本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
好的。
Okay.
我回来了，和Dylan Patel一起，他是Semi Analysis的创始人。
I'm back with Dylan Patel, founder of Semi Analysis.
我们每年感恩节家庭聚餐的版本，就是定期录一期播客。
Our version of a family Thanksgiving dinner is a regular yearly podcast.
我们其实没有血缘关系。
We are not actually related.
我们会告诉大家的。
We'll tell the people this.
这会打破那个神话。
It will destroy the myth.
基本上，世界经济的走向越来越取决于实验室经济的走向，以及计算市场的走向等等。
Walk me through, basically where the world economy is headed is more and more becoming a function of where like lab economics are headed, where like the compute market is headed, etcetera.
所以我想了解几年后那个疯狂未来会是什么样子，但让我们先从今天的情况开始吧。
So I want to understand where the crazy future ends up within a few years, but let's start with just where we are today.
那你现在给我讲讲实验室的计算能力和收入情况吧，顺便展望一下未来一两年。
So walk me through lab compute and lab revenue right now and maybe projecting out a year or two.
对，回顾去年，甚至到年底，美国GDP增长的大部分其实都来自AI基础设施。
Yeah, so when we go back to last year, even at the end of the year, most of GDP growth in America was just AI infrastructure.
而且随着
And as
我们展望今年，新增的计算能力中大约有三分之一是给OpenAI和Anthropic这些实验室用的。
we look towards this year, about a third of the compute coming online is for the labs for OpenAI and Anthropic.
虽然可能是别人建的然后租给他们，但最终客户是他们。
Now it may be built by others and then rented to them, but it's at the end customer, it's them.
展望未来，计算相关的数字会急剧膨胀，对吧，今年资本支出已经超过一万亿美元了。
As we go forward into the future, the the numbers for computer ballooning, right, we're at, you know, you know, a little bit over a trillion dollars of capex this year.
到2028年，这个数字会超过两万亿美元。
As we go out into '28, it's gonna be more than $2,000,000,000,000.
实验室在这其中的占比也在不断增加。
The labs are also taking an increasing percentage of this.
所以最终，你会看到一个非常有趣的情况：这些实验室公司从每年花费几百亿美元，到每年花费几千亿美元，再到预计到本十年末每年花费数万亿美元。
And so ultimately, you've got a very interesting situation where the labs are going from companies that spend, you know, tens of billions of dollars a year to hundreds of billions of dollars a year to forecasting to spend trillions of dollars a year even at towards the end of the decade.
这至少是他们已经开始与合作伙伴签订的一些合同。
And this is at least some of the contracts they've begun signing with their partners.
因此，这需要对其经济模式进行重大调整。
And so this requires a big reshaping of what happens with their economics.
对吧？
Right?
你知道，到目前为止，这些公司大多在亏损。
You know, so up until now, there have been companies that mostly lost money.
Anthropic在第二季度开始盈利。
Anthropic started turning a profit in q two.
据信，在第三季度的某个时候，OpenAI可能会开始盈利，即使Codex和5.6等大幅增长。
It's believed at some point in q three, OpenAI could potentially start turning a profit even with the big rise of codex and and five point six and all this.
但如果回到一年前，他们所有的钱都是风险投资支持的亏损。
But if we go back a year ago, everything that they all the money they had was venture funded losses.
对吧？
Right?
就算回到今年年初，也还是靠风投亏损撑着。
If we go back to even the beginning of this year, it was venture funded losses.
现在他们已经扭亏为盈，开始真正赚钱了。
They've now turned the corner and are actually starting to profit.
不过，这并不意味着他们不再吸收新资本。
Now, that doesn't mean they're not taking in new capital.
新资本还在进来，是为了进一步加速增长。
The new capital is still coming in to accelerate the growth further.
但归根结底，他们越来越多的业务是靠自己的收入来支撑，而不是靠外部注资。
But ultimately, there's more and more of their business is being funded off of their own revenue rather than capital injections into them.
在过去一年半里，他们的利润率真的飙升了。
Over the last year and a half, their margins have really skyrocketed.
基础算力成本大概是每兆瓦1000万到1300万到1500万美元。
The base cost of compute tends to be around 10 or 13 or $15,000,000 per megawatt.
现在最有趣的一点是，以前他们提供服务的时候，对吧？
The most interesting aspect about what's happening now is before, again, they were generating, if they served a model, right?
比如在英伟达的Hopper GPU上跑GPT-4，对OpenAI来说毛利率是负的。
GPT-four being served on, know, NVIDIA Hopper GPUs was generating negative gross margin for OpenAI.
但现在，OpenAI提供GPT-5.6服务，或者Anthropic提供Meet Opus 5或Mythos、Fable 5，收入已经远远超过了每兆瓦10亿到15亿美元的增量成本。
But now when OpenAI serves GPT 5.6 or Anthropic serves Meet Opus five or Mythos, Fable five, revenue generation has passed well beyond the sort of incremental $1,015,000,000 dollars per megawatt.
以Anthropic为例，收入已经高达每兆瓦5000万美元。
In the case of Anthropic, the the revenue has gone as high as $50,000,000 per megawatt.
这让他们现在能做到的是，如果我花10美元在推理能力上，实际上能产生50美元的收入，然后我可以把这些利润全部再投入到训练中。
And and what that now enables them to do is, hey, if I spend $10 on inference capacity, actually generate $50 of revenue, and then I can turn around and incrementally spend all of that profit on training.
我很想了解的是，你怎么看待计算能力在实验室的集中化，或者说，流向全球的计算能力与流向实验室的计算能力的相对比例。
One thing I'm very interested in understanding is how you see the centralization of compute happening at the labs or the relative ratio of compute that goes to the world versus goes to the labs.
如果说现在，边际计算能力的三分之一流向了实验室。
Where if you say right now, a third of marginal compute is going to the labs.
那么到什么时候，全球增量计算能力的一半以上会流向实验室？
By when is it over half of the incremental compute in the world is going to the labs?
到什么时候，实验室基本上会拥有世界上绝大部分的计算能力？
And by what point do the labs have basically a vast majority of the world's compute?
是的。
Yeah.
所以今年早些时候，你知道，今年年初，Anthropic和OpenAI，OpenAI是2，Anthropic不到2。
So earlier this year, you know, the beginning of this year, Anthropic OpenAI started at two for OpenAI and less than two for Anthropic.
到今年年底，两家都超过5了。
End of this year, they're both above five.
所以他们的计算能力整体增长了三四倍。
So they've three, four x compute as a whole.
你看看新增的计算能力，今年大约有30%是这些实验室增加的。
When you you look at the incremental compute added, that's about 30% of the compute added this year.
展望明年，考虑到已经签好的合同和协议，情况会更加惊人，对吧？
And as we step forward to next year, given what's already been signed and penned and inked, you've got something even more dramatic, right?
Anthropic和OpenAI明年会拿走多达40%到50%的计算能力。
You've got Anthropic OpenAI are taking as much as 40% to 50% of compute next year.
而且这种集中化的趋势看起来并没有放缓或停止的迹象。
And the centralization doesn't look like it's slowing down or stopping.
实际上，看起来反而在加速。
In fact, it looks like it's only accelerating.
对。
Right.
不过，为他们建造这些算力的公司会发生变化。
Now, who's building that compute for them will change.
你知道，明年会有一些大的新玩家，比如SpaceX正在建造大量的算力。
You know, next year, big at new entrants, for example, SpaceX is building a ton of compute.
而且他们很可能会把其中相当一部分租给Anthropic和OpenAI，因为他们是最有能力支付最高价格的人。
And they're actively going to lease quite a bit of it to Anthropic and OpenAI most likely because they're the ones who can who have the marginal capability to pay the highest price.
此外，OpenAI和Anthropic也开始自己建造算力。
In addition, OpenAI and Anthropic are also starting to build their own compute.
OpenAI用自己的芯片，Anthropic则使用从谷歌购买的TPU，并通过FluidStack进行部署。
OpenAI with their own chips, Anthropic with TPUs that they're purchasing from Google and deploying with FluidStack.
所以你会问，嘿，什么时候全球新增算力的一半会只流向OpenAI和Anthropic？
And so, you ask, hey, when does half of the world's incremental new compute go to just OpenAI and Anthropic?
我的意思是，到明年年底，实际上新增算力的一半已经流向Anthropic了，因为……
I mean, it's really by the end of next year, it's already half of the incremental compute is going to Anthropic Because and
算力增长这么快，新增算力基本上会占据大部分算力。
compute is growing so fast, incremental compute is going to be basically most of compute.
所以你说很快，可能一年半到两年内，全球大部分算力会被两家实验室拥有，或者至少服务于这两家实验室的需求。
So it's very soon you're saying maybe within a year and a half or two years and most of the world's compute is owned by two labs or at least is serving the demand from two labs.
你
How
觉得
long do
会有多久？现在有个趋势，全球算力可能每年翻倍，但前沿实验室的算力每年翻三倍。
you think, so there's this trend where maybe world compute in gigawatts doubles every year, but the compute at the frontier labs triples every single year.
但如果保持当前趋势，从今年年初的2倍增长到年底接近6倍，就是乘以3。
But if you keep the current trend going, it goes from like two at the beginning of this year to close to like six at the end of this year, just multiplying out by three.
到2027年底是18，到2028年底是54。
18 by the end of 2027, 54 by the end of twenty twenty eight.
但你是不是觉得，到那个时候，如果光靠世界算力的总量，他们根本没法继续三倍增长？那你觉得未来几年世界算力的情况会怎么发展？
But are you like, okay, at that point, if they simply can't continue tripling given the amount of world compute or how, well, how do you see the world compute situation over the next few years?
对。
Yeah.
所以，如果今年新增算力是30吉瓦，明年50吉瓦，后年大概70吉瓦，就会出现一个很有意思的现象：今年新部署的每一瓦特，效率都远高于两年前部署的瓦特。
So, if the incremental compute adds this year, 30 gigawatts next year, 50 gigawatts and the year after that 70, roughly, you end up with this really interesting phenomenon, which is, okay, well, a new watt deployed this year significantly more efficient than the watts deployed two years ago.
所以，实际上，你知道，今年部署的算力占了全球算力的很大一部分。
So, actually, you know, a humongous percentage of the world's compute was deployed this year.
虽然部署的瓦特数没有翻倍，但我部署的是GB300、TPU V7和Tranium 3，这些芯片每瓦特的性能是上一代的三倍、五倍，效率高得多。
Even though it didn't double the number of Watts deployed, I'm deploying GB300s and TPU V7s and Tranium 3s, which are way, way, way more efficient, 3x, 5x more performance per watt than the prior generation chips.
所以，最终，你面前有一个巨大的阶梯。
And so, ultimately, you've got a huge ladder here.
所以，如果Anthropic和OpenAI明年拿下45%的算力，那么到2027年12月，它们就已经拿下了全球新增算力的一半。
So, if Anthropic and OpenAI take on 45 of compute next year, You've got them in, let's say, December 27, they have taken on half of the world's incremental new compute.
但全球新增算力的一半，实际上性能比之前所有的算力都要高。
But that half of the world's new incremental compute is actually at a higher performance than everything else before it.
对。
Right.
所以这又是一个倍增的因素。
So you've got another multiplier on that.
所以等到2028年底，如果这个趋势继续下去——我看不到任何阻止它的因素——他们基本上就掌控了全球大部分可用的算力。
So by the time you're in like towards the end of 2028, if this trend continues, which I see nothing that's stopping it, you you've got them just controlling most of the usable, you know, flops in the world on their own.
我困惑的是，如果算力的价格和价值都大幅提升，你为什么认为2028年我们只增加80吉瓦？
The thing I'm confused about is why you think we only add 80 gigawatts in 2028 if we enter in a world in which the price, the value of compute increases so much.
顺便说一句，那是上限。
That's the upper bound by the way.
那是我非常看好的情况。
That's the that's the like, I'm so fucking bullish.
对。
Right.
好的。
Okay.
那咱们来梳理一下思路。
So let let let's let let's do some chain of thought here.
几个月前我采访你的时候，你说要造一个吉瓦的Vera Rubin，需要——等一下。
So when I interviewed you a few months ago, you said in order to make a gigawatt of, I think, Vera Rubins, you need one sec.
