第018集 - 别再声称2026年美国数据中心一半的容量被取消（数据中心、能源）| Jeremie Eliahou Ontiveros, Reyk Knuhtsen, Ellie Holbrook, Jordan Nanos

你好，

Hello,

大家好。

everyone.

欢迎回到《SemiAnalysis每周简报》。

Welcome back to SemiAnalysis Weekly.

这周我和数据中心、能源与工业团队一起带来分享。

I'm here this week with the data center, energy and industrials team.

他们是杰里米·埃利亚和雷克。

That's Jeremie Eliah and Reyk.

我们会聊聊他们最近发表的几篇文章。

We're going to talk about a couple of articles they put out recently.

第一篇是：别再说2026年美国数据中心容量有一半被取消了。

The first is stop saying half of 20 '20 six US data center capacity is canceled.

这标题真不错。

Nice title.

我们先来澄清一下这方面的一些问题。

We'll clarify some stuff there.

第二个标题也很直白。

And the second one, pretty descriptive.

是的。

Yeah.

第二个话题是关于机房内部的电力生成，到2028年将达到40吉瓦。

And the second one is about, behind the meter data center, power generation, 40 gigawatts by 2028.

欢迎两位来到节目。

Guys, welcome to the show.

谢谢，对。

Thank you, Yeah.

感谢邀请我们。

Thanks for having us.

这次我们聊的是三大洲的事情。

We're covering three continents on this one.

我们在新加坡有Reyk。

We got Reyk in Singapore.

延迟问题。

The latency.

我们现场试试。

We'll try on scene.

这是SemiAnalysis的内容，老兄。

Sets on SemiAnalysis, bro.

普通的一天。

Normal day.

是的。

Yeah.

好的。

Okay.

那我们就开始深入聊聊吧。

So let's let's dig in.

那篇第一篇文章的标题很棒，结论也一点都没藏着掖着。

The the first article had a great title and obviously the conclusion didn't bury the lead at all.

但我最近在媒体上到处看到，人人都说数据中心的容量有一半被取消了。

But I've been seeing this all over the media where everybody keeps saying that half of data center capacity is canceled.

你们之前分析过，为什么大家会这么讲，也澄清了实际情况到底是什么。

You guys walked through maybe reasons why people are saying this and then clarified some stuff as to like what's actually happening.

你能说说现在真实情况是怎样的，哪些是假的吗？

Can you give a lay of the land as to what's reality, what's fake?

对。

Yeah.

我先说说为什么会这样。

I can start with the why.

我觉得是彭博社先发的，搞了个大标题，说美国一半的数据中心容量都被推迟了。

I think Bloomberg started this, right, with this big headline of like half of US data center capacity is delayed.

之后其他人就跟风，都引用同一个数字，大部分文章也都指向同一个原始报告。

Everyone else started piling in and citing the same number and most of these articles point to the same underlying source, right, which is a report that's available out there.

你看看这些报告，上面说美国2026年计划上线的数据中心容量有12吉瓦，但目前只有5吉瓦在建。

And, you know, when you look at the reports, the report says there was 12 GW of data center capacity scheduled to go online in The US in 2026 and only five is under construction.

对吧？

Right?

我们看到这个的时候就纳闷了，为啥大家都拿这个当回事？

For us, when we saw this, we were like, why is everyone talking about this?

这根本不可能嘛。

Like, it's just not possible.

最搞笑的是，你根本不用搞什么复杂分析，光是看一眼就能戳穿这个说法。

And I think that the funny thing is that like you don't even need to like do anything fancy, you can just disprove this data so easily.

比如亚马逊明明公开说过，他们在2025年建了4吉瓦的数据中心。

Like, I don't know, man, Amazon announced publicly they built 4GW in 2025.

这个项目是会推迟，还是会按计划上线？

Is that going up or not in 2026?

明显是会上线的。

Obviously it's going up.

光是亚马逊一家，可能就要新增五吉瓦以上。

Amazon alone is probably gonna add like, you know, five gigawatt plus.

也就是说，他们觉得一家公司就能占满全部新增容量。

So that's basically all they think is gonna be going live like one company.

当然，你要是把所有超大规模运营商都算上的话。

Know, obviously you add all the hyperscalers.

CoreWeave 今年也会新增一吉瓦的容量。

CoreWeave, you know, is gonna add a gigawatt in 2026.

这些项目全都在建中。

All of that is under construction.

所以你是说，CoreWeave 这一吉瓦就占了整个市场的五分之一？

So you're telling me CoreWeave's one gigawatt under construction is like, you know, a fifth of the market?

这说不通吧。

Like, come on.

问题出在分母上，太离谱了。

There's a massive issue in the denominator.

图片显示错了。

Picture shows wrong.

别发这个。

Just don't publish that.

这是个大致的看法，Reyk，如果你想补充更多细节的话。

So, that's a high level take, Reyk, if you wanna add more deets.

但对我来说，我觉得这才是关键原因。

But for me, I think that's the, that was the trigger.

就是去看一下底层的数据。

It's just looking at the underlying source.

你就会觉得，天啊，你们完全搞错了。

You're like, oh, guys, geez, you're just off.

根本就是错的。

It's just wrong.

对。

Yeah.

我们早就看到这些东西被到处转发，人人都在社交媒体上反复 repost，搞得跟病毒传播似的。

We had seen this getting, like, paraded around quite a bit and, you know, I mean, like everybody would go viral just kind of reposting it.

每次刷推特动态看到这个，我都觉得简直让人头大。

I was like, my God, this is this is mind numbing to see every time on my Twitter timeline.

我们的客户其实都很聪明。

And obviously, like, our clients are pretty smart.

他们肯定会把能拿到的所有数据源拿来互相核对。

Our clients are gonna take every, like, data source they can and compare them against each other.

所以我们还得帮每个客户做这方面的工作。

So we also had to help every client on that front.

后来我们就想，干脆写个简报，直接说：嘿，这东西完全是假的。

And it was like, well, don't we just write a newsletter about this and say, yo, this this is just really fake.

现在我们没必要再反复解释这些了。

We don't have to really go into this as much anymore.

你们就去看看我们的简报吧。

Please just read the newsletter.

很多时候，这些基本事实其实根本不值得细说。

A lot of the kind of baseline here is not worth really talking about sometimes.

对。

Yeah.

所以我们才决定写这份简报。

So that's why we, that's why we ended up writing it.

是的。

Yeah.

