为什么更智能的AI模型可能将算力价格推高10倍

今天我想聊聊未来几年实验室的计算资源会是什么情况。

Today I wanna talk about what the compute situation for the labs will look like over the next few years.

过去连续三年，Anthropix 的收入每年都翻了十倍，今年很可能还会继续这样。

For the last three consecutive years, Anthropix revenue has 10xed year over year and it's likely to do so again this year.

所以去年他们的收入是 90 亿美元。

So they ended last year with 9,000,000,000 in revenue.

我觉得今年他们的收入大概会在 1000 亿到 1500 亿美元之间。

I think they'll probably end this year with somewhere between 100,000,000,000 to $150,000,000,000 in revenue.

如果这个趋势继续下去，Anthropic 明年年底就需要达到 1 万亿美元的收入。

Now for this trend to continue, Anthropic would need to make $1,000,000,000,000 in revenue by the end of next year.

当然，这并不一定非得发生，没有特别深层的理由。

Of course, there's no deep reason why this has to be true.

这个结论听起来很夸张，但归根结底还是取决于 AI 的能力。

It's a very wild conclusion and it's ultimately a question of AI capabilities.

到明年年底，AI 真的能变得那么有用吗？

Does AI get that useful by the end of next year?

但假设这个趋势确实持续下去。

But suppose the trend does continue.

我想思考一下在那个世界里会发生什么。

Well, I wanna think through what happens in that world.

现在，AI领域的另一个大趋势是，实验室的计算能力每年只增长三倍。

Now, other big trend in AI is that lab compute only three Xs year over year.

对于一个实验室来说，如果收入每年增长10倍，而计算能力只增长3倍，那么以下三种情况之一或它们的组合必须发生。

For a lab to keep 10 revenue year over year while compute only three Xs, one of the following three things needs to happen or some combination of the three needs to happen.

第一，实验室的利润率必须提高。

One, lab margins have to increase.

第二，计算的价格必须上涨。

Two, the price of compute has to increase.

第三，实验室用于推理而非训练的计算比例必须增加。

Or three, the percentage of compute that labs spend on inference rather than training has to increase.

据我所知，这三种情况基本上都已经在发生了。

My understanding is that basically all three of these things are already happening.

关于利润率，Anthropix的推理利润率据称从去年中期的40%上升到了现在的80%以上。

With regards to the margins, Anthropix inference margins reportedly went from 40% in the middle of last year to upwards of 80% now available.

关于算力，算力的现货价格比今年二月份的低谷时期高出超过40%。

With regards to compute, the spot prices for compute are more than 40% higher than they were in the February trough that we had earlier this year.

至于训练和推理的算力分配，根据Epoch的数据，2024年OpenAI只有四分之一的算力用于推理，而现在这个比例很可能接近50%，甚至更高。

And with regards to the share compute that goes to trading versus inference, in 2024 according to epoch, OpenAI was spending just a quarter of its compute on inference and that number is likely closer to 50% if not higher now.

实验室其实不太愿意走最后这条路，也就是增加推理算力的占比。

Now labs would prefer not to do this final thing of increasing the share of compute they spend on inference.

在实验室看来，推理收入的核心意义在于说服投资者继续投钱，好让他们去训练下一个更大、更好的模型。

The way the labs see the world, the whole point of inference revenue is to help convince investors to give you more money in order to train the next bigger, better model.

如果你把大部分算力都用在推理上，那基本上就等于在宣告AI的进步已经停滞了，你现在做的不过是个云服务提供商。

And if you're spending most of your compute on inference, you're basically declaring that AI progress has stalled and you're just now in the business of being a cloud provider.

这可比构建通用人工智能的吸引力差远了。

Now, this is a less compelling business than building AGI.

所以实验室不想干这个，也不认为自己已经走到了那一步。

So And the labs do not want to be in this business nor do they think they're in this world.

