Anthropic加入AI能源联盟，苹果推出服务器

欢迎收听本期播客。

Welcome to the podcast.

今天，我们想聊聊一系列突破性进展、AI新闻，以及AI硬件方面的建设情况。

Today, wanna talk about a bunch of breakthroughs, AI news, and build outs when it comes to AI hardware.

我觉得最近我们聊了很多关于是否应该放慢AI发展的话题。

I think we've had a lot of talk lately about should we slow down AI?

我们是否不应该放慢AI的发展？

Should we not slow down AI?

与Dario Amadeh一起探讨前沿发展的节奏。

Pacing the frontier with Dario Amadeh.

但有一件事，我觉得没人会争论，那就是AI将需要大量的电力、大量的数据中心、大量的芯片和大量的内存。

But the one thing that I think nobody is debating is the fact that this AI is gonna need a lot of a lot of power generation, a lot of data centers, a lot of chips, and a lot of memory.

而这些都不是你说“让我们放慢发展速度”就能做到的，因为一个数据中心可能需要十年才能接通合适的电力。

And all of those things aren't something you could say, let's pace the development of, because a data center might take ten years to get a proper electrical hookup to it.

这些基础设施需要很长时间才能建好，制造设施也需要很长时间，还有相关的规划工作。

These things, this infrastructure takes a very long time to get built out, the manufacturing facilities take a very long time, the planning that goes into it.

所以今天我想聊聊我们在哪些方面正在加速推进，以及我们在硬件领域看到的一些重大进展。

So today I wanna talk about some of the areas that we're accelerating on and some of the big advancements we're seeing in the hardware side.

首先我想说的是，苹果基本上要重新进入服务器市场了。

The first one I wanna talk about is that Apple is going to, is basically bringing back their server market.

他们要重新回到服务器市场。

They're coming back into the server market.

他们计划在2029年用M8 Ultra芯片来做这件事，并且会使用英伟达的互联技术。

They're gonna do this in 2029 with their m eight ultra, and they're gonna use Nvidia links.

这个市场他们之前退出了，我稍后会详细说说。

This is a market that they left, and I'll talk a little bit about that.

我还想提到，韩国的SK海力士正在和英特尔洽谈，要在美国生产内存芯片。

I also wanna talk about the fact that SK hynix out of South Korea is in talks with Intel to make memory chips in The United States.

如果你一直在关注AI的动向，你应该知道内存是最大的瓶颈之一，而目前只有三家公司能生产这种特定类型的内存。

If you have been following what's going on with AI, you know that memory is one of the massive biggest bottlenecks, and there are only three companies that are currently doing this specific type of memory.

与此同时，英伟达、谷歌和Emerald AI正在联合推出一个叫“面向电网灵活的AI数据中心联盟”的项目。

At the same time, we have Nvidia, Google, Emerald AI, which all launching something called the Alliance for Grid Flexible AI Data Centers.

这是AI能源管理联盟，实际上有18个成员。

It's the AI Energy Management Alliance, and there's actually 18 members.

所以Anthropic是其中一员，国家电网也是，AES、NRG，他们都联合起来打造一个互联电网。

So Anthropic is part of it, National Grid is part of it, AES, NRG, all of them are coming together to create an interconnected grid.

这些背后还有一些有趣的数据。

There's some interesting statistics that go along with all of that.

如果你讨厌给我刚才提到的那些公司付费订阅，比如Google、Anthropic、ChatGPT之类的，那我推荐你看看我的创业项目aibox.ai。

Now, if you're the kind of person that hates paying for subscriptions to all of those companies I just mentioned, like Google and Anthropic and ChatGPT and everyone else, I'd love for you to check out my startup aibox.ai.

我可以让你使用超过90个顶级AI模型。

I give you access to over 90 of the top AI models.

你可以用Anthropic、Gemini、Claude、Grok、DeepSeek、Perplexity，所有顶级模型都有，包括图像、音频、视频，你想要的所有顶级模型和模态，一共90个，起步价只要8.99美元，按token计费。

So you get Anthropic, you get Gemini, you get Clod, you get Grok, you get DeepSeek, you get Perplexity, you get all of the top models, image, audio, video, every top model and modality that you could want, there's 90 of them, it is $8.99 to get started and you use it by the token.