而且不仅如此，那篇文章的本意其实也不是故意造假，你能理解，也可能原谅别人把具体数字弄错了。

And in addition to that, like, it's not even like the spirit of the article was fake in a sense, like you can imagine, or you can forgive somebody getting the exact numbers wrong.

但要是我看看你们去年四月到今年五月期间发布的预测图表，发现整整一年过去，预测值居然连5%都没变过。

But if I look at this chart that you guys put out around how the outlook that you guys have forecasted from roughly the last year, April of last year to May of this year, it hasn't even changed by 5%.

我这就把这张图放上屏幕。

I'll put this chart up on screen.

你们能不能简单说说，这些预测背后的实际情况到底是怎样的？其实就是有些项目会延迟，但市场上真正取消的项目少之又少。

So maybe you can talk through a little bit about like what the reality of some of these forecasts are, which is that some stuff gets, delayed, but there's very few cancellations that are actually showing up in the market.

对吧？

Right?

我不同意这个说法。

Well, I would disagree.

我觉得取消的项目数量相当多。

I think there's a pretty substantial amount of cancellations.

我们一直说的是，这些取消的都是早期阶段的项目，对吧？

Just what we keep saying is that these cancellations are early stage projects, right?

我不确定你这边有没有看到这张图，但美国有个地图，标出了大型电力需求，虽然细节略有不同，但想说明的是同一个问题。

I'm not sure if you have this one on the screen, but like there is this map of The US where we show like the large load requests, which is like slightly different, but kind of goes to the same point.

截至今天，仅在美国，就有超过一太瓦的数据中心电力需求被运营商申请了，对吧？

As of today, you have over a terawatt of data center load that is requested by operators in The US alone, right?

这张图是六个月前的，现在需求已经翻了一倍多，总量超过了一太瓦。

This chart is as of six months ago, it's more than doubled now, so we have over a terawatt.

显然，整个美国的

Obviously, the whole U.

S.

S.

目前系统峰值负荷是750吉瓦，你不可能在短短几年内就把这个数字翻倍。

System right now, peak load seven fifty gigawatts, you're not going to double it right now in sort of just a couple of years.

这根本不可能，不现实，所以很明显这是假的。

It's just not possible, it's not a reality, so obviously this is fake.

有很多早期项目，这其实很合理。

There's a lot of early stage projects and it makes sense.

我觉得数据中心市场是在2023年底开始爆发的。

Essentially, I think the data center market started booming towards the end of twenty twenty three.

那时候最早的一批人开始意识到，这方面的资源很快就会吃紧。

That's when early folks started to realize, hey, there's going to be a big constraint on this.

第一个大单就是在那时候签下来的，从那以后，大家疯狂抢电，都在找哪里还有电力可用，这就导致了一种现象：因为人人都在抢，如果你不想掉队，就必须 aggressively 规划多个方案。

The first big deal started to get signed at that moment and since then there's been like this massive search for power and that means that everyone was trying to find where is power available and that leads to some behaviors where because everyone is doing it, then if you want to be successful, you also yourself have to be aggressive and plan multiple options.

这其实没什么新鲜的，超大规模运营商一直以来评估项目时都会准备多个备选方案。

And that is actually nothing new, like Hyperscalers have always had multiple options when evaluating projects.

在资源受限的情况下，这种情况更加突出。有时候，为了做出最终投资决策，云服务商可能会同时推进十几个项目，跟各个县、电力公司等多方接洽。这些项目并非全都可行，更多是在试探市场，看看哪个地方能支持他们建设大型项目。

It's even more critical in terms of constraints, and so you would see sometimes for one final investment decision, CIH perscaviors maybe have 10 different options, right, and so that would be 10 projects where they sort of talk to the local counties, they talk to like the utilities and so on and so forth, and these are not all realistic and they're sort of just testing the field and be like, okay, which county can enable me to build my large scale project?

哪里有足够的人力资源？供应链又在哪里？

Where do I have the workforce or do I have the supply chains?

到底哪里能建得起来？

Where can I build it?

所以，目前早期项目严重过剩。我们之前提到的观点——也是里克那句特别精辟的话——‘云编码项目’，我真的很喜欢这个说法。

So anyways, there's an oversupply of very early stage projects and you know our point and that was sort of Rick's brilliant phrasing, the cloud coded projects is that I just love this.

这太到位了。

This is amazing.

AI在筛选信息方面其实不太行，当有人参与判断时，很明显能看出，有些公告实在太激进了。

It's a kind of AI is, I guess, not very good at filtering, right, so when you have humans in the loop, yes, it's pretty obvious that some of these announcements are way too aggressive.

我们在文章里举了好几个例子，有人宣布要建一个10吉瓦的项目，说第一期500兆瓦明年就能上线，可你点进官网，只有个‘联系我们’，再没别的内容。再查许可证之类的信息，根本什么都没有。

We gave a bunch of examples in the article of folks where they announce a 10 gigawatt project, they say the first tranche of 500 meg is gonna be available next year, and then you click on the website, you see contact us, nothing more, and then you start digging into whatever permits and whatnot, you see nothing.

到最后你就会想，这些人可能在德克萨斯州手里攥着不少土地资源，但离真正落地的项目还差得远。

So at some point you're just like, okay, these people probably have a lot of NAT in Texas or something like that, but they don't yet have a real project, right?

任何人类判断都能过滤掉这些，但现在的AI还是很难做到这一点。如果你想预测这个行业的发展，就必须基于现实的分析和半结构性的评估。我们一贯的做法是深入研究贸易政策，不只是数据中心，还包括芯片，因为这会直接影响英伟达的下游表现，以及AI实验室的上游收入，这些环节都是环环相扣的。

Any human judgment would sort of filter that, but I guess for AI, it's still pretty hard these days, so if you want to build a forecast of the industry, have to base it on realistic forecasts or semi analysis, extensive trade regulation has always been our playbook, not just data centers but also chips because obviously this has downstream implications on NVIDIA, upstream on the AI labs, their revenue, all of that stuff connects to each other.

这就是SemiAnalysis的闭环逻辑，覆盖了每一个相关行业，构建出一个完整的体系——其他人很难做到这一点，也总在找数据、拼接项目产出方面碰壁。

That's the SemiAnalysis flywheel covering every single one of these industries and building a cohesive you know, others don't do that and easily, you know, struggle to find the ability to put the site outputs.

有道理。

Makes sense.

你们在文章里提到了三种不同类型的数据中心延期情况。

So in the article, you guys cover three different types of data center delays.

