Can Anyone Catch NVIDIA? | The Future of Chips and Infrastructure

随着像GPT-5这样的模型出现，或者开源模型持续压低成本，这些公司想赚到钱就越来越难了，很多时候它们的推理业务毛利率连50%都不到。

And as you get models like GPT-five or open source models continuing to drive it down, it's like the value capture is just harder and harder and harder for these companies because they're making 50% gross margin on inference if they're or less in many cases.

在这么多

In so many

换句话说，你的意思是我们在被商品化，所以赚不到钱，那也就该收敛一下在GPU上的投入预期了？

words, you're saying we're getting commoditized, and and therefore you can't capture the value, and thus you should temper your expectations of how much how much you can spend GPUs?

不，不是这样。

Well, no.

我觉得还是有办法能大幅提高价值回报的。

I think I think you can I think there's still ways to, like, inflect hugely on value capture?

对吧？

Right?

我刚才提到，广告是个巨大的价值提升点。

I mentioned the ads are a huge value capture inflection.

在大规模增长之前，这种情况还得先发生。

To happen before before we see massive increase.

不是。

No.

我觉得另一点是，还有很多资本还没花出去。

I think I think the other thing is, like, there's a lot of capital that's not been spent.

对吧？

Right?

像那些超大规模云服务商，明年资本支出还能再涨30%，比今年还高。

Like, the hyperscalers still can grow CapEx 30% next year, right, from what they're doing this year.

另外，像CoreWeave和Oracle这样的公司，因为能从资本市场融资，资本支出增幅能远超20%到30%。再往下看，全球最大的基础设施基金，比如布鲁克菲尔德和黑石，现在也都把目光转向了基础设施，尤其是AI基础设施，更别提各国的主权财富基金了，像G42、挪威的那家，还有新加坡的GIC。

In addition, companies like CoreWeave and Oracle, because they're tapping capital markets, can raise way more than 20 to 30% CapEx, and then you go down the list further and it's like, oh, the largest infrastructure funds in the world, like Brookfield and Blackstone, well actually they're turning all of their eyes to investing even more into infrastructure, AI infra, and then you're like, the sovereign wealth funds of the world, like the G42s, or you know, the Norway one or GIC in Singapore.

这些人对AI的投入才刚起步，所以我觉得，未来还有大量资本会涌入，而且不一定从一开始就冲着经济回报去的。

Like, these people have barely started touching AI, and so I think there's a whole lot more CapEx that can come without it being necessarily, like, economically motivated day one.

我想说的是，纯粹出于经济动机的资本支出，增长空间终究有限，但还有很多其他情况——比如你拿张表格算来算去，按实际业务来看，根本看不出该花这么多钱。

I'm more so saying, like, economically motivated CapEx can only grow, like, so much, but there's so much other, like like like, where it's not clear from you know, if you have a spreadsheet, you know, and you're basing it on real business that you should actually spend this much.

但人们还是会花，因为他们相信。

But people will because they believe.

我相信。

I believe.

我觉得你也相信。

I think you believe.

比如基础设施这块，大家觉得投了总能赚回来，但其实并没有什么百分百确凿的依据能说得通。

Like, infra you know, people believe that this will be you'll get profit out of it, but there's no, like, 100% certain, like, you know, way to argue it.

是的。

Yeah.

NVIDIA在多大程度上受到定制芯片的威胁？

How's threatened, if at all, is NVIDIA by by custom silicon?

我觉得这才是最关键的问题，你看谷歌、亚马逊，还有Meta下的订单。

I think that's the biggest thing, right, is when we look at orders from Google and from Amazon, right, especially, and and Meta.

他们的定制芯片，可不是微软那种。

Their custom silicon is not not Microsoft.

他们的定制芯片其实挺差的。

Their custom silicon kinda sucks.

但另外三家在过去一年里大幅增加了订单量。

But the other three, they're really upping their orders massively over the last year.

亚马逊正在生产数百万颗Trainium芯片。

You know, Amazon is making millions of Trainium.

谷歌也在生产数百万颗TPU芯片。

Google's making millions of TPUs.

TPU显然已经满负荷运转了。

TPUs clearly are, like, 100% utilized.

对吧？

Right?

是的。

Yeah.

他们目前还没做到，但我认为亚马逊迟早能搞定，Anthropic也会。

They didn't really not there, but I think Amazon will figure out how to do that, and Anthropic will.

所以我觉得对英伟达最大的威胁，其实是越来越多公司掌握了定制芯片的使用方法。

So so I think I think that's the biggest threat to NVIDIA is that people figure out how to use custom silicon more broadly.

所以如果人工智能的资源越来越集中，定制芯片的优势就会更明显。

And this sort of becomes the sort of, like, if AI is concentrated, then custom silicon will do better.

这还没算上OpenAI的芯片团队之类的事情呢。

And that's not even talking about, like, OpenAI's silicon team and stuff.

对吧？

Right?

要是人工智能真的被少数公司垄断，那定制芯片就会更有优势。

Like, if AI is really concentrated, then they'll do better, custom silicon.

但如果因为中国出现了大量开源模型，再加上NVIDIA在中国推出的开源软件库，让部署成本降到了极低，那情况就可能不一样了。

But if it gets dispersed broadly because there's all these open source models from China, and there's all these open source software libraries from NVIDIA in China, it makes the deployment costs rock bottom, then potentially

你听我说完。

Hear me out here.

理论上，如果谷歌的TPU真能和NVIDIA抗衡，它完全可以在公开市场上竞争。

If Google's TPU is able to compete with NVIDIA, in theory, it could do it on the open market.

现在谷歌的市值还不如NVIDIA高。

NVIDIA is worth more than Google these days.

谷歌难道不应该把他们的芯片卖给所有人吗？

Shouldn't Google start selling their chips to to everyone?

我的意思是，理论上他们完全有可能获得更高的市值。

I mean, in theory, they should be able to achieve a higher market cap.

我完全同意。

I I absolutely think so.

我觉得谷歌内部其实已经在讨论这件事了。

I think Google's even discussing it internally.

但这需要彻底重组企业文化，重新调整谷歌云、TPU团队，还有Jack软件团队和XLA软件团队的工作方式。

I think it would require a big reorg of culture and a big reorg of how Google Cloud works and how the TPU team works and how the Jack software team and XLA software teams work.

我坚信他们做得到。

I totally think they could.

只是他们得狠狠地自我革新一番才行。

It would just take them shaking themselves pretty hard to be able to do it.

对。

Yeah.

