Dylan Patel: GPT-5, NVIDIA, Intel, Meta, Apple

所以是三万亿美元。

So it's $3,000,000,000,000.

美元，没错。

Dollars Yeah.

对。

Yeah.

没错。

Right.

所以我们正在打造的技术，大概能为全球GDP增加三万亿美元的价值。

So so we we're probably building technology here, which adds $3,000,000,000,000 of GDP value.

理论上，这些钱都可以投到GPU上，因为那是主要成本，只是——

In theory, we could put that into GPUs because that's the main cost Just

从代码模型的角度来看，

from a coding model,

但也不能忽略其他所有应用场景。

not ignoring every other use case.

所以至少从理论上讲，这种价值创造还会继续增长。

So at least in theory, the value generation is here to keep growing.

对吧？

Right?

不过，这怎么转化成行业的实际影响，那就复杂多了。

Now how that translates to the industry is much more complicated.

我觉得我们已经看到AI创造的价值了，就是那个著名的三千亿、两千亿的问题，现在都成了六千亿了。

I think we've already seen AI's value creation ex so so so so there's, like, the whole, like, like, the famous, like, oh, 300,000,000,000 problem or 200,000,000,000 problem now.

是六千亿的问题。

It's 600,000,000,000 problem.

我敢肯定。

I'm sure.

用不了多久，Quiz肯定又要抛出个一万亿美元的问题。

So quiz gonna put out like a $1,000,000,000,000 problem right soon enough.

但说真的，这背后确实有现实依据。

But like, like, there is some like reality in that, of course.

但你知道吗，现在基础设施的投入相当于五年收入的总和，可不是一年的。

But you know, ignores that like infrastructure spend today is accounting for five years of revenue, not like one.

而且收入也不是平平无奇的，是往上冲的。

And the revenue looks like this not like flatline.

但我觉得最关键的是，AI创造的价值早就超过投入了。

But I think think the main thing is that AI is already generating more value than the spend.

问题在于，价值根本没被抓住。

It's that the value capture is broken.

对吧？

Right?

我真觉得，OpenAI 凭借聊天工具这一项，到现在连自己创造价值的十分之一都没捞到。

Like, I legitimately believe OpenAI is not even capturing 10% of the value they've created in the world already, just by usage of chat.

Anthropic、Cursor，还有其他那些公司，情况也差不多。

And I think the same applies to, know, Anthropic and Cursor and whoever else you're looking at.

我觉得，价值捕获这事儿，真的彻底脱节了。

Think the value capture is really broken.

就连内部来看，我们四个开发人员搞出来的自动化系统，用在Gemini API上的成本低得离谱。

Even like internally, I think like what we've been able to do with like four devs with in terms of like automation, our our spend on like Gemini API is absurdly low.

但我们却用AI处理了每一个数据中心的所有许可证和监管申报文件。

And yet we go through every single permit and regulatory filing around every single data center with AI.

我们还会拍下每个数据中心的卫星图，标记数据集，识别出对方用的发电机、冷却塔、施工进度和变电站等等。

And it's like and we we take satellite photos of every data center, and we, like, we're able to label our data set and then recognize what generators people are using, what like, cooling towers and the construction progress and substation.

这些全都是自动完成的，要不是生成式AI，根本做不到。

All this stuff is, like, automated, and it's only possible because of Gen AI.

我们用极少的开发人员就能干成这事，靠卖这些数据、提供咨询赚的钱多得惊人，可钱全让别人赚走了。

And we do it with very few developers, and then the value capture that I'm able to generate by selling this data, by consulting with it is so high, but the company is making it.

他们自己一点好处都没捞到。

They get nothing out of it.

这背后的价值分配问题，远远超过了技术创造本身。

Think there was a value capture challenge here that far out exceeds the sort of creation.

对吧？

Right?

而且随着像GPT-5这样的模型，或者开源模型不断进步，成本一降再降，这些公司想赚到钱就越来越难了——他们做推理的毛利，顶多也就50%，很多情况甚至更低。

And and and as you get models like GPT-five or open source models, like 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, you know, 50% gross margin on inference if they're you know, or or less in many cases.

说白了，你意思是

In so many words, you're

咱们正在被商品化，所以赚不到钱，那你就别指望在GPU上花大钱了。

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，不是这意思。

Well, no.