第一种，是新入行的数据中心开发商做出的过于激进的公告。

One, the aggressive announcement by a newer data center developer.

第二种，是进展较深的项目，但施工时间表过于乐观，可能只是推迟了而已。

Two, an advanced project with overly optimistic construction timelines that might just be a delay or pushed out or something.

第三种，是真正卡在审批和当地反对意见上的问题。

And then the third one is really something facing permitting and local opposition issues.

艾莉，你能来具体讲讲其中一个例子吗？就是我们去查审批文件时发现的那些情况。

So, Ellie, can you come in here and maybe explain one of those examples where we do look in and find something in permits?

好的，没问题。

Yeah, yeah, sure.

有一个项目正面临强烈反对，不只是新墨西哥州本地的抵制，还有来自监管机构的压力，那就是甲骨文公司在新墨西哥州的‘木星项目’。

So the project which is facing quite a lot of pushback locally, not just New Mexico local opposition, but also from jurisdictional regulatory kind of characters there is project, Oracle's Project Jupiter in New Mexico.

他们想建一条管道，为多纳·安娜县的场内数据中心供气，但这条管道还没建，规划的路线也没获批，连预设路线都没有，只有一条备用路线。目前所有能给站点供气的方案都显得不现实。我们每看到一份FERC的申请文件，都发现他们几乎没有进展，甚至完全没有进展，因为这事被归入了一种没有先例的监管流程，按规矩至少得两年才能走完，再加上本地反对声这么大，根本不可能加快。

So yeah, essentially they're trying to construct a pipeline to feed the behind the meter data center in Dona Ana Province and essentially this pipeline hasn't been built and the route that they are planning to, that they want to build on hasn't been approved, they haven't got a pre route, they have got a secondary approved route, and essentially all of their ways to kind of get gas to the site don't seem very feasible, so looking at, there's a pipeline issue which they're trying to make happen, they're seeing every FERC filing that I see that comes through on the docket, they haven't made much progress or any progress at all because essentially it's defaulted to a type of regulatory process which essentially, there is no precedent for it being done sooner than essentially two years and given the local opposition as well, it's very unlikely that it'll be sped up.

其实还有别的办法给站点供气，比如用卡车运输压缩天然气或液化天然气，这一点我们也提到了。

There are other options to getting gas to the site, like you could ship truck CNG and LNG to a facility as we point out.

但问题是规模，目前唯一被证实能大规模运输的压缩天然气，最大也就200兆瓦左右。

However, this scale, you know, getting to, I think the only proven, delivery of, CNG at scale is kind of around 200 megawatts.

200兆瓦已经是上限了，我觉得这个数字可能都偏乐观。

So like max 200 megawatts, I think maybe even that's optimistic.

当地生产这类设备的厂家很少，能提供服务的卡车也不够用。

And there aren't many, aren't many manufacturers, there aren't many producers in the local area and there are not enough trucks to facilitate that.

所以，不仅那条管道根本不存在，连可行的路线都没有，而用压缩天然气或液化天然气运输的方式，难度更大。

So essentially, yeah, there's not only a pipeline which doesn't exist and has no viable route of existing, but getting gas to the site via CNG and LNG is even harder.

根本行不通。

Just not viable.

所以，说实话，这暴露了一个潜在的瓶颈——随着更多‘表后’数据中心的建设，天然气管道基础设施的建设会变得特别困难，尤其是在新墨西哥州这种对这类设施不太友好的地区。

So yeah, no, it just, I think it brought it like it's, it demonstrates kind of an interesting bottleneck, which could occur with more behind the meter facilities, is like building gas pipeline infrastructure, especially in territories like New Mexico, that's quite not extremely, it's not an extremely friendly area towards that, that kind of infrastructure.

我们就是这么审阅这些申请文件的。

But yeah, I would, that, that's how we, we basically look through the filings.

我们已经把所有文件都看完了。

We are through all of the filings.

其实我们早就关注到他们了，因为他们最初想用的涡轮机出了问题，后来改用布卢姆燃料电池。在做这方面的调研时，我发现了这个时间线，当时就觉得这根本说不通。

You know, we looked, we already had it on our radar because they had issues with the initial turbines that they were trying to use at the site, so they ended up switching to Bloom fuel cells and which through kind of that research, I spotted this timeline and I was just like, this just doesn't make sense.

为什么会有这么多本地反对意见？为什么处处都是阻力？

Why are there so many local, all these all this pushback, all these all these local comments?

而FERC和其他监管机构还在反复斟酌这件事。

And then, yeah, FERC just, FERC and the other regulators are still thinking about it.

那这是不是个Claude编码的数据中心项目？

So is this an example of a Claude coded data center project?

所以我觉得这说明，我们必须重新思考如何建设数据中心，才能满足这样的需求。

So I think what this speaks to is rather that there's like, we have to rethink the way we build data centers in order to meet this demand.

但在此之前，别人也没建过一吉瓦规模的数据中心，对吧？

But it's not like other people had built one gigawatt sites before this, right?

我们还处在将首个吉瓦级数据中心推向市场的初期阶段。

Like we're still in the early innings of bringing the first gigawatt scale data centers to market.

所以我们不得不尝试各种不同的方法。

So we had to try a whole lot of different things.

没人有过这方面的经验。

No one has experience doing this.

也许现在有了，但一年前、两年前，真没人干过。

Maybe today, yes, but like as of one year ago, two years ago, no one had experience.

所以自然会尝试各种新办法，出错也在所难免。

So obviously, you try all these new things, there's going to be errors.

正因如此，我们在评估这种大型项目时，总会去思考一个合理的进度安排。

So that's why the way we analyze big projects like this, we always try to think of what is a reasonable timeline.

我觉得这个行业里，很多人会提出更激进的时间表，但这种情况其实普遍存在，不只是数据中心或承包商才这样。

I think it's pretty common in the industry that you have folks that pitch more aggressive timelines, but I think it happens and that's valid for every single industry, it's not data centers, contractors.

任何搞建设的人都会告诉你，他们总是过于乐观，供应商也一样，这根本不是什么新鲜事。尤其是当你在更大的规模上尝试新设计、新地点、新县市的管理时，遇到挑战几乎是必然的。

I think anyone that has a construction, we tell you they tend to be always over optimistic, suppliers are always over optimistic, it's not like this is anything new and especially when you try something much bigger at that scale with new designs and so on and so forth in new locations, new counties, know, new municipalities handling it, then obviously you're gonna face a few challenges.