如果你有大型项目，还有更高的订阅等级，但如果你只是想试试，你可以在同一个聊天线程里和所有这些不同的模型对话。

So if you have a massive project, there's higher subscription tiers, but if you just wanna test stuff out, you can chat with all of these different models in the same chat thread.

而且你还可以在内部连接它们，比如你经常用Claude，想让Claude生成图像，就用我们的MCP连接器，在Claude内部，图像可以由Chat生成，视频由Google VO3生成，或者你用的任何模型都可以。

And you can also connect them inside, so if you use Claude a lot, and you want Claude to generate images, you use our MCP connector, and inside of Claude, images can be generated by Chat or video by Google VO3, or whatever model you're working with.

这真的帮我省了很多时间，我希望这是个超级有用的产品。

It is something that saves me so much time, and I hope it's a super useful product.

就是 aibox.ai。

It is aibox.ai.

我会在描述里放个链接。

I'll leave a link in the description.

我们先聊聊苹果那边的情况。

Let's talk about what was going on with Apple first.

这距离他们砍掉 X Server 已经过去十五年了。

So this is fifteen years after they killed X Server.

如果你还记得，苹果正打算重返服务器市场，他们基本上会用四块 M8 Ultra 芯片加上英伟达的 NVLink Fusion 来搭建新的服务器，也就是他们新的服务器产品。

If you'll remember, Apple is looking to get back into servers, and how they'll essentially put this together is they will have four M8 Ultra chips and Nvidia's NVLink Fusion, and all of this is how they're going be building out their new server, basically their new server offering.

我觉得有意思的是，苹果一直是大家批评比较多的公司之一。

I think what's interesting here is we've seen a lot of Apple is basically one of the people that, one of the companies that people have criticized heavily.

他们会说，天哪，苹果错过了 AI 热潮。

They're oh my gosh, Apple's missing the AI boom.

他们太慢了。

They're so slow.

他们的新Siri，苹果智能，一切都太慢了。

Their new Siri, the Apple intelligence, everything was so slow.

然而，如果你看看苹果的股价，它并没有因此下跌。

And yet, if you look at the stock price of Apple, it has not declined over this.

他们实际上取得了一些令人难以置信的进展，而且股价似乎并没有下跌。

They have actually made some incredible progress, and it doesn't seem like the stock is, is declining.

其中一个重要原因是，苹果的硬件被大量使用，我不知道这是不是偶然的，或者怎么说。

One of the big reasons is because Apple, I don't know if it was accidentally or, or however you want to say it, but the hardware of Apple has been greatly used.

我认为很多人会指出的一个典型例子是，当Open Claw推出时，基本上每个人都跑去买一大堆Mac mini来运行OpenClaw代理。

And I think one of the big examples a lot of people will point to is when Open Claw came out, basically everyone was running out to go buy a whole stack of Mac minis that they were running OpenClaw agents on.

不过，我觉得Mac mini的热潮有点短暂。

Now, I think that Mac mini craze was kind of short lived.

那只是一瞬间的事。

It was kind of a quick moment.

显然，Mac mini 本身就是一个很棒的产品。

Obviously the Mac mini is just a great product.

所以我觉得很多人会把它当作一台好电脑来用。但我觉得苹果确实因此看到了非常真实、非常大的销量激增，他们也意识到了，嘿，你看，这基本上就是放在办公桌上的一台小服务器。为什么不为那些付着天价费用的公司制造真正的服务器呢？随着公司们想要拥有自己的虚拟机，想要在云端运行自己的智能体，这个需求正在增长和扩大。

And so I think a lot of people would just use it as good computer, But I think Apple did see a very real, very large explosion in sales because of that, and they've realized, hey, look, I mean, they're basically a little server on someone's desk, Why don't we make actual servers for companies that are paying, you know, crazy amounts of money as this is something that's growing up, as scaling up as companies are looking to have their own virtual machines, they want their own agents in the cloud.

我觉得苹果抓住了这个机遇。

I think Apple caught the vision.

他们可能尝到了一点甜头，所以这就是他们正在发展的方向。

They got probably a little taste of that, and so this is kind of the direction that they're they're moving towards.

再说一次，这要到2029年才会实现。

Again, this isn't gonna happen till 2029.

我们知道这一点，因为他们说过，这些服务器将配备四块 M8 Ultra 芯片。

We know this because they said that this is going to be the, they're gonna have four M8 Ultra chips.

