万亿美元AI竞赛：谁在领跑，谁注定出局

难以置信我竟会这么说，但在过去三十天里，已有超过1万亿美元被投入建设50至100千兆瓦的电力设施，以支持AGI的研发。XAI、OpenAI、Meta、微软、谷歌各自承诺投入数千亿美元建设新一代数据中心，这些设施将驱动我们见过的最强大、最具变革性的技术。现在我要断言——我认为埃隆会赢下这一局。尽管比竞赛晚了五年多，但在过去两年里，他打造出了全球最顶尖的前沿AI模型之一，并集结了最优秀的工程师和建设者来扩展计算力和数据以训练这些模型。不仅如此，埃隆还在排除一切不确定因素。

I can't believe I'm saying this, but over the last thirty days, over $1,000,000,000,000 has been committed to building out over 50 to 100 gigawatts of power to build AGI. XAI, OpenAI, Meta, Microsoft, Google have each committed hundreds of billions of dollars to building out the next generation of data centers that are going to power the most powerful and transformative technology we've ever seen. Now I'm just going to call I think Elon wins this one. Despite being over five years late to the race, in the last two years, he's built out one of the best frontier AI models the world has ever seen and assembled some of the best engineers and builders to scale compute and data to train these models. Not only that, but Elon's leaving nothing to chance.

他要掌控整个牌桌。特斯拉最初是家电动汽车公司，现在生产能为城市供电的超级电池组和储能系统。他还在造机器人。这永无止境。乔希，我想提醒观众，要构建最好的AI模型，需要价值数万亿美元的计算资源才能实现我们所认知的AGI。

He wants to own the entire table. Tesla, which started off as an electric car company, now produces Mega Blocks and Mega Packs that can power cities. He's building robots. It's never going to end. Josh, I wanna remind the audience that in order to build the best AI models, you need tons, trillions of dollars worth of compute to reach what we perceive as AGI.

我认为XAI会胜出。你觉得呢？

I think XAI is gonna win it. Do you?

目前看来确实如此。单就发展速度而言——我们常讨论这点——他们进展最快。他们拥有最顶尖的工程师团队，垂直整合程度也最高。

I think they are so far currently. Just based on, again, rate of acceleration, we always talk about this. They're moving the fastest. They have the most cracked team of engineers who are working on this. They have the most vertical integration.

得了吧。我是说，Colossus二代将会非常惊人。我最近真正意识到的是，这是人类史上最大的资本支出项目，而且各方都在独立推进。这就像曼哈顿计划，但规模超乎想象——不同之处在于，不是单一实体在做，而是全球最大的几家公司倾其所有甚至举债投入。

Come on. Yeah. Mean, Colossus two, it's gonna be pretty amazing. The thing that I've really come to grips with recently is that this is the largest capital expenditure project anyone has ever done before, and everyone is doing it independently. This is like the Manhattan project at a scale that we could never imagine, except instead of one entity working on this, there are several of them spending the largest companies in the world spending all of their additional money and even going into debt to pay for this.

有充分理由认为，这可能是人类最后的投资项目。Transformer架构开创了将电力转化为智能的途径，完全可以论证：我们能获取的能源越多，就越该将其转化为智能生成。这些公司正是这么做的——他们投入巨额资金。看起来XAI处于领先，我想你有些关于Colossus二代建设的趣事要分享，那是XAI的新基建项目。

And there's a strong case to be made that, like, this could be the last thing that we ever spend money on. The transformer kind of created this way of converting electricity into intelligence, and it there's a very strong case to be made that the more energy we can get, the more we should just funnel it into generating intelligence. And that's exactly what's happening with these these companies. I mean, they're spending a huge amount of money. It seems like XAI is in the lead, and I think you have some fun things to share about the classes two build out, is the new build out for XAI.

能请你详细介绍一下当前进展吗？

Can you please walk us through what's going on here?

先提醒听众们，计算能力是决定模型优劣的最关键因素。我想从XAI和埃隆·马斯克的战略说起。他们今年宣布了名为'巨像二号'的项目，正如这条推文所说，这将成为全球首个吉瓦级AI训练数据中心。简单来说，这将是我们前所未见的、用于训练AI模型的超大规模数据中心。据我所知，埃隆已投入超过100亿美元建设该项目，目标最终扩展至10至100吉瓦规模，足以供多个城镇使用。

So just to remind the listeners, compute is the most important thing to determine how good your model is gonna be. And I wanna start off with XAI and Elon Musk's strategy. They recently announced this year something known as Colossus two, which as this tweet says, is gonna be recognized as the world's first gigawatt scale AI training data center. Simple translation is it's gonna be the biggest data center to train AI models that we've ever seen before. And I think Elon has invested over $10,000,000,000 already into building this thing out with the goal to scale this to 10 and a 100 gigawatts eventually, which is enough to power multiple, multiple towns.

甚至在Terra Labs也是如此。我明白了。

And even at Terra Terra Labs. I see.

没错。我差点不敢说这个数字，心想'这能做到吗？'但要知道，目前投入已达数万亿级别。给听众们些背景信息：'巨像一号'团队部署了10万块GPU——这些正是训练AI模型所需的核心设备，耗时122天。而新版'巨像二号'规模远超初代...

Yeah. I I was almost scared to say that because I was like, can we do this? But, you know, trillions are being spent at this point. To give some context for the listeners, Colossus one, the team deployed a 100,000 GPUs, which is the thing that you need to kind of like train these AI models. And it took them one hundred and twenty two days with this new Colossus two, which is much, much bigger than the original one.

他们仅用19天就完成了首期建设，两位数天数！这太疯狂了。埃隆扩建数据中心的能力确实令人震惊。我想特别指出半导体分析机构关于基础设施竞赛的评论：XAI用六个月就完成了Oracle、Crusoe和OpenAI需要十五个月才能建成的规模。

It took them nineteen days, two digits. That's insane. To build out the first iteration of this. So Elon's ability to scale these data centers is just actually insane. And I wanna kinda point out this snippet from semiconductor, which do analysis on infrastructure races, saying x AI built in six months what took fifteen months for Oracle, Crusoe, and OpenAI to build.

不过乔什，我认为埃隆还有另一个赢得这场竞赛的秘密策略。你能说说吗？

But I think, Josh, Elon has another secret strategy to winning this war. Can you tell me about it?

当然。埃隆的战略之所以如此有效，关键在于他卓越的制造与生产能力。很多人低估了实体制造的难度——在物理世界造东西真的非常困难。他通过特斯拉、SpaceX数十年的实践，在多数公司只专注数字产品或手持设备的领域，积累了实体硬件制造的宝贵经验。

Yeah. Well, I mean, the the Elon strategy and the reason why all these things work so well is because he is amazing at manufacturing and production. And I think a lot of people overlook how difficult it is to actually make things in the physical world. It's really difficult to manufacture things. And he learned this over the decades through Tesla, through SpaceX, through actually creating physical hardware in the physical world in in a space in which a lot of these companies really only create digital goods or handheld products.

这很大程度上得益于垂直整合及其旗下企业的协同优势。比如现在屏幕上展示的特斯拉Mega Block储能系统，不仅能通过变压器连接电池组构建微电网，更能为模型训练系统提供稳定电力。训练大模型时面临的关键问题是'电力波动'——GPU每次启动都需要大量能源且启停频繁，这种波动对电网来说很难平抑。而Mega Block正是解决方案。此外，他们还在推进自研芯片计划，从AI5开始，最终将推出AI6。

A lot of this has to do with vertical integration and the unfair advantage of having this suite of companies that kind of complement each other. One of them, like we're seeing on screen now, is Tesla and the advent of the Mega Block, which basically not only allows you to create a micro grid by having a transformer connected to a bunch of batteries, but it allows you to smooth the electricity from the grid to your power systems that are training the models. So a big problem when you're training large models is there's this thing called, like, power jitter, where every time the GPUs spin up, they require a lot of energy and they spin down very quickly, and those jitters are very difficult for the grid to kind of smooth out. So these mega packs do that. They are also working on another effort where they're going to create their own in house chips, starting with AI five and then eventually AI six.