我觉得吧，还是有办法能大幅提高价值捕获的。

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

对吧？

Right?

我之前提过，广告就是个巨大的价值增长点。

Like I mentioned, the ads are a huge value capture inflection.

在大规模增长之前，得先发生点什么。

To happen before before we see massive increase.

不是。

No.

我觉得另一点是，其实还有很多资金没花出去。

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

对吧？

Right?

比如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.

还有像Core Weave和Oracle这样的公司，因为能从资本市场融资，资本支出增幅远不止20%到30%。

In addition, companies like Core Weave and Oracle, because they're tapping capital markets can raise way more than 20 to 30% CapEx.

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。

And then you go down the list further, and it's like, oh, the largest infrastructure funds in the world like Brookfield and Blackstone.

实际上，它们现在全都把目光转向了AI基础设施，打算投更多钱。

Well, actually, they're they're turning all of their eyes to investing even more into infrastructure, AI infra.

还有全球的主权财富基金，像G42、挪威的那家，或者新加坡的GIC。

And then you're like the sovereign wealth funds of the world, the g forty twos or, you know, the Norway one or GIC in Singapore.

这些机构才刚刚开始布局AI领域，投入的资金还只是九牛一毛。

Like, these people have barely started touching AI.

所以我觉得还有大量资本开支会陆续进场，而且这些投入不一定从第一天开始就要追求经济效益。

And so I think there's a whole lot more CapEx that can come without it being necessarily, like, economically motivated day one.

如果只看能产生实际收益的理性资本开支，增长空间其实有限，但还有太多非理性的投入会进来——你对着财务表格算来算去，按正常的业务逻辑根本不该投这么多钱。

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.

但大家还是会砸钱，因为他们看好这个赛道，坚信AI的前景。

But people will because they believe.

反正我是这么觉得的。

I believe.

而且你应该也认可这点。

I think you believe.

就说基建这块，大家都觉得投进去能赚到钱，但谁也没法拍胸脯打包票，说这个结论绝对站得住脚。

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.

没错。

Yeah.

受到定制芯片的威胁大吗？

How's threatened,

如果有的话，NVIDIA 到底被定制芯片威胁到什么程度？

if at all, is NVIDIA by by custom silicon?

我觉得这才是关键，你看谷歌、亚马逊，还有 Meta 的订单，尤其是他们三家。

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

他们的定制芯片可不是微软那种。

Their custom silicon is not not Microsoft.

他们那几款定制芯片，其实挺拉胯的。

Their custom silicon kinda sucks.

但另外三家，过去一年里订单量猛增。

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

亚马逊都在量产数百万颗 Trainium 芯片了。

You know, Amazon is making millions of Trainium.

谷歌也在造数百万颗 TPU。

Google's making millions of TPUs.

TPU的利用率明显已经到100%了。

TPUs clearly are, like, 100% utilized.

对吧？

Right?

是的。

Yeah.

虽然现在还没到高端水平，但我觉得亚马逊和Anthropic迟早能搞定。

Premium's 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，其实是大家越来越会用定制芯片了。

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.

如果AI技术越来越集中，那定制芯片就会更有优势。

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的芯片团队那些事呢。

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

对吧？

Right?

要是AI技术真的被少数几家垄断了，那定制芯片就更有优势了。

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

但如果因为中国出了很多开源模型，NVIDIA和中国那边又搞了大量开源工具，让部署成本低到不行，那情况可能就不一样了——

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, you know, NVIDIA and China, and it makes the deployment costs, like, rock bottom, then potentially Hear

你听我说完。

me out here.

如果谷歌的TPU真能在公开市场上跟英伟达一较高下，理论上是完全做得到的。

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

现在谷歌的市值还不如英伟达呢。

NVIDIA is worth more than Google these days.

那谷歌为啥不干脆把芯片卖给所有人？

Shouldn't Google start selling their chips to everyone?

理论上，他们这么做，市值说不定还能更高呢。

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

我完全同意。

I I, I absolutely think so.

我觉得谷歌内部其实已经在讨论这事了。

I think Google's even discussing it, internally.

这得彻底重组文化，还得重新调整云服务、TPU团队，还有Jack和XLA软件团队的运作方式。

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

我完全觉得他们做得到。

I totally think they could.