需要5.5万片N3晶圆、6000片N5晶圆，还有17万片DRAM晶圆。
You need 55,000 N3 wafers, six k n five wafers, and a 170 k DRAM wafers.
我不知道这些数字重不重要。
I don't know those numbers matter.
我要逗你一下，但你说“晶圆”的方式真的太书面了。
I'm gonna troll you, but the way you said wafers was so fucking in papers.
顺便说一句，我们刚搬到美国那会儿，我德语口音特别重，而且我还是个素食主义者。
By the way, when we first when we first moved to The US, I had the the VW thing pretty bad, and I was a vegetarian.
天哪，丹尼。
Oh my god, Danny.
我记得你跟我提过这件事。
And I remember you told me about this.
在北达科他州，我上小学的时候，我会
In North Dakota, I was in elementary school, and I'd
说，能给我来点YG吗？
be like Can I get a YG?
能给我来点YG吗？
Can I get some YGs?
总之，那就是一个吉瓦的用量，对吧？
Anyways, so that's for one gigawatt, right?
对。
Yeah.
然后我让一个语言模型运行了你的晶圆厂设备模型，算了一下每年生产一个吉瓦算力所需的工具成本大概是多少。
Now I had an LLM run your wafer fab equipment model and figure out how much how much the tooling cost to produce a gigawatt of compute basically every single year.
我们算出来大概是30到40亿美元。现在假设你把洁净室、厂房和其他设施都算进去。
And we said like 3 to $4,000,000,000 Now suppose you add in clean rooms and, shell and everything else at the fab.
所以每年60亿美元的晶圆厂资本支出能产生一个吉瓦的算力，而一个吉瓦现在能带来1000亿美元的收入。
So $6,000,000,000 of like fab capex produces every single year, a gigawatt and a gigawatt produces right now a $100,000,000,000 of revenue.
而且这60亿美元的资本支出每年都能产生一个吉瓦，这个吉瓦每年又能带来1000亿美元的收入。
But also that 6,000,000,000 in capex is producing a gigawatt every single year and that gigawatt is producing a $100,000,000,000 every single year.
所以即使按五年来看，第一个吉瓦能产生五年的利润，工厂生产的第二个吉瓦能产生四年的利润，以此类推。
So even over the course of five years, so you know, the first gigawatt is generated five years of profits, the second gigawatt that the fab has produced is generated four years of profits and so on.
工厂层面60亿美元的资本支出最终能带来超过一万亿美元的AI收入。
6,000,000,000 of capex at the fab level will have generated over a trillion dollars of end AI revenue.
对。
Yeah.
但过程中有很多运营成本。
There's lot of opex along the way.
还有很多其他资本支出，比如数据中心、电力。
There's a lot of other capex like the data center, the power.
而且你还得付钱给OpenAI，还有很多不同的……
And you had to pay like the OpenAI for There's the a lot of different
这里需要钱的人。
people who need money here.
但确实，这规模很大。
But yeah, it's huge.
但把一半分给这些中间商。
But take away half of it for all these middlemen.
那也意味着晶圆厂资本支出和最终收入之间有100倍的差距。
That's only means there's 100x discrepancy between fab capex and end revenue generated.
实际上还不止这些，我们只是非常保守。
More than that actually really, we're just being very conservative.
结果，这就是资本主义，对吧？
And as a result, this is capitalism, right?
你会想到人们会意识到，我们有一个巨大的差距，可以把1美元变成100美元，他们不会想办法造更多的镜子。
Like you would imagine that people are gonna figure out like we're gonna you have this huge discrepancy where you can turn $1 into a $100 and they're not gonna figure out a way to make more mirrors.
我的意思是，他们确实会。
I mean, they are.
对。
Right.
只是这些镜子需要时间才能造出来，对吧？
It's just these mirrors take some time to bake, right?
但需求太大了。
But the emergency are so big.
就像Anthropic和OpenAI，我们现在就能赚一万亿美元，但就是被ASML机器里的镜子卡住了。
We're like, Anthropic and OpenAI, we could make a trillion dollars right now, but we're just bottleneck on the mirrors that go into the ASML machines.
他们会想，好吧，如果我们花一千亿美元，怎么能造出更多镜子呢？
They'd spend, okay, how can we make more mirrors if we spend $100,000,000,000 on this, right?
我们很快就会面临这种情况。
That's a situation we're going to be in pretty soon.
我就觉得，我们没法解决这个供应瓶颈，这真的很难想象。
And I'm just like, we're not going be able to solve that supply constraint and that just seems quite hard to imagine.
不，确实有。你见过有人在这里做有趣的套利，比如买涡轮机然后转卖。
No, there's definitely You've seen people do funny arbitragers here where they buy like turbines and then they try and resell them.
对。
Right.
因为涡轮机的价值要高得多，因为它是数据中心瓶颈所在。
Because the value of a turbine is way more because it's the thing bottlenecking a data center.
你知道，我觉得如果有人有4亿美元，并且能说服ASML卖给他们一台EUV光刻机，他们就应该直接买一台，然后等着，等着，再以超过10亿美元的价格卖出去。
You know, I think like if anyone had like $400,000,000 and the ability to convince ASML to sell them an EUV tool, they should totally just go buy one and wait, wait, and then sell it for north of $1,000,000,000.
对吧？
Right?
但最终，资本主义会促使这些东西扩张。
But but ultimately, yes, capitalism will cause these things to expand.
但这就像鞭子一样，对吧？
But it's it's a whip, right?
鞭子信号传到末端需要很长时间。
It takes a long time for the whip signal to get to the tail end of that.
所以供应链不会立即做出反应。
And so the supply chain doesn't react immediately.
实际上，你去跟卡尔蔡司的人聊，他们就会说，对对对。
In fact, you go to talk to someone at Carl Zeiss, they're like, yeah, yeah, yeah.
我们到本十年末需要造出100台EUV光刻机。
We need to make a 100 EUV tools by the end of the decade.
我觉得今年早些时候我们做那期节目的时候，他们甚至都没想过要造那么多。
I think when we first had our when we had our episode earlier this year, they didn't even think they needed to make that many.
每年造100台EUV光刻机所需的足够多的反射镜。
Enough mirrors to make a 100 EUV tools a year.
所以现在他们就像，好吧，得这么干了。
And and so now they're they've like sort of they're like, okay, need to do that.
但实际上，你知道，考虑到所有经济因素，应该还要更多。
But in reality, you know, because of all the economics of what's going on, it it should be even more.
但这需要很长时间。
But it takes so long pill.
假设每一家公司，抱歉，在这个产业链里的每一家公司都得到了私募投资。
Suppose that every single company in the firm, sorry, in the stack got private equity.
比如有个超级AGI狂热分子进来，说我们要最大化生产。
Like somebody came in who was super AGI pilled and was like, we're going to maximize production.
那能有多快？
How fast is it?
你觉得制造更多东西的物理限制会是什么？
Like what do you think the physical constraints on making more things would be?
因为我问这个的原因是，我们很快会进入一个世界，实验室的收入或者AI现金流——因为加速器也有巨大的现金流——会大到你可以直接用这些现金流来资助所有生产的极端扩张。
Because we're the reason I ask is we're pretty soon gonna be in a world where the lab revenue or just AI cash flows, because obviously the accelerators also have these huge cash flows, will be so big that you can just fund extreme expansion of all this production from cash flows themselves.
对。
Yeah.
我大体上同意。
I I I do agree generally.
当然有一些物理限制。
There's obviously some physical constraints.
目前供应链扩张的方式来看，100台大概还是正确的数字。
The way the supply chain is expanding currently, the 100 is roughly still the right number.
到2030年？
For 2030?
2030年100台ASML设备。
A 100 ASML tools for 2030.
但是，你知道，如果你跟卡尔·蔡司说，这里有100亿美元，请你赶紧扩大生产。
But, you know, if you said, Carl Zeiss, here's $10,000,000,000, please fucking just expand production.
那情况就不一样了。
That would change things.
而且你得跟供应链上的每一家公司都这么做。
And you would have to do this with every company in the supply chain.
我觉得今年不会发生。
I don't think it'll happen this year.
明年也不会。
Don't think it'll happen next year.
我觉得后年会，因为全球资本是有限的。
I think it'll happen the year after because the world is capital constrained.
但在一个世界里，比如说，顶尖实验室明年加起来能赚一万亿美元，他们也没法拿出10%。
But in a world where, say, the top labs are generating, let's say, even combined a trillion dollars in revenue next year, they're not able to say 10%.
我觉得他们不会那么做，但是。
I don't think they're going do that, but.
对。
Yeah.
或者至少几千亿，对吧？
Or hundreds of billions at least, right?
对。
Yeah.
看来他们意识到世界的发展方向了。
Seems like they realize where the world is headed.
我觉得他们完全可以做到。
I feel like they could just make.
所以，问题是实验室可以花几千亿。
So, thing is the labs can spend hundreds of billions.
他们明年会创造数千亿美元的收入。
They're going to generate hundreds of billions of revenue next year.
但最终，明年的资本支出大约是两万亿美元。所以，这里有个巨大的不匹配，对吧？
But ultimately, capex next year is like $2,000,000,000,000 So, you've got this big mismatch, right?
晶圆制造设备供应链的规模大约会是两千亿美元。数据中心市场供应链的规模会更大。
The wafer fabrication equipment supply chain will do something on the order of $200,000,000,000 The data center market supply chain will do even more.
加速器供应链的规模会更大。
The accelerator supply chain will do even more.
能源供应链也会达到一个可观的数字。
The energy supply chain will do a number.
把这些加起来，资本支出总额将远超两万亿美元。
You sum all this up, it's going to be well north of $2,000,000,000,000 of capex.
所以，实验室还没有达到用现金流来支撑这些支出的地步。
So, the labs have not yet gotten to the point where their cash flows can fund this stuff.
当然，对，对，对。
Of course, yeah, yeah, yeah.
我是说，他们显然永远也到不了那个点，对吧？
I mean, obviously they will like never get to that point, right?
因为他们想保持，是的，你想让你的资本支出高于你的回报。
Because they want to keep Yeah, you want keep make your capex higher than your returns.
但我真正想理解的关键问题是，如果，是的，如果当前状况持续到2028年底，每个实验室超过50吉瓦。
But the key question I really want to understand is if, yeah, if the current condition continues to be like north of 50 gigawatts per lab by the end of twenty twenty eight.
所以他们之间会有100吉瓦。
So between them, they'd have a 100 gigawatts.
那些吉瓦，正如你所说，到2028年会比现在驱动多倍的吞吐量或性能，对吧，因为硬件变得更好了。
Those gigawatts, as you're saying, drive many fold more throughput or more performance by 2028 than they are now, right, because the hardware has gotten better.
所以不仅每瓦特的FLOPs增加了，而且硬件在处理AI工作负载方面也变得更好了。
So not only have like FLOPs for watt increase, but also the hardware gets better at working with AI workloads.
好的。
Okay.
所以到2028年底，实验室有100吉瓦。
So 100 gigawatts for the labs end of twenty twenty eight.
多少？
How much
大概是全球算力的多少？
is like world compute?
我觉得这有点难说，因为到2028年，他们可能已经拿下了新增算力的80%。
I think that may be a little difficult given 2028, you start to have they've taken 80% of incremental compute.
而且我不确定到时候市场会怎么样，对吧？
And I'm not sure what happens to markets then, right?
你知道，算力价格得涨多少，他们才能真的买下80%的算力？
You know, how much does the price of compute skyrocket for them to actually be able to buy 80% of compute?
谷歌、Meta或者亚马逊，他们愿意卖那么多吗？
Is, you know, Google or Meta or Amazon willing to sell even that much?
还有一点要注意，当我们讨论这些吉瓦数字的时候，比如亚马逊在跑Anthropic的Bedrock模型，这算作Anthropic的算力，也是我们世界观的一部分，因为最终这算作Anthropic的收入，尽管有收入分成和返点之类的。
Also, one caveat when we're sort of talking about these gigawatt numbers is, you know, when Amazon is serving bedrock Anthropic models, that counts as Anthropic compute and sort of our world view because it is effectively, at the end of the day, counted as revenue for Anthropic, though, like, there's a revenue share and credit back and all that.