但我真的觉得谷歌应该把TPU卖出去，不只是出租，而是真正地出售硬件。

But I I I totally think Google should sell TPUs externally, not just renting, but, like, physically.

挺讽刺的是，理论上这个副业的公司估值潜力反而比主业还高。

It's kind of funny if a side hobby in theory has a higher company value potential as you

尤其是当你考虑到搜索业务在核心业务中的衰退时。

move business, especially as you think about the degradation of search since the core business.

我的意思是。

I mean

是的。

Yeah.

我觉得如果你去问谢尔盖，他会怎么说。

I think but I think, like, if you were to ask, like, Sergei.

对吧？

Right?

比如说，嘿。

Like, hey.

你觉得卖芯片和机架，跟卖云服务或者Gemini相比，哪个更有价值？

Do you think selling chips and and racks is more valuable or cloud or or Gemini?

他肯定会说，不会。

He'd be like, no.

不会。

No.

不会。

No.

不会。

No.

Gemini未来的价值会高得多得多。

Gemini is gonna be worth way, way, way more.

只是现在还没到那一天。

It's just not yet today.

所以现在来看，你还是会说NVIDIA最厉害，这其实是个集中度的问题。

So I think today, you say NVIDIA is the most again, it's like a whole concentration thing.

对吧？

Right?

如果客户高度集中，那么英伟达就不会是全球最有价值的公司。

If the world is super concentrated in terms of customers, then NVIDIA will not be the most valuable company in the world.

但如果分布越来越分散，比如现在开源模型越来越强大，部署也越来越容易，那英伟达很可能在未来很长一段时间里，依然是全球最有价值的公司。

But if it gets dispersed more and more, which arguably we're starting to see with a lot of these open source models getting better and better and better, and with the ease of deploying them getting better, then you would see I think you could argue NVIDIA will remain the most valuable company in the world for a long period of time.

历史上看，不会。

Historically, no

这话有双关的意思——软件已经吞噬了大多数行业。

pun intended, software has eaten the world in most markets.

你看早期网络时代，思科曾经是全球最有价值的公司。

Mean, if you look at early networking days, Cisco was the most valuable company on the planet for a while.

但它现在不是了。

It's no longer.

那些在上面构建服务的公司，比如谷歌、亚马逊、Meta，最终都超越了它。

The guys that built services on top like Google or Amazon or Meta eventually eclipsed.

所以英伟达才在开发这么多软件库，对吧？

Which is why NVIDIA is making all these software libraries, right?

他们正试图让推理服务变得商品化。

And they're trying to commoditize inference.

对吧？

Right?

我觉得你们应该没有投资过专门做推理API的公司吧？

You guys don't, I think, even have an inference API provider investment, do you?

我们确实和很多模型提供商都有合作。

Well, we have all kinds of model providers.

模型提供商。

Model providers.

我说的是纯粹做API服务的公司，我觉得是这个意思。

I'm talking about a pure API provider investment, I think.

对吧？

Right?

这么说对吗？

Is that correct?

我记得我跟团队里某个人聊过，问过为什么你们没投资像 Together 或 Fireworks 这样的平台。

I think I talked to one of the team members, maybe maybe or someone about, like, why you guys don't didn't invest in, like, you know, like a together or, like, a fireworks.

他们的说法是，光做模型服务而不自己开发模型，迟早会变成红海。

And sort of the argument was, like, well, we think just serving models alone without making them will sort of be commoditized.

是的。

Yeah.

对吧？

Right?

我们确实投了一些

We have some in

比如在 Stable Diffusion 生态里。

the stable diffusion ecosystem.

像那种 FOUL FOUL 的项目。

Like, with, like, a FOUL FOUL.

是的。

Yeah.

逐渐变成是的了。

Decay to Yeah.

我觉得那边的情况有点不一样。

It's a little bit different dynamics there, I think.

他们更倾向于打造更复杂的模型，而不是像语言模型团队那样。

They tend to be make much more compound models than the the LM folks.

对，没错。

Think Yeah.

是的。

Yeah.

不一样。

Different.

但你们自己并没有类似Base10那样的API投资，是因为你们觉得，这种基础设施会因为软件和视频的发展而被商品化？毕竟像VLM和SD这样的开源项目，最初来自伯克利，现在已经有自己的生态体系，还得到了广泛支持。

But but, like, you guys don't have one of these, like, know, base 10 or any of these, like, sort of, like, API investments because you think this is from someone on the infra team that you guys think it'll get commoditized because of software and video is making, because VLM and SD LAN, which is like open source software coming out of Berkeley and now sort of has their own environments now and supported by many.

这意味着API提供商的价值可能没那么高，这大概是你的观点吧。

This being commoditized means that API providers aren't necessarily worth a ton, right, is sort of your argument maybe.

我觉得这跟整个话题都相关，你知道的，为什么呢？

And I think that's relevant to this whole thing, is, you know, why?

对吧？

Right?

那为什么会有

Like, why would

人这么做呢？

you do this?

换个话题，那硅芯片领域的初创公司呢？

Shifting gears, what about the silicon startups?

你怎么看这些公司？

What's your take on those?

我的意思是，现在有大量资金涌入这个领域，对吧？

I mean, there's a ton of capital flowing into that, right?

我们看到，虽然没有具体数字，但芯片初创公司已经吸引了数十亿美元的投资。

We've seen, not numbers, but probably billions being invested in chip startups?

是的，确实如此。

Yeah, for

当然，毫无疑问。

sure, for sure.

你看，像Etched、Rivos，还有Maddox等一批公司，连芯片都没发布，就拿到了这么多融资，这真的很了不起。

I mean, whether you're looking at companies like I think it's pretty impressive that a few companies like Etched and Revos and a number of other companies, Maddox and others, have gotten the amount of funding they've had without even launching a chip.

过去，半导体公司要想融资或盈利，至少得先把芯片推出来。

In the past, yes, silicon companies would make money or raise money, but they would at least launch a chip before they get a big round.

但Etched和Rivos没公开发布过任何芯片，就融到了巨额资金，这说明一点：硅芯片领域确实烧钱，尤其是做加速器这种结构复杂的芯片，牵涉的环节太多了。

But Etched and Rivos have raised a lot of money without ever launching a chip publicly, which I think is I mean, it speaks to well, yes, silicon is super capital intensive if you're building a chip, especially an accelerator, which has so many moving pieces.

现在市面上至少有十家新兴的AI加速器公司，都是最近几年才冒出来的。

And there's there's, like, there's, like, 10 different AI accelerator companies out there, right, like, are newish in the last few years.

