克伯格在今年一月初曾表示——虽然我可能会记错原话——但他高度自信地宣称2025年某个时间点他们将拥有最优秀且性能最佳的人工智能。

Second, I think it's really important that Meta, you know, Mark Zuckerberg at the beginning of this year in January said, I'm highly confident I'm gonna get the quote wrong, that at some point in 2025, we're gonna have the best and most performant AI.

我不确定他是否排得进前100名。

I don't know if he's in the top 100.

所以他的判断可以说是错得离谱。

So he was as wrong as it was possible to be.

我认为这是一个非常重要的事实，因为这表明这四家公司所做的事情确实非常困难，因为Meta投入了大量资金却失败了。

And I think that is a really important fact because it suggests that what these four companies have done is really hard to do because Meta threw a lot of money at it and they failed.

Yama Koon不得不离开。

Yama Koon had to leave.

他们不得不为那些人才支付了十亿美元。

They had to have the famous billion dollar for AI researchers.

顺便说一句，微软也失败了。

By the way, Microsoft also failed.

他们并未做出如此明确的预测，但他们确实是翻译中。

They did not make such an unequivocal prediction, but they bought Inflection AI.

而且他们有很多评论和预期表明我们的内部模型会迅速改进。

And there were a lot of comments from them that we anticipate our internal models quickly getting better.

我们将在内部模型上运行越来越多的人工智能。

And we're going to run more and more of our AI on our internal models.

不。

Nope.

亚马逊收购了一家名为Adept AI的公司。

Amazon, they bought a company called Adept AI.

他们有自己的模型叫Nova。

They have their models called Nova.

我认为它们排不进前20名。

I don't think they're in the top 20.

显然这比人们一年前预想的要困难得多。

So clearly it's much harder to do than people thought a year ago.

这其中有许多许多原因。

And there's many, many reasons for that.

比如要维持大型GPU集群的协同运作实际上非常困难。

Like it's actually really hard to keep a big cluster of GPUs coherent.

很多这些公司过去习惯于运行以成本优化而非性能优化的基础设施。

A lot of these companies were used to running their infrastructure to optimize for cost instead of performance.

复杂性与性能。

Complexity and performance.

在大规模集群中保持GPU的高利用率及其复杂性，这实际上非常困难。

Complexity and keeping the GPUs running at high utilization rate in a big cluster, it's actually really hard.

而且不同公司在GPU运行效率上存在巨大差异。

And there are wild variations in how well companies run GPUs.

考虑到物理定律的限制，也许最多只能让二三十万个黑井保持一致性。

If the most anybody, because of laws of physics, you know, maybe you can get two or 300,000 black wells coherent.

我们拭目以待。

We'll see.

但如果你的集群运行时间只有30%，而竞争对手有90%，那你根本没法竞争。

But if you have 30% uptime on that cluster and you're competing with somebody who has 90% uptime, you're not even competing.

所以第一点，人们在GPU运行效率上存在巨大差异。

So one, there's a huge spectrum in how well people run GPUs.

第二，我认为这些AI研究人员喜欢谈论品味，我觉得这很有趣。

Two, then I think there is, you know, these AI researchers, they like to talk about taste.

我觉得这非常滑稽。

I find it very funny.

你知道，哦，为什么你能赚这么多钱？

You know, oh, why do you make so much money?

我的品味非常好。

I have very good taste.

明白吗？

You know?

所谓品味，就是你对要进行的实验有良好的直觉判断力。

What taste means is you have a good intuitive sense for the experiments to perform.

这就是为什么你要付给人们高薪——因为事实证明，随着这些模型规模扩大，你无法先在千级GPU集群上做实验，再在十万级GPU集群上复现。

This is why you pay people a lot of money because it actually turns out that as these models get bigger, you can no longer run an experiment on a thousand GPU cluster and replicate it on a 100,000 GPUs.

你必须直接在五万级GPU集群上运行实验，这可能需要耗时数天。

You need to run that experiment on 50,000 GPUs and maybe it takes, you know, days.

因此机会成本极高。

And so there's a very high opportunity cost.

你必须拥有一支真正优秀的团队，能够正确决策该进行哪些实验。

You have to have a really good team that can make the right decisions about which experiments to run on this.

然后你还需要在训练后阶段做好所有强化学习，并优化推理时的计算资源。

And then you need to do all the reinforcement learning during post training well, and the test time compute well.