但最终，到2028年，如果他们合计达到100吉瓦，那他们真的对市场做了颠覆性的事情，因为现在每兆瓦算力10到1500万美元，任何人都能赚钱。
But ultimately, in 2028, it's it's, you know, if they get to 100 gigawatts combined, they have done really disruptive things to the market because anyone can make money off of 10 to $15,000,000 per megawatt compute today.
真的，我没开玩笑，其实没那么难。
You literally like, I kid you not, it's not that hard.
去弄一个GB 300机架。
Go get a GB 300 rack.
去下载Kimi的权重。
Go download the Kimi weights.
去下载VLM或者SGLANG。
Go download VLM or SGLANG.
把它设置好。
Set it up.
你知道，Codecs和Fable其实能帮你搞定这个。
You know, codecs and Fable can actually help you do this.
挺简单的。
It's pretty simple.
我的意思是，虽然不是特别容易，但也不是什么高深莫测的事，然后放到Open Router上就行。
I mean, it's not like it's, you know, it's not trivial, but it's not like rocket science and go put it on open router.
这很简单。
It's very simple.
而且你会开始产生比你支付的计算成本更多的收入。
And and you'll start generating more revenue than you're paying for the compute.
所以这已经导致了计算价格开始上涨，每兆瓦1000万到1500万美元。
And so this is this is sort of already led to this compute pricing 10 to $15,000,000 per megawatt start to inflect up.
要到2028年达到100吉瓦，你必须相信实验室能支付得起计算费用，因为任何人都能在10到15这个价位赚钱。
And to get to that 100 gigawatts in 2028, you have to believe that the labs can outpay for compute because anyone can make money at 10 to 15.
你知道，计算价格现在会涨到每兆瓦2500万美元吗？
You know, does compute now get to $25,000,000 a megawatt?
会涨到每兆瓦4000万美元吗？
Does it get to $40,000,000 a megawatt?
但正如你
But as you're
所说，实验室每兆瓦产生的收入已经远超其他所有人了。
saying, it's already the case that the labs are generating way more revenue per megawatt than everybody else.
如果他们能一直保持现在这么大的领先优势，那这种情况应该会持续下去。
If they stay as far ahead as they are currently, you would expect that to be the continuing case.
如果出现某种递归式的自我提升，比如AI实验室相对被提升，或者他们内部有未对外发布的模型，用来帮助改进下一代模型，那这种情况会更明显。
If there's like some kind of recursive self improvement where the AI labs are like relatively uplifted or they have models internally they're not releasing externally that are helping them make the next model better, you'd expect that to be even more the case.
其实现在已经能看到这种迹象了，比如SpaceX或者其他稍微落后一点的公司，如果内部不能像那些实验室一样把算力变现，他们就会把算力卖给出价最高的人。
I don't you're already seeing this where like SpaceX or whoever's like slightly further behind will just sell compute to the highest bidder if they can't internally monetize it as well as the labs.
所以我觉得，可以预期他们会继续吞下越来越多的算力份额。
So you I feel like it's continue expecting them to be able to gobble up like, bid for larger and larger shares of the compute.
我也这么认为，他们确实会继续抢占更多的算力。
I think that is my worldview that they will continue to gobble up more of the compute.
对。
Yeah.
但最终，他们不可能以现在的价格或者接近现在的价格做到这一点。
But ultimately, they can't do it at at current pricing or anywhere close to it.
确实。
For sure.
他们确实得开始支付每兆瓦2500万到5000万美元，才能在2028年真正吞下全球70%的计算资源。
They they do have to start paying $25.30, $50,000,000 a megawatt to really gobble up 70% of the world's compute in 2028.
对。
Yeah.
要在2028年达到100吉瓦，这确实是个相当激进的目标。
To get to that 100 gigawatts by 2028, which is a very sort of aggressive goal.
这件事另一个很棘手的方面是，我们已经看到AI实验室出现了明显的放缓，对吧？
The other aspect of this that's really challenging is we've already seen a huge slowdown for the AI labs, right?
他们倡导的这些监管措施，实际上对实验室的拖累远大于对开源中文语言模型的限制。
This this regulation that they advocate for is actually slowing down the labs a lot more than it slows down, you know, sort of the open source Chinese language models.
你知道，OpenAI没有发布Astra。
You know, OpenAI not releasing Astra.
OpenAI暂停训练两周，Anthropic没有公布他们的安全评估集Model 2，而大家普遍认为那会是Mythos的下一代版本。
OpenAI stopping training for two weeks, Anthropic not releasing what their safety assessment set is Model two, which is widely believed to be the next version of Mythos.
他们显然没有发布他们最好的模型。
They're clearly not releasing their best models.
在这种情况下，每兆瓦的营收就会停滞甚至开始下滑，因为其他模型又变得有竞争力了。
And in which case, the revenue per megawatt stalls or even can start to decline again because other models are competitive again.
所以并不是说他们落后了。
So it's not that they're falling behind.
只是他们没有发布最好的模型。如果有一些监管影响阻止他们发布最好的模型呢？
It's just that they're not releasing their best What if there is some regulatory impact that prevents them from releasing their best models?
那么他们的每兆瓦营收就不会增长得那么快，他们以高于其他人的价格购买额外算力的能力就会开始减弱。
Now, their revenue per megawatt does not climb as fast, then their ability to buy that incremental compute for a higher price than everyone else starts to diminish.
然后也许他们无法达到100吉瓦。在一个安全不重要的世界里，我相信这正是会发生的情况，对吧？
And then maybe they can't get to that 100 gigawatts is sort of in a world where safety doesn't matter, I do believe that's exactly what happens, right?
他们可以开始产生每兆瓦1亿美元甚至更多的收入，并且可以支付每兆瓦5000万美元。
They can start generating $100,000,000 per megawatt or more, and they can pay $50,000,000 a megawatt.
其他任何人都没有合理的理由去处理他们的算力，除了说，拜托，Dario，把我所有的东西都拿走吧。
And no one else has any logical reason to do anything with their compute besides say, please, Dario, take everything off of my hands.
但是，你知道，有一些我们无法描述的力量在起作用，可能会减缓这个过程。
But there are, you know, forces at play that which we cannot describe that that would potentially slow this down.
对对对。
Yeah, yeah, yeah.
我觉得一个很好的思考实验就是，如果AI模型真的能像完全自动化的软件工程师一样厉害呢？
I mean, I think a good intuition pump is just what if the AI models were literally as good as a fully automated software engineer?
它们现在还没到那个水平，对吧？
They're not currently there yet, right?
我觉得它们离完全自动化一个白领工人的工作还差得远。
Like I think they're far from just being able to fully automate the job of like a full white collar worker.
但是白领工人一年能挣六位数或者更多。
But white collar workers earn 6 figures or north of that a year.
如果你有一个吉瓦的算力，能支撑大约一百万个白领工人的工作量，那粗略算下来，那就是一千亿美元。
And if you have a gigawatt that can sustain a population of like, say a million white collar workers, let's say roughly, right, that's like, you could then off the back of that, that would be a 100,000,000,000.
这其实低得有点出乎意料。
That's actually surprisingly low.
是的。
Yeah.
每人10万美元，一百万人。
A 100 k per person, million population.
对。
Yeah.
对。
Yeah.
对。
Yeah.
我不知道。
I don't know.
但如果实现完全的通用人工智能，那将是数千亿美元。
But it'd be many hundreds of billions of dollars if you get like full AGI
我觉得另一方面是，我们一直在看到，大部分价值并没有被捕获，对吧？
per I think the other aspect of this is and we've continued to see this, most of the value capture is not happening, right?
这些模型产生的大部分价值并没有给到OpenAI和Anthropic。
Like most of the value that these models generate does not get given to OpenAI Anthropic.
好在到目前为止，这些价值大部分都给了用户。
Thankfully, so far, it is mostly just being given to the users.
对，对。
Yeah, yeah.
对吧？
Right?
比如Jane Street，他们和OpenAI签了GPT 5.6超快模式的独家合同，或者Jane Street作为Anthropic最大的客户之一，他们从这些token里创造的价值远远超过Anthropic从中获得的利润，对吧？
Jane Street, with their exclusive contract with OpenAI for GPT 5.6 ultrafast mode or Jane Street, where they're like one of Anthropic's biggest customers, is generating way, way, way, way more value out of the tokens they're paying for than Anthropic is generating in terms of profit, right?
因为他们可以利用这些在市场里赚钱。
Because they get to, you know, make money off of the market.
还有Meta，曾经据说占了Anthropic业务的10%。
Or Meta, who at one point was, you know, rumored to be, you know, as much as 10% of Anthropic's business.
他们通过优化广告算法之类的，把用户参与时间延长了5%，效率大大提升。
You know, they're generating way more efficiencies by optimizing their ad algorithms or what have you and getting engagement time 5% longer.
所有这些公司，他们靠使用这些模型赚的钱都比Anthropic多得多。
And, you know, all these things, they're making way more money off of using these models than Anthropic.
所以最终，你知道，这是必须的。
And so the ultimately, you know, and that's what's required.
当然，如果你有一百万个新软件工程师，每个软件工程师的成本也会下降。
So sure, if you had a million new software engineers, the cost per software engineer would also fall.
我困惑的一点是，市场会达到均衡吗？
One thing I'm confused about is does the market come into equilibrium?
如果达到均衡，你会不会预期计算的价格等于Anthropic和OpenAI能从中赚到的钱，或者非常接近，但Anthropic和OpenAI只加一点价。
And if it comes into equilibrium, would you just expect the price of compute to equal whatever Anthropic and OpenAI can generate from it or be very close to it, but like a small amount of markup for Anthropic and OpenAI.
就像现在，计算的价格和Anthropic能从中赚到的钱之间有4倍甚至更大的差距，这真的很奇怪。
Like right now it's really weird that there is a 4X or more difference between what compute sells for and how much money Anthropic can make from it.
在一个每吉瓦收入持续增长的世界里，如果Anthropic从一吉瓦中赚钱的能力翻倍或翻三倍，那如果差距继续扩大，就奇怪了。
And in a world where the revenue per gigawatt continues to increase, if Anthropic's ability to monetize a gigawatt doubles or triples or something, it'd be weird if then the gap continued to increase.
所以Anthropic仅仅通过拥有一些软件、一些权重，就能把花10美元的东西变成100美元。
And so Anthropic just by like having some software, having some weights can take something that cost them $10 and then turn it into $100
对。
Yeah.
所以这个问题总是很有趣，对吧？
So there's a bit of this is always a fun question, right?
就是AI的价值到底流向了哪里。
Which is where does the value go in AI?
AI正在创造所有这些价值。
AI is generating all this value.
你看，有最终用户，我想我们都同意他们创造的价值比任何人都多。
You've got, you know, the end user, which I think we all agree is generating more value than anyone else.
所以他们愿意为这些模型付很多钱。
Hence, they're paying a lot for these models.
但接下来还有应用层。
But then you have, you know, the app layer.
嗯，到目前为止，应用层创造的价值非常少。
Well, so far the app layer has generated very little value.
还有模型层，直到一年前还在产生负毛利，现在却产生了巨大的正毛利。
And you've got the model layer, which again, up until a year ago was generating negative gross margins and is now generating massive positive gross margins.
而且看起来正在朝着每兆瓦创造一亿美元价值的方向发展。
And looks like it's on the path to generating, you know, $100,000,000 per megawatt.
所以就像你说的，把10.15美元变成100美元。
So turning, know, $10.15 dollars into a $100 as you said.
但如果我们再回到一年前，硬件供应链正在创造所有这些毛利润，而其他所有人都在亏钱。
But if we go back again a year ago, the hardware supply chain was generating all this gross margin while literally everyone else was losing money on it.
OpenAI和Anthropic只是在大量投入风投资金。
OpenAI and Anthropic were just plowing VC money in.
许多其他初创公司也是如此，而且许多超大规模企业都在建设基础设施，却不知道是否会得到回报。
And as were many other startups and many of these hyperscalers are building infrastructure without knowing if there was going to be a payoff.
所以，最终，模型层几乎是在创造负价值，因为他们以低于基础设施成本的价格出售令牌。
So, ultimately, you had this like, you know, negative value being created on the model layer almost, if you will, because they were selling the tokens for less than it cost them on the infra side.
而所有创造的价值都流向了芯片和晶圆厂。
And all the values being created used at the chip, the fab.