所以这类项目特别适合那些已经推进得比较靠前的，它们已经拿到了融资，也获得了Oracle和OpenAI的订单——抱歉，是Oracle转向了OpenAI，现在已经有巨额合同在手，只是执行上遇到了困难。

So that's a great fit for a category of projects that are already well advanced because they already got financing, they got it for release from, Oracle, OpenAI, but sorry, Oracle charted to OpenAI, so there's already a massive contract on this, facing execution challenges now.

对。

Yeah.

说得通。

Makes sense.

我们不妨聊聊‘表后’的情况。

Maybe we could talk about, behind the meter now.

你们在第二篇文章里公布了一些相关数据，杰里米，你们图表里那个Y轴的数据挺有意思的。

So you guys in the second article published some, some numbers on this, The first y axis on one of your charts in a little while there, Jeremie, which is pretty cool.

当我们讨论这些项目的规模时，表后新增容量和现有电网容量之间的差距确实相当显著。

So when we talked about the size and scale of some of these projects, I guess the behind the meter net additions when compared to, like, the available grid capacity is a significant difference.

你特别预测，到2028年，数据中心的户内新增装机容量将超过40吉瓦。

And you specifically forecasted over 40 gigawatts of net additions of data center capacity behind the meter by 2028.

这个数字现在几乎可以忽略不计，对吧？

That number is effectively like it rounds to zero right now, right?

目前确实有少数户内项目在推进，但大家才刚开始下单采购涡轮机，供应链也才刚起步。

Like there's a few behind the meter projects, but people are just getting going ordering turbines, the supply chains ramping up.

要实现这个目标，到底需要什么条件？

What does it actually take to get there?

你觉得，这些户内项目会因为艾莉刚才提到的那些原因更容易被取消，还是因为不依赖电网而更不容易被取消？

And do you, do you expect that behind the meter projects are more at risk of cancellation based on some of the reasons Ellie just described or less at risk of cancellation because they don't depend on the grid?

问得真好。

Excellent question, man.

这个问题问得非常好。

Excellent question.

首先，这个预测主要基于我们对美国电网的分析，重点看的是供需缺口。

So, I guess the first thing is, okay, this, this forecast sort of depicts our analysis mostly of The US grid, right, and looking at the gap.

背景是，

Context,

我们在电网上的新增发电量

we're

根本跟不上需求的增长。这个需求每年都是几十吉瓦的量级，而且根据我们看到的各种信号，每年都在持续增加。

not adding enough generation on the grid to meet that demand that is gonna be in the tens of gigawatts per year and just keeps increasing every single year based on all of the signals that we keep seeing.

所以现在，到2028年新增40吉瓦的离网数据中心，这个预期现实吗？

So now is it realistic to assume there's going to be 40 gigawatts of new behind the meter data centers added by 2028?

你得看看现在都在规划些什么。

You have to look at what is being planned right now.

我跟你说，过去两周里，已经有三吉瓦的数据中心项目签了合同，都是做表后计量的，对吧？

You know, I can tell you in the last two weeks, three gigawatts of data center deals were signed for behind the meter purposes, right?

所以交易确实是在推进。

So deals are happening.

我觉得选址工作应该早就完成了。

Site selection, I think happened beforehand.

我觉得在2024年的一部分时间，尤其是2025年，大家开始把能否接入天然气管道作为选址的关键条件之一。

I think you saw a massive move in part of 'twenty four, mostly 'twenty five, where folks really started to have like access to a gas pipeline as one of the main site selection criteria.

你开始看到一些变化，传统数据中心运营商更倾向于选择像弗吉尼亚北部这样的核心市场，因为他们需要稳定的电网供电；而一些新入局的玩家，特别是那些押注AI的公司，反而更早预判了这个趋势。

I think you started seeing sort of diversions where maybe some of the traditional data center operators were more focused on, let's be in a tier one market like Northern Virginia, you know, let's find grid connected sites because we need the Five Nights, and you saw sort of more newer operators, maybe more, you know, AI failed that sort of try to foresee this trend.

Crusoe 就是个很好的例子，他们在这方面一直走在前面，已经签下了不少大单，比如今年第一季度公开宣布的与微软在阿比林达成的672兆瓦的场内供电协议，就是个典型例子。

I think Crusoe is a great example, You know, they've been quite ahead of the curve on this and they've been able to, you know, sign massive deals in, you know, like, Abilene's six seventy two gigawatts with Microsoft announced publicly by Crusoe in Q1 that's behind the meter, for example.

所以总的来说，我觉得供应端正在跟上，因为已经有大量开发商锁定了场地。

So anyways, I think we're getting there in terms of is the supply ready for it because there's like many, many developers that have secured sites.

至于设备制造商方面，艾莉，你可能能给出一份非常详细的分析。

In terms of manufacturers, Ellie, I think you can probably do an extremely detailed rundown.

关于项目延期的可能性，我觉得场内部署的延期风险更低，因为电网约束的机制是这样的：开发商跟电力公司沟通后，对方根本没有任何义务必须遵守他们承诺的时间表。

In terms of probability of delays, I think the probability of delays is lower with behind the meter because of the way these grid constraints play out is that as a developer, you you talk to a utility, they have no binding obligation to abide by the schedule that they provided you.

我们经常听到开发商说，本来以为2027年能拿到500兆瓦的供电，跟电力公司也口头达成了协议，结果几个月后对方突然说：‘抱歉，兄弟，这事办不了了。’

And what keeps happening time and time again is, you know, with the anecdotes we hear from developers say, I thought I was gonna get 500 megawatts by 2027, I had a handshake with the utility or whatever and then, you know, a couple months later they tell me, actually, sorry bro, can't do it.

最后变成2028年只能给100兆瓦，2032年才给500兆瓦，因为他们得先升级整个电网，结果发现当初自己根本没算清楚。

It's gonna be, you know, a 100 by 2028, you're 500 by 2032 because I gotta do bigger network upgrades because, hey, actually I didn't consider my analysis.

另外一个人也想要这个。

Other guy also wants it.

发电能力根本跟不上，诸如此类的问题。

You don't have enough generation coming and so on and so forth.

所以，电力公司逐渐意识到，他们遇到的延误比预想的要多得多。

And so, I think utilities are sort of realizing they're faced with more delays than they thought.

因此，如果你的策略完全依赖电网，那很可能要失望了。

And so in many cases, if your strategy is only grid, you're extremely likely to be disappointed.