这真的很难做到，但所有人都以为很容易。这些事情就像...我早年做零售分析师时常说，随便选美国任何一个垂直领域。

It's really hard to do and everybody thinks it's easy, but all those things, you know, I used to have this saying, like, I was a retail analyst long ago, pick any vertical in America.

如果你能在50个州运营上千家门店，保持店面整洁、照明良好、商品陈列合理、价格优惠，员工友善且不偷窃，你就能成为200亿甚至300亿美元市值的公司。

If you can just run a thousand stores in 50 states and have them clean, well lit, stocked with relevant goods at good prices and staffed by friendly employees who are not stealing from you, you're going to be a $20,000,000,000 company, a $30,000,000,000 company.

但全美能做到这点的公司不超过15家。

Like 15 companies have been able to do that.

这真的很难。

It's really hard.

道理是相通的。

And it's the same thing.

要完美做好所有这些事情真的非常困难。

Doing all of these things well is really hard.

而随着这个飞轮效应持续运转，它正在形成行业准入壁垒。

And then reasoning with this flywheel, this is beginning to create barriers to entry.

更重要的是，这些实验室——XAI、Gemini、OpenAI和Anthropic——每个内部都拥有更先进的模型检查点。

And what's even more important, every one of those labs, XAI, Gemini, OpenAI and Anthropic, they have a more advanced checkpoint internally of the model.

检查点基本上就是持续优化这些模型，然后发布一个阶段性版本。

Checkpoint is just kind of continuously working on these models and then you release kind of a checkpoint.

这就是为什么这些模型能快速迭代的原因，是的，

And then the reason these models get fast Yeah, the

他们内部使用的版本是专门为

one they're using internally is for a

球形优化设计的，他们正在用那个模型训练下一代模型。

spherical And they're using that model to train the next model.

如果你没有那个最新检查点，追赶会变得极其困难。

And if you do not have that latest checkpoint, it's getting really hard to catch up.

中国的开源项目简直是天赐给Meta的礼物，因为你可以利用中国的开源资源。

Chinese open source is a from God to meta because you can use Chinese open source.

那就能成为你的检查点。

That can be your checkpoint.

你可以将其作为一种启动方式。

And you can use that as a way to kind of bootstrap this.

我确信他们以及其他人都在尝试这么做。

And that's what I'm sure they're trying to do and everybody else.

最大的问题和关键变数在于，我认为中国在稀土问题上犯了个严重错误。

The big problem and the big, a giant swing factor, I think China's made a terrible mistake with this rare earth thing.

因为中国有华为昇腾芯片，虽然性能尚可但相比已淘汰的热保存技术优势不大。

So China, because, you know, they have Huawei, Ascend, and it's a decent chip versus something like the deprecated hot preserving something that looks okay.

所以他们试图强制中国开源项目使用国产自主研发芯片。

And so they're trying to force Chinese open source to use their Chinese chips, their domestically designed chips.

问题是 Blackwell 即将面世，这将导致美国前沿实验室与中国开源之间的差距急剧扩大。

The problem is Blackwell's going to come out now and the gap between these American frontier labs and Chinese open source is going to blow out because of Blackwell.

实际上深度求索在最新技术论文 V3.2 中坦言：'难以与美国前沿实验室竞争的原因之一是我们的算力不足'。

And actually DeepSeek in their most recent technical paper, V3.2, said one of the reasons we struggle to compete with the American frontier labs is we don't have enough compute.

这是他们用非常政治正确（但仍带风险）的方式在暗示：因为中国表态拒绝 Blackwell 芯片。

That was their very politically correct, still a little bit risky way of saying, cause China said we don't want the black wells.

没错。

Right.

他们说的是，

And they're saying,

能不能把Blackwell给我们？

won't you please give us the black wells?

这可能是个重大失误。

That might be a big mistake.

如果这样发展下去，美国四大实验室与中国开源之间的差距会进一步拉大，这让其他追赶者更加困难，因为差距正在扩大。

So if you just kind of play this out, these four American labs are gonna start to widen their gap versus Chinese open source, which then makes it harder for anyone else to catch up because that gap is growing.

所以你无法利用中国开源来启动。

So you can't use Chinese open source to bootstrap.

而从地缘政治角度看，中国本以为他们掌握了筹码。

And then geopolitically, China thought they had the leverage.