最初在2023年，内存厂商并没有从HBM或AI内存中赚到钱，尽管理论上他们提供的价值是巨大的。
Initially, in 2023, the memory guys were making no money off of, you know, HBM or memory for AI, even though theoretically their value they were delivering was humongous.
现在呢，其实台积电赚的价值比存储厂商少多了。
Now, you've got well, actually, TSMC makes way less value than the memory guys.
他们确实捕获的价值更少，所以价值捕获发生了很大的转移，是的。
Is that actually how much know, they're capturing less value, So, you the value capture shifted around a lot Yeah.
这对跟踪市场或参与市场的人来说很有趣，比如Jane Street。
Which is very fun for people tracking the market or participating in the market like Jane Street as an example.
有很多
There's lots
你得把它们插进去，有很多很多，你得使劲插。
of you gotta plug them There's lots lots you plug them that hard.
所以，
So,
你知道，接下来会发生什么？
you know, what happens, you know, going forward?
Anthropic和OpenAI，它们慢慢开始大幅增加价值捕获。
Does Anthropic and OpenAI, they've slowly started to balloon in value capture.
他们会膨胀并拿走所有的价值吗？
Do they balloon and take all the value capture?
嗯，当时是这么想的。
Well, that was the thought.
然后埃隆证明了，其实不是。
And then Elon showed, actually, no.
我可以把我的算力以每兆瓦2500万或4000万美元的价格卖给Anthropic和谷歌。
I can sell my compute for $25,000,000 a megawatt or $40,000,000 a megawatt to Anthropic and Google.
即使是短期的事情，我也已经以这个价格卖出去了，一年内就能收回全部资本支出。
Even if it's a short term thing, I've sold it for this price and I'll recoup my entire capex in a year.
那么，你
So, what's your
预测一下，相关的算力部分，比如P300或者类似的，以每吉瓦400亿美元的价格卖出去，会怎么样？
prediction of how much the relevant tranche of compute like p three hundred's or whatever that sold for 40 b a gigawatt.
SpaceX以每吉瓦400亿美元的价格卖给了谷歌。
The SpaceX sold for 40 b a gigawatt to Google.
到明年年底，这个能卖多少钱？
What does that sell for at the end of next year?
我觉得大部分算力还是会以低于每吉瓦200亿美元的价格成交。
I think most compute will still continue to transact at sub $20,000,000,000 a gigawatt.
即使到明年年底也
Even at the end of next
是这样吗？
year?
因为所有的算力都需要融资。
Because all of it has to be financed.
对于不需要融资就能建设的算力，对吧？
For compute that you can build without financing, right?
如果Meta、微软、亚马逊、SpaceX能自己建算力，不需要先找客户，就直接说，管他呢，我就要建这个算力。
If Meta can build compute, Microsoft, Amazon, SpaceX can build compute without finding a customer just saying, fuck it, I'm going to build this compute.
然后等建好了再说。
And then turn around and wait till it's already built.
现在他们掌控了局面。
They now control what's going on.
所以，大部分算力在建成之前就已经被预定了。
So, most compute is contracted well before it's built.
对。
Yeah.
这就是埃隆在市场上利用的一点——他手里确实有这么多算力，然后他说，嘿，Anthropic，我知道你们每吉瓦能赚600多亿美元。
And so, this is sort of what Elon took advantage of in the market is, he actually had all this compute and he was like, Hey, Anthropic, I know you're making like 60 plus billion dollars per gigawatt.
不如你们花大价钱把我的东西买了吧？
Why don't you just buy my stuff for a crazy amount of money?
当然，这也不是埃隆或Anthropic单方面决定的，市场自己就找到了平衡。
And obviously, it's not like Elon decided this or Anthropic decided this, sort of markets figured itself out.
其他人呢，你去随便一个云服务商那儿，他们会说，好吧，我要建一吉瓦或者100兆瓦的算力。
Other people, you go to a random cloud, they're like, okay, I'm going to build a gigawatt of compute or a 100 megawatts of compute.
我要投入这笔资本开支。
I'm going to spend the capex.
我得赶紧回头找个客户。
I need to turn around and find a customer.
如果我想找客户，就得先找到资金。
If I want to find a customer, I need to find the capital.
谁会给我资金和客户呢？客户得先签合同，然后我拿着客户的承诺去信贷市场融资。
Who's going to give me the capital and the customer, the customer has to sign a deal and then I take the customer's commitment to the credit markets and I raise the capital.
所以这里有一种完全不同的权力结构，Meta实际上在囤积算力，他们和SpaceX可能是第三名。
And so there's this sort of like completely different power structure where Meta who is effectively hoarding compute, them and SpaceX are plausibly the like number three.
他们能成为第三名，唯一的原因就是他们在囤积这些算力。
The And only plausible number three is because they're hoarding all this compute.
他们用自己的资产负债表和能力来建设算力，在没有终端客户的情况下，以巨大的规模进行变现。
They're using their balance sheets and capabilities to build compute, to build compute without end customer that's monetizing at a huge degree.
而且他们有实际的资产负债表，所以可以去信贷市场说，嘿，各位，我让你们建一个兆瓦，你们可以赚取利润，虽然不是暴利，但利润不错，而我现在拥有了所有这些算力。
And they have an actual balance sheet so they can go to the credit market and being like, Hey guys, I have you build a megawatt, you can make your margin, not a crazy margin, but you can make a good margin and I now have all this compute.
现在Meta和SpaceX有了这种选择权，可以四处看看，心里想：我的内部用例能让我赚更多钱，还是我应该把这些算力以疯狂的高价卖给Anthropic或OpenAI？
Now Meta and SpaceX have this optionality of looking around and being like, Is my internal use case going to make me more money or should I go out there and sell it to Anthropic OpenAI at crazy margins?
对。
Yeah.
所以，现在我们进入了一个新阶段，SpaceX和Meta都在说，其实我要自己建算力，而且我可以开始出租，价格不是13。
So, now we've sort of entered a regime where SpaceX and Meta are saying, actually, I'm going to build the compute and I can start to rent it out for not 13.
我可以卖到25.5美元，甚至更高。
I can sell it for $25.50, and more.
所以，在我做视频论文和其他内容的时候，我一直在想为播客招一个新编辑。
So as I've been doing video essays and other formats, I've been looking to hire a new editor for the podcast.
但主动找编辑真的很费时间，因为绝大多数候选人都不是我想要的那种。
But actively searching for editors has been quite time consuming because the vast majority of candidates don't fit the profile that I'm looking for.
所以我在Grok Bot里建了一个招聘机器人，看看有没有用。
So I created a recruiter in Grok Bot to see if it would help.
我给了它一大堆背景信息，基本上把我想要的全说了一遍，然后它又生成了四个机器人，分别负责搜索的不同方面。
I gave it a huge context dump where I monologued basically everything that I wanted, and then it spun up four other bots to narrow in on different parts of the search.
其中一个去翻了我过去一年的邮件，找相关的主动投递。
One went through the last year of my email for relevant inbound.
展开剩余字幕（还有 480 条）
一个搜索了我的X动态和私信。
One searched my x feed and DMs.
一个翻看了我喜欢的一些纪录片的片尾字幕。
One went through the end credits on various documentaries I like.
最后一个找了我关注的一些YouTuber的编辑。
And the last one looked for editors who work for some of the YouTubers that I follow.
然后Grok Bot把那些子代理找到的所有候选人都汇总起来。
Grok Bot then took all these different candidates that these sub agents had found.
它根据我的标准筛选了一遍，最后给我列了一个简短的候选名单让我过目。
It filtered them against my criteria and then delivered for me a final short list to review.
说实话，第一批里确实有几个我想看的候选人，但大部分都不怎么样。
To be honest, the first batch had a few good candidates I wanted to see, but mainly a bunch of duds.
不过在我给Grok Bot反馈了它遗漏的地方之后，它又给我列了一份新名单，这次我真的非常期待。
But after I gave Grok Bot some more feedback about what it was missing, it came back with a new list of candidates that I'm actually extremely excited about.
最后我把整个工作流程保存成了一个例行程序。
I ended up saving this whole workflow as a routine.
所以现在每周，Grok Bot都会检查我的收件箱和X私信，寻找可能值得采访的新候选人。
So every week now, Grok Bot checks my inbound email and x DMs for promising new candidates to potentially interview.
如果你想自己试试Grok Bot，可以去x.ai/bot。
If you wanna try Grok Bot yourself, go to x.ai/bot.
你觉得到2027年底，他们每吉瓦的营收会是多少？
What do think their revenue per gigawatt is by the end of twenty twenty seven?
你是说像Anthropic或OpenAI到2027年底的情况？
Like, for Anthropic or OpenAI by end of twenty twenty seven?
我觉得这很大程度上取决于谁拥有最好的模型，如果他们被允许继续发布最好的模型的话，但我不觉得为什么不会达到每兆瓦5000万美元以上。
I think I think it's highly dependent on who has the best model if they're allowed to keep releasing their best models, but I don't see why it wouldn't be 50 plus million dollars a megawatt.
到27年底。
By the end of twenty seven.
哦，到27年底？
Oh, by the end of twenty seven?
对。
Yeah.
这就更难说了。
That's where it gets more challenging.
但我觉得，可能会比那个数字更高。
But I think I think it could get to, you know, higher than that.
大概每兆瓦70.8亿美元，全公司平均下来，甚至可能更高。
It's like $7,080,000,000 dollars a megawatt blended across the company, if not higher.
对。
Yeah.
对。
Yeah.
所以我想，如果真是这样的话，那算力的价格会怎么变呢？
And so I think if if that's the case, right, then what happens to the price of compute?
如果我是Anthropic，增加算力是值得的。
Well, if I'm Anthropic, incremental compute is worth it.
也许我会在SpaceX的算力上每兆瓦花4000万美元。
Maybe I spend $40,000,000 megawatt on SpaceX compute.
如果我是SpaceX，你知道，我会看供应链。
And if I'm SpaceX, you know, I look to the supply chain.
我会想，嗯，我和Jensen达成了协议，他现在突然开始用Twitter了。
I'm like, well, you know, I've struck this deal with Jensen where he's now all of a sudden using Twitter.
而且，你知道，Elon说他们只跟NVIDIA合作，但Jensen为什么不涨价呢？
And and, you know, there's Elon saying they're exclusive to NVIDIA, but why doesn't Jensen raise his prices?
然后，SK海力士、美光和三星看着NVIDIA，也会想，他们为什么不涨价呢？
And then, you know, SK Hynix and Micron and Samsung looked at NVIDIA and were like, well, why don't they raise their price?
所以，我觉得价值捕获这里有个牛鞭效应，对吧？
So, think I think the value capture, there's a bullwhip effect here, right?
就是，仅仅因为有人涨价，并不意味着整个供应链会立刻重新平衡。
Where just because someone has risen the prices doesn't mean the entire supply chain rebalances immediately.
对。
Yeah.
但随着时间的推移，供应链会重新平衡，成本会越来越高。
But over time, the supply chain will rebalance and things will cost more and more.
而要获得那额外的产能，你基本上不得不这么做，对吧？
And to get that incremental capacity, you sort of have to, right?
所以，台积电涨价很慢，但内存公司涨价很快。
So, TSMC raising prices very slowly but memory companies raising prices very quickly.
你知道，基板公司涨价也很快，供应链的不同环节都在涨价。你看，如果只是15%，埃隆可能不会卖，但他卖是因为已经超过25%了。
You know, substrate companies raising prices very quickly, different parts of supply chains raise you know, Elon wouldn't have sold if it was 15, but he's selling because it's 25 plus.
对。
Yeah.
对。
Yeah.
所以，显然，他很快就提价了。
So, obviously, he rose his prices really quickly.
对。
Yeah.
我有点惊讶，你觉得每吉瓦的营收增长不会远远超过100%甚至更多。
I I'm sort of surprised you think like revenue per gigawatt doesn't increase way more than even like a 100 per
到明年年底，每吉瓦的营收什么时候会实现？
When gigawatt by the end of next does RSI happen?
什么时候会起飞？
When does take off?
对吧？
Right?
我觉得
I think
我觉得即使没有实现，目前的进展速度也会继续下去。
I think Even if RSI doesn't happen, the current rate of progress continues.
你看看过去一年半的进展，比如一年前的模型是什么样的？
If you just look at how much progress have you made in, let's say the last year and a half, Like, what was the model from a year and a ago?