我觉得实际数量还要更多。

I think there's a lot more.

是的

Yeah.

那些公司确实如此

That that are like yeah.

是的

Yeah.

是的

Yeah.

这很

That's

合理

fair.

还有那些老牌企业，也在持续融资

And then and then and then there's the old guard, which continues to raise money.

是的

Yeah.

对吧？

Right?

比如Grok、Cerebras，还有那个。

Like, Grok and Cerebras and Mhmm.

还有SemiNova、Tensor，诸如此类。

And SemiNova and and Tensor and so on and so forth.

对吧？

Right?

比如Graphcore被软银收购了，软银也在往这个项目里砸钱。

Like, or Graphcore getting bought out by SoftBank, and SoftBank dumping money into this effort as well.

对吧？

Right?

现在有大量资金涌入，想打破英伟达的垄断地位。

There's there's a lot of capital being invested to disperse dispel sort of NVIDIA's top dollar or top position.

但这变得很有挑战性。

But it becomes challenging.

对吧？

Right?

那你怎么才能超越英伟达呢？

It's like, how do you beat NVIDIA?

对吧？

Right?

我觉得那些超大规模云服务商挺幸运的，因为他们基本上能做和英伟达差不多的事。

Like, the hyperscalers, I think, are, like, kind of lucky in that they can they can do mostly the same thing as NVIDIA.

对吧？

Right?

他们是自己的客户，完全被自己绑定了。

They're they're captive customer, which is themselves.

对吧？

Right?

没错。

Right.

这可是个巨大的优势。

It's a huge asset.

而且这

And it's

他们完全可以靠供应链取胜。

they can they can just win on supply chain.

对吧？

Right?

比如，我用的是更便宜的供应商。

Like, I'm using cheaper providers.

说白了，这就是个压缩利润的过程。

So it's a margin compression exercise, essentially.

是的。

Yeah.

没错。

Yeah.

对于某些特定的工作负载，比如Meta的推荐系统，他们可能能做得更好，更专注一些。

And maybe maybe for certain workloads, like Meta for recommendation systems, they'll have a better you know, they can specialize more.

但大多数情况下，答案是否定的。

But for the most part, it's like, no.

我们针对的其实是相同的工作负载。

We're we're targeting the same workloads.

我们只要简化供应链，或者把大部分环节内部化，压缩一下利润空间，就能搞定。

We can just simplify supply chain or in house a lot of it and compress margin, and it'll be fine.

但像这些其他公司，问题在于他们没有自家的固定客户。

But in the case of these other companies, it's like, well, they don't have a captive customer.

所以现在你得面对一个问题：我用的是同样的生态，要么找一家定制芯片供应商，他们照样要抽一层利润，这会压低我的售价；要么自己从头做起，但说实话，这难度太大了。

So now you have to contend with, well, I'm using the same ecosystem, and either I can use some custom silicon provider who's gonna take a margin anyways on top, and that's gonna compress my like, what I can sell for, or I can try and in house everything, but then it's like, this is really hard, right?

我要自己做全部的软件设计。

Like, I'm gonna do all the software design.

还要自己做全部的芯片设计。

I'm gonna do all the silicon design.

我得自己开发所有这些不同的知识产权。

I'm gonna build all this different IP.

我还得管芯片、机架这些所有环节的供应链，对吧？

I'm going to manage the supply chain on chips, racks, on everything, right?

最后需要的团队规模大得惊人。

Ends up being a huge effort in terms of team size.

归根结底，我像英伟达一样能拿到75%的毛利。

All in the end, hey, I make a 75% gross margin as NVIDIA.

AMD的GPU毛利只有50%，他们工程能力很强，却还是很难超越英伟达。

AMD sells their GPUs for 50% gross margin, and they have a hard time out engineering NVIDIA, and they're great at engineering.

他们用的芯片面积更大，内存更多，才能达到同样的性能，但售价还更低，利润就被压得厉害。

They're they but yet they still take more silicon area, more memory to achieve the same performance, and they have to sell for less, so their margin gets compressed.

这倒说得通。

That makes some sense.

说实话，从历史上看，新进入者通常不是靠小幅改进现有产品而胜出的。

Look, the I think historically, if you look at it, typically, new entrants in markets didn't win by marginally improving on something existing.

这种情况偶尔会发生，但更常见的是，新进入者抓住了某种颠覆性的技术突破，采用了完全不同的思路和技术。

That happens sometimes, but more likely, they jumped on some kind of disruptive technology leap, where it's like, we have a different approach, we have different technology.

这里有可能出现这种情况吗？

Is that possible here?

这么说吧，虽然有点简化，但我认为Transformer模型胜出的部分原因，是因为它在GPU上运行得实在太出色了。

Mean, to some degree, this over simplifying a little bit, but I think part of the reason why the transformer model won was because it runs so incredibly great on GPUs.

对吧？

Right?

比如循环神经网络，性能看起来也差不多，但在GPU上跑起来却糟糕透顶。

Like, a, recurrent neural network is similarly performing, it looks like, but it it runs terribly on a on a GPU.

所以我们是不是因为架构选了这个模型，现在就很难再想出一个真正……

So so did we sort of pick the model for an architecture, and now it's it's hard to come up with an architecture that, you know, really

这其实是硬件和软件协同设计的结果。

Well, it's it's hardware, software co design.

对吧？

Right?

现在不是到处都在吹神经形态计算吗？

Like, there's, like, there's all this hype about neuromorphic computing.

对吧？

Right?

理论上确实很厉害，效率也超高。

Like, theoretically, it's amazing and super efficient.

说白了，就是这么回事。

It's like, okay.

挺好的。

Great.

但硬件生态根本不存在。

Like, there's no ecosystem of hardware.

软件生态也一样没有。

There's no ecosystem of software.

要证明这玩意儿值不值得搞，光是硬件、软件和模型这三个方面，就得让现在最顶尖的成千上万名AI专家全身心投入才行。

It would take, like, you know, tens of thousands of people who are the best AI today focusing on that to even prove out if it's worthwhile or not, right, on a hardware side, on a software side, on a model side.

所以你看Grok、Ebris、SambaNova这些公司，设计芯片时都过度押注了当时最主流的模型。

And so, like, you look at, like, Grox or Ebris, SambaNova, they all, like, sort of over indexed to the models that were leading at the time when they designed their chips.

因此他们做了一些权衡。

And so they made certain trade offs.