他们刚刚在八月份发布了 M6 系列芯片，所以 M8 系列至少还要两个芯片周期。所以我们知道，在 M8 上发布这个，很可能要到2029年了。

They just announced the M six generation of chips in August, so the M eight line is at least two silicone cycles out, and so we know that the, you know, announcing it on the M eight, it's gonna probably be 2029.

这些服务器可能还会采用英伟达的 NVLink Fusion 技术。

These servers are also going to probably incorporate Nvidia's NVLink Fusion.

这是英伟达用来把多个芯片拼接成一个逻辑单元的同一技术。

This is the same technology that Nvidia uses to stitch multiple chips into one logical unit.

你知道，他们有自己的芯片，但可能会用英伟达的技术把它们拼在一起。

You know, they have their chips, but they're gonna probably using Nvidia to stitch them all together.

这很有意思，因为苹果和英伟达的关系曾经一度很紧张，但现在他们又回头找英伟达，用他们的芯片来支持升级版的Siri服务器。

This is interesting because Apple is already kind of, they had a pretty rocky relationship with Nvidia for a while, but they've turned back to Nvidia and turned back to their chips to help power the revamped Siri servers.

所以他们的关系已经紧张了一段时间。

So their relationship has been on the rocks for a while.

看起来他们正在缓和关系，重新合作，这将成为他们新服务器业务的重要支撑。

It looks like they are de thawing that and getting back into it, and this is gonna be a big part of what powers their new server offering.

我们有一些来自韩国的重大消息。

We have some big news out of South Korea.

SK海力士正在与英特尔洽谈，计划在美国生产存储芯片。

SK hynix is in talks with Intel to make memory chips in The United States.

这家韩国存储巨头将首次使用英特尔在美国的晶圆厂。

The South Korean memory giant would use Intel's US fabs for the first time.

这可能是HBM在AI加速器领域以及其生产地点上的一个重大转变。

This is probably a big shift in where HBM is going for AI accelerators and where they're gonna get built.

我觉得这事有意思的原因在于，只有三家公司能生产这种特定类型的存储芯片。

The reason why this isn't interesting to me is because there's only three major companies that are building this specific type of memory.

顺便说一句，SK海力士生产的存储芯片正是英伟达很多技术所依赖的。

And by the way, SK hynix is what's making the memory that powers a lot of Nvidia's tech.

所以这是一家非常重要的公司。

So this is a major company.

我们有SK海力士，有三星，还有一家在爱达荷州的公司也在生产这种特定的存储芯片。

We have them, we have Samsung, and then we have one other company out of Idaho that is building these specific memory chips.

所以这确实是个全球性的瓶颈，但我觉得他们看到，大部分需求其实来自美国国内。

And so, this is really kind of a global bottleneck, but I think they're seeing that there's, you know, most of the demand is coming inside of The United States.

显然中国也会有大量需求，但可能面临制裁，或者我不确定韩国在那里的情况如何。

Obviously China's gonna have a lot of demand, but there's probably, could be sanctions, or I'm not sure what South Korea's situation is there.

我觉得他们看到了来自美国的需求，想离这个市场更近一些。

I think they see the demand coming out of The United States, and they wanna move closer to it.

他们还在做这个，有趣的是他们现在和英特尔的关系。

They're also making this the the inter the interesting part is their relationship they have with Intel right now.

英特尔现在和英特尔代工基本上一直在寻找一个核心客户，来证明大规模在美国建厂是值得的。

Intel right now and Intel Foundry has basically been searching for an anchor customer to justify a massive US manufacturing build out.

他们谈这件事已经很久了。

They've been talking about this for a long time.

我们知道英特尔股价一直在跌，虽然我不想说崩盘，但英特尔确实挣扎了很久。

We know that Intel stock was sliding and, you know, kind of, I don't wanna say cratering, but Intel has struggled for a long time.

它真的应该是我们经常谈论的大公司之一。

It really should be one of the big companies we're talking about all the time.

之前老CEO下台，我们换了新CEO，英特尔最近一段时间表现相当不错。

And we had the old CEO lead, we brought in a new CEO, and Intel has been on kind of a generational run for the last little bit.

他们和美国政府达成了很多合作。

They've made a lot of partnerships with the United States government.

所以感觉英特尔正在找回状态，而且动作频频。

And so it feels like Intel is kind of getting its mojo back and they're making a lot of moves.