就拿产品来说吧
Let's take, like, prod
我对此的问题是，世界上最好的模型是在二月份训练的。
my problem with this is the best model that exists in the world was trained in February.
好的。
Okay.
你是说我们可能没法发布那个，比如，
You're saying maybe we we just won't be able to release the, like,
最好的模型。
best model.
他们说他们两周内不训练模型了，天哪。
Says they're not training models for two weeks, man.
搞什么
What the
鬼？
hell?
对。
Yeah.
对。
Yeah.
对。
Yeah.
我的意思是，还有一件事，从内部来看，他们用得够多吗？
I I mean, there's another there's one thing, internally, are they getting enough use for it?
他们只会把计算机的价格炒高。
It's just they'll, like, bid up the price of computer.
另一个就是，整个AI进展放缓会让监管变得……
Or another is, like, this AI progress as a whole slowdown makes the regulation.
对。
Yeah.
但他们甚至都不允许用这个新模型，比如Astra在公司内部都没有广泛部署。
But they're not even allowed to use this, like, new model in like, Astra's not widely deployed internally
甚至。
even.
对。
Yeah.
对。
Yeah.
但还是，我也不确定。
But still, I don't know.
就是那个，嗯。
Just like the the yeah.
如果你有一个模型，比如去年年初发布的那个模型？
If you go if you have, like, a model that is what was the model released, like, let's say, the beginning of last year?
像GPT-4o？
Like GPT Four o?
那是4o吗？
Is that four o?
对。
Yeah.
那就像是从4o到Fable大小或者Mythos 2大小的飞跃，到2027年底的时候。
That that's like you're talking about a four o to Fable size or mythos two size leap by this point, again, by the end of twenty twenty seven.
对，Mythos 2 还没发布。
Yeah, Mythos two is not out.
是啊，甚至 Mythos 也是，对吧？
Yeah, or like even Mythos, right?
就是那种跳跃式的进步。
Like that leap again.
连 Mythos 都不允许发布，对吧？
Even Mythos is not allowed to be out, right?
他们把它阉割了。
They've neutered it.
对对对。
Yeah, yeah, yeah.
比如我们不能用它来优化推理性能。
Like we can't use it to optimize inference performance.
我们不能用它来优化各种东西。
We can't use it to optimize all sorts of things.
对。
Yeah.
也许AI的进展或者AI的部署速度会有所放缓。
Maybe there's like some slowdown in AI progress or the deployment of AI.
这意味着每吉瓦的营收可能会更低。
That means that the revenue per gigawatt can be lower.
但我觉得，只有在这种情况下，到明年年底每兆瓦的价格才会只有100。
But I I that that's the only way I could see it being only a 100 per megawatt by the end of next year.
对。
Yeah.
我的意思是，只要模型变得更好，产生的价值显然会增加，但谁能捕获这个价值还有待讨论。
I mean, as as long as the model gets better, the value generated out of Obviously, it gets who captures the value is still up for debate.
但最终，每个人都会提高价格。
But ultimately, everyone's going to raise their prices.
对。
Yeah.
对。
Yeah.
对。
Yeah.
因为他们能这么做。
Because they can.
而且这非常具有通胀性，尤其是目前的监管方式，到目前为止，只是不发布模型。
And it's super inflationary, especially if the method of regulation is right now, so far, it's just don't release the models.
但越来越多的监管方式是纽约禁止数据中心。
But more and more, the method of regulation is New York's banning data centers.
德克萨斯州在暂停审批。
Texas is holding moratoriums.
俄亥俄州要求，或者至少试图要求，你得支付一定范围内所有人的房产税。
Ohio saying you have to or at least trying to say you have to, like, pay everyone's property tax in a certain radius.
这类事情会减少供应、增加成本，而这些成本也会转嫁出去。
These sorts of things are going to decrease supply and increase cost, and that's going to get passed on as well.
所以最终你会陷入一种局面，进步确实会放缓，至少从外部来看是这样。
So you start to end up in a spot where progress does slow, at least in the external sense.
对。
Yeah.
即使模型内部在不断提升，越来越好。
Even if the models internally keep getting better and better.
我觉得没有理由不这样，比如在爆发式增长的情况下，为什么Anthropic不会因为安全和监管的原因，同时也为了竞争优势，而把最好的模型比外部能看到的提前六个月呢？
I see no reason why, like, you know, again, like in a takeoff scenario, why would Anthropic not have their best model six months ahead of what is externally available because of safety and regulation, but also, you know, the competitive advantage.
而且如果进步加速，那六个月的差距实际上会变成更大的差异。
And then that six month difference, if progress accelerates, is actually a bigger differential.
所以，这就是为什么每兆瓦的收益增长会远低于今年上半年的水平。
So, that's the thing that would cap revenue per megawatt gains to a much lower growth than we've seen in the first half this year.
有件事我非常感兴趣。
Here's something I'm very interested in.
当这些公司上市，对投资者负责，比如说明年底，大概有接近20吉瓦的规模。
As these companies go public and they're accountable to investors let's say end of next year, have, I don't know, close to 20 gigawatts.
所以大概10%的算力，就是两个吉瓦。
So like 10% of the compute, two gigawatts.
假设他们想把训练算力的占比从60%提高到70%。
Let's say they want to go from 60% compute to, training to 70% compute to training.
然后他们的投资者就会说，如果你每吉瓦能赚1000亿美元，那你为了增加训练算力，基本上就是在放弃2000亿美元的收入。
And their investors are like, well, if you're able to generate a $100,000,000,000 per gigawatt, you're basically saying no to like $200,000,000,000 of revenue in order to increase your training compute.
投资者就会觉得，这搞什么鬼？
As investors are like, what the fuck?
你们已经在训练上花那么多钱了。
You're already spending so much on training.
为什么还要在训练上花更多的钱？
Why are you spending even more on training?
作为一家上市公司，你觉得，如果他们就是说，不，我们会继续增加训练算力的占比，来抵消每吉瓦算力带来的收入增长，那会发生什么？
As a public company, do you think, yeah, what do think would happen if they're just like, no, we will keep increasing the share of compute we spent on training to offset the increase in revenue that each gigawatt of compute is giving us?
对。
Yeah.
所以，我个人认为实验室会逐渐减少推理计算的比例，这我觉得是挺非主流的观点。
So, this is sort of what I personally believe that the labs are going to allocate less and less compute to inference over time, which I think is very non consensus.
对吧？
Right?
大多数人的普遍看法是，大部分计算都会用在推理上。
Everyone's sort of the standard belief of most people is, oh, most compute will go to inference.
大部分会用于训练的前向传播，而不一定是产生收入的推理。
Most of it will go to forward passes for training, not maybe necessarily revenue generating inference.
但最终，如果他们现在每兆瓦能产生30.4亿美元的收入，你会把40%分配给推理。
But ultimately, end up with if they're generating, you know, $3,040,000,000 dollars per megawatt today, you allocate 40% to inference.
如果现在每兆瓦能产生60.7亿美元，你还会把40%分配给推理，赚取所有利润，然后分红和回购股票吗？
If you now get to generating $6,070,000,000 dollars per megawatt, do you still allocate 40% to inference and generate all this profit and then do dividends and share buybacks?
还是去构建通用人工智能？
Or do you go build AGI?
对。
Yeah.
对。
Yeah.
而且我觉得，Anthropic 和 OpenAI 的答案很明显，不只是高层，连董事会也是，那就是去搞 AGI，因为那更赚钱。
And I think the obvious answer from Anthropic and OpenAI, and not just at the executive level, but also their board, is go build AGI because it's way more profitable.
没错。
Right.
所以你会看到他们把越来越多的算力投入到训练上。
And so, you're going to see them ratchet up their percentage of compute dedicated to training.
而每一份算力如果用在推理上，都能带来越来越多的利润。
While each increment of compute is getting more and more profit generating, if they had dedicated to inference.
对。
Right.
所以关键问题是，如果我在卖 token，那 OpenAI 推出的极速模式是只给外部用，还是内部也在用？
And so the whole point is, well, okay, if I'm generate if I'm selling tokens, is Anthrop is OpenAI releasing ultra fast mode for just external or are they doing it internally too?
结果发现，不是的，实际上我会同时分配给内部和外部，因为我内部用超快 AI 或最好模型产生的价值，远比外部用户带来的多。
And it turns out, no, actually, I'm going to allocate it to internal and external because my internal, you know, value that I'm generating from super fast AI or the best AI model is way more than what someone externally is.
对。
Yeah.
所以最终，当然，我每兆瓦能赚一亿美元。
And so ultimately, sure, I could generate a $100,000,000 per megawatt.
但如果我把这些算力投入到AI研究中，我能获得多少增量进展呢？
But if I turn that towards AI research, what is the incremental progress that I get?
然后这对我的未来收入潜力、我做的任何事情的折现现金流又有什么影响，对吧？
And then what does that do towards my future earnings potential, the discounted cash flows of whatever the hell I've done, right?
对。
Yeah.
所以，你知道，他们不会去算这笔账。
And so, you know, they're not going through that calculation.
但最终，把越来越多的算力内部化使用更有意义。
But ultimately, it makes more sense to dedicate more and more compute internally.
而推理算力之所以要这么大，唯一的原因就是为了扩大你的训练集群。
And the only reason to, you know, have inference compute be so large is so you can grow your training fleet.
对，对，对。
Right, right, right.
我觉得这是个很有意思的经济学问题，我觉得我们可以让模型来消化一下，在一个他们减少推理计算占比的世界里，必须得有什么样的条件才能成立？
I think this is an interesting economics question that I feel like we can have the models digest of what is the what would have to be true about a world where they reduce fraction of compute spend on inference?
我觉得过去三个月他们已经在这么做了。
I think they have been over the last three months already.
有意思。
Interesting.
我觉得今年有些月份他们是在增加计算占比的。
I think parts of this year they were increasing fraction of compute.
那我们就一个月一个月来看吧。
So let's just take month by month.
你会同意，Anthropic每个月增加的计算量都比上个月多。
You would agree that every month Anthropic has added more compute than the prior month.
可能他们签SpaceX那种大单的时候会有一些波动，但总的来说，计算量是呈上升曲线的。
There might be some noise when they like sign a SpaceX deal or whatever, but in general, the amount of compute is a curve up.
所以一月份增加的算力比十二月份少。
And so in January, added less compute than December.
然而，收入增长却突飞猛进，然后趋于平稳。
And yet, the revenue adds skyrocketed and then they've plat sort of plateaued.
他们现在并不是每个月都增加250亿美元的年度经常性收入。
They're only adding they're not adding $25,000,000,000 of ARR every month now.
所以这意味着他们新增的每兆瓦电力，用于研发的比例比用于推理的比例更高。
And so that means the marginal megawatt they're getting is going higher percentage to R and D than it is inference.
有意思。
Interesting.
对。
Yeah.
所以事实上，他们现在确实在把更多的算力投入到研发上。
And so they are factually increasing their compute towards R and D today.
对。
Yeah.
对。
Yeah.
我觉得这很明显，如果你仔细看他们在做什么的话。
I think this is like self evident if you if you like look at what they're doing enough.
对。
Yeah.
所以如果我看你刚才说的那些数字，关于全球算力增长的速度，我有几个问题想搞清楚。
So if I look at the numbers you said of like how fast world compute grows, here's some things I want to understand.
所以看起来，如果我把你刚才说的数字加起来，到2028年底全球算力会超过200吉瓦，对吧？
So it seems like if I added the numbers you just said, it would be over 200 gigawatts of world compute by the end of twenty twenty eight, right?
对，全球范围内。
Yeah, globally.
好的。
Okay.
那这个增长能持续多快？
And how fast can that continue growing?
2028年之后的全球AI算力？
Like global AI compute after 2028?
对。
Yeah.
所以今年是30，明年50，28年是70。
So 30 this year, 50 next year, 70 in '28.
29年大概在90到100左右。
'29 should be, like, on the order of 90 to a 100.
那之后就是每年增加100多，差不多这样。
Like, then just a 100 more every single year or something.
我觉得这个增长斜率可能会继续往上走。
I think I think the the slope could continue to go upwards.
我的意思是，超过四年的预测很难说，谁知道呢，我们是不是已经进入RSI阶段，或者世界经济什么时候能每年增长10%？
I mean, it's hard to predict anything more than four years out, given who knows what's, you know, are we an RSI regime or, you know, when is the world economy growing at 10% a year?