相比之下，电力供应方面的问题更突出，而场内发电则带来新的风险——比如能否按时建成电厂，能否顺利拿到许可，这些都成了新的难题。

If anything, that's more so on the power side, behind the meter, I guess, adds a new risk, which is more of the execution layer, like can you build a power plant on time, can you get the permitting as a new set of complexities?

我们肯定会看到一些备受关注的项目延期。

We're gonna see some high profile delays.

根据我们的分析，新墨西哥州的这个项目是最引人注目的，基本上是OpenAI和甲骨文之间一笔价值一千亿美元的交易，堪称最严重的延期案例。

Obviously the New Mexico one, based on our analysis, is the highest of profiles, right, basically a $100,000,000,000 deal for OpenAI with Oracle, so the highest type of delays that you can have.

希望他们能按时解决，祝他们顺利。

Hope they're going to solve it on time, wish them the best.

我觉得这类情况会越来越多，但从绝对规模来看，现在开发者的选择多得惊人，成功案例也会大量出现。

I think you're going to see more of that, but I think in terms of raw volumes, there's like so many options now for developers, you're going see a lot of success stories as well.

Colossus一号和二号就是很好的证明，说明这种事完全可以规模化做起来。

And you know, Colossus one, sorry, Colossus one and two are proof that you can do this at scale.

不过我不确定它们是不是最好的例子，毕竟在审批上出了不少问题，但至少它们证明了这事能成，而且还有其他人也在做。

Now, I don't know if they're the best examples because of the issues they have with regards to permitting, but they're demonstrating that it can be done and there's others doing it.

Crusoe就是另一个例子。

Crusoe is another example.

我刚想补充一点，我们看到设备制造商也在积极应对分布式能源这个趋势。

I was just going to add to that, like we're seeing the OEMs really reacting to the behind the meter meter story as well.

大型老牌厂商比如西门子、Jeeve都在扩大产能，同时整个产业链上也涌现出不少新玩家、新技术，甚至还有旧技术的重新利用——我们之前写过好几份报告提到这些。这其实是个明显的信号，说明市场正在快速演变。至于审批问题，我们也注意到有些公司转用Bloom燃料电池，主要是为了加快供电进度，一方面因为Bloom的生产 footprint 更小，另一方面它的氮氧化物和硫氧化物排放更低。比如新泽西的Nubius项目，之前用的是Bergen涡轮机，后来就换成了Bloom。

So, we're seeing obviously the major OEMs, the established players like Siemens, Jeeve, those guys expanding manufacturing capacity, but then throughout the stack, you're seeing new players, new entrants, new types of technology and then recycled all types of technology, which we've written about in several different notes, but yeah, I think that that's a really like increasing, that's a good tell for like, you know, how this market is developing at speed and also I think on the permitting thing, that's something which could delay, they sense that we've seen, you know, companies switch to bloom fuel cells in order to kind of get speed to, well, possibly speed to power because of manufacturing footprint of Bloom fuel cells but also its lower NOx and SOx emissions, so yeah, so there, yeah, and I think that was the Nubius New Jersey facility, for example, is they switched to Bloom, I think was Bergen turbines prior.

这主要是为了应对审批拖延的问题。

So, yeah, in order to combat the permitting issues there.

所以未来我们可能会看到技术路线的调整，甚至有些项目可能从东海岸转向得州这类地方。

So we might see, you know, shifting around of technology, maybe relocating to more to other states like to Texas rather than in on the East Coast.

不过，没错，目前BTM市场的情况就是全面快速增长。

But, yeah, no, that's a that that that would be kind of the the story going on in the BTM market right now, full full growth.

说得通。

Makes sense.

对。

Yeah.

杰里米，我们能不能回头聊聊你提到的关于项目规模的事？

Jeremie, maybe we could go back to one thing you said related to the like size of some of this.

你刚才说上周签了大约三吉瓦的项目，而我们刚才讨论的这篇文章提到，到2028年，BTM发电总量将达到40吉瓦。

So at this point, when you say that you signed, know about three gigawatts signed in the last week, and we're talking about in the context of the article, 40 gigawatts of behind the meter power generation by 2028.

能不能给听这个播客的普通听众简单讲讲，从签约到开工，再到能为第一批GPU供电，最后到一个吉瓦级大型项目完全建成，整个时间线大概是怎样的？比如，2026年中期签下的这三吉瓦，到2028年真能有项目上线吗？

Can you just high level for the general audience that listens to this podcast, describe a rough timeline for when something gets signed versus when construction starts versus when you can power the first tranche of GPUs on versus when a large site, like a gigawatt scale site might actually be completed is three gigawatts signed in mid twenty twenty six actually going to have anything by 2028?

是到2029年、2030年才会有结果？

Is it a 2029, 2030 story?

大概需要一年半？两年半？三年？四年？还是五年？

Like is this year and a half, two and a half years, three, four, five?

是的，这得看具体签的是什么类型的协议，实际情况差别很大。

Yeah, I would say it tires a lot depending on what signed actually, what actually is signed.

现在市场上有不少协议在谈。

There's a bunch of deals out there.

目前大部分项目都是交钥匙租赁模式，比如数字房地产公司或QTS为微软建数据中心，签的是全包租赁合同，QTS负责一切，微软只管租用。

The bulk of the volumes these days would be turnkey leases where, let's say, a company like Digital Realty, QTS, built a data center for Microsoft, they sign up for lease, turnkey lease where sort of QTS takes on everything and Microsoft just rents.

但也有一些是纯电力供应协议，这种模式下开发商承担的责任更少，租户需要投入更多资本支出。

But you have powerchills, which are a bit different, which is sort of a lower bound on the developer and sort of more CapEx from the tenant.

还有纯粹的购电协议，比如甲骨文和VoltaGrid在德克萨斯州查科县的项目，这纯粹是电力采购，根本不涉及数据中心建设，所以不同协议类型，时间线也完全不同。

You have just power deals, right, where it's basically a PPA, let's say Oracle with VoltaGrid in the Chacophore County, Texas site, essentially a PPA, there's no data center involved, right, so depending on the deals, the timelines can vary.