假设是，最终这些芯片将强大到足以直接嵌入他们自己的数据服务器和数据中心，实际使用自主研发的芯片进行训练。这不仅展现了他们在垂直整合上的卓越能力，更体现了他们在产品研发上的极致专注。正如你之前提到的EJAS，他们部署这些技术的速度远超同行。无论是工作文化的体现，还是与像OpenAI这样将大量基础设施外包给甲骨文、建设任务交给第三方公司XAI相比，XAI大部分工作都是内部完成，我认为这正是他们能如此迅速领先的关键。你觉得XAI成立多久了？

And the assumption is that, well, eventually, these chips will be powerful enough that they could just embed them in their own data servers, data centers, and actually use their own in house chips to train this. So not only are they really great at the vertical integration, but they're just really hardcore at creating these products. I mean, like you said earlier, EJAS, the speed at which they're able to deploy this stuff is just faster than everyone else. And whether it's a testament to the actual work culture or them doing it all themselves versus a company like OpenAI who's outsourcing a lot of the actual infrastructure to Oracle and a lot of the construction stuff to a third party company, XAI is doing a lot of this in house, and I think that's going a long way in actually making them move so much faster than everyone else. He just how old do you think XAI is?

其实他们成立时间并不长。

Like, they haven't been around that long.

大概三年左右吧？相当年轻。

Has to be, like, what, three years? They're quite young.

约两年半。不到三年，而OpenAI已存在九年之久。虽然早期多年主要用于AI领域的研究和铺垫，但XAI的发展速度确实惊人。我认为这些因素共同作用形成了他们的优势。

Like two and a half years. It's less than three years, and OpenAI has been around for nine. That is a long time. Granted, a lot of the a lot of the earlier years were spent just doing research and preparing the world for the world of AI, but XAI moves so fast. And I think all of these things kind of converge together to come to their advantage.

随着我们在通用人工智能（AGI）道路上加速前进，这种指数级增长的速度开始产生巨大差异。当我思考这些数以万亿计的惊人数字时，这堪称史上最大规模的资本支出——如同全球顶尖企业同时开展七八个曼哈顿计划。某种程度上这是合理的，因为这可能是最后需要巨额投入的领域。我正考虑将这个概念纳入我们新推出的时事通讯（大家应该订阅），拟定为三条定律：第一定律是能源终被捕获或利用，即当我们创造更多能源时，总会有人找到用途。

And as we move faster and further along on this path to AGI, that rate of acceleration on an exponential curve starts to really make a big difference. And, you know, just the more I think about this, like, the these outrageously large numbers of trillions of dollars, I mean, this is this is the biggest capital expenditure ever made in history. It's like seven or eight Manhattan projects all going at once from all the largest companies in the world, and it kind of makes sense. Like, it seems like this is the last thing you'll need to spend money on. And I was thinking about it, and I'm probably gonna include this in our newsletter, which we recently started, everyone should subscribe to, but into, like, three laws where the first law is energy gets captured or energy captured disused, which means when we make more energy, someone always finds a use for it.

就像杰文斯悖论。第二定律是智力开发是额外能源的最高回报用途。不过这个我打算留到通讯里详细阐述，现在先卖个关子。

It's like Jevan's paradox. The second one is that intelligence is the highest return use of extra energy. I don't think I'm gonna do this now. I'll save this for the newsletter because I haven't written it. I'm just gonna write cool blitz.

没关系。但我喜欢你最初的观点。你能...

That's okay. But I like what you said initially. Are you able

在此之前切断

to cut it off before

你解释一下法律？因为我想补充一点你的观点，我觉得这会很有共鸣。是的。

you explain the laws? Because I wanna add something to your point, which I think would be relatable. Yeah.

我就直接切了。因为我在想，等等，这些会变得太长了。

I'll just cut. Because I was thinking, was like, wait, these are gonna get it's gonna get way too long.

对。就是。不。我懂你。对。

Yeah. Just Yeah. No. I got you. Yeah.

我是说，你在自我评估，这他妈太棒了。好吧。所以让我就

I mean, you're you're self assessing, which is fucking great. Okay. So let me just

所以点击Keygen two的屏幕，然后

So tap the screen of Keygen two, and then

我来滚动

I'll roll

直接进入正题。关于Keygen。

right into that. Of Keygen.

好的。明白。

Yeah. Okay.

对，我不会分享展示它

Yeah. I won't share show it

现在我在添加这个东西的时候？

now while I'm adding this thing?

你现在可以分享，因为这样我就只展示我们俩的脸。完美。好的。对。然后在合适的地方切入就行。

You could share it now because then I'm just gonna show our two faces. Perfect. Okay. Yeah. Then just hop in wherever it makes sense.

好吧。酷。确实。这让我想起扎克上周在采访中说的话，他说，如果意味着我有10%到20%的机会构建出超级智能，我愿意在重大失误上损失数千亿美元。关键在于，现在大家都在玩一场如此高风险的博弈，谁都输不起。

Alright. Cool. Yeah. It actually reminds me of something Zuck said in an interview last week, which was, I am willing to lose hundreds of billions of dollars on a massive mistake if it means I have a 10 to 20% chance of building out super intelligence. The point being is it is such a high stakes game that everyone's playing right now that they can't afford to lose.

假设AGI在五年内实现。假设这是你的预估时间，但你在三年内就实现了。那两年的滞后会让你赔上整个公司，这就是他要表达的观点。你说得完全正确。这是有史以来最重要的竞赛。

Let's say AGI is achieved in five years' time. Let's say that's your projected estimate and you achieve it in three years' time. That two year lag is gonna cost you your entire company, and that was the point that he was making. You're absolutely right. This is the most important race ever.

没错。事物沿着那条指数曲线飞速发展。如果你在那条垂直线上偏差几个月，你就麻烦大了。我们稍后会讨论Meta、扎克伯格、OpenAI以及其他疯狂新闻。但首先，我们必须提一下我们的赞助商Keygen。

Yeah. Things move so fast along that exponential curve. If you are off by a couple of months on that vertical line, you are just you are in trouble. And we're going to talk about Meta and Zuck and OpenAI and all the other crazy news. But first, we do have to make mention of our sponsor, Keygen.

Keygen正在构建全球最大的验证分发协议，简称VERIFI。你可能会问，VERIFI是什么？它专注于确保只有真实高质量用户参与数字平台，解决虚假账户、机器人和欺诈活动等问题。它是如何运作的？

Keygen is building the world's largest verified distribution protocol, aka VERIFI. What is VERIFI? You may ask. It focuses on ensuring only real high quality users participate in digital platforms, addressing issues like fake accounts, bots, and fraudulent activity. How does this work?

该协议采用先进的生物识别和防欺诈技术来拦截虚假用户。对于需要高质量无欺诈数据来训练模型并全球扩展产品的开发者而言，这套系统对AI和消费类应用都至关重要。目前已有200多家客户在AI游戏、DeFi和消费应用中采用，报告显示虚假账户减少了99.8%以上。若你需要用真实数据训练AI，不妨看看Keygen。节目备注中会附上链接。

Well, this protocol uses advanced biometrics and fraud prevention technologies to block fake users. The system is essential for AI and consumer applications alike for developers who require high quality, fraud free data and engagement to train models and expand their products globally. They are already used by over 200 plus clients across AI gaming, DeFi, and consumer apps, and has reported for them over 99.8% reduction of fake accounts. So if you are need to train your AI on real data, check out Keygen. We'll have a link in the show notes.

感谢Keygen。EJS，交回给你。我们来聊聊OpenAI吧，他们本周宣布的消息似乎在数据和训练领域掀起了不小波澜。

Thank you, Keygen. EJS, back to you. Let's talk about OpenAI because this seems like a pretty big thing that's happening in the world of data and training that they just announced this week.

好的。我们有xAI和埃隆，如你所说是这场竞赛的新入局者，却莫名其妙地跑在最前面。但他们与萨姆·奥尔特曼有些历史渊源——这位OpenAI创始人曾与埃隆共同创立了OpenAI。

Okay. So we have x AI and Elon, which, as you pointed out is the newcomer to this race, but somehow leading the leading the front. But they have some history with Sam Altman, who is, of course, the founder of OpenAI. They go way back. They actually founded OpenAI together.