对吧？

Right?

他们在芯片上集成的内存多得多。

They put a lot more memory on chip.

而NVIDIA却说，我们不这么干。

And and NVIDIA was like, well, we're not gonna do that.

至少速度更快一些。

A lot faster, at least.

对吧？

Right?

更准确地说，如果你对比NVIDIA芯片上的SRAM内存容量，那真的低太多了。

Well, more like like, if you compare the amount of memory of SRAM on NVIDIA's chips, it's much, much lower Yes.

没错。

Correct.

用的是SRAM，而不是DRAM。

Than SRAM instead of DRAM.

但确实如此。

But Right.

所以他们的DRAM容量通常更小，这也是一种权衡。

Then they usually have less DRAM, so there's a trade off there as well.

对。

Right.

DRAM更少。

There's less DRAM.

SRAM更多。

There's more SRAM.

因为芯片上SRAM多了，计算单元就得减少，结果他们反而输了。

And because there's more SRAM on the chip, you have to have less compute on the chip, and so they ended up losing.

对吧？

Right?

因为模型变得太大了，各种原因吧。

Because the model sizes got too big and all this.

对吧？

Right?

于是就出现了一个特别奇怪的局面：他们押注的其实是更优的方案。

And so you have this super weird dynamic where they bet on something that was actually better.

毫无疑问，Cerebras 在运行某些类型的模型时，比 NVIDIA、Grok，甚至 Dojo 都更出色。

Have no doubt that Cerebras would run certain types of models better than NVIDIA or Grok or, hey, Dojo.

特斯拉的 Dojo 在运行某些特定模型时，性能远超 NVIDIA 芯片，因为它们就是为此优化的。

Dojo runs certain Tesla's Dojo would run certain types of models way better than NVIDIA chips, because they're optimized to that.

但问题是，就连视觉任务现在也普遍用视觉变换器了。

But then it's like, oh, well, actually, even in vision tasks, he's vision transformers now.

所以嘛，好吧，挺有意思。

So it's like, okay, cool.

因为模型规模越来越大，还有各种其他因素。

Because model sizes grew, and all these things.

所以这就成了一个进退两难的境地，你为了某个目标去优化。

So it ends up being a catch 22 in that you optimize for something.

所以现在，当今时代出现了一批新的AI加速器公司，它们说，好吧，我们要针对Transformer进行优化。

And so now, today, you have this new age of AI accelerator companies that are like, okay, we're gonna optimize for transformers.

但在它们开始设计的时候，情况是这样的，好吧，Transformer是这种体量的稠密模型。

But the time they started designing, they're like, okay, transformers are dense models that are this big.

大家知道，隐藏层维度是八千，批次大小和序列长度也都很大，所以干脆搞个超大的脉动阵列，好实现最高效率。结果发现，看看 DeepSeek，或者去查查那些实验室都在做什么。

What's the best know, the hidden dimension is eight k, and your batch sizes are this big, and your sequence lengths are this big, so let's just make a super large systolic array so you can, you know, create the maximum efficiency, and that turns out, oh, look at DeepSeek or, you know, go look at what the labs are doing.

实际上，他们的形状要小得多。

Actually, their their shapes are much smaller.

实际上，你需要做很多小矩阵乘法，而不是每层都搞那种巨大的、单一的矩阵乘法。

Actually, you need to do a bunch of small matrix multiplies, not massive, massive, massive, you know, singular matrix multiplies per layer.

结果你会发现，你为那个场景设计的芯片，其实效率并不高。

And then it ends up, you know, oh, well, that chip you're designing for that is actually not super effective for

是的

Yeah.

就是这样。

That.

软件一直在不断演变，因为要适应英伟达硬件上表现最好的方式，你看DeepSeek、阿里巴巴，还有各大实验室内部的做法，都是这样。

And And so the software is evolving constantly because of what because of what works best on NVIDIA, and you see that with, you know, whether it be what DeepSeek's doing or Alibaba's doing or what the labs are doing internally.

就连谷歌也不例外。

And you even see this, like, for Google.

对吧？

Right?

他们的开源Gemma模型做出的不同选择，就是因为TPU的结构和GPU不一样。

Like, their open source Gemma models make different decisions because the shapes of a TPU are different than a GPU.

其实GPU和TPU的差距并没有那么大。

And those the GPU and the TPU are actually not that far apart.

对吧？

Right?

比如说，它们确实很不一样，但Blackwell和TPU其实正在朝相似的设计方向靠拢。

Like, would say, yes, they're very different, but, like, Blackwell and TPUs are are very, very they're converging on similar designs actually.

要想超越NVIDIA，光靠供应链优势是不够的。

Whereas, like, to beat NVIDIA, you can't just have the supply chain, you know, win.

对吧？

Right?

你没有那种锁定的客户群体。

You don't have this captive customer.

所以你现在得做出点东西，比如在特定工作负载上把硬件效率提升五倍，然后祈祷这个负载别突然变了。

So now you need to do something, you know, that will give you five x advantage, right, in hardware efficiency for a certain type of workload, and then pray the workload doesn't shift.

对吧？

Right?

因为NVIDIA也在不断优化他们的架构迭代。

Because NVIDIA is also optimizing their architecture generation.

他们加了很多功能，让芯片对现有模型的支持好了很多。

They've added a lot of stuff to make their model their chips way better for the existing models.

但他们每年或每两年都在稳步前进，而你却得跑到很远的偏门里去赌，希望模型别跑偏。

But it's like they're taking, you know, large steps every year, every two years towards something, whereas you have to, like, go way over there in left field and hope that models stay over there.

对吧？

Right?

因为你必须比他们强五倍，毕竟NVIDIA在供应链效率上压着你。

Because you have to win by five x because NVIDIA is gonna have supply chain efficiency over you.

他们在新制程、新内存，或者任何新技术的上市时间上，都能比你快一步。

They're gonna have time to market over you in terms of, like, a new process node or new memory or, you know, whatever whatever technology.

对吧？

Right?

就连AMD也做到了。

Even AMD.

对吧？

Right?

他们比NVIDIA更早推出了2纳米工艺。

They got to two nanometer before NVIDIA.

他们的HBM密度更高。

They had higher density HBM.

他们用了三维堆叠技术。

They used three d stacking.

在供应链的这些方面，他们本该比英伟达更强，可还是输了。

All these things on supply chain that should be better than NVIDIA, and yet they still lose.

还有软件这一块。

There's still the software angle.

对吧？

Right?