一般来讲，协议签完大约十八个月能有首批电力容量，但具体能多快建成，还得看施工进度。不过现在普遍预期是，从首期投产到满负荷运行的周期会越来越短。我们那篇文章强调的是，对买家来说，现在分布式发电比电网供电更有吸引力，因为你掌握了主动权。

Rule of thumb would be say, deal signed eighteen months, you have capacity, but you have the first tranche and then depending on how fast you sort of can build it, but I think it's pretty clear that these days the expectation is that the ramp from sort of first capacity, first phase to full ramp is expected to be faster and faster, and the point we make in that article is that from a buyer's perspective, behind the meter is now becoming much more attractive than grid because you're in control of your destiny, right?

比如，你很清楚自己在某个时间点前一定能拿到多少吉瓦的电力——当然，前提是得通过审批等流程，但从电力供应的角度，你至少能确切知道现场什么时候能用上多少电，只要供应商那边别出太大延误。

Like, hey, you know you're gonna have gigawatts by X amount of time, now obviously it has to be permitted and so on and so forth, but at least from a power standpoint, you know exactly what you're gonna have on-site and when, provided that there's delays from the suppliers and so on and so forth.

所以关键在于，你有了这些电力，还得建数据中心，不然电力再充足也是浪费。

So the idea is, hey, if you have all of this power, then you also need to build a data center because otherwise it's useless.

所以大家预期，这三吉瓦的产能建设不会花五年，会快得多，对吧？

So the expectation is that ramping up this three gigawatts is not gonna take five years, it's gonna be much faster than this, right?

通常市场上，一旦有吉瓦级别的交易敲定，比如甲骨文在新墨西哥州或与万通在德克萨斯州的合作，整个项目的融资很快就会跟进，对吧？

And generally, you observe on the marketplace is that when there's a gigawatt scale deal being signed, let's say, you know, Oracle with a stack in New Mexico or with Vantage in Texas, you see that financing for the whole project takes place shortly after, right?

就像我们看到的，数字桥梁在万通的租赁协议签署后不到一个月就完成了融资，金额高达二百五十亿美元。

Like we saw like, you know, Digital Bridge raised money for Vantage, just maybe a month after, after the lease was signed, and we're talking about, you know, $25,000,000,000 of, financing.

所以资金早就已经到位了。

So the whole the money's already secured.

没错。

Right?

只要合同签了，钱一搞定，接下来就是抓紧时间把项目建起来。

Once you have that deal, you secure all the money and then you just try to bill as fast as possible.

对。

Mhmm.

根本用不着五年。

It's not gonna take five years.

那可要久得多。

It's gonna be much longer than that.

是的。

Yeah.

所以不是一年，也不是五年，而是在这两者之间。

So not one, not five, somewhere in between.

有道理。

Makes sense.

根据这个模型。

By the model.

根据这个模型。

By the model.

确实如此。

Indeed.

是的。

Yeah.

所有听众都是根据这个模型来判断的。

Everybody listening by the model.

好的。

Okay.

我想给大家看一张图表，听听你们的看法，我觉得这张图对理解市场份额特别有意思。

I wanna show you guys one chart and and get your reaction here because, I thought this one was fascinating in terms of understanding the market share.

我觉得很多购买这个模型的人，是用你们提供的数据来帮公司做决策的。图例里，AWS是橙色，Google是蓝色，Meta是这种粉紫色，我们就这么叫吧，微软是绿色。

I think a lot of people who buy the model are using the data that you guys provide in order to make decisions for companies on the legend here, AWS in orange, Google in blue, Meta in this fuchsia pink color, let's call it, and then Microsoft in green.

这些就是超大规模云服务商。

So these these are the hyperscalers.

它们在建的兆瓦规模占了市场很大一部分。

They make up a large percentage of the market in terms of how many megawatts they have under construction.

这些公司未来的市场份额，说实话，看着它们互相竞争，真的挺有意思的。

The market share that these guys represent going forward is, I don't know, just really interesting as we see them kind of compete with each other.

所以，Reyk，你怎么看这些超大规模云服务商之间的竞争？还有它们跟市场上其他所有玩家之间的差距？现在大家的规模到底有多大？

So I don't know, Reyk, what's your take on the relationship between the Hyperscalers competing each other with each other and then maybe Hyperscalers versus everyone else in the market in terms of just how big everybody's going right now?

说到彼此竞争的话。

In terms of competing with each other.

我觉得从产能竞赛来看，主要推动力其实是各个AI实验室。

Well, I mean, I think when we take a look at kind of the capacity race, you're going to notice that it's mostly driven by the respective AI labs.

比如AWS的产能扩张，很多时候就是由Anthropic推动的。

So most of the time, like an AWS capacity build out will be kind of fueled by Anthropic, right?

Meta则通常是为MSL提供支持。

Or Meta will typically be for MSL.

微软则主要支持OpenAI。

Microsoft typically open AI as well.

我们知道Anthropic那边也有一些合作，但不管怎样，目前数据中心的建设主要就是由这些公司驱动的，对吧？

We know that there's some Anthropic deals going on there as well, but by the model, in any case, these are going to be, in any case, these are going to be, yeah, like the, the drivers of the data center build out right now, right?

因为只有它们才有足够的资金来建设AI实验室所需的数据中心规模。

Because they're the ones who have the capital able to actually build out the data center capacity that's needed for the AI labs.

它们还具备良好的信用评级，能获得融资，顺利启动并推进这些项目。

They're the ones who have the kind of investment grade that like investment grade financing, able to start these projects and get them going forward.

大多数项目都卡在融资这一关上。

Most projects struggle with, you know, yeah, actually getting the financing.

如果你是家新兴云服务商，而且正在听这段话，也许你已经看过那篇别的文章了。

Like if you're a Neo cloud, which if you're listening to this later on, maybe you already have read the other article.

但作为一家新兴云服务商，想筹到钱去建数据中心、采购GPU、投入这么多资金，真的很难，毕竟总成本可能高达数十亿美元。

But if you're a Neo cloud, it's a bit hard to get the financing to start and go ahead and like get a data center and go ahead and buy the GPUs and expend the capital for all of this like build out, Cause it's going come out to probably billions of dollars.

所以本质上，现在推动AI基础设施建设的，要么是大型云服务商，要么就是资金极其雄厚的新兴云服务商。

So in essence, you're mainly left with Hyperscalers driving the AI build out, or it's going to be very, very well funded or well capitalized neo clouds.

比如那些背后有黑石或凯雷支持的公司，他们从这些投资方那里能拿到充足的资金来推进自己的建设。

Like if we look at any of the kind of Blackstone backed guys, any of the KKR backed guys, they've got plenty of money coming in from them to sort of port their own build out.