对吧？后来两人因商业理念不合分道扬镳，竞争就此展开。萨姆想将组织从非营利转为营利性质，希望以某种埃隆不认可的方式开发AI。于是埃隆离开创立了xAI，这两家公司就这样各自发展至今。

Right? But then a rivalry ensued when the two of them kind of broke up because they didn't align on the same types of business interests. Sam wanted to take it from a nonprofit to a for profit. He wanted to start building AI in a certain type of way that Elon didn't agree with. And so Elon broke off and founded XAI, and that's kind of like how those two companies kind of have have progressed since.

但萨姆在建设数据中心训练顶级AI模型方面同样激进——甚至更甚。乔什，记得今年初他们宣布的'星际之门'项目吗？这其实就是他们对全球各地训练不同模型的数据中心的命名。这些星际之门数据中心地理分布极广，美国计划建5到10个，阿联酋和欧洲多地也在建设中，旨在为全球OpenAI用户提供支持。

But Sam has been equally, if not more aggressive in building out data centers to train the best AI models. If you remember, Josh, I think earlier this year, actually the start of this year, they announced their Stargate project, which is basically the name that they're giving to the different types of data centers that they're building to train these different models. These Stargate data centers exist all over the world. Geographically, I think they have the highest concentration. I think they've planned for five to 10 of these Stargate sites to be in The US, but they're building one in The United Arab Emirates, as well as a multiple in Europe, just to have this presence and support for open AI users anywhere in the world.

星门项目最初宣布将在未来四年内投资5000亿美元，作为与美国政府共同承诺的一项更大规模AI基础设施法案的一部分。但过去一周他们又接连放出重磅消息，Josh。首先是他们与甲骨文达成的3000亿美元算力协议，甲骨文将提供GPU和算力来协助训练OpenAI的模型。这项合作将持续五年，直接推动甲骨文股价暴涨42%，导致许多批评者质疑他们是否真具备这样的算力储备。

So the Stargate project initially announced to invest $500,000,000,000 over the next four years as part of a larger AI infrastructure bill that they committed in tune with the American government as well. But they then had some banger announcements over the last week, Josh. The first of two being their deal with Oracle, which is a $300,000,000,000 compute deal where Oracle will supply the GPUs and the compute to help train OpenAI's models. And this commitment is gonna be had over five years. This in tune results in Oracle's stock surging 42%, which led to a lot of critics saying, okay, so they don't have the compute.

协议公布后甲骨文股价飙升42%，他们用这笔资金购买英伟达GPU，某种程度上形成了循环效应。但关键在于OpenAI承诺每年采购600亿美元算力，五年总计3000亿美元，这将消耗4.5吉瓦电力——约相当于两座胡佛水坝的发电量，Josh，足够为400万户家庭供电。

They announced this deal. Oracle stock jumps up 42% and they use the money to buy Nvidia GPUs and it kind of results in this kind of like cycle. But the point is OpenAI commits to buy $60,000,000,000 of compute per year, 300,000,000,000 total over the next five years, which results in 4.5 gigawatts of power, roughly the output of two Hoover dams, Enough for about 4,000,000 homes, Josh.

没错。我确实...等等。我仔细核算过数据，还专门向Gruk查询了一吉瓦到底相当于多少能源作为参考。

Yeah. You do. Okay. So I I went through the math. I, like, actually, I queried Gruk about how much energy one gigawatt is just for reference.

因为我们总笼统地说一吉瓦能供应83.4万户家庭，但具体量化就很惊人。以太阳能板计算，每吉瓦需要约330万块标准住宅太阳能板，乘以五就是超过1700万块。每吉瓦也等同于400台大型风力涡轮机...

Because we always talk about, like, okay. It's about 834,000 homes, but it's like kind of outrageous think about exactly how much it is. So with solar panels, each gigawatt is equal to approximately 3,300,000 standard residential solar panels. So multiply that by five, you are like over 17,000,000 solar panels. It's equivalent each gigawatt is 400 large wind turbines.

再乘以五就是200万台风机。这些数字庞大得令人眩晕，将占据总能源产出的显著比例。这个趋势很可能持续下去：我们投入AGI研发的能源越多，其他领域的效率提升就越快。这些数字只会不断膨胀。

So multiply that by five, you're at 2,000,000 wind turbines. So the numbers here are staggeringly large. Like, this is a significant this will become a significant percentage of the total energy output. And it's probably a likely trend that continues where the more energy we actually throw, the greater percentage energy we throw towards solving AGI, the more efficient everything else gets. It makes sense that these numbers are just gonna keep growing larger and larger.

正如Elon近期提到的太瓦级目标，这些公司正朝此迈进。虽然能源来源尚不明确，但这场竞赛的连锁效应很酷——如果有公司能调动一太瓦能源，意味着我们取得了某种技术突破，不仅能优化数据中心，更能大幅降低用电成本。当每千瓦时价格下降时，世界总会变得更好。

About a terawatt that Elon mentioned earlier in the post. That's what these companies are aiming for. Where they get that energy, I don't know, but I think the downstream effects of this war are really cool. Like, if if a company is able to conjure up one terawatt of energy, the assumption is that we've gotten some sort of technological breakthrough that allows some sort of downstream effect that will not only allow these data centers to get better, but also, like, our cost per electricity goes down a lot. And the world always gets better when the cost per kilowatt decreases.

全球范围内皆是如此：不存在能源匮乏的富裕国家。

And we see this across the board. There's no energy rich energy poor countries that are rich.

但是，乔希，有些数字根本对不上啊。你看这个——OpenAI每年要花600亿美元，是当前收入的六倍。所以我只能想象OpenAI正在策划某种交易，我们很快会谈到与英伟达的一笔交易，他们要么是在为未来融资，要么是与软银和甲骨文达成协议：'你们先提供这批GPU，到时候我们就能筹到启动资金'。在甲骨文公告之后，他们紧接着宣布将与甲骨文和软银合作新建五个星际之门基地，这种操作模式已经初现端倪。

But, Josh, like, some of the math isn't mathing, dude. Like, look at this. OpenAI will spend $60,000,000,000 a year, six times its current revenue. So I only have to imagine that OpenAI is structuring some sort of deal, and we're gonna talk about one with NVIDIA actually soon, where they're kind of raising money in the future or they're agreeing with like SoftBank and Oracle that, okay, you give me this amount of GPUs and by that time we'll have raised the money to to begin with. And we're kind of seeing this with kind of follow on announcement that they made after this Oracle announcement, which is they're gonna build five new Stargate sites in cooperation with Oracle and SoftBank.

孙正义先生也投入巨资支持这个计划。细节相当有趣，当然涉及大量天文数字——未来三年承诺投入4000亿美元。选址包括俄亥俄州、得克萨斯州、新墨西哥州，以及一个未命名的中西部地点。看，我们又要把资源高度集中在美国本土了。

Masayoshi san is putting up a ton of money to support this as well. And the details are are are pretty interesting and, of course, includes a lot of large numbers. $400,000,000,000 pledged over the next three years. Locations in Ohio, Texas, New Mexico, plus an unnamed Midwestern side. So, again, we're concentrating very much in The US.

有意思的是，AI算力正在成为国家间的军备竞赛，基本上谁拥有最强算力谁就能胜出，这大概就是激化中美竞争的原因。山姆有句话让我印象深刻：'只有建造足够的算力，AI才能兑现其承诺。如果受限于算力，我们将被迫做出优先级选择。没人愿意做这种抉择，所以让我们去建造吧。'

It's funny. AI computer is pretty much becoming like a a war chest literally between nations and whoever has the most power basically wins, which is, I guess, what's instigating a lot of The USA versus China rivalry. And a statement from Sam I found really interesting is he goes, AI can only fulfill its promise if we build the compute to power it. If we are limited by compute, we'll have to choose what to prioritize. No one wants to make that choice, so let's go build.

这就是他豪掷千金的理由和正当性依据。

And that's his reasoning and justification for spending this amount of money.