英伟达的软件确实厉害。

NVIDIA is fantastic.

没错。

Yeah.

软件方面也是个关键因素。

And and then there's software as well.

对吧？

Right?

但NVIDIA的网络性能还是会比你强。

But it's like NVIDIA is gonna have better networking than you.

他们的HBM也会更优。

They're gonna have better HBM.

他们的制程节点也会更先进。

They're gonna have better process node.

他们上市速度会更快。

They're gonna come to market faster.

他们产能爬坡也会更快。

They're be able to ramp faster.

无论是在台积电、SK海力士这些芯片和内存供应商，还是机架、铜缆这些环节，他们的议价能力都会更强，成本效率也更高。

They're gonna have better negotiations with whether it's TSMC or SK Hynix and the memory and silicon side or all the rack people or, like, copper cables, everything, they're gonna have better cost efficiency.

所以你得比他们强五倍才行。

So you have to be, like, 5x better.

但说真的，如果真有哪个竞争对手能拿出性价比稍微好一点的产品，我猜很多大型GPU用户都会立刻把一部分订单转过去，就为了有个第二选择。

But but to be fair, if if somebody had a viable competitor, which would even be marginally cost competitive, my guess is many of the big consumers of GPUs would immediately shift some revenue there just to have a number two.

对吧？

Right?

就只是为了

Just to just to

等一下。

Hold on.

这不就是今天AMD的情况吗？

That's that's AMD today.

对吧？

Right?

微软已经退出了。

And Microsoft stopped.

没错。

Yeah.

我是说，那个

I mean, like, the the

不过，目前他们的市场接受度还是很有限。

There's still pretty limited traction, though.

对吧？

Right?

没错。

Sure.

但Meta还在持续采购，微软之前也买过一批，后来就停了，因为虽然AMD确实提供了不少优势，但在能效比上还是不如NVIDIA，而且他们的软件团队规模大得多。

But Meta continues to buy from them, and Microsoft did buy a bunch, and then they stopped because it's like, well, yes, they're you know, AMD is giving you all these advantages, but ends up still not being better on a performance per watt basis, and they have a way bigger software team.

在我说的那些方面，他们确实有一定程度的竞争力。

They're somewhat competitive on, like, all these dynamics that I mentioned.

对吧？

Right?

所以你不能简单地照搬NVIDIA的做法。

So you can't just, like, do the same thing as NVIDIA.

你得真的做得更好，或者像AMD那样执行得更到位，对吧？

You really and and do it better, right, or try and execute better like AMD.

你必须在别的方面实现真正的突破。

Like, you have to really leap forward in some other way.

但设计周期太长了，等你做出来，市场趋势早就变了。

But that's the the design cycle takes so long that models will shift.

对吧？

Right?

因为他们会想，下一代TPU和GPU会是什么样子？

Because they're like, oh, what's the next generation TPU and GPU look like?

明白了。

Okay.

那就朝着那个方向优化吧。

Let's optimize for that.

至于研究路径，确实挺不错的。

And the research path is, you know, like, great.

没错，类脑计算可能是我们最该走的方向，但没人干，因为你得在自己选的这条技术路线上往前推进。

Like, yes, neuromorphic computing could be the most optimal thing for us to do, but no one's working on that because you have to advance in the tech tree you've chosen.

对吧？

Right?

要是你重开一条技术路线，那结果肯定惨不忍睹。

If you restart the tech tree, you're gonna be like, well, this sucks.

所以，一旦技术路线分岔，而你偏偏选了这边，那就完了，因为你必须做到

And so, like, if it branches this way and you're over here, you're screwed because you have to be

五倍

five x

差不多得

kind of

有个模式。

There's a mode.

因为供应链的问题，所谓的五倍实际只能做到两倍半，而如果你真有竞争力，英伟达还能稍微压缩一下利润空间。

Because the supply chain stuff means that five x actually turns into a two and a half x, and then NVIDIA can compress their margin a little bit if you're actually competitive.

然后这2.5倍的提升，实际上就变成了50%的改进。

And then that two and a half x becomes like a 50% better.

接着呢，

And then yeah.

所以到最后，事情变得太复杂了，尤其是软件方面。

So it's like it ends up being way too difficult to and the software stuff.

对吧？

Right?

所有这些因素都会把你原本5倍的提升，拉回到只有50%的改进。

Everything, like, takes your five x and makes it like, oh, you're actually only 50% better.

国防供应链肯定也是。

And defense supply chain for sure.

没错。

Yeah.

国防供应链。

Defense supply chain.

然后他们就明白了。

And then, like, they get they get that.

对吧？

Right?

卢特尼克自己也说过，我们必须为稀土矿物这么做。

Like so it's like and and Lutnick himself said we had to do the this for rare earth minerals.

我觉得这挺有意思的。

I was like, interesting.

中国有些省份有规定，说H20能效不够，不能部署，这简直太奇怪了，因为这明明是中国最好的AI芯片。

China there's, like, provinces in China that have, like, rules that say the h 20 is not efficient enough to be deployed, which is, like, super bizarre because it's clearly the best AI chip China has.

华为还是稍微落后一点。

Huawei is still a little bit behind.

有意思的是，效率在中国根本没那么重要，因为他们电力基础设施足够强大，完全撑得起来。

Well, what's interesting is that, you know, efficiency is just not is is so much less of an issue in China than here because they just have the power infrastructure to be able to support.

就算用性能稍弱的芯片，也没太大关系，毕竟中国的电力供应几乎是无限的，他们根本不在乎。

So even if they're running less powerful chips, you would imagine that it doesn't really matter because China has just such an infinite supply supply of power that they'd sort of be okay with it.

这挺有意思的。

So it's interesting.

这在美國是個大問題。

Which is it's a big challenge in America.

对吧？

Right?

詹森老说H20在美国白送都没人要，可我现在真听说有公司说，是啊，我真不想要，因为我电就这么多。

There have been companies that were like, wouldn't Jensen keeps saying he couldn't give away H20 in America for free, but I've literally heard companies now say, Yeah, no, I I wouldn't because I only have this much power.

接下来一年，我那些准备就绪的数据中心该怎么安排？

How am I going you know, in data centers ready to go over the next year?

我要是买了H20，算下来算力反而更少，那我不是亏了？

If I bought an h 20, I'd literally have less compute capacity, and then I'd lose.

对吧？

Right?

就算白送也不行。

Even if it was free.

这根本说不通。

Like, it doesn't make sense.

而中国根本不在乎。

Whereas China doesn't care.