就是得狠狠地自我革新一把，才有可能实现。

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

对。

Yeah.

但我真觉得谷歌该把TPU卖出去，不光是出租，而是直接卖硬件。

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更大，尤其是考虑到搜索业务的下滑。

It's it's kinda funny if a side hobby, in theory, has a higher company value potential as you move

尤其是当你想到搜索业务在走下坡路的时候。

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

所以我是说做生意。

So I business.

我的意思是

I mean

是的。

Yeah.

我觉得，如果你去问谢尔盖的话。

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

对吧？

Right?

比如说，嘿。

Like, hey.

你觉得卖芯片和服务器机架，跟卖云服务或者Gemini，哪个更有价值？

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

他肯定会说，不是。

He'd be like, no.

不是。

No.

不是。

No.

不是。

No.

我的意思是，Gemini 将来会值钱得多得多。

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

只是现在还没到那一步。

It's just not yet today.

对吧？

Right?

所以我觉得，就现在来说，英伟达还是最牛的，这完全是聚焦的问题。

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

对吧？

Right?

如果客户市场太集中，那NVIDIA就不会是全球最有价值的公司。

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

对吧？

Right?

但如果分布越来越分散——说实话，我们现在已经开始看到，开源模型越来越强，部署也越来越容易——那NVIDIA很可能会长时间保持全球最有价值公司的地位。

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 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.

从历史上看，软件早就吞噬了大部分市场，这可不是开玩笑。

Historically, no pun intended, software has eaten the world in most markets.

对吧？

Right?

你想想早期网络时代，思科曾经是全球最有价值的公司，对吧？

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

但现在不是了。

It's no longer.

对吧？

Right?

他们才是搞上层服务的，比如谷歌、亚马逊，还有脸书，最后都超越了他们。

They're they're the guys that build services on top, like Google or Amazon or or Meta eventually eclipsed.

所以NVIDIA才在拼命搞这些软件库啊。

Well, which is why NVIDIA is, like, making all these software libraries.

对吧？

Right?

他们想把推理服务变成标准化产品。

Like, that's that's and they're trying to commoditize inference.

对吧？

Right?

我觉得你们连推理API这块都没怎么投吧？

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

我们跟各种模型提供商都有合作。

Well, we have we have all kinds of model providers.

模型提供商。

Model providers.

我说的是纯粹的API服务提供商投资，我觉得。

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

对吧？

Right?

我说得没错吧？

Is that correct?

我记得我跟他们团队里的人聊过，可能是谁吧，问过为啥你们没投像Together或者Fireworks这样的公司。

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, together or, like, a fireworks.

他们的说法是，光做模型部署服务，不自己搞模型研发，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?

我们在Stable Diffusion生态里倒是投了一些。

We have some in the stable diffusion ecosystem.

比如那个FOUL。

Like, with, like, the FOUL.

对。

Yeah.

慢慢就变成这样了。

Decay to Yeah.

那里的运作模式有点不一样，我觉得。

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

他们通常会做出更复杂的模型，不像搞语言模型的那些人。

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

我...

I

是的。

think Yeah.

对。

Yeah.

不一样。

Different.

但你们根本没搞过这种基础十型或者类似的API投入吧？你们觉得这东西是基建团队搞的，迟早会被软件和视频技术给标准化了——毕竟像VLM和SD-LAN这种，本来就是伯克利出来的开源项目，现在都自成体系了，还有好多公司在支持。

But but, like, you guys don't have one of these, like, you 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's making, because VLM and SD LAN, which is like open source software coming out of Berkeley and now, you know, sort of has their own environments now and supported by many.

所以你的意思是，一旦这玩意儿被标准化了，API提供商就没啥价值了，对吧？

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

我觉得这跟整个问题都相关，就是为啥？

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

对吧？

Right?

为啥会有人……

Like, why would

这么做呢？

you do this?

换个话题。

Shifting gears.

那硅基初创公司呢？

What what about the the silicon startups?

你对这些公司怎么看？

What's what's your take on those?

我的意思是，现在有大量资金涌入这个领域。

I mean, there there's there's a ton of capital flowing in into that.

对吧？

Right?

我们虽然没具体数字，但估计有几十亿美元投进了芯片初创公司。

We've seen I have not numbers, but probably billions, being invested in in in chip startups.