有趣的是，山姆已经两次公开提及——一次在博客文章，一次在与黄仁勋CEO的访谈中谈到即将提到的交易——他们可能要在'全民教育'和'治愈疾病'之间做选择。我觉得这种销售话术很有意思：'各位，除非获得更多电力，否则你们将面临这个艰难抉择。'

And it's funny. Sam has mentioned twice publicly now, once in a blog post and once in an interview with CEO Jensen Huang about the deal that we're about to mention, that they can choose between education for all or curing diseases. And I think that's a very interesting way of approaching the sales pitch to this is like, hi, guys. You have to choose. Unless we get more electricity, you're gonna have to make this very difficult choice.

我对此并不认同。我开始注意到这些CEO们在传播策略上的手法——比如Anthropic的达里奥本周公开贬低开源项目，现在山姆·奥特曼又向公众抛出这种情感绑架式的抉择。观察他们如何包装信息，使其合理化如此庞大的能源消耗和AI算力部署，倒是很有意思。

And I don't I don't love that. And I'm starting to see the tactics that some of the CEOs are using in terms of messaging. Like, Dario of Anthropic recently came out this week, and he was talking down on open source. Sam Altman is now making this, like, critical emotional decision to the public. So it's interesting to see how they actually deliver this and try to get the messaging across so it becomes appropriate to raise and deploy this much energy, this much power and AI compute?

是啊。我想他的挫败感可能源于人们（包括他自己）一直在承诺通用人工智能（AGI），但注意到吗？在你引用的那封信里他一次都没提AGI。他们至今未能实现消费级AGI，目前成果主要停留在编程领域。

Yeah. I I mean, I think the frustration that he's probably facing is people have been promising AGI, including Sam, and notice how he didn't mention AGI once in that letter that you're you're referencing. Mhmm. And they haven't really been able to deliver it at the consumer level. They're achieving it with coding.

他们通过数学和其他一些极客的、非常小众的东西来实现这一点，但尚未让更广泛的受众接受，让大众开始相信它。因此，我认为他试图用一个非常强烈且有目的的愿景来补充这一点，这某种程度上传达了故事或观点。对吧？当他说，我们要么治愈疾病，要么为孩子们提供最好的教育。你选哪个？

They're achieving it with math and a bunch of other nerdy, very niche things, but they haven't achieved it to the wider audience where he gets the masses to start believing in it. And I think he's trying to therefore supplement that with a very strong and purposeful vision, and it it kinda gets the story or the point across. Right? When he says, we either cure diseases or we have the best education for children. Which one do you want?

我们不想被迫做选择。这让你有点思考。不过，继续谈谈最近的公告，我认为这简直疯狂。英伟达正在向OpenAI投资1000亿美元。

We don't wanna have to choose. And it makes you kind of think. But yes, moving on to the most recent announcement, which I actually think is the craziest. NVIDIA is investing a $100,000,000,000 in OpenAI.

1000亿美元。

100,000,000,000.

1000亿美元。这还是在已经承诺了50亿美元之后，不管是在他们的哪一轮融资中，字母代号是什么。这次合作将为OpenAI的数据中心增长提供10吉瓦的GPU支持。这是数量惊人的GPU。我认为这是战略性的。

A $100,000,000,000. This is after already committing, what was it, like, $5,000,000,000 in one of their series, whatever the hell, whichever letter they did. This partnership will supply 10 gigawatts of GPUs to fuel OpenAI's data center growth. That is a staggering amount of GPUs. I think this is strategic.

我认为Sam想要某种保证，确保Jensen和英伟达会向OpenAI交付GPU，不会在承诺上动摇，因为Jensen也在供应XAI、Meta，某种程度上还有Google。所以，在这些依赖Jensen的竞争对手之间，形成了一种信任问题。显然，Jensen只是坐享其成，笑看这一切，但我认为这是一种战略联盟。

I think that Sam wants some sort of guarantee that Jensen and NVIDIA are gonna deliver OpenAI GPUs, and they're not gonna kind of falter on their promise because Jensen's supplying XAI. Jensen's supplying Meta. Jensen's supplying Google as well, right, to an extent. So there's kind of like this trust issues forming between the competitors with so much reliance on Jensen. Obviously, Jensen's just sitting back laughing and enjoying all of this, but I think this is a strategic alignment.

你有什么看法吗？

Do you have any comments? I

我是说，这里10吉瓦，那里10吉瓦。突然间，你谈论的就是一些真正的电力、真正的基础设施和真正的资金了。这些钱从哪儿来，我不知道。很明显，这一切都建立在依赖一个人——Jensen Huang的基础上。

mean, it's like 10 gigawatts here, 10 gigawatts there. Suddenly, like, you're talking about some serious power Yeah. Some serious infrastructure, some serious money. Where this is all coming from, I don't know. There's very clear that there is a reliance on a single man being Jensen Huang that all of this is built on top of.

所以我的意思是，这又是另一个重大事件，我们将拭目以待其发展。目前还没有人真正突破过千兆瓦级别。要在物理世界中实现10倍增长，一个完整的数量级跃升，需要实际建造物理基础设施，这确实是个挑战。好吧，这很棒，去做吧。

So it's just I mean, it's another big deal that we will see how it plays out. We have yet to see anyone really eclipse more than a gigawatt. So to go 10 x, a full order of magnitude, in the physical realm where you actually have to build physical infrastructure, like, okay. This is great. Go do it.

就像，让我们看看当你真正去做时会发生什么，因为到目前为止，还没有人能够解决这个问题。

Like, let's see what happens when you go do it because so far, no one's been able to figure this out yet.

如果你正在收听这期节目，开始觉得这听起来有点像金字塔骗局或庞氏骗局，你不是一个人。很多人都在关注这些最近的公告，他们正在串联线索并意识到：等等，如果OpenAI投资1000亿美元给Oracle购买云计算服务，然后Oracle又投资1000亿美元给NVIDIA购买显卡，但NVIDIA刚刚宣布他们将再投资1000亿美元给OpenAI构建AI系统。这不就是我们谈论的同一笔钱吗？这笔钱甚至还不存在，尚未筹集，甚至尚未正式承诺，但所有这些公司的股价都在大幅飙升。所以，这就像是先有鸡还是先有蛋的问题？

Now if you're listening to this episode and this is starting to sound like a bit of a pyramid scheme or a Ponzi scheme, you wouldn't be alone. A lot of people are looking at all these recent announcements and they're kind of connecting the dots and they're realizing, hang on a second. If OpenAI invests a $100,000,000,000 in Oracle to buy cloud computing services, and then Oracle invest a $100,000,000,000 in NVIDIA to buy the graphic cards, but then NVIDIA just announced that they're reinvesting a $100,000,000,000 in OpenAI to build AI systems. Isn't it just the same money that we're talking about, which doesn't even exist, which hasn't even been raised, which hasn't even officially been committed just yet, but all of their stock prices are massively soaring. So like, which comes first, the chicken or the egg?

这只是一个我觉得值得一提的有趣观点。乔什和我对这样的增长同样感到兴奋，我们认为AGI和投资计算资源非常重要，但我们也不想本末倒置。我们承认这可能感觉有点泡沫化，实际上可能确实存在一些泡沫，但这就是我们愿意承担的风险，或者说投资者愿意承担的风险，以实现这个目标。

And it's just kind of like a funny point that I think is worth mentioning. Josh and I are equally as excited about this growth, and we think AGI and investing in compute is super important, but we're also not trying to put the cart before the horse. And we're admitting that this might feel a little bubbly and it might actually be a little bit of bubbly, but that's the a little bit bubbly, but that's the risk that we're willing to kind of take or investors are willing to take to build this out.

是啊。你知道吗，EJI？我不再关心那些大数字了。它们对我来说毫无意义。1000亿美元对我来说什么都不是。

Yeah. You know what, EJI? I don't I don't care for the big numbers anymore. They mean nothing to me. A $100,000,000,000 means nothing to me.