他们能建得出来。

They can build these things.

他们有这个实力。

They have the muscle.

我很想知道最后会怎样。

I'm I'm curious how this all shakes out.

你知道吗，中国现在态度特别强硬。

You know, China's posturing really hard.

他们甚至放出话说，要调查H20有没有后门。

They even, like, put out something that was like, we're investigating to see if there's backdoors in the h 20.

根本就不存在什么H20的后门。

It's like, there's no backdoor in the h 20.

别紧张。

Like, chill.

你知道，GPU 本来就不直接连着公网，本来就有多重防护。

You you know, it's like, you know, GPU GPUs are usually, like, firewalled from the public Internet anyways.

在接触到 GPU 集群之前，早就过好几道关了。

Like, you you step through stuff before you get to the GPU clusters.

所以，说什么后门，根本没意义。

So, like, a back wall backdoor wouldn't even matter.

我不太确定。

I don't know.

我觉得挺值得观望的，因为中国一旦下定决心，立刻就能把海量资源砸到 AI 上。

I think I think it'll be it'll be interesting to see because China can definitely deploy way, way, way more power to AI the moment they decide to.

但这里面牵扯着各种不同的利益诉求。

But there's these, like, there's, like, competing interests.

对吧？

Right?

就是

Like

因为他们希望华为能超越英伟达。

Because they want Huawei to be better than NVIDIA.

对。

Yeah.

这就是英伟达向政府游说的理由。

And then this is how NVIDIA argued to the administration.

他们说，如果我们不这么做，其实这是一个非常有力的论点，比如在Triton这个常用的机器学习库里。

They're like, if we don't do this actually, I think it's like a very, like, powerful argument that, like for example, within Triton, which is a common ML library.

总之，字节跳动已经开源了一些能接入这个系统的很棒工具，还有其他很多类似的库。

Anyway, ByteDance has open sourced some stuff that plugs into this that is super awesome, and there's all these other libraries.

中国开源的不只是模型。

It's not just models that China open sources.

还有中国公司为英伟达开发并开源的软件。

It's software for NVIDIA that Chinese companies open source.

从某种角度说，英伟达卖GPU本身就是在论证：他们成功阻止了华为构建自己的软件生态，而西方的生态确实更强大。

In a sense, like, by NVIDIA selling GPUs is NVIDIA's argument again, like, was like, they were able to, you know, stop Huawei from building up a software ecosystem, and the Western ecosystem is better.

但反过来想，如果你认为模型给社会带来的经济价值比硬件更大——我个人就这么觉得——那现在的问题其实是价值分配不均。如果你真这么想，那给中国卖降配版的H20，甚至像特朗普说的那样，给个缩水版的Blackwell，其实比直接卖芯片给中国带来的好处要大得多。

But then on the flip side, it's like, again, if you believe the models deliver more economic value to society than the hardware, which which I actually think they do, it's just there's a value capture problem today, then you're giving China way more by giving them h twenties and soon a version of Blackwell that's cut down, like like Trump said, right, versus versus, you know, selling them the chips.

对吧？

Right?

卖芯片带来的经济收益，远不如通过提供AI服务来获取的收益大。

The economic value derived from selling them the chips is not as large as, you know, being able to somehow sell them AI services.

那中国是在掌控AI的主导权吗？

So is is China gatekeeping power for AI?

我不这么认为。

I don't think so.

我觉得，即便现在H20已经卖到中国，未来还有H20E和其他芯片，我们还是能看到像阿里这样的中国公司，宁愿到海外租GPU，因为海外的GPU在性能性价比上实在强太多了。

I think I think, again, there's a lot of what we see is that even with h 20 being sold to China into China and and and future versions of the chip, the h 20 e and other chips, we still see, like, Chinese companies like Alibaba renting GPUs outside of China because the GPUs they can get outside of China are just so much better on a dollar spend per performance basis.

他们要么直接租，要么通过像新加坡的公司——其实本质上还是中资背景——建数据中心，把芯片装进去。

Renting them or even going through sort of like a Singaporean company that is effectively a Chinese company and building data centers and putting chips in them.

所以我觉得中国并不是在主动限制算力。

So it's like I don't think China's limiting the power per se.

关键是，你知道，明年中国企业的资本支出增长率会远超美国企业。

It's that it's you know, you can only if you can spend like, Chinese companies are growing their CapEx way more than US companies on a percentage basis next year.

从绝对金额来看，美国公司目前在AI上的投入还是更多。

The absolute dollar number is you know, obviously, The US companies are spending more still on AI.

但按增长率算，明年中国企业增长更快，可问题在于，这些芯片性能较差，每美元投入换来的AI输出——不管是token还是别的——都会更低。

The percentage basis, Chinese companies are growing more next year, and you you still have the problem of, well, dollars spent to AI output in tokens or in whatever is gonna be lower because these chips are worse.

所以算力并不是关键瓶颈。

So power is not the gating factor.

真正卡脖子的一直是资金。

It's always capital.

对吧？

Right?

至少现在是这样。

At least today.

对吧？

Right?

现在中国要是想多花钱，完全能花得起。

Now China can spend a lot more capital if they wanted to.

他们通过国企、非营利性资本支出等方式，每年补贴半导体行业高达一千五到两千亿美元。

They're subsidizing the semiconductor industry to the tune of, like, 150, $200,000,000,000 a year through SOEs, through CapEx that's not generating revenue, etcetera.

所以他们不是不能把同样力度的钱投到AI生态里，毕竟Meta一年的资本支出才六百亿，谷歌才八百亿。

So it's not like they couldn't do this to the AI ecosystem, right, given, you know, Meta's CapEx is, like, 60,000,000,000.

对吧？

Right?

谷歌的资本支出大概八百亿。

And Google's CapEx is, like, 80,000,000,000.

对吧？

Right?

他们完全有能力在单一项目上花更多钱，只是还没下定决心而已。

Like, they could totally spend way more than that on a single effort, they just haven't decided to.

我觉得美国这边的建设主要受限于电力供应。

And I just think for The US, our build outs are constrained by power.

谷歌有大量TPU芯片闲置着，就等数据中心通电上线；Meta的GPU也是一样。

Google has a ton of TPUs sitting, waiting for data centers to be powered and ready, as does Meta with GPUs.

我们之前提到过，Meta现在正在搭这种临时性的棚屋式数据中心。

We posted about how Meta is now building these effectively tents.

这在多大程度上也跟他们不愿把算力开放给更广泛的生态有关呢？

Is this to some degree also coupled to their unwillingness to sell them to a broader ecosystem?