他们会意识到：哦，这下搞砸了。

They're going to realize, oh, whoopsie daisy.

我们确实需要Blackwell芯片

We do need the Blackwells.

不幸的是这对他们来说很可能是

And unfortunately they'll probable for them.

他们可能会在2026年底意识到这一点

They'll probably realize that in late twenty six.

到那时，将会付出巨大的努力

And at that point, there's an enormous effort underway.

DARPA有各种非常酷的项目，包括国防部的计划，旨在激励针对稀土资源的创新技术解决方案。

DARPA has, there's all sorts of really cool DARPA and DOD programs to incentivize really clever technological solutions for rare earths.

而且有很多稀土矿藏位于对美国非常友好的国家，这些国家并不介意用传统方式进行提炼。

And then there's a lot of rare earth deposits in countries that are very friendly to America that don't mind actually refining it in the traditional way.

所以我认为稀土问题会比任何人预想的都更快得到解决。

So I think rare earths are going to be solved way faster than anyone thinks.

要知道，它们显然并不那么稀有。

Know, they're obviously not that rare.

它们只是名字起错了。

They're just misnamed.

它们之所以稀有是因为提炼过程非常复杂。

They're rare because they're really messy to refine.

所以从地缘政治角度看，我认为Blackwell确实很重要，随着这个差距的扩大，它将给美国带来很大的筹码。

And so geopolitically, I actually think Blackwell's pretty significant and it's going to give America a lot of leverage as this gap widens.

然后在这种背景下，回到这些公司之间的动态关系，xAI将率先推出首个Blackwell模型，很可能也会成为第一个大规模使用Blackwell进行推理的公司。

And then in the context of all of that, going back to the dynamics between these companies, xAI will be out with the first Blackwell model and then they'll be the first ones probably using Blackwell for inference at scale.

我认为这对他们来说是个重要时刻。

And I think that's an important moment for them.

顺便说一句，这很有趣。

And by the way, it is funny.

比如你去open router看看，就会发现他们占据主导份额。

Like if you go on open router, you can just look, they have dominant share.

现在open routers不管具体是什么，它只占API令牌的1%，但这表明他们处理了1,350,000,000,000次请求。

Now open routers, whatever it is, it's 1% of API tokens, but it's an indication they process 1,350,000,000,000.

谷歌大概处理了80到9000亿次。

Google did like 8 or 900,000,000,000.

这大概是过去七天或上个月的数据。

This is like whatever it is last seven days or last month.

Anthropic处理了7000亿次。

Anthropic was at 700,000,000,000.

xAI的表现确实非常出色。

Like xAI is doing really, really well.

而且他们的模型非常棒。

And the model is fantastic.

我强烈推荐，你会看到xAI推出这个。

I highly recommend it, But you'll see xAI come out with this.

OpenAI会更快推出。

OpenAI will come out faster.

OpenAI试图通过Stargate解决的问题是因为他们需要为算力支付额外成本。

OpenAI's issue that they're trying to solve with Stargate is because they pay a margin to people for compute.

也许管理他们电脑的人并不擅长运行GPU。

And maybe the people who run their computer are not the best at running GPUs.

他们是高成本的token生产者。

They're a high cost producer of tokens.

我认为这在一定程度上解释了他们很多

And I think this kind of explains a lot of their

最近的Code Red。

Code Red recently.

是的。

Yeah.

嗯，那1.4万亿美元的支出承诺。

Well, the $1,400,000,000,000 in spending commitments.

我想那就像是，他们知道自己需要筹集大量资金，尤其是如果谷歌继续保持当前吸走整个行业经济氧气的策略。

And I think that was just like, hey, they know they're gonna need to raise a lot of money, particularly if Google keeps its current strategy of sucking the economic oxygen out of the room.

而且，你知道，从1.4万亿美元的粗略感觉很快变成Code Red状态，速度相当快，对吧？

And, know, you go from 1,400,000,000,000.0 rough vibes code red, like pretty fast, you know?

他们发布红色警报正是因为所有这些动态因素。

And the reason they have a code red is because of all these dynamics.

所以他们虽然会推出新模型，但在单位token成本上仍无法与XAI、谷歌和Anthropic相抗衡。

So then they'll come out with a model, but they will not affix their per token cost disadvantage yet relative to both XAI and Google and Anthropic at that point.

Anthropic是家优秀的公司。

Anthropic is a good company.