不过说起来，我刚才忘了你第二个问题是什么了。

But in any case, I forgot your second question actually.

你能再重复一遍吗？

So if you could say that again.

没问题。

No problem.

是的。

Yeah.

对。

Yeah.

挺有意思的。

Is interesting.

要不我换一个问题吧，听你这么一说，我突然想到了。

Maybe let me ask a different one because I'm thinking of it now listening to you talk.

这篇文章开头提到，有公共研究称，2026年美国一半的数据中心容量都被取消了。

So at the very beginning of this article, were, commenting on the public research that has driven all of these articles in the media about how half of twenty twenty six U S data center capacity is canceled.

但我刚才在屏幕上放的那张图，还有你刚才说的关于超大规模云厂商为AI实验室建设的情况表明，哪怕只算那四家超大规模厂商中的两家，目前在建的容量就超过五吉瓦了，更别说另外两家和其他所有玩家了。

But that chart that I put on screen and roughly what you're saying about the hyperscalers building for the AI labs says that, you know, even if you just take two of those four hyperscalers, we've got more than five gigawatts under construction, And then there's the whole rest of the market, the other two hyperscalers and then everybody else.

所以，你能谈谈吗？为什么做数据中心建设研究的人，会完全漏掉这么多吉瓦的在建容量？

So can you opine on where the gap is in terms of people trying to do research on data center capacity under construction and just missing entire gigawatts worth of capacity.

这事儿明年会变得特别明显。

That's a real thing for next year.

是的。

Yeah.

对。

Yeah.

幸运的是，杰里米之前用三分钟就把整篇文章讲清楚了，所以我能接着他的话展开说说。

Luckily Jeremy was able to explain the whole article in three minutes earlier, so I can kind of expand on what he said.

这种沟通方式实在太高效了。

It's a bit too efficient with this communication.

基本上，像我们这样的模型，做事都是自下而上的，对吧？

Basically when you have a model like ours we do everything kind of in a bottoms up way, right?

比如，每个数据中心对我们来说就是Excel表格里的一行数据。

Where we have like each individual data center is like, for example, a row in an Excel file for us.

没错。

Right.

所以你可以分别预测每栋建筑的用电需求，知道租户是谁，比如OpenAI或者Anthropic这些终端用户，然后就能准确锁定每栋建筑的建设时间表。

And so you can kind of get the forecast for each individual building, the tenant, if the end users and OpenAI are in Anthropic, you already have that information for that building in specific, then you can go ahead and find the timeline for the building.

没错。

Right.

很多人并不具备评估这些时间线的专业能力。

A lot of people don't, a lot of people don't have the expertise to judge timelines on these things.

没错。

Right.

我们在做数据中心研究的时候，我常开玩笑说，我和杰里米大概看过十万到二十万张卫星图像。

When we go ahead and do our data center research, I like to joke that Jeremy and I have probably seen, probably seen over 10 to 20,000 satellite images.

对吧？

Right?

没几个人能懂，那个棕色地面上的一个灰色像素到底意味着什么。

So not a lot of people can say that they understand like what that single gray pixel on a brown floor means.

那其实是混凝土垫层开始施工了。

It's like, oh, that's actually the concrete padding coming in.

这可是个好迹象。

That's going to be a good sign.

希望我们能看到垂直施工的迹象，就是那边角落那个更深的灰色像素，大多数人根本注意不到。

Hopefully we see vertical construction, which is the darker gray pixel in the corner over there that most people wouldn't get.

卫星图对我们来说特别有用，不过这只是其中一个特别有意思的话题。

So the satellite images are super useful for us, but like, that's just one that's really fun to talk about.

其他工作就比较枯燥了，我们要查遍全球每个县、每个城市、每个州、每个国家的许可系统。

The other stuff is a bit more boring, which is where we go through every kind of permit portal for every county, city, state, country in the world.

对吧？

Right?

我们总爱聊我们用了多少Claude。

We like to talk about how much we use Claude.

就像文章里说的，一周就用了十七万美元。

As you see in the article, 170 k in a week.

说到Claude这部分，人人都感兴趣。

The Clod part, everybody loves the Clod part.

总之，我们弄了个挺漂亮的图表，放在页面最前面，写着‘一周花了十七万美元’，对吧？

Anyway, so we got, you know, our lovely chart on the kind of front of the page where it's, hey, a $170,000 spent in a week, right?

我们现在的token使用量已经爆表了，每个员工用的token比Meta还多，不管你怎么叫它，反正我们主要用Claude，还专门搭建了一整套流程和工具来支持我们的智能代理。

We're we're token maxing, you know, we're token mugging meta or whatever you want to call it, more tokens per employee than, you know, yeah, meta is one of the, but in any case, we tend to use Claude and we've actually built out basically an entire harness and workflow for our agents.

我们特别依赖这些代理，它们会自动扫描并抓取全球各地的许可申请平台，从国家、州、市到县一级都不放过。

We love our agents that go ahead and scan and scrape every permit portal basically around the world down to state, municipality and city and county level.

通过这些数据，我们能获取大量公开的申报信息，进而估算出各种项目的进度时间线。

And with this, we get a ton of public filing information that allows us to basically estimate timelines with that.

不过就不多说了，这只是我们方法论中的一个环节，总共还有两个部分。

And without going into too much more detail, that's one part of basically three in our methodology.

至于另外那两个部分，你也可以直接买现成的模型来解决。

And for the other three, you can also buy the model to figure that out.

雷克，我把Claude的使用费用图表放到屏幕上来了。

Reyk, I put the, put the Claude code spend chart on screen here.

为了让大家清楚，这个图表的颜色是按每个用户分开统计的，我觉得是这样。

So we've got, to be clear to describe this, the color coding is a per user breakdown, I believe.

对。

Yes.

正在看的人能清楚地看到底部那个大大的黄色区块。

And people who are looking can see the really big yellow block at the bottom.

对。

Yeah.

没错。

Yeah.

九成把握。

90% sure

是的。

Yeah.

那就是

That's

问题可能出在这里。

culprit may be on this.

culprit 现在正在打电话。

The culprit is on the phone right now.

是的

Yeah.

那是杰里米，正让Anthropics的GPU在底下全速运转呢。

That's Jeremy just abs, just melting Anthropics GPUs at the bottom there.

是的

Yeah.

是的

Yeah.

是的

Yeah.

他正在看的那些

The ones that he's looking at

来自这人的请求

from this guy are

也在处理他的某些任务。

also, cooking up some of his requests there.