是的。

Yeah.

当然。

For sure.

当然。

For sure.

我的意思是，你看像Etched、Rivos，还有Maddox这些公司，连芯片都没发布呢，就融到了这么多钱，这真的挺厉害的。

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

以前呢，硅芯片公司要想融资，至少得先把芯片做出来，才能拿到大笔投资。

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.

但现在Etched和Rivos连公开的芯片都没推出，就融了这么多钱，这说明什么呢？没错，做芯片，尤其是加速器这种复杂玩意儿，确实烧钱，零件多、环节多，成本高得吓人。

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来家做AI加速器的初创公司，都是最近几年冒出来的。

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

我觉得还不止呢。

I think it's a lot more.

是啊。

Yeah.

确实挺多的。

That that are like yeah.

对。

Yeah.

是的。

Yeah.

这话说得在理。

That's fair.

而且那些老牌公司还在继续融资。

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

是的。

Yeah.

对吧？

Right?

比如Grok、Cerebras这些公司，

Like, Grok and Cerebras and

对。

Mhmm.

而且

And

SemiNova、Tensor，还有这一类的公司。

SemiNova and and Tensor and so on and so forth.

对吧？

Right?

比如Graphcore被软银收购了，软银也在往这上面砸钱。

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，但真要做起来挺难的。

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

对吧？

Right?

问题是，你怎么才能赢过英伟达？

It's like, how do you beat NVIDIA?

对吧？

Right?

像这些超大规模云服务商，我觉得他们挺幸运的，因为他们能做大部分和英伟达一样的事情。

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?

他们自己的客户就是自己，完全被自己掌控。

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

对吧？

Right?

是的。

Yeah.

这可是个巨大的优势。

It's a huge asset.

而且他们可以

And it's they can

他们可以

they can

就是靠供应链取胜。

just win on supply chain.

对吧？

Right?

我用的是一些更便宜的供应商。

Like, I'm using cheaper providers.

说白了就是个压缩利润的游戏。

Margin compression exercise, essentially.

对吧？

Right?

没错。

Yeah.

是啊。

Yeah.

而且对于某些特定的工作负载，比如Meta的推荐系统，他们能做得更专、更到位。

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

但大部分情况下，其实不行。

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

我们瞄准的还是同样的工作负载。

We're we're targeting the same workloads.

我们只要简化供应链，或者把大部分东西自研了，压缩点利润，也就够了。

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

但像那些其他公司，他们没有自家的客户群体。

But in the case of, you know, these 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：我用的也是同一套生态，要么找家定制芯片供应商，他们照样要抽一成利润，最后我卖的价格就被压得更低；要么就干脆所有东西自己搞。

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.

但问题是，这太难了。

But then it's like, this is really hard.

对吧？

Right?

我得从头做起，自己写所有软件。

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

我得自己搞定所有芯片设计。

I'm gonna do all the silicon design.

我得从头开发一堆不同的知识产权模块。

I'm gonna build all this different IP.

从芯片到机架，所有供应链我都得自己管。

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

对吧？

Right?

最后搞这么大阵仗，团队得扩得多大。

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

归根结底，像NVIDIA这种能拿到75%的毛利。

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

AMD的GPU毛利才50%，他们工程实力那么强，还是干不过NVIDIA。

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

对吧？

Right?

他们虽然也是做同样的东西，但用的硅片面积更大、内存更多，性能却跟我们一样，价格还得压得更低，利润空间就被挤得没剩多少了。

Like, 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.

但你听我说。

But, hey, look.

说实话，回顾历史，新进入者一般都不是靠一点点改进现有产品就能赢的。

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

偶尔也有这种情况，但更多时候，他们是靠某种颠覆性的技术突破杀出来的。

That happens sometimes, but but more likely, they they jumped on some of to on some kind of disruptive technology leap.

对吧？

Right?

就是说，我们走的是另一条路。

Where it's like, we have a different approach.

我们用的是不一样的技术。

We have different technology.

这种情况有可能吗？

Is that possible here?

我的意思是，也许这么说有点简化了，但我觉得Transformer模型胜出的部分原因，是因为它在GPU上跑得实在太溜了。

I mean, to to to some degree, maybe 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上跑得一塌糊涂。

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GPU友好的模型，现在就很难再搞出一个真正……

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

这其实是软硬件协同设计的结果。

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?