因为如果每年增加100多吉瓦，那GDP增长和收入增长就非常惊人了。
Because if you're at 100 plus gigawatts a year, you're you're at absurd revenue, GDP growth.
对。
Right.
如果你认为2028年全球有200吉瓦的算力，那到时候中国会占多少？
If you think there's 200 gigawatts globally in 2028, how much is in China by that point?
还有，中国的算力在整个趋势中是怎么持续增长的？
And how does like, yeah, how does Chinese compute continue increasing through this whole trend?
因为如果西方先出现RSI的情况，而中国还没有大量算力，那我们可能就生活在另一个世界里了，跟中国有大量算力的情况完全不一样。
Because if, if the RSI stuff kicks off in the West before China has a large amount of compute, maybe we're living in a different world than when it doesn't.
是的。
Yes.
所以中国现在……如果我们回到2022年来看，美国当时贡献了全球算力的45%到50%。
So China today So if we sort of level set back to 2022, The US was adding about 45% to 50% of the world's compute.
中国贡献了大约30%到35%，剩下的由世界其他地区承担。
China was adding about 30% to 35% of the world's compute and the rest being taken up by the rest of the world.
自2022年以来，我们对中国实施了严格的管制，而美国的算力则大幅增加。
Since 2022, we've had big regulations against China and a dramatic increase in America.
所以现在，70%的电力都部署在美国。
So today, 70% of Watts are being deployed in America.
而中国其实占比非常小。
And China is really a very small number.
中国数据中心AI计算所用的电力不到10%。
It's sub 10% of Watts being deployed for data center AI compute is in China.
而且往前看，他们的规模依然很小。
And as we step forward, they're still at a very small number.
他们的国内产量相当有限。
Their domestic production is quite small.
他们从英伟达的采购量也依然很少。
Their purchasing from NVIDIA is still quite small.
而且其中很多还流向了其他地方，对吧？
And a lot of that ends up in other places as well, right?
比如马来西亚之类的。
You know, Malaysia or what have you.
所以最终，中国国内新增算力仍然只占10%。
So ultimately, China domestically still continues to have 10% of incremental new compute.
所以到2028年，我觉得可能会开始上升。
So in 2028, it might start to inflect up, I think.
但很容易说中国会有大约30吉瓦或更少的AI算力。
But it's pretty easy to say China will have like 30 gigawatts of AI compute or less.
到2028年？
By 2028?
是的，到2028年。
Yeah, in 2028.
好的。
Okay.
那他们的增长曲线会上升得多快呢？
And then how fast does their hockey stick go up?
我确实认为到2028年，他们能部署的算力会有大幅提升。
I do think in 2028, they have a big uplift in what compute they're able to deploy.
2026年，他们主要还是靠很多走私的芯片，你知道的，很多台积电给那些他们认为不是华为的公司造的芯片，结果最后还是华为的，还有三星运出来的很多HPM，但在27年，晶圆厂开始建起来了。
2026, they're still mostly relying on a lot of the smuggled chips, you know, you know, a lot of the chips that TSMC made for companies that they thought weren't Huawei but ended up being Huawei or a lot of HPM that Samsung is shipping, you know, sort of but in '27, fabs start to go up.
到了28年，尤其是中芯国际和长鑫存储这些厂开始投产，国产芯片的年产量真的能达到几百万颗。
In '28, especially, fabs start to go up from SMC and CXMT and such, where domestic production is actually reaching many millions of units a year.
然后他们就在2028年这一年里，逐步增加国产芯片的产能，大概五到十个吉瓦。
And now they're incrementally adding, you know, five, ten gigawatts in just 2028 of domestically produced chips.
对。
Right.
而且这些芯片肯定比不了英伟达、谷歌或者OpenAI在28年用的那些芯片。
And what Those chips are definitely worse than the chips that NVIDIA will have in '28 or Google will have in '28 or OpenAI will have in 2028.
所以你是说，就算用吉瓦这个数字来衡量，也还是高估了。
So even even the gigawatt number overstates things you're saying.
就像30吉瓦，但芯片质量差远了。
It's like 30 gigawatts, but it's really much worse chips.
但接下来，嗯，对。
But then how yeah.
要是你觉得明年全球算力需求会再增加一百吉瓦之类的，我知道你说过很难预测那么远。
How does it if you think the world is gonna add a 100 gigawatts of following year or something, you know, I know you said you can't really say that far out.
那中国接下来一年能增加多少呢？
How much is China able to add the subsequent year?
我就是想搞清楚，他们的增长曲线是不是在开始大规模出货算力芯片后突然陡峭上升的，还是说即便这样，总量还是比美国加上盟友的总和要少。
Basically, wanna know, did they just hockey stick at the point at which they are able to start shipping large amounts of compute or is it still gonna be less than US plus
以及盟友那边的情况。
allies?
还有很多不确定因素，比如美国会不会通过《芯片法案》，出口管制工具会不会继续收紧，以及中国能多快建立起他们开始具备国产化能力的新设备。
There's a lot left to, you know, whether or not The US passes the Match Act, whether or not tools continue to get export controlled, how fast China can build their new equipment that they're starting to be able to domestic produce domestically.
但归根结底，如果真有什么趋势的话，中国的产能肯定会呈指数级爆发式增长。
But ultimately, you know, China China is definitely gonna hockey stick if there's anything.
中国最擅长的就是极其迅速地扩大制造规模。
China is really good at is is scaling manufacturing really, really quickly.
我想，中国会逐渐把更多原本购买外国芯片的需求转向国产芯片，或者至少缩小美国允许英伟达卖给他们的芯片与国产芯片之间的差距。
And, you know, I imagine, you know, China's China will start to be able to extract more and more purchasing of even foreign chips into into domestic China or at least close the gap in what The US is allowing the you know, NVIDIA to sell them or what have you.
但你觉得中国到2029年能新增50吉瓦吗？
But but do you think China could do adding 50 gigawatts by 2029?
2029年的边际新增吉瓦数？
Marginal incremental gigawatts in 2029?
我觉得这完全合理。
I think that's I think that's completely reasonable.
对。
Yeah.
而且其中一部分也可能来自国外采购。
And part of that could also be purchased from foreign.
对。
Yeah.
但我认为中国在2029年达到50吉瓦是完全合理的。
But I think it's completely reasonable that China in 2029 can do 50 gigs.
但如果大部分是国产芯片，那这50吉瓦实际上可能只相当于美国的20吉瓦，对吧。
But if most of those are the domestic chips, there is some factor there where that 50 gigawatts is really worth as much as 20 gigawatts in America or Yeah.
我明白了。
I see.
来自美国芯片。
In from from American chips.
对。
Right.
对。
Right.
对。
Right.
所以，是的，你实际上在预测一个世界，到2028年，领先的实验室可能拥有比中国在2029年甚至2030年整个国家拥有的计算能力还要多的计算能力。
So, yeah, you're you're actually projecting a world where maybe the leading lab in 2028 has more compute that China will have in, like, all of China will have in '29 or even 30.
如果你按质量来权衡吉瓦的话。
If you if you weighed gigawatts by their quality.
这意味着没有什么能减缓美国实验室的发展。
Implying that there's nothing to do to slow down The US labs.
对。
Yeah.
很明显，政府和政客们已经开始这么做了。
Clearly, That's the government is starting and politicians are starting to do that.
对。
Yeah.
对。
Yeah.
对。
Yeah.
而中国不会放慢AI的发展。
Whereas China is not gonna slow down AI.
事实上，他们唯一会做的就是加速它。
In fact, the only thing they're gonna do is accelerate it.
说实话，当我跟黄仁勋重新确认出口管制的事情时，我其实是个自由意志主义者。
So honestly, I when I renewed Jensen and asked about expert controls, I am a libertarian person.
我当时就想，我其实不太确定自己对这个问题的看法。
I'm like, I wasn't, like, genuinely sure what I thought about this issue.
我是在为他相反的观点做辩护，因为我觉得把想法理清楚很重要。
I was steel manning what what is, like, the opposite view that he has and it because I think it's important to hash out ideas.
但我也在想，也许如果我们和中国合作，对我们反而更好，尤其是他们控制了那么多供应链，还有我们做机器人和其他东西需要的资源。
But I'm like, yeah, maybe there's a world where we if we just cooperated with China, it would be better for us, especially since they control so much of the supply chain and the other things that we needed for robotics and other things.
但我之前没意识到算力的情况像你说的那么糟糕。
But I didn't realize the compute situation was as fucked as you're saying.
实际上，如果出口管制真的像你说的那样限制出货量，那差别可就大了。
Like, actually, export controllers do seem to have, really, if they ship the amount that you're saying, that's a huge difference.
等到我们有自动编码器，甚至进入自动研究员的阶段，中国在算力储备上已经落后很多了。
By the time we have automated coder and even getting into like automated researcher, China is like way far behind on the compute stock.
如果真是这样，那这个策略就奏效了。
And so if that ends up being the case, that would have worked.
我觉得这其实是个不小的成功。
I think that's actually a notable success.
我觉得唯一的保留意见是，这里面有一部分是出口管制，但另一部分其实也是金融体系的问题，对吧？
I would say the only caveat there is some of it is export controls, but some of it is also just financial systems, right?
美国的金融体系比中国的更愿意冒险投资初创公司。
American financial systems are more willing to yolo into startups than Chinese financial systems.
但一旦中国的金融体系选定了一个行业重点扶持，它们会给予多得多的补贴。
But once Chinese financial systems choose an industry to focus on, they'll subsidize it a hell of a lot more.
对。
Yeah.
所以中国半导体行业获得的补贴比全球其他半导体行业加起来还要多得多，这说明如果技术起飞不像你暗示的那么快，而是需要更长时间，那么最终中国在半导体方面会大幅追赶上来，而半导体最终就是算力。
And so the Chinese semiconductor industry has significantly more subsidies than the rest of the world's semiconductor industries combined, which which points to, like, if takeoff is not as fast as sort of you're implying, but actually takes longer, then ultimately China will catch up drastically on the semiconductor side, which then is compute at some point.
我觉得另一个值得注意的方面是，今天中国公司在AI模型方面并不落后太多，至少从公众感知来看，相对于他们拥有的算力来说。
The other aspect of this that I would think is like I think is like noteworthy is Chinese companies today are not that far behind in AI models, at least percept perceivably by the public relative to the amount of compute they have.
对吧？
Right?
中国领先的实验室总共最多只有一两百兆瓦的算力。
The leading Chinese labs have a hundred, two hundred megawatts total of compute at most.
字节跳动的Seed是唯一的例外，他们的算力远超这个数。
ByteDance seed seed being the one outlier where they have significantly more than that.
但你知道，Kimi并没有运行一千兆瓦或接近这个规模的算力。
But, you know, Kimi is not running, you know, a gigawatt or anywhere close to it.
对。
Yeah.
而Anthropic到年底差不多有近五千兆瓦，对吧？
Whereas Anthropic is, you know, nearly five gigawatts by the end of the year, right?
或者更多，抱歉。
Or more, sorry.
所以，问题就是，这有关系吗？
And so, you know, the question is sort of, well, does it matter?
我觉得目前算力差异影响不大，因为当我们分析一个实验室的算力分配时，历史上大概是60%用于训练，40%用于推理。
And I think right now, it doesn't matter that much, difference in compute, because, when we break down the compute ratio or budget of a lab, historically, it's been, let's say, far it's been like 60% training, 40% inference.
但训练部分还可以进一步细分。
But that training gets broken down further.
实际上，算力的大约50%用于研究，10%用于开发，然后40%用于推理。
And as it's actually like 50% of the compute is research, like 10% of the compute is development, and then 40% is inference.
我说的研究和开发是指，研究人员在产生想法、测试新的架构、测试新的数据组合、测试新的超参数，不管他们在做什么，新的注意力机制等等，诸如此类。
And what I mean by research and development is, researchers are generating ideas, testing new architectures, testing new data mixes, testing new hyperparameters, whatever it is they're doing, new attention techniques, blah, blah, blah.
但最终，当他们进行训练时，比如Anthropic训练Mythos，功耗低于200兆瓦，对吧？
But ultimately, when they do the training run, when Anthropic trains Mythos, it's sub 200 megawatts, right?