是的

Yeah.

对，没错

Yeah, exactly.

他们看到杰里米大笔花钱，马上就把SpaceX的算力全抢光了

They saw, they saw Jeremy spend and they went and bought out the SpaceX capacity right after.

这还得感谢他

So thanks to him.

预测得很准

Good forecast.

我们现在就这么影响市场了

We're moving markets now like that.

你先说

Hey, go ahead.

艾莉，你在想什么呢？

Ellie, what you thinking about over there?

没有，我本来就想开个玩笑，说杰里米的论文把教皇变成了先知。

No, I was just gonna, yeah, make a joke about Jeremie's thesis and making a prophet good of a pope.

有意思，确实，接下来值得看看这些AI实验室到底有多少能源设施会建在用户侧，而不是完全依赖主电网。

Yeah, interesting, yeah, it's gonna be interesting to see how much of these, yeah, how much is gonna be built behind the meter from the main AI labs as well.

而且我觉得，‘用户侧’只是个说法，其实对分布式能源的定义有很多不同版本，比如在本地发电、不全靠电网供电这些方式。

Yeah, and I think also they're kind of very, you know, behind the meter is, that's one term, there are kind of various different kind of definitions of kind of moving energy, the moving kind of energy on-site and not taking all of it from the grid.

你可以把发电设备和数据中心放在一起，也可以采用净计量方案，这样和电网之间是互动的关系。

I mean, you can co locate your energy, you can have a net metering kind of solution, which is more of a kind of interactive relationship with the grid.

你甚至可以为数据中心打造完全离网的独立供电系统。

You can have fully islanded off grid energy supply for your data center.

不过我觉得，未来随着这类设施越来越多，真正值得关注的是哪种模式会被优先选择——是完全离网，还是保留部分电网连接？这可能对未来并网能力形成一种对冲。

But yeah, I think that's gonna be the kind of thing that is gonna be interesting to observe in the future as the more of them are built, seeing what's preferred, I mean, which state, fully islanded, bit of grid connection, maybe that might hedge some interconnection in the future.

另外，这些设备，比如涡轮机，以后如果接入电网，它们本身也可能变成电网的一部分。

Also then there's a, you know, what happens to these assets, these turbines, once, you know, once maybe they do connect to the grid, they might end up being part of the grid themselves.

所以这些涡轮机将来很可能会作为燃气调峰机组，帮助缓解电网本身的供电缺口。

So like a lot of these turbines could be used as gas peakers and help with solving the actual lack of power on the grid itself.

我的意思是，输电问题最终会理清楚，但我觉得实际情况比其他媒体描述的要更相互依存，不是非此即彼的关系，而且我觉得这种非此即彼的状态也不会一直持续下去。

I mean, we'll see this transmission to sort out, but yeah, so I think it's more symbiotic than it's kind of portrayed by a lot of other outlets, I think it's more of, yeah, it's not, I don't think it's one or the other, it will be, well, I don't think it'll be one or the other forever.

有道理。

Makes sense.

对。

Yeah.

我这儿有个图表，是第二篇文章里的，觉得特别棒，正好跟你们说的分布式发电有关，我放屏幕上看看。

Well, let me throw this chart on screen that I thought was really awesome from the second article about behind the meter power generation, kind of based on what you're talking about.

这是今年各厂商的供电能力上限分解图。

This year's a breakdown of the supply ceiling by OEM per year.

我们可以看到，这些负责分布式燃气发电的供应商，他们的产能增长趋势有多明显。

So we can see that increase of some of these suppliers that would be responsible for the power generation in behind the meter gas settings and how much capacity they're increasing in terms of how much they can produce.

我一看这张图，有两个想法，第一就是增长幅度真大。

And I mean, two things jump to mind when I look at this, first of all, is just how much it's increasing.

看起来增长速度正好跟我们未来对电力需求的增长节奏吻合。

So it seems to be going up at roughly the rate at which we're going to increase the power generation requirement.

这是好事。

That's a good thing.

第二点是，现在可供选择的供应商数量真不少。

And then the second thing is just how many different suppliers there are available.

我们之前讨论过市场瓶颈，比如GPU、晶圆、内存这些，一般来说，一旦出现瓶颈，往往就只依赖一两家甚至三四家供应商。

We've looked at stuff that's a constraint in the market, whether this is like GPUs or wafers or memory or, anything else roughly generally speaking, when something's a constraint, it depends on like one, two, three suppliers.

但你看这里，屏幕上这八个主要供应商，颜色标得清清楚楚，加起来才刚过一半，而图表底部的备注里还列了另外二十多家潜在的供应商，都能生产电力设备。

But when you have this diversity where the top eight that are on screen and color coded make up just more than half, let's say, but then your footnote at the bottom of this chart shows like 25 more names that can possibly produce power.

我们以前也聊过这个，你们能不能详细说说，目前电力供应的选择到底有多多样？供应链有哪些是共用的？这样当有人需要分布式发电、又不确定该找谁的时候，能有更多备选方案。

I mean, we've talked about this previously, but maybe you guys can walk through like just how diverse the options are available for power generation is and what sort of shared supply chain there is that benefits people who are looking for alternatives when they need power and they're not sure where to get it for behind the meter?

当然可以。

Yeah, sure.

就像我之前说的，那些老牌企业早就建立了庞大的供应链，经营多年，全球都有制造能力。他们财报里经常提到，瓶颈主要在铸件、涡轮叶片这些部件，不管是传统涡轮、气动机还是往复式设备，问题都差不多。但奇怪的是，这些瓶颈似乎并没有拖慢进度，他们还在不断宣布扩大产能，而且明确说这波增长主要来自AI需求的上升。

So, I mean, as I was saying, like there would be kind of established players who have like a large supply chain where, you know, and they've been in the business for years and years and years, they have manufacturing capacity globally, they quite often say in earnings that, you know, the bottlenecks are kind of the castings, the bottlenecks are the blades for whatever kind of turbine or aeroderivative or recip kind of, whatever solution they're offering, but it seems that like this is not slowing anything down because they're still announcing capacity increases they're still saying that these capacity increases are driven by AI demand growth.

所以，去年那种关于铸件和上游零部件限制的说法，现在看来更像是个叙事，而不是真实制约。

So it seems to be actually a little bit of a narrative talking about the constraints to castings and, you know, could could kind of upstream constraints to to building a building these turbines, as a bit of a narrative from last year.