理论上讲，这玩意儿简直太牛了，效率也超高。

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

就是说，嗯，挺好的。

It's like, okay.

不错。

Great.

根本就没有配套的硬件生态。

Like, there's no ecosystem of hardware.

软件生态也完全没有。

There's no ecosystem of software.

要证明这玩意儿值不值得搞，光是硬件、软件、模型这三个方面，就得让现在最顶尖的上万人专门盯着干才行。

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、Rebis、SambaNova这些公司，全都是冲着当时最火的模型去设计芯片的。

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

所以他们做了一些权衡。

And so they made certain trade offs.

对吧？

Right?

他们在芯片上塞了更多内存。

They put a lot more memory on chip.

而英伟达呢，他们觉得——

And NVIDIA and NVIDIA was like, well,

我们可不这么干。

we're not gonna do that.

要快，最要紧的是快。

Faster, least.

对吧？

Right?

其实吧，你要是比比英伟达芯片上的SRAM内存容量，那可低多了。

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

是的。

Yes.

没错。

Correct.

用SRAM代替DRAM。

Than SRAM instead of DRAM.

对。

But Right.

所以他们的DRAM通常更少，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更少了。

There's less DRAM.

SRAM更多了。

There's more SRAM.

因为芯片上SRAM多了，计算能力就得减下来，结果他们就输了。

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?

因为模型越来越大，各种问题都来了。

Because the model sizes got too big and all this.

对吧？

Right?

所以就出现了这种特别奇怪的局面——他们押注的其实是更优的方案。

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

对吧？

Right?

我敢说，Cerebras在跑某些类型的模型时，肯定比NVIDIA或者Grok、Dojo强。

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

对吧？

Right?

特斯拉的Dojo在运行某些特定模型时，比英伟达的芯片强多了，因为它是专门优化过的。

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

但后来发现，就连视觉任务现在也普遍用视觉Transformer了。

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

所以嘛，也就这样了。

So it's like, okay.

挺好的。

Cool.

因为模型越来越大，各种因素都变了。

Because model sizes grew and all these things.

结果就陷入了一个两难境地：你为某样东西做了优化，结果它却不再主流了。

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

所以现在，新一代AI加速器公司都开始说：行吧。

And so now, like, today, you have this new age of AI accelerator companies that are like, okay.

我们就专攻Transformer了。

We're gonna optimize for transformers.

但他们刚开始设计的时候，心想：好吧。

But the time they started designing, they're like, okay.

Transformer 是那种规模很大的密集模型。

Transformers are dense models that are this big.

隐藏层维度是八千，批量大小和序列长度也都这么大，那干脆做个超大的脉动阵列，最大化效率好了。

What's the best you know, the hidden dimension is eight k, and your batch sizes are this big, and your sequence lengths are this big.

结果你去看看 DeepSeek，或者看看那些实验室在干嘛，

So let's just make a super large systolic array so you can, you know, create the maximum efficiency, and it turns out, oh, look at deep seq or, you know, go look at what the labs are doing.

其实他们的模型结构小得多。

Actually, their their shapes are much smaller.

真正需要的是做一堆小矩阵乘法，而不是每层都搞一个超级巨大的单次矩阵运算。

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

这么一来，你专门为那种需求设计的芯片，其实并不太管用。

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

真的。

That.

所以软件一直在变，因为什么在NVIDIA上效果最好，就跟着怎么改。

And so the software is evolving constantly because of what because of what works best on NVIDIA.

你看深求、阿里，还有各大实验室内部，都是这样。

And you see that with, you know, whether it be what deep seeks doing or Alibaba's doing, or what the labs are doing internally.

连谷歌也不例外。

And you even see this, like, for Google.

对吧？

Right?

他们的开源Gemma模型做的选择就不一样，因为TPU的结构跟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也没差那么远。

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

对吧？

Right?

你说它们差别很大？其实不然，Blackwell和TPU的设计正越来越接近。

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

但你要知道，作为英伟达，光有供应链可不行。

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

对吧？

Right?

你没有那种锁定的客户。

You don't have this captive customer.

所以你得搞出点东西，让硬件在特定任务上效率提升五倍，然后祈祷这任务别突然变了。

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?