他们最近宣布了一个大规模的建设计划，在美国的制造设施上大举扩张，而这就需要客户。

They recently announced a massive build out, US manufacturing build out, and with that they need customers.

所以看起来SK海力士可能会成为大客户之一，开始在他们的美国工厂里生产。

And so it looks like SK hynix might be might be one of the big customers that's gonna start building in their US fabrication facilities.

这些都是根据路透社的一些报道。

This is all according to some reports by routers.

这笔交易如果达成，将是SK海力士首次在美国生产内存。

This particular deal would be the first time SK hynix produces memory in The United States.

传统上，这些都是在韩国完成的。

Traditionally, it's all been done in South Korea.

SK海力士基本上也是英伟达AI加速器所需高带宽内存的主要供应商。

SK hynix is also basically the dominant supplier of high bandwidth memory to Nvidia's AI accelerators.

所以这其实是在加速很多供应链的环节。

So it's, you know, just kind of speeding up the supply chain on a lot of that.

另外，英伟达、谷歌和Emerald AI正在组建一个联盟，推动他们所谓的“电网灵活AI数据中心”。

We also have Nvidia, Google, and Emerald AI, which are launching a alliance for what they're calling grid flexible AI data centers.

这个联盟叫做AI能源管理联盟。

This is called the AI Energy Management Alliance.

里面大概有18家不同的公司。

There's about 18 different companies that are inside of this.

有意思的是，这个联盟这个月刚成立，创始成员包括Emerald AI、谷歌和英伟达，他们都估计，灵活的数据中心可以在美国现有的电网中释放出100吉瓦的电力。

What's interesting with this, it was just launched this month and founders Emerald AI, Google, and Nvidia are all estimating that flexible, flexible data centers could unlock a 100 gigawatts in the existing US grid.

他们说的意思是，美国电网的平均利用率只有大约50%。

So what they're saying is that essentially The United States, The US power grid is only running at about 50% utilization on average.

你可以想象，有高峰时段，比如在加州或得州，有时候会听到停电或限电的消息，或者遇到大风暴时部分地区电网瘫痪，各种问题都有。

And I mean, you can imagine there's peak hours and sometimes when that happens in California or Texas, hear about like blackouts or brownouts or, you know, when there's a big storm and their parts of the grid go down, like there's all sorts of issues.

但总的来说，在普通的一天里，我们的利用率大概在50%左右。

But generally speaking, on the average day, we're we're sitting at about 50% utilization.

可能晚上是高峰，大家都回家了，给电动车充电，而白天人们上班时用电量低一些，半夜更是低谷，对吧？

It's probably like peak hours in the evening time when everyone gets home and is charging their electric cars or, and, you know, in the middle of the day when people are at work, it it it's lower in the middle of the night, it's off, right?

所以电网有起伏，但平均利用率大约是50%。

So there's there's ups and downs in this, but it's on average about 50% utilization.

他们的观点是，有个叫Brattle Group的机构，他们估计电网利用率每提高10%，电价就会下降大约3.4%。

Their thesis on this, there's actually something called Brattle Group, and they're estimating that each 10% gain in grid utilization is going to lower electrical rates by about 3.4%.

所以他们说，你看，我们有能力发双倍的电。

So they're saying, look, we have the capacity to generate double the electricity.

如果我们从50%提高到60%，每个人的电费都会下降3.4%，因为电力更充裕了，所以就会，我觉得，很大程度上，在用电高峰期，你就不用付高峰价格了。

If we went from 50% to 60%, everyone's electrical costs would drop 3.4% because there'd be more abundance in electricity and so it would just, I mean, I think it's, it, a lot of it gets down to like at peak usage times, you're not gonna be paying peak prices.

所以如果电力更多，剩余更多，高峰期的价格可能就会更低，这样你就能省钱。

And so if there's just more electricity, there's more surplus, the peak I guess would be smaller, and so you'd save money.

这就是他们的估计。

So that's their estimation.

Nvidia、Digital Reality和Emerald AI即将在弗吉尼亚州启动一个约100兆瓦的灵活AI工厂，今年晚些时候就会启动，所以他们显然在关注这个。

Nvidia Digital Reality and Emerald AI are gonna switch on, they're about to switch on about a 100 megawatt power flexible AI factory in Virginia later this year, so there's obviously something they're looking at.

为这些数据中心和很多项目获取电力简直是一场噩梦。

Getting power for these data centers and for a lot of these projects is an absolute nightmare.