他们只想把资源锁在自己的数据中心里，如果没及时扩大自己的超大规模数据中心建设，那确实会限制自己。

I mean, they want to be confined in their own data centers, and they didn't ramp data center build out for their own hyper scale use cases quickly enough, right, then yes, that constrains them.

如果他们面向公开市场，我们还会受这种限制吗？

They were on the open market, will we still be constrained?

是的，毫无疑问。

Yeah, yeah, for sure.

比如CoreWeave这样的公司，为什么它会有价值呢？

Because, like, companies like CoreWeave you know, why is CoreWeave valuable?

主要是因为他们建设基础设施的速度特别快。

It's really because they build infrastructure really fast.

对吧？

Right?

没错。

Yeah.

他们的软件确实不错，但很多客户用的都是裸金属服务器。

And and their software is nice, I think, but, like, a lot of their customers are bare metal.

对吧？

Right?

坏了就换GPU，网络也得跟上。

Replace the GPUs whenever they're broken and network They they

他们扩张得更激进，我觉得不是这样。

grew more aggressively, and I I think And no.

别走远了。

Don't go anywhere.

詹森也很喜欢他们

Jensen likes them as well,

公平地说。

to be fair.

是的。

Yeah.

他们会的。

They'll go yeah.

是的。

Yeah.

这也很重要，但他们这么做是因为这样能让生态更分散，对英伟达更有利。

That's very important as well, but they'll, like, go because it's it it it it unconcentrates the ecosystem, which is better for NVIDIA.

对吧？

Right?

我们以前在英特尔工作过吗？

Have we worked at Intel?

我完全明白Chip当时在想什么。

I know exactly what Chip was going through his mind.

是的。

Yeah.

我觉得真正重要的是，CoreWeave根本不在乎，对吧？

I think what's really important is that like, Core Weave doesn't care, right?

他们就想，哦，加密货币数据中心？我把它转成AI数据中心就行，对吧？

They're like, Oh, crypto data center, I will convert it to AI data center, right?

他们花了一百亿美元收购了一家做加密挖矿的公司，而那家公司几年前估值才二十亿美元——不是因为他们的比特币挖矿业务在增长，而是因为他们拥有电力数据中心。

They bought a company for like $10,000,000,000 that's doing crypto mining, which is worth $2,000,000,000 a couple years ago, and it's not because their Bitcoin mining business is growing, it's because they have power data centers.

无论哪里，大家都在建电力数据中心，像CoreWeave和甲骨文这样的公司都在往这个方向走；对了，今天谷歌不是也收购了加密挖矿公司TeralWright的8%股份吗？

Anywhere and everywhere, people are trying to build power data centers, and companies like CoreWeave and Oracle are moving to the actually, today, Google just didn't bought 8% of a crypto mining company called TeralWright?

并不是因为他们想进军加密挖矿。

Not because they're getting into crypto mining.

当然不是。

No.

因为他们需要数据。

Because they need the data.

想要电力。

Want the power.

他们需要电力。

They need the power.

对吧？

Right?

所有超大规模云服务商都说了，别管什么可持续发展承诺了，他们急需电力。

And it's like, all the hyperscalers have, like, said, screw off to my sustainability pledges because they need power as fast as possible.

是的。

Yeah.

对吧？

Right?

他们确实做了一些事，虽然改变方向没那么快，但即使不在自己建的数据中心里做，公开市场上也还有很多难题。

They're, you know, they're they're doing things that are not that they take a little bit longer to move the ship, but even if you didn't do it in your own self built data centers, there's still a lot of challenges in the open market.

现在存在缺口。

There's a deficit.

这严重限制了美国芯片的部署进度。

And that's that's constraining American chip build outs heavily.

没错。

Yes.

其他人可能能快一点，比如Core Weave之类的公司。

Others could maybe do it a little bit faster like Core Weave or or others.

对吧？

Right?

甲骨文的态度也挺开放，但还是严重拖慢了美国这边的建设速度。

Oracle's got an open mind as well, but it's still constraining US build outs heavily.

就算资金已经投了，芯片也占了整个集群成本的六到八成，具体看用的是哪种芯片。

Even though the capital's been spent, the chips are 60 to 80% of the cost of the cluster depending on what chips you're getting.

所以他们其实已经把芯片买好了。

So it's like they've already bought the chips.

他们根本没法把芯片装进去，因为数据中心还没准备好。

They just can't put them anywhere because the data centers aren't ready.

这问题不仅影响谷歌，微软、Meta，还有好多公司都一样。

Supplies to Google, applies to Microsoft, applies to Meta, applies to a lot of folks.

说实话，在美国建电力基础设施真的特别难。

I mean, it's really hard to build power infrastructure in The US.

所以呢。

So

电力供应、电网接入、输电线路、变电站，所有这些环节都卡住了。

Power, grid interconnections, transmission, substations, all of this stuff.

在德克萨斯，如果你愿意当流动电工，收入简直跟石油行业一样高。

Electrical contractors, electricians in Texas, if you're willing to be a travel electrician, it's like oil pay.

以前身体好点的人去西德克萨斯能挣一百美元，可谁愿意干这活儿啊？

It used to be that if you're physically adept, you could go make $100 in West Texas, but who the fuck wants to do that?

现在不一样了，你离达拉斯两百英里，找个还算像样的小镇，建个数据中心，干干布线、搞搞输电，收入直接翻倍，跟几年前完全没法比。

Now it's like, well, you could go 200 miles away from Dallas in what's still a reasonable town and build a data center and work on the wiring within the data center and all this other stuff, the transmission stuff, and your pay is up like 2x now versus what it was just a few years ago.

劳动力问题也是一个挑战。

This labor problem's a challenge too.

我觉得中国那边根本没这些麻烦，只是还没把钱花出去而已。

I think in China, they don't have any of these problems, but they just haven't spent the capital yet.

但资金也是个问题，毕竟投入的规模太大了。

But capital is an issue as well because of the scale of what's being spent.

对吧？

Right?

比如英伟达今年的收入就要超过两千亿美元，明年预计还要超过三千亿，再加上谷歌打算花五百亿美元建TPU数据中心。

Like like, NVIDIA's revenue this year is gonna be, like, over $200,000,000,000, and next year expects over 300,000,000,000, plus Google's gonna spend, like, $50,000,000,000 on TPU data centers.

对吧？

Right?