因为英伟达本身也在不断优化他们的架构迭代。

Because NVIDIA is also optimizing their architecture generation.

他们加了不少功能，让芯片跑现有的模型更高效。

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才迈一大步，而你得直接冲到偏门那边去赌，指望模型别跑偏。

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?

因为你得比他们强五倍，毕竟NVIDIA在供应链上比你有优势。

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速度上，都能抢在你前面。

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都做到了。

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密度也更高。

They had higher density HBM.

他们用了三维堆叠技术。

They used three d stacking.

在供应链这些方面，他们本该比英伟达强，可还是输给了对方。

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

但问题还在——

There's still

是软件层面。

the software angle.

对吧？

Right?

英伟达的软件确实厉害。

NVIDIA is fantastic.

没错。

Yeah.

软件方面也是关键。

And and and there's software as well.

对吧？

Right?

但NVIDIA的网络性能肯定比你强。

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

他们的HBM也会更先进。

They're gonna have better, HBM.

他们的制程节点也会更领先。

They're gonna have better process node.

他们上市速度会更快。

They're gonna come to market faster.

他们产能爬坡也更快。

They're be able to ramp faster.

不管是台积电、SK海力士这些芯片和内存供应商，还是机架厂商、铜缆供应商，他们谈判能力都更强。

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.

他们的成本效率也会更高。

They're gonna have better cost efficiency.

所以你得比它强五倍才行。

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

但说真的，如果真有哪个竞争对手能站得住脚，哪怕只是成本上稍微有点优势，我猜很多大公司都会把一部分订单转过去，就为了有个第二选择。

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

对吧？

Right?

就为了表明自己……

Just to just to show that to

这不就是AMD现在的情况吗。

to That's AMD today.

对吧？

Right?

微软已经退出了。

And Microsoft stopped.

没错。

Yeah.

我的意思是，比如说这个

I mean, like, the

不过，目前他们的市场接受度还是很有限。

the There's still pretty limited traction, though.

对吧？

Right?

没错。

Sure.

但确实如此。

But Yeah.

Meta还在继续买他们的产品，微软当初也买了一大批，但后来停了——毕竟，AMD确实给了你这么多优势，可最终在能效比上还是没超过他们，而且他们的软件团队规模大得多。

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，他们确实有那么点竞争力。

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

对吧？

Right?

所以你不能照搬英伟达的做法。

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

你得真真正正做得比他们更好，或者像AMD那样执行得更到位。

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

你必须在别的方面有真正的突破。

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

但问题是，产品开发周期太长了，等你做出来，市场风向早就变了。

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.

那我们就往那个方向优化吧。

Let's optimize for that.

研究方向嘛，说实话挺棒的。

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

没错，类脑计算可能是我们最该走的路，但没人干，因为你得在自己选的这条技术路上往前冲。

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?

你要真把技术路线重来一遍，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，你偏偏选了这边，那就完蛋了，因为你得做到五倍才行。

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

这中间有护城河。

There's a moat.

因为供应链的问题，本来五倍的差距，实际可能就剩两倍半；要是你真有竞争力，英伟达还能压一压利润，结果两倍半又缩水成只强50%。

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

然后呢，就是这样。

And then yeah.

所以到最后，事情变得难到不行，尤其是软件这块。

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

对吧？

Right?

本来能有五倍的提升，结果一折腾，就只剩一半的改进了。

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.

是的。

Yeah.

国防供应链。

Defense supply chain.

然后他们就能拿到这些优势。

And then, like, they get they get that.

对吧？

Right?

比如卢廷克自己也说过，我们得为稀土矿物这么做。

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

我当时就觉得，这挺有意思的。

I was like, interesting.

中国有些省份居然有规定，说H20能效不够，不能部署——这简直太离谱了，毕竟这可是中国目前最强的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.

华为现在还稍微落后一点。

Huawei is still a little bit behind.

有意思的是，在中国，能效根本不是大问题，因为他们电力基础设施太强了，撑得住。

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

就算用效率低一点的芯片，也没关系，毕竟中国电力多到用不完，他们根本不在乎这点损耗。

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 infinite supply of power that, you know, they'd sort of be okay with it.

所以这事儿挺耐人寻味的。

So it's interesting.

这在美国可就成大难题了。

Which is it's a big challenge in America.

对吧？

Right?