他们认为，一年之内就能造出模型，让现在的模型看起来非常糟糕，但他们需要投入大量算力，大部分算力，用于训练和实验，这是打造下一代模型所必需的。

They think that within a year they'll have built models that make the current ones look extremely shitty, but they need to invest a lot of their compute, the majority of their compute into doing the training and experiments that are necessary to build the next model.

这就只剩下两种办法，能解决实验室算力每年只增长三倍，但收入却增长十倍之间的差距。

So that leaves only two options for how you can get out of this gap between the fact that lab compute only increases 3x year over year, but revenue increases 10x.

要么实验室的利润率必须提高，这样他们才能获得盈余；要么算力的价格必须上涨，这样实验室下游的所有环节才能获得盈余。

Either the lab's margins have to increase so that they get the surplus or the price of compute has to increase so that everybody in the stack below the lab gets a surplus.

我不确定最终会是哪种情况。

It's not clear to me which world we end up in.

如果实验室利润率效应占主导，我们会不会进入一个世界，顶级模型的利润率从80%涨到90%以上？

Do we end up in a world where we go from 80% margins for some of the top models to greater than 90% margins if the lab margin effect dominates?

嗯，那需要领先模型远远甩开竞争对手，因为利润率之所以在市场经济中存在，是因为你提供的东西比市场上别人能拿到的替代品好得多。

Well, that would require the leading model to be so far ahead of the competition because the nature of margins, why they exist in a market economy is that the thing you are serving is so much better than what somebody else could go get and replace you on the market.

但对我来说，想到智能这种东西的利润率会超过90%，而且在这个水平上还不会被竞争蚕食，这真的很疯狂。

But it's just really wild for me to consider that the margins for something like intelligence will be greater than 90% and they don't get competed away at that level.

所以，要解决这两种趋势之间的差距，只剩下另一种可能性，那就是算力的价格必须上涨。

So that leaves only one other possibility of this escape valve between these two trends, which is that the price of compute has to increase.

就像我刚才说的，这已经开始发生了。

As As I mentioned, this is already starting to happen.

而且，当你看看前沿实验室真正需要积累的那部分算力时，这种效应就更明显了，因为他们不能直接去买一个按需实例。

And the effect is even stronger when you look at the tranche of compute that the frontier labs actually need to accumulate because they can't just go out and buy a spot instance.

他们需要确保获得足够的规模，以实现极高的效率和灵活性，同时还要拥有那种能保障自身安全和客户信息安全的算力。

They need to make sure that they get enough scale to get really good efficiency and flexibility and also that they have the kind of compute that lends itself to the security they need for their own ways and for their customers information.

所以，我觉得这里有个相关的案例值得看看，就是谷歌和Anthropic从SpaceX租用的算力。

So, I think a relevant case study here is to look at the compute that Google and Anthropic are renting from SpaceX.

比如，谷歌每月支付9亿美元，租用11万块GPU，这些GPU是GB200和GB300的混合体。

Google, for example, is paying $900,000,000 a month for a 110,000 GPUs that are a blend of GB two hundreds and GB three hundreds.

谷歌支付的价格是这些GPU按需价格的每小时两倍。

The price that Google is paying here is two x the spot price per hour for those GPUs.

而那个按需价格本身，已经比二月份高出40%以上了。

And that spot price itself is more than 40% higher than it would have been in February.

我想强调一个关键结论：随着AI模型越来越智能，它们将能更好地利用同样的算力来变现。

I wanna emphasize the key conclusion here that as AI models get smarter, they'll be better able to monetize the same amount of compute.

如果一名真正达到人类水平的软件工程师能在一台H100上运行，那么按目前软件工程师的薪资水平，一台H100的年租金应该超过25万美元。

If a true human level software engineer could run on an H100 equivalent, then at today's prices for software engineers, that h 100 should rent for over 250 k a year.

这比目前H100的现货价格高出15倍以上。

That's over 15 x the current spot price for an h one hundred.

这还没算上你的AI可以夜以继日地工作。

And this is not even accounting for the fact that your AI can work nights and weekends.