要知道，他们的现金消耗速度远低于OpenAI，增长速度却更快。

You know, they're burning dramatically less cash than OpenAI and growing faster.

所以我认为必须给予Anthropic高度认可。

So I think you have to give Anthropic a lot of credit.

这很大程度上得益于他们与谷歌和亚马逊在TPU及Trainium芯片上的合作关系。

And a lot of that is their relationship with Google and Amazon for the TPUs and the Trainiums.

因此Anthropic得以共享谷歌享有的相同发展红利。

And so Anthropic has been able to benefit from the same dynamics that Google has.

我认为这盘大棋局很能说明问题。

I think it's very indicative in this great game of chess.

你可以看看达里奥和詹森的公开言论，他们之间可能有过一些交锋。

You can look at Dario and Jensen, maybe have taken a few public comments, you know, that were between them.

较量。

Jousting.

一点小小的较量。

A little bit of jousting.

Anthropic刚刚与英伟达签署了50亿美元的协议。

Well, Anthropic just signed the $5,000,000,000 deal with Nvidia.

这是因为达里奥是个聪明人，他了解Blackwell和Rubid相对于TPU的这些动态。

That is because Dario is a smart man and he understands these dynamics about Blackwell and Rubid relative to TPU.

所以英伟达现在从拥有xAI和OpenAI两个战士，变成了三个战士，这有助于英伟达与谷歌的对抗。

So Nvidia now goes from having two fighters, xAI and OpenAI, to three fighters that helps in this Nvidia versus Google battle.

如果Meta能迎头赶上，那将非常重要。

And then if Meta can catch up, that's really important.

我确信英伟达正在尽其所能帮助Meta。

I am sure NVIDIA is doing whatever they can to help Meta.

你正在这样运行那些GPU。

You're running those GPUs this way.

也许我们应该这样拧螺丝或者那样调节旋钮。

Maybe we should twist the screw this way or turn the dial that way.

如果Blackwell回归的话——看起来它很可能会回归——那也将非常好，因为届时中国的开源生态也会回归。

And then it will be also if Blackwell comes back to which it seems like it'll probably happen, that will also be very good because then Chinese open source will be back.

我总是对事物的两极充满好奇。

I'm always so curious about the poles of things.

一极是你关注的其它突破性进展，比如我们之前讨论过的数据中心里非芯片类的东西就是一个例子。

One pole would be the other breakthroughs that you have your mind on, things in the data center that aren't chips that we've talked about before as one example.

我认为未来三到四年内，世界上最重要的事情将是太空中的数据中心。

I think the most important thing that's going to happen in the world, in this world, in the next three to four years is data centers in space.

这对地球上所有建设发电厂或数据中心的人都有着极其深远的影响。

And this has really profound implications for everyone building a power plant or a data center on planet earth.

明白吗？

Okay?

而且现在正掀起一股巨大的淘金热。

And there is a giant gold rush into this.

我还没听说过这件事。

I haven't heard anything about this.

所以请赐教。

So please educate.

是啊。

Yeah.

要知道，就像大家都觉得，嘿，AI有风险，但你猜怎么着？

Know, it's like everybody thinks like, Hey, AI is risky, but you know what?

我要建一个数据中心。

I'm going build a data center.

我要建一个能运行数据中心的发电厂。

I'm going to build a power plant that's going to do a data center.

我们会需要这个的。

We will need that.

但从第一性原理来思考，数据中心应该建在太空。

But if you think about it from first principles, data centers should be in space.

运营数据中心的基本要素是什么？

What are the fundamental inputs to running a data center?

就是电力和冷却系统。

Their power and their cooling.

然后还有芯片。

And then there are the chips.

如果从总成本的角度来考虑。

If you think about it from a total cost perspective.

没错。

Yeah.

还有让这些神奇机器产出代币所需的投入。

And just the inputs to making the tokens come out of the magic machines.

在太空中，卫星可以24小时持续接受阳光照射，而且太阳光强度要高出30%。

In space, you can keep a satellite in the sun twenty four hours a day and the sun is 30% more intense.

你可以让卫星持续接收光照。

You can have the satellite always kind of catching light.

太阳光强度增加30%，这导致外太空的辐射量是地球表面的六倍。

The sun is 30% more intense and this results in six times more irradiance in outer space than on planet earth.

所以你能获得大量太阳能。

So you're getting a lot of solar energy.