有些公司根本就没人要H20，连Jensen都说在美国白送都没人要。

Like, there have been there have been companies that were like, they wouldn't they've like, you know, Jensen keeps saying he couldn't give away h 20 in America for free.

但我真听过一些公司说：‘不是我不想拿，是我电不够用啊。’

But I've literally, like, heard companies like, say, like, now say, like, yeah, no, I mean, I wouldn't because, like, I only have this much power.

接下来这一年，我的数据中心电都快不够用了，怎么办？

How am I going to, 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，那我不是亏大了？

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

对吧？

Right?

就算白送也没用。

Even if it was free.

这根本说不通。

Like, it doesn't make sense.

但中国根本不在乎。

Whereas China doesn't care.

他们能造得出来。

They can build these things.

他们有这个实力。

They have the muscle.

我很好奇最后会怎么样。

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

你知道吗，中国这阵子可真能造势。

You know, China's posturing really hard.

他们甚至又搞出点别的名堂。

They even, like, put out something else.

说要查查H20里有没有后门。

Like, we're investigating to see if there's backdoors in the H 20.

根本就没什么后门。

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

别紧张。

Like, chill.

你知道，GPU 本来就不直接连公网，本来就有一层层防护。

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

想访问 GPU 集群，前面得过好几道关。

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

所以什么后门不后门的，根本没意义。

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

我也不确定。

I don't know.

我觉得挺有意思的是，中国一旦下定决心，分分钟能把算力堆到天上去。

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.

但问题是，里头还有各种利益在拉扯。

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

对吧？

Right?

就是说

Like

因为他们希望华为能超越英伟达。

Because they want Huawei to be better than NVIDIA.

对啊。

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，其实有个非常有力的论点，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, like, for example, within Triton, which is a common ML library.

总之，字节跳动开源了一些能接入这个系统的工具，特别厉害。

Anyway, like like, ByteDance has open source some stuff that plugs into this that is, like, super awesome.

还有好多类似的库。

And there's, like, all these other libraries.

其实就是中国开源的那些模型。

It's just models that China open sources.

其实就是中国公司开源的、给NVIDIA用的软件。

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

从某种角度说，NVIDIA的论点是，他们卖GPU能阻止华为建立起自己的软件生态，而西方的生态更强大。

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，如果你觉得模型给社会带来的经济价值比硬件更大——我确实这么认为——那现在的问题其实是价值捕获不到位。你要是真这么想，那还不如直接卖给他们降配版的H20，甚至未来降配版的Blackwell，就像特朗普说的那样，这比直接卖芯片给中国带来的好处大多了。

But then in 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，根本没法跟卖AI服务相比。

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?

我不这么觉得。

I don't think so.

我觉得吧，就算H20和未来的H20E这些芯片卖到中国，我们还是看到中国企业比如阿里，宁愿到中国境外租GPU，因为国外的GPU性价比实在太高了——不管是直接租，还是通过像新加坡那种本质上是中国背景的公司建数据中心、装芯片，都比在境内用强。

I think I think again, like there's a lot of like we what we see is that like even with h 20 being sold to China in 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, renting them or even, like, going through sort of, like, a Singaporean company that is effectively a Chinese company and building data centers and and putting chips in them.

所以我觉得中国本身并没有在限制能力。

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

关键是，你看，明年中国公司的资本支出增长率会远超美国公司。

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

绝对金额上，美国公司现在还是花得更多，这点没得说。

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

但按比例算，明年中国公司增长更快，可问题还是在于，花同样的钱，产出的AI算力——不管是token还是别的——就是少，因为芯片性能差。

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

所以真正卡脖子的不是算力。

So power is not the gating factor.

一直都是资金。

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15.2万亿美元。

They're subsidizing the semiconductor industry to the tune of, like, $152,100,000,000,000 dollars a year through SOEs, through CapEx that's not generating revenue, etcetera.

所以他们要是想把这笔钱砸进AI领域，也不是做不到，毕竟Meta一年才花600亿美元，Google才800亿。

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?

Google的资本支出一年大概800亿美元。

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得比这多得多。

Like, they could totally spend way more than that on a single effort.

他们就是没下这个决心。

They just haven't decided to.

我觉得美国这边，扩建受限于电力供应。

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

对吧？

Right?