当然，你可能会想，如果经济中突然多出1000万名软件工程师，那么软件工程师的边际价值就会下降，因此那台H100能产生的收入就不会比现在高出15倍。

Of course, you might expect that if we had 10,000,000 extra software engineers suddenly appear in the economy, the marginal value of a software engineer would decrease and thus the revenue that that H100 would be able to generate would not be 15x higher than it is right now.

但我其实不确定这是否成立。

But I actually don't know if this is true.

如果把这个论点套用到人身上而不是AI身上，那这就是经典的劳动力总量谬误。

If we apply this argument to people instead of AIs, then this would be the classic lump of labor fallacy.

举个例子，经济学家普遍认为，从长远来看，高中移民并不会降低工资，因为创新和专业化提升了劳动力的价值。

For example, economists generally believe that high school immigration does not decrease wages in the long run because of how innovation and specialization increase the value of labor.

也许这次劳动力供给冲击规模太大、速度太快，以至于我们不能指望这个通用经验法则了。

Maybe this labor supply shock will be so big and so fast that we can't count on this general heuristic anymore.

但如果你认同标准经济学的观点，那么劳动的边际价值，进而算力的边际价值，应该会保持在令人惊叹的高位。

But if you believe what standard economics says, then the marginal value of labor and thus the marginal value of compute should stay astonishingly high.

所以咱们来想想，在这样的世界里会发生哪些变化。

So let's think about what changes in such a world.

首先，随着头部实验室在算力变现方面越来越强，以及算力成本不断上升，其他竞争者很难跟他们抗衡，因为他们得跟那些能更高效利用算力的人竞价争夺资源。

Well, one of the things that would happen is that as the top labs get better and better at monetizing compute and the cost of compute increases, it becomes harder for anybody else to compete against them because they have to bid for this resource against somebody who is basically able to make better use of it.

另外还会发生一件事，我觉得这才是整个思想实验里最有意思的推论：如果你能训练出最好、最高效的模型，你就能获得比现在高得多的利润率。

Another thing that will happen, and I think this is actually the most interesting implication of this whole thought exercise, is that if you can train the best, most efficient model, then you'll be able to charge much higher margins than you can today.

这就是经济学里的阿利金-艾伦效应。

This is the Alkin Allen effect in economics.

它的基本意思是，租用一块H100的成本大概是每小时20美元。

And what it's basically saying is that it would cost $20 an hour to rent an H100.

要是用那种性能更弱、效率更低的模型，那就太傻了，因为为了得到同样的结果，你得在昂贵的算力上消耗更多的token。

That would be extremely stupid to use a weaker, less efficient model because it's gonna burn more tokens on your expensive compute to get the exact same result.

所以，如果能训练出能更高效利用这种稀缺资源的模型，实验室就能收取更高的溢价。

So labs will be able to charge a much larger premium if they can train a model that better economizes this scarce input.

基本上，如果你有一个能用更少计算资源达到同样效果的模型，那在某种意义上你就创造了更多的计算资源，而计算的价值也会随之上升。

Basically, if you have a model that can get the same result by using less compute, then you've in some sense created more compute and the value of compute is gonna increase.

另一个会发生的事情是，现在很多流行的人工智能应用可能会因为价格过高而被淘汰。

Another thing that will happen is that a lot of current popular applications of AI will probably get priced out.

目前人工智能相对便宜的原因，是因为它还有很多事情做不到顶尖人类能做的水平。

The reason AI is relatively cheap right now is that AI just can't do a lot of things that top humans can do.

但这种情况迟早会改变。

But this at some point will no longer be the case.

到那时，谷歌、Anthropic或OpenAI愿意为自动化人工智能研究的计算资源支付更高的费用，而我们普通人却不愿意为生成更多低质量的人工智能内容花那么多钱。

And at that point, Google or Anthropic or OpenAI will be willing to pay more for the tokens to automate AI research than you or I will be willing to pay to make more AI slop talk.

说实话，我有点担心这种分析方式，跟过去人们对稀缺性的错误判断模式很像。

I'm a bit worried that this kind of analysis honestly pattern matches a lot onto the ways that people in the past have been wrong about scarcity.