去建设基础设施吧。去启动它。给我展示50万个协同训练的GPU。这才是我现在感兴趣的，因为每个人都在谈大项目，每个人都说有无数千兆瓦级的资源即将到来。

Go build the infrastructure. Go launch it. Go show me a half a million coherently training GPUs. Like, that's what I'm interested in now because everyone has a big deal. Everyone has a gazillion gigawatts incoming.

拜托，去建那个该死的数据中心吧。就像我之前提到的，XAI似乎正在这条路上快速前进，但他们不是唯一一家。我们还有更多公司在为此努力。下一个是谁？微软。

Go build the damn data center, please. And I mean, like I mentioned earlier, it seems like XAI is very much on their way, but they're not the only ones. We have more companies that are working on this. Do we got next? Microsoft.

不，是另一个巨头萨提亚。

No. Another giant Satya.

有家小公司叫微软。上周萨提亚宣布了一个名为Fairwater的项目，本质上是微软的数据中心，他们之前尚未加入这场竞赛。我想指出这一点。他们此前一直在部分办公场所或邻近区域进行大量相关研发，但现在他们决定全力投入建设自己的人工智能超级计算集群。他说，如果智能是算力的对数，那么一切都始于庞大的算力基础。

Have a tiny company called Microsoft. So Satya last week announced this thing called Fairwater, which is basically Microsoft's data center, which they hadn't entered the race yet. I wanna point that out. They were working on a lot of this on-site at some of their offices or slightly just adjacent to their offices, but now they're going fully committed into building out their own AI super cluster. He goes, if intelligence is the log of compute, it starts with a lot of compute.

这就是为什么我们的GPU集群扩张速度远超任何竞争对手。他继续解释微软已经建成了10吉瓦的算力容量，这相当可观，但他的目标是再增加10吉瓦。乔希，我注意到所有这些公司都在吹嘘10吉瓦——这现在成了新的标准单位。如果我没记错的话，之前的标杆还是500兆瓦。

And that's why we're scaling our GPU fleet faster than anyone else. He goes on to explain how Microsoft had basically already built out 10 gigawatts of capacity, which is pretty big, but he wants to go much, much harder than that targeting another 10 gigawatts. I'm noticing that Josh, all of these companies are touting 10 gigawatts. It's the new gigawatt now, dude. And if I remember prior to that, it was 500 megawatts.

所以我们进步神速。现在几乎每个季度都以数量级增长。虽然看起来有点泡沫化，但这本质上是微软尝试建立超级计算集群来训练自己的模型。乔希，我觉得有趣的是，微软原本与OpenAI关系紧密——事实上他们一直是OpenAI的主要云服务与算力供应商，直到大约一个月前双方出现了微妙的分歧，现在微软更多地向Copilot用户提供Anthropic的模型和谷歌的Gemini模型。

So we've come a long way. We are going by orders of magnitude almost every quarter now at this point. Again, seems kind of bubbly, but this is basically Microsoft's attempt at building out a super cluster to train their own models. Now, what I find is interesting here, Josh, is Microsoft was very closely aligned with OpenAI. In fact, they were the premium cloud and compute provider for OpenAI up until I think about a month ago, where they sort of gone through a bit of a subtle breakup, where Microsoft is now kind of like supplying more of Anthropix models to their Copilot users and Google's Gemini models.

而OpenAI已宣布在2030年2月之后，他们将不再如此依赖微软。我认为这明确释放了信号：微软准备自立门户，甚至可能训练自己的模型。

And OpenAI has announced that after 02/1930, they're basically not really gonna be relying on Microsoft as much. So this I see as a clear signal that Microsoft is gonna be going out on their own and maybe even training their own model.

嗯。确实。这简直就是人工智能版的《权力的游戏》。规则就是如此——所有人都在为自己谋利。

Mhmm. Yeah. This is I mean, it's very much the this is the AI Game of Thrones. This is how it goes. Everyone is out for themselves.

有利可图时就结盟，无利可图时就毁约。在这场游戏中，真正的王者只有黄仁勋和他的GPU。没有这些GPU，一切都不可能发生。你现在开始看清这些博弈的运作方式了——微软只是在OpenAI还有价值时加以利用，一旦出现阻力失去价值，就立刻分道扬镳。

You have deals when it's convenient. You destroy those deals when it's not convenient, where the only person that matters in this, the king of them all, is Jensen and those GPUs. Because without those GPUs, none of this happens. So you you start to see how these dynamics work and play where I mean, Microsoft was using OpenAI to the extent that it was valuable. The second there became resistance and it lost value, see you later.

我们接下来要讨论另一个重磅话题。我认为这种趋势会持续下去。但还有更多消息——英特尔最近上了新闻，这家公司究竟是做什么的？

We're moving on to the next big one. And I think that's gonna be the nature that we continue to see as we go through. But there's more. Intel is in the news. What does Intel do?

自从英特尔停止为MacBook生产芯片后，我想我很久没接触过他们的产品了。

Intel I don't think I've touched Intel products since they stopped making the chips for the MacBooks.

那么他们现在到底在做什么呢？好吧，我们经历了漫长的过程——这里说的'漫长'可不是什么好事。过去三年间，英特尔的股价和整体基础设施，或者说他们生产的芯片质量，都出现了大幅下滑，老实说已经到了...

So what what are they doing here, please? Okay. So we're we've come a long way. And by a long way, I don't mean a good way. Intel's stock price and general infrastructure or quality of chips that they were producing declined pretty massively over the last three years, I would say, to the point where they I'm sorry to call it this, but it is.

他们需要政府救助的地步，Josh。你知道吗？美国政府现在持有英特尔10%的股份。

They needed to get a government bailout, Josh. Did you know that the US government currently owns 10% of Intel?

是的，你听说了？事实上我们一两周前的节目就提到过，这可是大事——政府正在投资私人市场，还持有股权。

Yes. You know that? In fact, we mentioned it on an episode a week or two ago, which was a huge deal. Like, the government is now investing in private markets, and they own a claim.

我从未见过这种规模的国家化行为，可能自工业革命以来都没有。美国政府对核心技术的介入程度前所未有——他们持有英特尔10%股份，英伟达的黄仁勋还投资了50亿美元。而本周爆出的新闻是：英特尔正请求苹果进行类似投资。Josh，我觉得有趣的点不在英特尔，而在苹果——我们正在分析各大科技公司的AI战略，但差点忘了提苹果...

I've never seen nationalization on this scale until probably the, until since the industrial revolution or to some extent, I've never seen nationalization of a core technology so aggressively as the US government today. And yeah, they own 10% of Intel, and, actually, Jensen Huang, NVIDIA invested $5,000,000,000 into Intel as well. And the breaking news this week is Intel has asked Apple to make a similar investment into them. Why I found this funny, Josh, isn't to do with Intel at all, but it's to do with Apple because we're giving a breakdown of the top companies and their strategy towards AI. And I'd be remiss if we didn't announce our friends or talk about our friends.

本来我根本不打算提苹果，因为他们既没有AI战略，也没有数据中心，更没有算力基础设施。但我想给他们个机会——如果真进行这笔投资，或许会成为明智的战略布局。如你所说，MacBook仍在使用英特尔芯片，未来苹果的AI产品可能会采用英特尔的设计芯片。可能我想得太乐观了，但我确实看到了这种可能性。

Wasn't I gonna mention Apple at all because they have zero strategy, zero data centers, zero gigawatts. But I I wanted to throw them a bone. If they do end up going forward with this investment, I do think it will end up being a really good strategic move for themselves. As you mentioned, they use a bunch of Intel chips for their MacBooks, and potentially they might end up using a design chip that Intel makes for whatever Apple AI product. Maybe that's me being too optimistic, but I see it in the future.

英特尔最近还与英伟达签署了一项非常紧密的合作协议，双方同意将共同开发特定的GPU架构和GPU产品，以实现其产品线间的无缝协作。总的来说，我认为苹果这次确实有机会一试身手。

Intel also signed a really close partnership with NVIDIA recently, which agrees between both of them that they're gonna be building out very specific GPU architecture and GPUs that work cohesively amongst their product suites. So I think all in all, Apple has a chance to to take a shot here.