第二点，由于卫星24小时处于日照中，不需要电池。

Point number two, because you're in the sun twenty four hours a day, don't need a battery.

这是成本中很大一部分。

This is a giant percentage of the cost.

因此太阳系中最低成本的能源就是太空中的太阳能。

So the lowest cost energy available in our solar system is solar energy and space.

其次是冷却问题。

Second for cooling.

在这些机架中，大部分质量和重量都来自冷却系统。

In one of these racks, a majority of the mass and the weight is cooling.

这些数据中心的冷却系统极其复杂。

The cooling in these data centers is incredibly complicated.

包括暖通空调、冷却分配单元、液体冷却等。

The HVAC, the CDUs, the liquid cooling.

看起来非常酷。

It's very cool to see.

令人惊叹的景象。

It's amazing to see.

在太空中，冷却是免费的。

In space, cooling is free.

只需在卫星背阴面安装散热器即可。

You just put a radiator on the dark side of the satellite.

这简直太棒了。

It's fucking gold.

而且

And it's

几乎可以接近绝对零度。

as close to absolute zero as you can get.

所以所有这些都不需要了，能省下巨额成本。

So all that goes away and that is a vast amount of cost.

好的。

Okay.

让我们想想，也许每个卫星就像一个机架。

Let's think about how these maybe each satellite is kind of a rack.

这是一种思考方式。

It's one way to think of it.

也许有些人会造更大的卫星，相当于三个机架。

Maybe some people make bigger satellites that are three racks.

那么，你打算如何连接这些机架呢？

Well, how are you going to connect those racks?

有趣的是在数据中心，机架之间超过一定距离就会用光纤连接。

Well, it's funny in the data center, the racks are over a certain distance connected with fiber optics.

这其实就是指激光通过电缆传输。

And that just means a laser going through a cable.

唯一比激光通过光纤电缆更快的是激光在绝对真空中传输。

The only thing faster than a laser going through a fiber optic cable is a laser going through absolute vacuum.

因此，如果能用激光将这些太空中的卫星连接起来，实际上就拥有了比地球数据中心更快、更连贯的网络。

So if you can link these satellites in space together using lasers, you actually have a faster and more coherent network than in a data center on earth.

好的。

Okay.

至于训练过程，那将需要很长时间。

For training, that's going to take a long time.

因为规模太庞大了。

Because it's so big.

没错，正因为规模如此庞大，训练最终总会完成的。

Yeah, just because it's so big, training will eventually happen.

但说到推理环节，让我们考虑下用户体验。

But then for inference, let's think about the user experience.

当我向Grok询问关于你的事情，它给出了一个很好的回答时，整个过程是这样的。

When I asked Grok about you and it gave the nice answer, here's what happened.

无线电波从我的手机传输到基站。

A radio wave traveled from my cell phone to a cell tower.

然后信号到达基站，进入光纤电缆，传到纽约某个城域汇聚设施，大概离这里就10个街区左右。

Then it hit the base station, went into a fiber optic cable, went to some sort of metro aggregation facility in New York, probably within like, you know, 10 blocks of here.

那里有个小型城域路由器，把数据包路由到某个大型XAI数据中心。

There's a small little metro router that's routed those packets to a big XAI data center somewhere.

完成计算后，结果沿着原路返回。

The computation was done and it came back over the same path.

如果卫星能直接与手机通信（星链已展示过直连手机技术），整个过程就变成咻-咻两下。

If the satellites can communicate directly with the phone and Starlink has demonstrated direct to cell capability, you just go boom, boom.

这种用户体验更好，成本也更低。

It's a much better, lower cost user experience.

所以从第一性原理来看，太空数据中心在各方面都优于地球上的数据中心。

So in every way, data centers in space, from a first principles perspective, are superior to data centers on earth.

如果我们能瞬间实现这个构想，这部分我能理解。

If we could teleport that into existence, I understand that portion.

为什么这不会实现呢？

Why will that not happen?

是因为发射成本吗？

Is it launch cost?

是因为发射可用性吗？

Is it launch availability?

是因为容量问题吗？

Is it capacity?

需要大量这种设备。只有星舰能经济高效地实现这个目标。

Need a lot of those Like stationary the Starships are the only ones that can economically make that happen.

我们需要大量这样的星舰。

We need a lot of those Starships.

也许中国或俄罗斯能够实现火箭着陆。

Maybe China or Russia will be able to land a rocket.