你是指预训练还是整个过程？
You mean the pre train or the whole thing?
预训练。
The pre train.
对。
Yeah.
大约两个月，功耗低于200兆瓦。
It's sub 200 megawatts for call it two months.
然后强化学习用的算力更少。
And then the RL is even less.
但你觉得强化学习的计算量比预训练少吗？
But you think the RL was less to compute than the pre trained?
至少从单一站点的推理来看是这样。
At least in terms of single site inference.
我是说，预训练的站点，对。
I mean, site of pre Yeah.
总计算量是
Total compute was
可能更高，对吧？
probably higher, right?
总计算量。
Total compute.
但它是顺序进行的，对吧？
But it's like sequential, right?
对。
Yeah.
所以，他们最多在某个时间点同时使用的电力可能只有200兆瓦。
So, at most, the most they ever used at one point in time was maybe 200 megawatts.
但实际上，他们拥有多个吉瓦的电力。
And then, in reality, they had multiple gigawatts.
所以，他们大部分的计算资源都用于研究，而不是开发。而且这其中有原因，对吧？
So, most of their compute was going to the research, not the development of And a there's reasons for this, right?
你没法——协调所有这些集群是很困难的。
You can't It's hard to coordinate all these clusters.
把它们都放在同一个地方也很难。
It's hard to co locate all of them.
进行多地点训练也很困难。
It's hard to do multi site training.
进行强化学习也很困难。
It's hard to do RL.
你知道，在强化学习过程中生成更多的 rollout 并不一定会让结果更好。
You know, generating even more rollouts during RL does not necessarily make it better.
有很多原因可能导致你没法把所有的两吉瓦算力都用在训练上。
There's all sorts of reasons why you may not be able to leverage, you know, all two gigawatts that you have for training onto training.
对吧？
Right?
实际上，我只能用上两百兆瓦。
Actually, I can only leverage 200 megawatts.
随着我们越来越深入，实现自动化编码、自动化研究，我其实预计用于研究和训练的算力比例会开始变得模糊，甚至训练的比例会更高。
As we get closer and as we get further and further down, implement automated coding, automated researcher, I actually expect the percentage of the compute budget that goes to research versus training to start to and to start to like become a lot more fuzzy or even higher for training.
还有像持续学习这样的东西。
Also things like continual learning.
对吧？
Right?
所有这些都意味着越来越多的算力实际上会用在训练模型上。
All of these things start to mean that more and more is actually going to training the model.
如果你最终处于每年消耗一百吉瓦的情况，按当前价格计算，那每年就是五万亿美元的资本支出。
If you end up in a world where you're doing 100 gigawatts a year, at current prices, that would be 5,000,000,000,000 of capex every single year.
再加上你得提前很久就建好发电厂，而且那是个三十年的资产。
And then stack on the fact that you have to build the power plants way before then slash it's a thirty year asset.
再加上数据中心也是，你知道，十五到二十年的资产，而且那时候也得建好。
You stack on the fact that the data centers are a, you know, fifteen, twenty year asset, and you have to build that then too.
所以那五万亿，你知道，你得考虑到未来几年的增长。
So the 5,000,000,000,000, you know, you have to account for future years growth.
所以实际上可能会是七万亿到十万亿。
So it's actually going to be more like 7 or 10,000,000,000,000
我明白了。
of I understand.
因为你没算进去的是，那不包括发电基础设施或者数据中心本身那些东西。
Because you're not including the fact that like that doesn't include the fact that there's not the infrastructure for the power generation or whatever in the data center itself.
对。
Right.
没错。
Exactly.
对。
Yeah.
对。
Yeah.
而且数据中心本身，说到AI资本支出，大家说四万零五百亿美元，但那其实只是关键的IT部分。
And and the data center itself is when when you talk about AI capex, people are saying $4,050,000,000,000 dollars, but that's really just the critical IT.
对。
Yeah.
对吧？
Right?
服务器、网络、光纤、收发器、光通信，所有这些。
The servers, the networking, the fiber, the transceivers, optical communications, all this sort of stuff.
这还没算数据中心本身或者发电厂本身。对。
It doesn't account for the data center itself or the power plants themselves Yeah.
而这些都要提前建设。
Which are being built ahead of time.
对。
Yeah.
对。
Yeah.
对。
Yeah.
如果我今年建100吉瓦，明年建150吉瓦，那么这150吉瓦的所有建筑都需要在今年计入资本支出。
If I'm building a 100 gigawatts this year and a 150 gigawatts next year, well, then all of the buildings for that 150 gigawatts need to be built in capex this year.
如果后年建200吉瓦，那些电厂的投资也得提前花出去。
And if I'm building 200 gigawatts the year after that, all those power plants need to be spent.
今年就得买好涡轮机。
You have to buy the turbines this year.
对。
Yeah.
对。
Yeah.
对。
Yeah.
对吧？
Right?
所以你看，这其实比五万亿美元还要大得多，如果你要建一百吉瓦的话。
And so you've got this like this like actually, it's much bigger than even $5,000,000,000,000 if you're building a 100 gigawatts.
对。
Right.
所以很可能，每年的增量资本支出都快接近十万亿美元了。
So very plausibly, incremental capex every year is getting close to $10,000,000,000,000
到本十年末。
By the end of the decade.
对。
Right.
那将接近全球经济总量的十分之一了。
Which is gonna be like close to a tenth of the world economy.
而且如果所有这些都发生在美国，那差不多相当于美国经济的三分之一。虽然美国经济本身也会增长，但按目前规模来看，差不多四分之一到三分之一的经济都要投入到数据中心上。
And like a third of, if all of it's going up in The US, it's like, well, The US economy will have grown as well, but still, the current size of The US economy will be like a third to a quarter of The US economy would just be going towards data centers.
我这么一说，我自己都觉得，也许你说得对。
And as I say that out loud, I'm like, maybe maybe you're right.
我是说，就是不允许这样。
Mean, just won't allow it.
这就是为什么这事不会发生。
And that's the reason this doesn't happen.
对吧？
Right?
因为要让这种指数增长继续下去，就得把美国经济的四分之一都用来建数据中心。
Because like for this exponential continue, just like a quarter of the world a quarter of America's economy is just building data centers.
对。
Yeah.
我相信资本主义，相信资源会流向最赚钱的地方。
I mean, I believe in capitalism and reallocation of resources towards the most profitable thing.
但与此同时，政治因素也是存在的。
But at the same time, politics exist.
对。
Yeah.
而且信贷市场和资本市场也都是存在的。
And credit markets exist and capital markets exist.
所以，要实现比如说2030年100吉瓦的目标，或者甚至缩小到2028年，那时所有这些项目的资本支出加起来会有三四万亿美元。
So to enable, let's say, that 100 gigawatts by 2030 or let's even like let's even like pare it down to 2028 where it's like 3 or $4,000,000,000,000 of capex across all of these items.
你知道，其中大概两万五千亿是用于计算和IT方面的资本支出，另外一万到两万亿用于数据中心、能源以及下游的供应链，比如半导体等等。
You know, a couple, you know, over, you know, two and a half towards compute IT capex and then another one to two on data center and energy and all the supply chain downstream like semiconductors and all that stuff.
那么，如果资本支出达到三四万亿美元，这些钱都从哪里来呢？
So, if you're at 3 or $4,000,000,000,000 of capex, where does all this cash come from?
目前还没有哪家企业能从业务中产生这么多现金。
No one is generating that much cash from the business yet.
对吧？
Right?
超大规模云服务商到目前为止支撑了所有的增长。
Hyperscalers, they funded all of the growth up until now.
谷歌、微软、亚马逊、Meta。
Google, Microsoft, Amazon, Meta.
他们承担了其中很大一部分。
They funded a huge percentage of it.
他们占了计算资源的一半以上。
They were more than half of compute.
但他们现在不产生现金流。
But they now don't generate cash.
实际上他们把所有的钱都花在了资本支出上。
They actually spend everything on capex.
而且，他们还举债，把所有的钱都花在资本支出上，对吧？
And in addition, they raise debt and spend everything on capex, right?
你已经看到Meta这么做了。
You've seen Meta do it.
连亚马逊、谷歌也一样。
Even Amazon, even Google.
你知道，微软很快也会这样。
You know, Microsoft will be there soon.
大家都在举债来支付资本开支。
Everyone is raising debt to pay for their capex.
那么现在，谁会是那个以前没参与、现在要掏钱的新买家呢？
So now who is the incremental person to pay for this that was not doing it before?
比如说谷歌，他们很容易就停止回购股票，或者像Meta那样停止回购，转而购买计算机基础设施。
Well, in the case of like Google, it was pretty simple for them to stop doing buybacks or Meta stopped doing buybacks and turn around and buy computer infrastructure.
这对市场影响不算大，但确实有一些影响。
And that doesn't have a huge effect on the market, but it does have some effect.
但展望到2028年，超大规模云服务商现在要举债数千亿美元，他们的整个供应链也要举债数千亿美元，那谁来买单呢？
But as you step forward to 2028, where the hyperscalers are now raising hundreds of billions of dollars of debt, and then all of their supply chain is raising hundreds of billions of dollars of debt, who pays for this?
所以有几种不同的情况。
And so there's a few different ways.
你知道，像英伟达、博通这些半导体公司，还有存储公司，他们也在考虑出资支持一部分这些资本开支。
You know, there's the semiconductor companies like NVIDIA and Broadcom and the memory companies turning around and deciding to fund some of this capex.
还有传统的基础设施投资者，他们也在筹集资金，投资基础设施，不过不是桥梁，而是数据中心。
There's the traditional infrastructure investors who are turning around and gathering capital and investing in infrastructure and instead of bridges, it's data centers.
最后，经济中的每个人都在意识到，也许我不该买房，或者我不该为帮助别人买房的信贷投资，或者我不该买政府债券。
And then lastly, there's everyone in the economy who's realizing maybe I shouldn't buy a home or maybe I shouldn't invest in credit for a home that's helping people buy homes or maybe I shouldn't buy government debt.
我应该直接买超大规模数据中心运营商的债务，或者直接买这个数据中心的债务，或者买Anthropic的债务，因为Anthropic愿意支付20%的利率，来获得那额外的十亿美元建设产能，因为他们知道从中获得的收入会非常可观。
I should just buy hyperscaler debt or I should just buy this data center's debt or I should buy Anthropic's debt because Anthropic's willing to pay 20%, you know, rates for, you know, the incremental billion dollars to build their capacity because they know their revenue from it's going be huge.
他们愿意支付20%，因为这仍然比从SpaceX租用每吉瓦500亿美元要便宜。
And they're going to pay 20% because it's still better than renting it from SpaceX for $50,000,000,000 a gigawatt.
对。
Yeah.
所以，你就看到了所有这些竞争。
So you've got all of this contention.
但如果现在这样做，整个世界经济真的会发生巨大转变。
But if you now do this, the whole world economy is like really shifted around.
Antithesis 是一个确定性的软件测试平台，能够实现完美的可复现性。
Antithesis is a deterministic software testing platform that enables perfect reproducibility.
它还解锁了一些非常厉害的调试方法，比如时间旅行。
It also unlocks some pretty insane approaches to debugging, like time travel.
使用 Antithesis，你可以跳到轨迹中的任意一点，然后从那里开始。
With Antithesis, you can jump to any point in the trajectory and start from there.
所以当发生崩溃时，你可以倒回到出问题的确切时刻，并冻结整个系统。
So when there's a crash, you can rewind to the exact moment that something went wrong and freeze the entire system.
包括应用程序、数据库，甚至环境本身。
The application, the database, even the environment itself.
这让你能够做到原本不可能的事情，那就是在完全相同的瞬间观察分布式系统的每一个部分。
This lets you do something that would otherwise be impossible, which is to observe every part of a distributed system at the exact same instant.
时间旅行还允许你为已经发生的事件添加遥测和日志记录。
Time travel also allows you to add telemetry and logging to an event that has already happened.
例如，你可以倒回到崩溃前五秒，然后决定捕获所有网络流量。
For example, you can rewind to five seconds before a crash and decide to capture all the network traffic.
最厉害的是，Antithesis 给你一个系统的实时终端，你可以用它来随意干扰任何东西。
Most powerfully, Antithesis gives you a live terminal into your system that you can use to perturb anything you wish.