比如，我就想到了著名的西蒙-埃利希赌约。

I'm for example thinking of the famous Simon Ehrlich bet.

保罗·埃利希是个著名的悲观主义者，他预测人口增长会导致资源短缺，于是跟人打赌，说一篮子商品的价格在1990年之前的十年里会上涨，而不是下跌。

Paul Ehrlich was this famous doomer about population growth and he made this bet that a basket of commodities would increase in price rather than decrease in the decade preceding 1990.

这个赌局非常有名，因为它本意是用来证明埃利希那种马尔萨斯式的世界观是错误的，说明他没有预见到市场信号和人类的创造力能够找到更好的方式来节约稀缺资源。

And this is a very famous bet because it's supposed to illustrate how Ehrlich's Malthusian worldview was wrong and how he did not anticipate the way in which market signals and human ingenuity can find better ways to economize scarce inputs.

我猜这个类比可能不太对。

I'm guessing that the analogy to this bet is probably wrong.

有其他分析显示，如果这个赌局放在不同的十年里，埃利希很可能会赢。

Other analysis has shown that if that bet had been made in a different decade, Ehrlich might well have won.

但更广泛地说，我认为算力的供给弹性要小得多，吸收大规模需求冲击的能力也弱得多，而且很难像开采不同金属那样通过使用替代品来应对。

But more generally, I think the supply of compute is much less elastic and much less capable of absorbing large demand shocks and much less capable of being accommodated by using different substitutes than the extraction of different metals.

为了说明为什么我认为每年三倍的算力增长很难被撼动，甚至可能难以维持，是因为我看不出构成这个三倍增长的三个要素中，有哪一个能大幅加速。

To illustrate why I think this 3x in compute capacity year over year is hard to budge or potentially even sustain is that I don't see how any of the three elements that constitute that three x can be much accelerated.

其中1.4倍来自摩尔定律。

So 1.4 x of that is coming from Moore's Law.

别说加速了，我觉得能再维持几年就已经是奇迹了。

Far from increasing it, think it'll be a miracle if you can just keep it going for a few more years.

1.2倍来自建设新的晶圆厂。

1.2 X is coming from building new fabs.

这个过程最终会受限于制造新的ASML EUV光刻机，这种情况可能会持续到2030年甚至更久。

This process is ultimately gonna be bottlenecked up to 2030 and potentially even beyond by just building new ASML EUV machines.

迪伦几个月前上播客的时候，详细讨论过这个问题。

Dylan, when he was on the podcast a few months ago talked about this in great detail.

而1.8倍的增长来自于AI正在吸收原本用于智能手机和个人电脑的大量晶圆产能。

And 1.8 x comes from the fact that AI is absorbing a lot of wafer allocation that was previously going to smartphones and PCs.

这很可能在明年年底前遇到瓶颈，届时在台积电最先进的N3节点上，AI的占比将从60%上升到86%。

This is probably gonna hit a wall by the end of next year when at the leading edge N3 nodes at TSMC, AI will have gone from 60% to 86%.

到了某个点，所有最先进的晶圆产能都会被AI占满，这个数字就没法再继续增长了。

At some point, you have just absorbed all leading edge wafer capacity for AI and you can't keep increasing this number.

所以我不知道未来几年我们怎么能继续维持每年三倍的计算规模增长，更别提超越这个速度了。

So I don't know how we get even to continue to do three x compute scaling year over year for the next few years, much less go beyond that.

月底的时候，我会按照老传统，把账目结清。

At the end of the month, I go through the time honored tradition of closing my books.

我先打开我的银行平台Mercury，确保所有交易都归类正确。

I start by opening Mercury, which is my banking platform, to make sure that all my transactions are properly categorized.

自动分类规则能很好地处理那些常规的支出，但我经常要跟新的合作方打交道，比如家教、研究员、视频制作人，还会试用各种新工具。

Auto categorization rules handle the predictable stuff pretty well, but I'm constantly working with new contractors, you know, tutors and researchers and videographers, and I'm also trying new tools.