亚马逊也要在Tranium数据中心上砸大笔钱。

And it's like, and Amazon's gonna spend tons and tons on Tranium data centers.

这笔钱的规模正在迅速攀升到国家层面。

It's like the scale of dollars is is quickly growing to nation state level stuff.

更重要的是，得能决定把钱花在哪儿，哪些才最划算。

And and what's more important is being able to decide to spend the dollars and what's cost effective.

某种程度上，中国还是受这限制，但他们能偷偷把芯片带进来。

And so to some extent, China's still constrained by that, but they can smuggle chips in.

他们可以在国外建数据中心。

They can build data centers outside of China.

他们也可以在国外租用数据中心，用上最划算的Blackwell芯片之类的。

They can rent data centers outside of China and be more and have the most cost effective, you know, Blackwell chips or whatever.

对吧？

Right?

字节跳动之所以成为谷歌云最大的客户，或者第二大客户，是有原因的。

ByteDance is, you know, either the biggest or the second biggest customer of Google Cloud for a reason.

对吧？

Right?

他们从谷歌那边拿到了大量Blackwell芯片。

And they're getting, you know, over you know, they're getting many, many black well from them.

对吧？

Right?

甲骨文、微软，还有其他这些公司，也都把大量芯片租给中国，因为这么干比自己建更划算。

And the same with Oracle and the same with Microsoft and all these other companies are renting tons of chips to China anyways because it's more cost effective to do that than build it yourself.

所以中国并不是所有企业都抱定‘必须自建’的想法——政府可能这么想，但像阿里、腾讯、字节这样的基础设施公司并不会这么干。

So it's not like China has this mentality where we only have to well, the government does, but the infrastructure companies don't, like Alibaba, Tencent, ByteDance, etcetera.

那数据中心的最终方向会是什么？

So what's the endgame for data centers?

我的意思是，我们确实需要更多电力。

I mean, like, we we need more power.

我们还需要更强的散热能力。

We need more cooling.

以后所有的数据中心都会建在核反应堆旁边，或者大量用太阳能，又或者靠近深海冷水来散热？

Will the m p every all data systems will be next to a nuclear reactor or lots of solar, you know, next to deep level, like deep sea water that we use for cooling or something like that?

你觉得会怎样？

What's the think

数据中心的散热其实没那么夸张，大家总说它耗电厉害，但事实并非如此。

that cooling is like, the physical cooling of a data center are like you know, there's this whole narrative about, oh, It uses so much power, it's like, not really.

种苜蓿用的水，比AI数据中心多出一百倍。

You know, farming alfalfa uses, like, 100x the water of of AI data centers.

就算到十年底，情况也不会变，而苜蓿本身价值很低。

Even by the end of the decade, it'll be the same, and it's, alfalfa's, like, worth very little.

至于散热，有人试过把数据中心建在海底来省冷却成本，但其实……

So it's, anyway there's, like, it's, like, cooling is, like, not that you know, people have, like, experimented with, like, undersea data centers to reduce the cooling cost, but it's

这根本说不通。

That doesn't make sense.

省下的成本也就10%，可代价却大得多。

It's like 10% savings, but then It's

把数据中心放进海里容易，可怎么把海水排出去呢？

hard to get the water out of the ocean then then put their data center into the ocean.

要是想维护，那可就麻烦了。

It's like if you wanna service it, like, yeah, you're screwed.

对吧？

Right?

电力也是同样的道理。

So the same with power.

我们总说电力其实没那么贵。

It's like we talk a lot about, like, the power is not actually that expensive.

只是建造起来太难了。

It's just hard to build.

对吧？

Right?

我得

I have to

找到合适的地方。

get to the right place.

还得找到合适的场地，把电压降到芯片需要的水平，还要处理各种配套问题。

And I have get to the right space and convert it down to the voltages and and all the stuff that that chips need.

所以关键不在于电力的总量，而在于它在哪里、怎么流动。

So it's less the magnitude of power and more where it is how it moves.

但总量也很重要。

Well, the magnitude too.

对吧？

Right?

就像说，它将会是

Like, it's gonna be

从全球能源消耗总量来看，AI数据中心所占的比例仍然微不足道。

In terms of total worldwide energy consumption, AI data centers is still the full

简直可以忽略不计。

It's like like nothing.

是的。

Yeah.

只占不到百分之一。

It's a fraction of a percent.

它只是

It's

不是那样。

not Yeah.

就算到本十年末，美国的数据中心耗电量也只会占我们总用电量的10%左右，但这仍然

Even by the end of the decade, you know, The US will be, like, 10% of our Power BI data centers, which is still

只是用电量的一部分。

like Of electricity.

我们总用电量的一部分。

Of our electricity.

从能源总量来看，这个比例还要更小。

And in terms of energy, that's even a smaller fraction.

对吧？

Right?

是的。

Oh, yeah.

是的。

Yeah.

因为你觉得

Because you think

转向电动汽车也能带来更大的改变，比我们建的所有AI数据中心加起来影响还大。

about Shifting to electric vehicles also, you can probably make a bigger swing than than, you know, with all the AI data centers we can build in

第一季度。

the first quarter.

比如在欧洲，这个数字增长得并不快，其他很多国家确实需要建设更多电力设施，但也没多到离谱。

Outside, like, it's like in Europe, like, that number's not moving up that fast, and all these other countries I think we need to build a lot more power, but it's not some crazy, crazy amount.

只是做好这件事很难，确实。

It's just doing it properly is the hard Yeah.

我同意。

I agree.

再说电力成本，你去看看那些企业签的合同，还在不断签下新单。

And again, the cost of power, you go look at these deals people are signing, they're still signing.

尽管这些大规模采购的电价从每千瓦时几美分飙升到了10美分，但你想想整个TCL集群的总成本——GPU本身、网络设备这些，加起来远比电费高得多。

Even though the price has skyrocketed from a few cents a kilowatt hour for these massive, massive purchases to 10, It's still you know, when you think about the full TCL cluster, you know, the GPU cost and networking, all of this stuff far outstrips the power.

是的。

Yeah.

制冷也是同样的情况。

And same with cooling.

如果我算一个四年摊销的GPU数据中心，电力成本大概占多少比例？

But what what percentage is power from like, if you if I do four year amortized GPU data center, what percentage will be power?

如果你在建Blackwell这样的数据中心，80%的成本都花在硬件采购上。

80% of the cost of a GPU data center if you're building Blackwell is capital.

对。

Yeah.

没错吧？

Right?

主要就是GPU的购置费用。

It's the GPU purchases.