杀掉一个节点或者禁用某个功能，然后按下播放键，看看会发生什么。
Kill a node or disable a feature and then hit play and see what happens.
然后再回去试试别的。
Then go back and try something else.
在生产环境中，这种破坏性分析通常只有一次机会。
In production, you often only get one shot on goal with this sort of destructive analysis.
比如，如果你重启一个死锁服务，你本来想研究的那个死锁就消失了。
If you restart a deadlock service, for example, the exact deadlock you needed to study disappears.
但是有了 Antithesis，原始时间线总是可以重现的，所以你可以想测试多少假设就测试多少。
But with Antithesis, the original timeline is always reproducible, so you can test as many hypotheses as you need.
如果你不想自己搞这些时间旅行，你也可以让你的代理通过 Antithesis API 帮你做。
And if you don't wanna do all this time traveling yourself, you can just have your agents do it for you via the Antithesis API.
去 antithesis.com/dwarkesh 了解更多。
Go to antithesis.com/dwarkesh to learn more.
所以，过去几天我们私下一直在争论，AI会不会导致主权债务危机。
So you and I have been debating off air for the last few days, whether there will be a sovereign debt crisis as a result of AI.
逻辑是这样的。
And the logic is this.
AI的情况是，就像我们刚才提到的，很少的投资就能变成很多钱，对吧？
AI is, have a situation where, as we were mentioning, very little investment turns into a lot of money, right?
所以，回报率——
So, the rate of return-
这真是个该死的问题，兄弟。
What a fucking problem, dude.
是啊。
Yeah.
真不敢相信。
Can't believe it.
没错，这对其他那些不能把少量资金变成大量资金的人来说，是个巨大的问题，对吧？
No, it is a huge problem for everybody else who can't turn a little money into a lot of money, right?
所以回报率非常高。
So the rate of return is incredibly high.
即使在数据中心层面，如果你建一个数据中心，然后以折旧成本十倍的价格租给Anthropic或OpenAI，那也很疯狂。
Even at the data center level, if you build a data center and you're like get rented out to Anthropic or OpenAI for like 10x what it costs you on a depreciated basis to build it.
这简直太疯狂了。
It's fucking crazy.
所以一年下来，你的一美元能变成两美元、十美元甚至更多。
And so you turn $1 to like $2 or $10 or something at the end of the year.
这会推高利率。
That raises the rate of interest higher.
现在，如果利率上升，并且这种情况影响到整个经济，人们借越来越多的钱，他们与政府、其他公司或你作为消费者或房贷买家本来会进行的贷款竞争，那么这基本上会让其他所有人的借贷成本更高。
Now, if the rate of interest goes higher and it does that for the entire economy and people are borrowing more and more money, they're competing against the other lending that the government would have done or the other companies would have done or that you as a consumer or a mortgage buyer would have done, then that's just making it basically more expensive for everybody else to borrow.
这对很多人都有巨大的影响。
This has huge implications for tons and tons of people.
抱歉，我可能要说个不停了。
Sorry, I'm going go on a bit of a monologue here.
但我们一直在共同思考这个问题。
But we've been thinking about this together.
所以，我觉得美国最终会没事的，因为如果数据中心建在美国，基本上可以直接对数据中心征税。
So, I think The US will be fine at the end of the day because they can If the data centers are built in America, you can fundamentally just like tax the data centers.
但目前的税收体系下，企业收入占联邦收入的不到10%，而80%以上来自工资税和所得税，随着自动化程度越来越高，这部分会减少。
But the way the current tax system is set up, corporate income is like less than 10% of federal revenues and 80% plus is payroll taxes and income taxes, which as more and more automation happens will shrink.
同时，在支出方面，目前20%的税收收入用于偿还债务，基本上就是支付债务利息。
At the same time on the spending side, currently 20% of tax revenue spending goes towards paying servicing the debt, basically paying interest payments on the debt.
现在很多债务是短期的。
Now a lot of the debt is short duration.
所以每五年就会滚动一次。
So refurbishes every five years it rolls over.
你为什么笑成这样？
Why why are you fucking laughing?
因为，你知道，这就像我们最近一个月学到的东西。
Because, you know, it's like things we've learned you've learned in the last month.
对。
Yeah.
这对你来说有什么不同吗？
Like, it's any different for you?
我没有。
I didn't.
你不是拿了金融经济学学位吗？
Like, you got a degree in fucking financial economics?
我没有。
I didn't.
拿了。
Did.
网上的人以为我是养蜂的。
The Internet thinks I'm a beekeeper.
就几个月。
Few months.
几个月。
Few months.
几个月。
Few months.
这是我们的生意，迪伦。
This is our business, Dylan.
对。
Yeah.
对。
Yeah.
对。
Yeah.
抱歉。
Sorry.
抱歉。
Sorry.
所以现在我自己都觉得不好意思了。
And so now I'm self conscious.
靠。
Fuck.
不会啊。
No.
挺好的。
It's good.
你做得不错。
You're doing good.
我就是觉得挺搞笑的。
I just think it's funny.
上百万人。
Million people.
听听这个这个月才刚知道死亡这回事的家伙。
Listen to this guy who just learned about death this month.
所以，假设利率上升1%，那么在五年期内，用于偿还债务的税收收入比例基本上会从20%上升到25%。
So, you go from 20% of Suppose interest rates rise 1%, then over a five year basis, fraction of tax revenue that goes towards servicing the debt basically goes from 20% to 25%.
如果利率上升5个百分点，那这个比例就会超过40%。
If there is five percentages, that would go towards like north of 40%.
但如果考虑到政府每年借款2万亿美元，那么这个比例就会从大约40%上升到超过60%。
But if you take into account the fact that the government is borrowing $2,000,000,000,000 every single year, then that goes from like 40% to like north of 60%.
所以，基本上60%的税收收入都用来支付债务利息了。
So, 60% of tax revenue basically just goes towards paying interest payments on the debt.
不过，我觉得美国会没事的，因为如果让数据中心在美国建起来，税基也会增加。
Now, I think The US is going to be fine because also the tax base will increase if we let data centers get built in America.
在我看来，其他国家绝对要完蛋了。
Other countries are absolutely fucked in my opinion.
我刚才在看哪些国家债务多、税收收入少，而且债务偿还频率高。
I was just looking at which countries have a lot of debt, have very little tax revenue, and also a lot of their debt is serviced quite often.
像巴基斯坦或尼日利亚这样的国家，我觉得在新的利率环境下会非常惨。
And those countries, like Pakistan or Nigeria or something, I think are just going to be very fucked in this new interest rate regime.
所以这种挤出效应其实就是我一直在说的，也是为什么不是随便就搞个十亿吉瓦的疯狂计划。
So this crowding out effect is actually like the thing that I've like is the reason it's not like YOLO 1,000,000,000 gigawatts.
对，对。
Yeah, yeah.
对吧？
Right?
你看，有这么多行业和国家都大量举债，不管是那些你之前提到的贫穷国家，它们肯定会违约。
You've got you've got all these industries and countries that use a lot of debt, whether it's, you know, all these impoverished countries that you mentioned earlier that are just going to default.
还有消费品行业，对吧？
You've got like consumer packaged goods, right?
就像那些在Trader Joe's或者别的地方能看到的产品公司，它们都借了很多钱。
Like all of these like companies that make things you see at Trader Joe's or wherever use a lot of debt.
所有电信公司也借了很多钱，银行也是。
All these telecom companies use a lot of debt and banks use a lot of debt.
所以如果市场利率上升——不一定是政府设定的利率，而是市场上的利率——政府公布的联邦利率和其他人实际收取的利率之间会有差距，因为亚马逊明年可能要借一千亿美元，或者不管具体数字是多少。
And so if interest rates go up in the market, not necessarily the government set interest rate, but in the market, the spread of interest rate between what the government says their federal rate is versus what everyone else is charging because Amazon wants to raise a $100,000,000,000 of debt next year or whatever the hell the number is.
你知道，可能还少一些。
You know, probably less.
但最终你会遇到一个非常棘手的问题：钱从哪里来？
But you end up with this, like, really challenging problem of where does the cash come from?
有一部分是由现金流支撑的，而且现金流还在不断增长。
There is some level that is funded by cash flows and cash flows keep going up.
但合理的做法是投资远超你现金流的部分，因为未来几年的回报会非常可观。
But the logical thing to do is to invest way more than your cash flows because then the returns in the future years will be amazing.
所以就有了这个差额。
So you have this delta.
而压低这个差额的，是所有这些其他因素，对吧？
And then what's pushing down on the delta is all of these other things, right?
有对数据中心的监管，有消费者不满、政客不满，还有对AI的监管，以及AI实验室因为安全原因不发布最新模型。
There's regulations against data centers, regulations consumers getting mad, politicians getting mad, regulations against AI, the AI labs not releasing their latest models because of safety reasons.
所有这些因素，加上利率上升，都在影响这些事情。
All of these things and interest rates going up are an influence on all of these things.
所以，所有这些因素都会让曲线从资本主义在纯粹简单经济学下想要的结果，弯曲到我们所处的这个复杂系统想要的结果，而且越弯越低，最终导致实际建造的千兆瓦数量远低于应该建造的数量。
So, all of these things bend the curve from what does capitalism want in terms of just pure simple economics to what is the complex system that we have want and bends it lower and lower and lower to where not as many gigawatts as should be built will be built.
嗯，利率本身也是资本主义的一部分，对吧？
Well, the interest rate is part of capitalism, right?
对。
Yeah.
但就像，你知道，简单的经济模型和我们实际面对的复杂情况之间是有区别的。
But like, you know, like in the simple economic model versus like the more complex what we have.
你觉得明年亚马逊或Anthropic之类的公司发行债券的利率会是多少？
What is the rate at which you think Amazon or Anthropic or whatever will be releasing bonds for debt next year?
他们可是要借几千亿美元的债。
They do hundreds of billions of dollars of debt.
利率是多少？
What is the rate at?
平均利率是多少？
What is the average rate?
我觉得亚马逊不会借几千亿美元的债。
I don't think Amazon will do hundreds of billions of dollars of debt.
总共来说，比如说
To total, let's say the
大型科技公司，超大规模数据中心，还有所有云服务商加起来。
big tech The hyperscalers in total will rate and all the clouds.
对，对。
Yeah, yeah.
在我们的模型里，从2024年到2029年，资本支出大约是11万亿美元。
In the modeling that we do, we have about $11,000,000,000,000 of capex from 2024 to 2029.
总共。
Total.
总共。
Total.
如果你尽量用现金流来融资，最后还是需要发行超过5万亿美元的信贷来支撑这11万亿美元以上的支出。
And if you do you know, if you fund a lot of this with cash flows and as much as you can, you still end up with north of $5,000,000,000,000 of credit that need to be issued for this $11,000,000,000,000 plus You
你觉得AI收入不会继续三倍增长吗？
don't think the AI revenue continues even three x ing
同比三倍？
year over year?
AI收入确实在增长。
AI revenue does go up.
但我觉得不可能永远增长下去。
I don't think it can go up forever.
你知道，就像，如果没有遇到某些限制，我觉得某些实验室会有一些动力。
I don't you know, like, just like without, like, certain constraints being hit, I think certain labs will have certain incentives.
而且很多时候，实验室并不是建设所有算力的那方，尽管他们越来越想自己搞。他们会有这些现金流。
And and labs are not the ones building all the compute in many cases, even though they're increasingly trying to go They'll have all these cash flow.
比如，如果他们的收入一直保持，那也无所谓。
Like, if their revenue keeps whatever.
那没问题。
That's fine.
那你刚才说收入会是多少来着？
But if you how much did you say the revenue will be?
你觉得他们的收入不会有那么多吗？
You think they'll not have that much revenue?
对。
No.
我只是说，到2029年，大概会有11万亿美元的资本支出。
I'm just saying till 2029, there's, you know, something on the order of $11,000,000,000,000 of capex.
其中6万亿是用现金支付的。
And six of that is funded with cash.
另外5万亿是靠债务融资的。
And five of that is funded with debt.
如果是这样的话，整个生态系统要借5万亿美元的债，确实会让利率上升。
And if that's the case, $5,000,000,000,000 of debt being raised across the whole ecosystem does make interest rates go up.
那什么能阻止这种情况发生呢？
And then what prevents that?
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。