好的。这确实说得通——苹果拥有自研芯片，并且一直在为其添加神经引擎核心来进行AI训练。在我看来，这则新闻意味着英特尔正逐渐成为国家安全问题，我们需要保住本土芯片制造能力，政府也介入表示：你们不能倒下，国家需要你们。黄仁勋，快去支援你的伙伴们。

Okay. It it makes sense that I mean, Apple has their own silicon that they make, and they've they've been adding these, like, the neural cores into it to do the AI training. To me, it seems like my interpretation of this news is is Intel is kind of becoming a national a matter of national security where we need to just onboard our ship manufacturing, and the government is getting involved and saying, like, you cannot die. We need you. Jensen, go help your friends.

蒂姆，也去帮帮你的同行。我们需要这些企业存活下去。这就是当前局势的实质。我不确定苹果对他们有多大需求——最近我刚看过iPhone的拆解报告。

Tim, go help your friends. Like, we need them to live. And that's kind of what we're seeing here. I'm not sure Apple, like, has much of a use for them. I was looking at an iPhone teardown recently.

他们使用苹果自研芯片，高通处理器负责蜂窝网络，基本就这些了。所以这可能更像是政府在说：伙计们必须去支援，我们会减免关税之类的——至少这是我的解读。考虑到英特尔未来可能获得的大量补贴，我对他们持乐观态度。

They have the Apple chips. They have Qualcomm processors for cellular, and that's like, that's kind of it. So this might just be a like, hey, guys. You gotta go help them, and, like, we'll spare you on tariffs and things like that is my take at least. I'm kinda bullish on Intel just through all the subsidy that I presume they will be getting going forward.

那位前OpenAI研究员叫什么来着？利奥波德·阿本施赖伯？对，这家伙赚翻了。

What was the name of that ex OpenAI researcher? Leopold Abenschreiber? Oh, yeah. Dude made a killing. Yeah.

没错。他离开OpenAI后成立基金，据说募集了10亿美元，专门投资符合他撰写的AI/AGI发展宏论的企业。该基金最大的一笔投资就是英特尔，当时还没人看好英特尔，如今这笔投资已浮盈120%。

Yes. He left OpenAI, and he started a fund. I think he raised a billion dollars to invest in companies that were in accordance to this massive thesis that he wrote on how AI and AGI was gonna pan out. And one of the biggest investments that he made with that fund was in Intel. This was before anyone was talking about Intel, and he is now up a 120% since then.

干得漂亮。

Bravo.

所以其中之一

So one

是有史以来回报率最佳的投资人之一。

of the best investors in returns ever.

EJS，我最后还有一个问题要问你。关于Warrior's Anthropic。他们似乎是家大公司对吧？就是那个开发Claude、长期引领前沿代码模型的Anthropic。

EJS, I have one last question for you. Warrior's Anthropic. They seem to be a big comp yeah. Right? Like, Anthropic, the ones who make Claude, the ones who, like, leading code model with the frontier for a long time kind of.

哦，你是说那个天天唱衰AI，自己却是个AI公司的家伙？抱歉Dowdy，我必须直说。好吧，让我用马斯克本人的推文告诉你他们的现状——我刚好把这条调出来了：'Anthropic从一开始就注定不可能成功'。

Oh, you mean the one that's fudding AI every day, despite them being an AI company? Sorry, Dowdy. I gotta call it out, right? Okay, so I'll tell you exactly where they are through the words of Elon Musk himself, which is a tweet that I have pulled up here. Winning was never in the set of possible outcomes for Anthropic.

这句话够狠的——Anthropic根本没有稳定的算力供应商。他们在数据中心建设上投入不足，资金链快断了，背后支持者也撤了。Josh，我有个观点你可能不认同，不知道算不算劲爆。

An absolutely savage sentence to describe that Anthropic hasn't really got any compute providers. They haven't really made much effort or as much investment in data centers, and they're running out of funding. They're running out of backing. Josh, I have a take that you may or may not agree with. I don't know how hot this is.

听着。我认为你最喜欢的苹果公司会收购Anthropic。

Okay. I think your favorite company Apple acquires Anthropic.

这观点够劲爆。那会挺疯狂的。我不确定苹果收购后会让Speedy怎么办？

That's a hot take. That'd be kind of crazy. I'm not sure what Apple would do Speedy with No?

他们需要一个

They need a

能够自主掌控的模型。

model that they can own themselves.

但他们需要

But they they need

一个小型模型。他们需要一个能在iPhone上运行的模型，一个能提升我MacBook性能的模型。我不确定Anthropic是否值得，因为苹果正在开发相当不错的本地模型。说实话，这并不难。

a small model. They need a model that can run on iPhones. They need a model that can, like, improve my MacBook. And I'm not sure anthropic is worth that because, like, Apple's making pretty good local models. And to be honest, it's not that hard.

事实上，未来一年的开源模型可能会足够优秀，他们可以直接选用一个，根据工作流程定制，然后就能用。我不认为苹果会开发最前沿的大语言模型。

In fact, open source models in the next year are probably gonna be good enough where they can just grab one, make it custom for their workflow, and then it's good. Don't think Apple's going to be making bleeding edge LLMs.

好吧。那也许它们只是两家失败的公司，因为那个领域不知为何缺乏资金支持。

So Okay. So fine. Then maybe they're just two loser companies because that topic, for some reason, doesn't have the funding.

是的。在我看来，当前AI竞赛有三个主要玩家：谷歌、OpenAI和XAI。其他没提到的都不太有意思。没错。

Yeah. I the way I see the AI race currently is there's three players. It's Google, it's OpenAI, and it's XAI. Everybody not listed there is not really interesting. Yep.

我们最终可能会面临某种消亡，或者干脆在技术前沿不复存在。这包括微软，包括Perplexity，包括Anthropic，包括任何其他公司。

And we'll probably have some sort of eventual demise or just will not exist at the frontier. That includes Microsoft. That includes Perplexity. That includes Anthropic. That includes any any anybody else.

这确实极具挑战性。随着这些数字越来越大，基础设施不断扩张，情况只会变得更困难。这是我们从未尝试过的巨大规模，真的会非常艰难。所以现在能在这个领域占有一席之地已经非常了不起。光是未来六个月里这些基础设施建设的进展，就足够引人入胜了。

It's it's really challenging. And as these numbers get bigger, as the infrastructure grows, it's only going to get harder. This is a, like, a tremendous scale that we've never tried to do before, and it's really going to be difficult. So to even have a seat at the table now is very impressive. It's gonna be really interesting to see, I mean, just over the next six months, how all these build outs are gonna go.

因为如果你在建设进度上落后，如果无法真正完成这新增的10吉瓦算力，你就完蛋了。就这么简单。技术前沿会抛下你继续前进。

Because if you are late on a build out, if you cannot actually complete this new 10 gigawatts of compute, you are cooked. That's it. The frontier has moved on without you.

说到完蛋，我想花点时间聊聊欧洲。我们刚才讨论的都是美国的公司。美国！美国！美国！

Well, speaking of being cooked, I wanna take a second to talk about Europe. What we've what we've just discussed are all the companies that are in America. USA. USA. USA.

你可能在想：世界其他地区在做什么？我们大洋彼岸的朋友们情况如何？答案是并不乐观，但他们正在庆祝极其微小的胜利——我说的是4000块GPU那种微小。我们刚谈到多家公司投资数百亿美元建设数据中心，部署数百万块GPU。

And you might be wondering, well, what's the rest of the world doing here? How's our friends over across the sea? The answer is not very well, but they're celebrating incredibly tiny wins. And I mean, 4,000 GPU tiny. So we just spoke about a bunch of data centers that are being invested in by separate companies to the tune of hundreds of billions of dollars for millions of GPUs.

德国正在庆祝部署了4000块GPU来帮助他们实现数字主权。

Germany is celebrating setting up 4,000 GPUs to help them become digital digitally sovereign.