谷歌有大量TPU芯片闲着，等数据中心通电上线；Meta的GPU也是一样，堆着没用。

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

对吧？

Right?

我们之前提到过，Meta现在干脆搭起了这种临时帐篷式的机房。

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

这在某种程度上，是不是也跟他们不愿把芯片开放给更广的生态有关？

Isn't this to some degree also coupled to their unwillingness to sell them to a broader ecosystem?

我的意思是，如果他们只想把芯片锁在自己数据中心里，又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，对吧？

I mean, if they want to be confined in their own data centers and they're, you know, didn't ramp data center build out for their own hyperscaler use cases quickly enough, right, then yes, that constrains them, right?

如果他们能进入公开市场，我们还会受限吗？

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

对，没错，

Yeah, yeah,

肯定的。

for sure.

像CoreWeave这样的公司，为什么值钱？就是因为它们建基础设施特别快。

Because, like, companies like CoreWeave you know, why is CoreWeave valuable is really because they build infrastructure really fast.

对吧？

Right?

是啊。

Yeah.

他们的软件确实不错，但我感觉很多客户其实用的是裸金属服务。

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

对吧？

Right?

就就是

Just just

坏了就换GPU，网络也一样，他们

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

增长得更激进了，我觉得，但也不是。

grew more aggressively, and I I think And no.

别走远了。

Don't go anywhere.

詹森也挺喜欢他们的，

Jensen likes them as well,

说真的。

to be fair.

是的。

Yeah.

他们会的，没错。

They'll go yeah.

是的。

Yeah.

这一点也很重要。

That's very important as well.

但它们会走，因为这会分散生态系统的注意力，对英伟达反而更有利。

But they'll, like, go because it's it it it it unconcentrates ecosystem, which is better for NVIDIA.

对吧？

Right?

你以前在英特尔工作过吗？

Have you worked at Intel?

我太清楚他心里在想什么了。

I know exactly what's what's going through his mind.

所以我觉得真正重要的是，Core Wave根本不在乎。

So so I think what's really important is that, like, Core Wave doesn't care.

对吧？

Right?

他们就说：哦，这个加密货币数据中心，我转成AI数据中心就行了。

They're like, oh, crypto data center, I will convert it to AI data center.

对吧？

Right?

几年前他们花了一千亿美金收购了一家做加密挖矿的公司，那家公司实际就值两百亿左右。

They bought a company for, like, $10,000,000,000 that's doing crypto mining, which is worth, like, $2,000,000,000, like, a couple years ago.

可不是因为他们的比特币挖矿业务在增长。

And it's not because their Bitcoin mining business is growing.

纯粹是因为他们手上有数据中心的电力资源。

It's because they have power data centers.

对吧？

Right?

现在到处都在建电力型数据中心，像CoreWeave和Oracle这样的公司都在往这方向跑，对了，今天谷歌不是刚收购了家叫TeralWright的加密挖矿公司8%的股份吗？

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

可不是因为他们想进军加密挖矿。

Not because they're getting into crypto mining.

不是。

No.

因为他们需要的是数据中心。

Because they need the data center.

对吧？

Right?

他们

They

需要的是电力。

need the power.

对吧？

Right?

所有超大规模云服务商都撂下环保承诺了，只求赶紧搞到电。

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?

你知道吗，他们确实有些做法，推进起来没那么快，但就算你不在自己建的数据中心里搞，公开市场上也有一堆难题。

They're, you know, they're they're doing things that are not that they're they take a little bit longer to move the ship, but, like, 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.

供不应求。

There's a deficit.

对吧？

Right?

这严重拖慢了美国芯片的产能扩张。

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

没错。

Yes.

其他人可能快一点，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?

甲骨文那边想法也挺开放的，但就算钱已经花出去了，美国的建设还是被卡得很死。

Oracle's got an open mind as well, But it's still constraining US build outs heavily even even though the capital's been spent.

对吧？

Right?

芯片呢，占整个集群成本的六到八成，看你用什么芯片了。

The chips are, you know, 60 to 80% of the cost of the cluster depending on what chips you're getting.

所以他们其实芯片早就买好了。

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

就是没地儿放，数据中心还没建好。

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

这事儿对谷歌、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，在美国建基础设施、电力设施真的太难了。

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

电力、电网接入、输电线路、变电站，这些全都是问题。

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