如果手动给所有这些交易分类，每个月得多花好几个小时。

Manually categorizing all of these transactions would add a couple of hours of overhead every single month.

所以我不再一笔一笔地处理，而是让Mercury自带的AI工具Command一次性搞定所有分类。

So instead of going through them one by one, I have Command, which is Mercury's built in AI, take a stab at all of them at once.

Command会为每笔交易建议一个类别，并给出它的判断依据。

Command proposes a category for each transaction and provides its rationale.

我只需要过一遍，把不对的地方改掉，然后确认就行了。

I just review, I fix anything that's off, and I approve.

等Mercury里这些工作都完成后，它会自动和QuickBooks同步。

And once all this work is done in Mercury, it syncs everything with QuickBooks.

而且Command的判断确实挺准的。

And Command's judgment calls are genuinely good.

它不光会看商家名称这种明显的信息，还会查一下是我团队里谁花的钱，再看看备注和说明，尽可能多地了解背景情况。

It does the obvious things like looking at the vendor, but it also investigates who on my team made the purchase and looks at notes and memos to build up as much context as possible.

这只是用Command来自动化你业务后端的一种方式。

This is just one of the ways you can use Command to automate the back end of your business.

想了解更多，请访问mercury.com/command。

To learn more, go to mercury.com/command.

Mercury是一家金融科技公司，并非FDIC承保的银行。

Mercury is a fintech company, not an FDIC insured bank.

银行服务由Choice Financial Group（NA分行）提供，该机构为FDIC成员。

Banking services provided through Choice Financial Group in column NA, members FDIC.

AI生成的回复和建议可能会有所不同，且不保证其准确性。

AI generated responses and suggested actions may vary and are not guaranteed.

现在，我想说明一点：在未来某个时候，计算成本会再次变得便宜。

Now, I want to clarify that at some point in the future, compute will get cheap again.

总有一天，我们会有机器人能把沙滩上的硅砂和矿山里的铜直接转化为新的电脑芯片。

At some point, we'll just have robots that can convert shores of silica sand and mines of copper into new computer chips.

到那时，计算成本基本上就只剩下原材料和加工所需的工具了。

And then the price of compute is basically the raw inputs and the tools required to do this processing.

我现在说的是当前这个前奇点阶段，AI算力每年只增长三倍，这不足以抵消AI价值随时间增长的速度。

I'm just talking about this current presingularity regime where AI compute merely three Xs year over year, which is not enough to offset how much more valuable AI is becoming over time.

顺便说一句，Anthropic的收入每年增长十倍，而他们的算力只增长了三倍，我认为这正好说明了模型业务中规模经济效应有多强。

By the way, the fact that Anthropic revenue has been 10x in year over year, whereas their compute has only been 3x in year over year, I think illustrates how strong the economies of scale are in the model business.

从逻辑上讲，这很合理。

And logically, this makes sense.

训练一个模型时，只需要一次性投入成本来学习各种不同的技能，这些技能随后可以共享给所有用户。

When you train a model, just had to spend this one time cost to learn all these different skills that then get to be shared across all your users.

这和人类劳动力完全不同，每个个体都需要从头开始重新训练。

This is very unlike human labor where each instance has to be retrained from scratch.

我真希望我们不是生活在这样一个智能领域规模经济效应如此强大的世界里，因为我担心权力集中，但现实似乎就是这样。

I wish we didn't live in a world with such strong economies of scale for intelligence because I'm worried about power concentration but it seems we do.

好了，这是我博客文章的一个配音版本，文章也发布在我的网站dwarkesh.com上。

Okay, this was a narration of a blog post that I also released on my website at dwarkesh.com.

去看看其他文章吧，或者关注一下，这样我以后发新文章时你就能收到通知。

Check it out for other posts or to be notified when I release a post in the future.

不然的话，我们下期完整节目再见。

Otherwise, I'll see you for the next full episode.