4000块？我以前做过黄牛，会用机器人抢购新发布的英伟达显卡转卖。感觉我个人积累的数量都快赶上他们全国了，这太离谱了。4000块GPU？我们现在讨论的可是百万级规模的连贯GPU集群啊。

4,000. I I used to do, like, a reselling reselling where I would buy the NVIDIA the new hot NVIDIA GPUs when they first came out with using bots and, like, resell them. And I feel like I I have accumulated, like, a comparable amount to the entire country. That's, like, very disturbing. 4,000 GPUs is like we are talking about a million coherent GPUs.

哇哦，真为你高兴，但你们欧洲的朋友们还有些工作要做呢。

So Oh, wow. I'm happy for you, but you have some work, my European friends.

这工作还很严肃。他们会不断加码监管——这始终是我对欧洲最大的不满。任何炫酷新技术诞生后，他们总是过度监管到令人窒息，逼得创业者想逃离欧洲来美国发展。我不怪他们，但乔什，确实存在一个对我们构成实质威胁的外国对手，可能影响美国赢得这场算力竞赛。

This serious work too. Keep they're gonna keep over regulating. That's the biggest gripe that I've had with Europe in general. Whatever new fancy technology takes place and gets created, they just overregulate the hell out of it so much so that founders wanna leave the country or the region, and they wanna come over and build an American. I don't blame them, but there is a foreign adversary, Josh, which presents themselves as a real threat to The US winning this compute race.

实际上，我认为这不是美国内部的竞争，而明显是中美对决。有两点特别值得强调，这个话题足够单独做期节目：第一，据Grok和OpenAI数据，中国过去十年秘密储备了超过3300吉瓦的能源供应——

In fact, I wouldn't actually say that it's America versus America. I would frame it as it's very much America versus China. There there are two real things that I wanna highlight here, because there's a lot we can cover here. It probably deserves an entire episode to itself. So point number one is China has secretly amassed over 3,300, that's according to Grok and OpenAI, gigawatts of energy supply over the last decade.

这些建设早在AI成为主流话题前就完成了。众所周知中国长期是基建和制造王者，擅长打造可规模化、高效节能的廉价技术。现在他们正将全部能源导向训练顶级AI模型，但有个致命短板，乔什——

This is before AI became a mainstream topic and use for this energy. They just built out this infrastructure. And as we know, China has been kind of the infrastructure and manufacturing kings for a while now. There are the kings of building out cheap tech that can scale, that's efficient, cost efficient, energy efficient, all of the efficiencies, right? And they're now funneling all this energy towards training some of the best AI models, but they had a linchpin, Josh.

他们缺乏顶级硬件芯片，特别是GPU。所以依赖英伟达和黄仁勋。这条推文引出第二个重点：2022年美国禁止向中国出口高端英伟达GPU以延缓其AI发展，结果中国三年内就实现了国产AI芯片的赶超。

They didn't have the best hardware chips, GPU specifically. So they relied on our friends at NVIDIA, on Jensen Huang. And this tweet kind of highlights another major concern, which is the second topic. In 2022, The US banned high end NVIDIA GPU exports to China, wanting to slow down AI development in China. So China invested in homegrown AI chip production and caught up in three years.

现在中国反过来禁止采购英伟达AI芯片。这带出第二个故事：中国禁止科技公司购买英伟达芯片。我深入调查过此事，乔什，上周的新闻简报就有详细报道，建议大家去看看。

Now they are banning buying NVIDIA AI chips. And this highlights the second story, which is China's banning tech companies from buying NVIDIA AI chips. And I dug into this, Josh. In fact, I covered this in our our most recent newsletter, which we released last week. You guys should definitely go and check this out.

报道对比了阿里和华为研发的芯片，发现其性能已接近英伟达上一代GPU（非最新款）。他们以难以置信的速度构成了实质威胁。

And it compared the chips that Alibaba had created and Huawei. And they are as good as not the latest iteration of NVIDIA GPUs, but the iteration just before this. So they have fought up so unfathomably quickly that they present themselves as a real threat.

嗯，是的。关于中国的情况，我们在之前与DeepSeek及其他类似开源公司的多期节目中讨论过，他们非常善于在资源有限的情况下发挥创造力，将资源约束视为最大的优势，并能在限制范围内进行优化。因此，一旦他们获得额外资源，这些优化就会爆发式增长。我们现在看到的是，许多大型语言模型公司和AI实验室已经习惯了在训练模型时面临巨大限制，但他们依然处于前沿，因为他们在软件优化上做得非常出色。

Mhmm. Yeah. The the thing about China, and we talked about this in a lot of previous episodes with DeepSeek and other open source companies like that, is they are so good at being resourceful where they can take the limited resource constraint as the biggest benefit to them, and they're able to optimize within the confines. So the second they unlock additional resources, those optimizations bubble out into this, like, huge explosive growth. And what we're seeing now is is they've been, I mean, a lot of these large language companies, these AI labs, they've been accustomed to having these tremendous amounts of constraints on how they can actually train these models, but they've still been at the frontier because they've optimized the software so much.

现在中国开始自主研发优化的GPU，这确实是一件大事。因为回顾他们过去制造GPU的方式，虽然不算顶尖，但他们一直很善于在资源有限的情况下创新，毕竟他们没有英伟达那样的资本支出和资源。如今，他们正在加速发展，并且显然已经在软件方面取得了突破，现在又加上硬件方面的进步，这确实值得密切关注。

So now that China is starting to build their own in house optimized GPUs, well, that's that's a really big deal because if you think about the way they've been building GPUs in the past, I mean, haven't been that good, but they've been kind of being resourceful because they don't have all the the CapEx that NVIDIA has. They don't have all the resources NVIDIA has. Now that they are starting to really ramp that up and they have they clearly have the software side. Now they're getting the the hardware side in addition. That becomes a serious thing to look out for.

英伟达的垄断地位突然看起来有些动摇，因为中国的发展速度实在太快了。中国快速发展的好处是，美国可以拥有更多英伟达的资源。而由于美国有更多英伟达的资源，我们的所有项目也能更快推进。所以我认为这对中国和美国都是好事。

Like, NVIDIA's monopoly suddenly starts to look a little shaky because, I mean, at the same time, China's moving really quick. The benefit of China moving very quick means there is more NVIDIA for The US. And because there's more NVIDIA in The US, all of these projects that we have grow much quicker. So I think it's it's good for China. It's good for The US.

我们只需要确保这对美国没有负面影响，并且我们继续保持快速发展的势头，因为天啊，中国的速度实在太快了。如果我们不能保持这种高强度的发展速度，我们也会陷入困境。正如之前提到的，如果你落后于这条曲线，它变化得非常快。如果我们落后一个周期，被甩在后面，那将是巨大的差距，我们会明显感受到。

We just need to make sure that it is not bad for The US and that we still continue to move quickly because my god, China is moving very fast. And if we don't keep up this hardcore rate of acceleration, we are also going to be in trouble because, I mean, like we mentioned earlier, if you are late on that curve, it moves very quickly. And if we are one cycle behind, if we get left in the dust, that's a huge difference that we're going to notice.

有趣的是，我眼中的美国与中国对比是这样的：美国和西方会砸大量钱解决问题，而中国没有那么多钱，所以他们必须更善于利用资源，正如你所说。而且这不仅限于芯片或能源。

It it's funny. The the way I picture America versus China is America and the West throw money, tons of money at the problem. China doesn't have as much money, so they need to be more resourceful as you point out. And it's not just chips. It's not just energy.

对吧？全球超过50%的顶尖AI研究人员来自或居住在中国。所以他们有高密度的人才，也有制造和规模化能力。

Right? Over 50% of the world's top AI researchers reside and are produced in China. Right? So they have the talent density. They have the manufacturing and scaling capability.

他们有能源，现在可能还会有芯片。如果2026年我们看到一些最优秀的AI模型来自中国，我一点也不会感到惊讶，但我们——

They have the energy, and now they might have the chips. I wouldn't be surprised if we saw some of the best AI models in 2026 come out of China, but we'll

等待一些信号出现，对吧？比如第一个信号是中国禁止外部GPU，这表明他们对自己的GPU技术已有足够信心，将强制所有人转向国产。下一个可能出现的信号——也是我们该警惕的时刻——是他们停止开源这些模型。他们一直开源所有模型，因为这是推动迭代发展的方式。

wait to some signals. Right? Like, this this first signal is China's banning external GPUs, where now they feel confident enough in their own GPUs that they will force everyone to go in house. The next signal we probably get, which is when we should start raising some red flags, is when they stop making these models open source. They've been making all the models open source because that's how you kind of get this iterative development.