这非常棘手。

It's it's very tricky.

显然，美国的数据中心目前在某些现有的并网流程下，要等上十年甚至更久才能接入电网。这就是为什么埃隆·马斯克在建立他的孟菲斯设施时，会那样做。

Apparently US data centers currently wait up to ten years or more for a grid connection under some of the existing interconnection processes, And this is why you had Elon Musk when he stood up his Memphis facility.

他真的是直接去把所有能买到的便携式柴油发电机都买下来，运到那个设施，然后不接入电网，而是用柴油发电机给整个设施供电。这让很多人不满，还引发了一堆关于污染之类的诉讼。

He literally went and just, you know, purchased every, portable diesel generator he could get and shipped it over to the facility, and instead of plugging into the grid, just powered the whole thing off of diesel generators, you know, much to the chagrin and a whole bunch of lawsuits about, like, pollution and that kind of stuff.

但这就是他让数据中心快速运转起来的方式。

But this is how he got the data center up and spinning.

除非你觉得这有点偶然，我是说，那个数据中心特别赚钱，他正在建更大的项目，但那个数据中心每个月能赚十亿美元，大约九亿到十亿美元一个月，主要是把他们的额外空间和算力租给Anthropic。

Unless you think it's kind of random, I mean, that data center in particular, he's building much bigger projects, but that particular data center is making a billion dollars a month, about 900,000,000, a billion dollars a month roughly, renting their, their extra space, their extra compute to Anthropic.

所以Anthropic每个月付给他们十亿美元。

So Anthropic's paying them a billion dollars a month.

他们能赚这么多钱，唯一的原因就是他们用一种非传统的方式快速搭建了这个设施。

The only reason they can make that much money though is because they kind of threw this together in in a, you know, in a sort of not traditional way.

所以很多数据中心并不打算去租用所有便携式发电机。

And so a lot of these data centers aren't looking to go and rent every portable generator.

我觉得是SpaceX，或者现在应该叫SpaceX了，但在建这个设施的时候，它叫xAI。

I think SpaceX or I guess it's SpaceX now, it was XAI at the time when they built this thing.

我相信他们肯定加入了一些更可持续的发电方式，希望如此。但归根结底，这事儿非常复杂、棘手，所以这个新联盟的目标就是搞清楚怎么接入电网，并提高电网的利用率。

I'm sure they've kind of, they've built in some more sustainable, I'm hoping they've built in some more sustainable power generation methods, but at the end of the day, these things are very hard, very tricky, so the goal of this new alliance is to figure out how to plug into the grid and just increase the grid's utilization.

这样大家的成本都能降下来，而且这些大型数据中心项目也能顺利接入电网。

The costs should come down for everyone, but they're also gonna be able to plug in a lot of these big data center projects.

这是个我们不得不去啃的硬骨头。

It's a it's a problem we're gonna have to wrestle with.

在发电方面，中国正在以惊人的速度加速超越我们。

China is absolutely out accelerating us to an insane degree when it comes to power generation.

如果我们想为所有在建项目供电，就必须解决这个问题。

And if we wanna power all of the projects that are currently under construction, we need to solve this.

看来这也是很多人正在考虑的方案。

So this appears to be the solution that is being looked at by many people.

非常感谢大家今天收听我们的播客。

Guys, thank you so much for tuning into the podcast today.

如果你喜欢这期节目，记得去访问 aibox.ai。

If you enjoyed this episode, make sure to go check out aibox.ai.

无论你在哪里收听播客，都请给节目留下评分和评论。

Leave a rating and review on the show wherever you get your podcasts.

如果你不介意的话，这对节目帮助很大。

It helps the show a ton, if you wouldn't mind.

另外，如果你想将各种AI模型接入ChatGPT、Clot或你工作的其他平台，比如Cursor等，一定要去看看aibox.ai。

And make sure to go check out aibox.ai if you wanna get all of the different AI models plugged into ChatGPT or Clot or wherever you do your work, cursor, etcetera.

你可以在那里接入90种不同的AI模型，这样你就能在同一个工作区内完成音频、视频、图像生成等所有工作。

You can get 90 different AI models plugged into that, so you can do audio, video, image generation, all inside of the same all inside of the same workspace.

非常感谢大家收听本期播客，我们下期再见。

Thanks so much for tuning to the podcast, and I'll catch you guys all in the next episode.