这样才能吸引更多开发者。一旦这些模型转为闭源，一旦中国开始使用自主研发的GPU，事情就会变得非常可怕。这意味着他们已经迎头赶上，意味着我们将势均力敌，他们有信心将所有技术封锁在内部。这是一场高风险竞赛，很可能成为新冷战。

That's how you access more developers. Once these models start to become closed source, once they start to have their own in house GPUs from China, that's when things get really scary. That means that they have caught up, and that means that we are then we are the neck and neck, and they're confident enough to keep all of this now behind closed doors. And this is a very high stakes race. This is probably the new Cold War.

对吧？因为在我们真正实现AGI之前，谁也无法预知其连锁反应，但可以想象它极其重要且影响深远。

Right? Because, I mean, you never really know what the downstream effects of reaching AGI will be until we get there, but you can assume it's probably pretty important and pretty impactful.

以上就是顶级AI公司及其算力扩张的全景分析。你可能注意到我们没提谷歌——因为他们对很多事守口如瓶。虽然大量采购英伟达芯片和GPU，但他们也有自研TPU用于模型训练。等我们挖到更多信息，节目里一定会讨论。

So that is the breakdown of all the top AI companies and their scaling efforts. You might notice that we didn't mention Google. That's because Google keeps a lot of stuff under wraps. They buy a hell of a lot of NVIDIA chips and GPUs, but they also have their own in house TPUs, which they use to train these models. So once we dig up more information, we'll we'll definitely talk about that on the show.

但Josh，我想用最后一个问题结束本期节目，这问题你并不陌生：我们是否处于AI资本支出泡沫中？有趣的是上期我就提过——嗯哼——但自那之后发生太多事，我必须再问你一次。

But Josh, I wanna kind of round this episode up with one final question, which isn't one that you're gonna be unfamiliar with. Mhmm. Are we in an AI CapEx bubble? And it's funny because I mentioned this in the last episode, but Uh-huh. So much has happened since that last episode that I need to ask you again.

是我们疯了，还是这确实合理？

Are we crazy or is this valid?

我的答案可能和上期一样：从长期看，计算力将是唯一且最后的投资重点；中短期来看...我不确定。在资金枯竭前，在无法向投资者证明其收益能覆盖成本前，还能砸多少钱？方向上看，这绝对不算泡沫。

I think I probably have the same answer as I did last episode, which is Okay. Compute will be the only and last thing we will need to spend money on at a long time scale, at a short to medium term time scale. I don't know. How how much money is left to throw at this before it starts to run dry, before you can no longer prove to investors that it is worthy, it is able to generate the revenues required to offset the costs? It seems I mean, directionally, this is absolutely not a bubble.

这才是真正重要的事。这将是我们永远唯一需要花钱的地方。将能量转化为智能的潜力是巨大的。但在这个过程中，我也说不准。我们进展得很快。

This is the real thing. This will be the only thing we spend money on forever. The conversion of energy to intelligence is tremendous. But along the way, I don't know. We're moving quick.

而且，就像我们说的，10吉瓦的规模非常困难。这比我们现在的水平高出一个数量级。如果他们能建成，那么我们可能会延长这个泡沫期，甚至可能根本看不到它的破裂。但如果我们在开发过程中每隔几个月到一年就要实现数量级提升时开始遇到瓶颈，那时情况可能会变得有些不稳定。不过这些数字确实在迅速变大。

And, I mean, like we said, 10 gigawatts is very hard. That's an order of magnitude bigger than where we are. If they can build it, then we are probably going to extend the length of this bubble, maybe not even see it. But in the world where we start to run up against walls in this development of these these order of magnitude gains every couple of months to a year, that's when that's when things can get a little bit shaky. But the numbers are getting big quickly.

所以我得问问，你怎么样？你有

So I have to say, what about you? Do you have

什么特别强烈的倾向吗？

any strong feelings either way?

没有。我...我觉得如果你不把大部分时间、金钱和精力都用在训练最好的AI模型上——其中很大一部分其实就是花钱在计算基础设施上——那你肯定会输。你甚至从一开始就没打算参与竞争。所以我认为这个策略是合理的。但最终这会是个制胜之举吗？

No. I I I think if you're not spending most of your time, money, and effort on training the best AI models, a large part of which is just simply spending money on compute infrastructure, you're gonna lose. You're not even trying to play in the first place. So I think it's valid. Will this end up as a winning move?

说实话，我觉得会有很多失败案例。事后我们会发现，这些钱有很多都花错了地方。我们还没讨论过可能出现的新情况：如果诞生了全新的AI架构，可能会进一步降低训练前沿模型的算力成本。当然这有点像是戴着锡箔帽子的异想天开。但目前来说，我认为这个方向是对的。

I think there'll be a lot of failure, if I'm being honest with you. I think we'll find out retroactively or in hindsight that a bunch of this money was misspent. We didn't talk about the potential red herring of having a completely new AI architecture be produced, which drives the cost of compute down even more to train a frontier model. And of course that's kind of like not wishful thinking, but kind of with my tinfoil hat on. But I think right now, it's the right move.

嗯。

Mhmm.

是的，听起来差不多是这样。我们拭目以待吧。我们将站在前沿报道一切，保持大家的同步更新。我非常期待了解更多关于谷歌的信息。

Yeah. That sounds about right. And again, we'll see. We'll be here right on the frontier covering it all, keeping everyone up to date. I am very excited to learn more about Google.

我对谷歌非常乐观，但他们在处理一切事务时非常低调。公平地说，Anthropic也相当保密。所以还有很多未知之处，但一个10吉瓦的数据中心能隐藏的方式毕竟有限。从太空都能看见它。

I am so optimistic on Google, but they are very private in how they handle everything. You And know what? To be fair, Anthropic is fairly private too. So there's a lot still left to be known, but there's only so many ways you can hide a 10 giga watt data center. Like, you can see that from the stars.

所以人们会发现真相的。我们将获得所需的答案，并在这个节目中为大家报道。不过，我想就到这里吧，Ejaz。结束前还有什么想说的吗？

So people will find this out. We will get the answers we need, and we'll report them right back here on the show for everyone to hear. But, yeah, I guess that's it, Ejaz. Any parting thoughts before we wrap here?

没有了。希望这次概述能帮助大家理解建设这些数据中心所投入的巨大努力和资金，以及它们为何要这样做。我认为这是最值得关注的竞赛，我们会在后续节目中紧密追踪。另外，我们收到了大量听众反馈。

Nope. That's it. I hope that this was a useful overview in understanding the pure effort and money that is going into building out these data centers and why they're building these out. I think it's the most important race to watch, and we're gonna be covering it very closely as we produce more episodes on the show. Separately, we've been getting so much feedback from you guys.

每期结尾我们征集反馈时，总想着或许能多几条评论。结果YouTube评论和私信都堆满了，这些都是最有价值的建议——从节目使用的图表、结束方式、讨论话题，到我们可能遗漏或错误的内容。感谢所有指出问题的人，也感谢为我们加油的观众。

We ask for feedback at the end of every episode thinking, oh, maybe we'll get a few more comments. Our YouTube comments and our DMs are piled up, and it is some of the most useful feedback that we've got. And it's everything from the graphics that we're using on these episodes to the way that we sign off these episodes to the topics that we discuss, to things that we might have missed and things that we might be wrong on. So I appreciate all of the people that are calling us out. I appreciate all the people that are cheering us on.

如果还没给我们的节目点赞订阅，请支持一下。如果你认识可能对我们话题感兴趣的人，请分享给他们。下期节目再见，祝安好。

If you aren't liked and subscribed to any of our episodes, please like and subscribe. And if you have anyone that you think might be excited by the episodes and stuff that we talk about, please share it with them. And we will see you on the next one. Peace.