谷歌AI的800亿美元之问

谷歌的母公司字母公司正在尝试通过股票发售筹集800亿美元。

Alphabet, the owner of Google, is trying to raise $80,000,000,000 in a stock sale.

其中100亿美元将来自伯克希尔哈撒韦，这简直让人难以置信，而这一切都是为了支持他们的AI研发。

10,000,000,000 of that is gonna come from Berkshire, which is pretty crazy, and all of this is to fund their AI build out.

与此同时，特朗普签署了一项范围更小的AI行政命令。

At the same time, Trump is signing a narrower AI executive order.

他之前受到了AI行业的强烈反对，所以做了调整。

He got a lot of pushback from the AI industry and has adjusted.

我会聊聊这项命令的利与弊。

I'll talk about the pros and cons of that.

GitHub Copilot新推出的按使用量计费模式，一天就花光了用户的整月预算。

GitHub Copilot's new usage based pricing is burning through people's entire monthly budgets in a single day.

Opal公司正在转向AI硬件领域。

Opal is pivoting to AI hardware.

他们从OpenAI获得了4000万美元的投资。

They received 40,000,000 from OpenAI.

他们现在的估值达到2.75亿美元，而优步在员工不到四个月就花光了全年预算后，决定将每个AI工具的月度支出上限设为1500美元。

They're now valued at $275,000,000, and Uber is capping their employees' AI spending at $1,500 per tool after they burnt through the entire annual budget in under four months.

如果你想每天收到这些AI新闻摘要，可以直接在aichatdaily.com的订阅页面上订阅我的通讯。

If you wanna get all of these news stories straight into your email inbox every day, I have a newsletter you can subscribe to on the subscribe tab of aichatdaily.com.

我刚刚才看过。

I was just looking at it.

各位，我的每日AI通讯打开率高达40%。

Guys, I have a 40% open rate on a daily AI newsletter.

也就是说，我发了10封邮件，每一封都有40%的用户会打开——这可是有成千上万订阅者的情况。

Meaning, I sent out 10 emails, 40% of people are gonna open every single one for something with thousands of subscribers.

我真的很佩服。

I'm super impressed.

这份通讯是专门为真正动手做项目的人打造的。

It is built for people that are actually shipping stuff.

我能看到订阅者的邮件地址。

I can see the emails of the people that subscribe to it.

我的读者里有来自谷歌、OpenAI、Meta、Databricks、微软、亚马逊、Techstars、Copy AI、Eleven Labs、TikTok、思科、DeepMind，还有好多其他顶尖公司的人都在看这份通讯。

So I have people from Google, OpenAI, Meta, Databricks, Microsoft, Amazon, Techstars, Copy AI, Eleven Labs, TikTok, Cisco, Google, DeepMind, and tons of other incredible places that all read this newsletter.

如果你想每天掌握所有AI领域的最新动态，赶紧去 aichatdaily.com/newsletter 订阅吧。

So if you wanna get everything that's happening with AI, go check out aichatdaily.com/newsletter.

链接在简介里，或者直接点 aichatdaily.com 页面右上角那个醒目的订阅按钮。

There's a link in the description or it's the big subscribe button on the top right hand corner of aichatdaily.com.

第一条新闻是， Alphabet 通过股票发售筹集了800亿美元。

The first story is that Alphabet is raising $80,000,000,000 through stock sales.

这次发售的核心是，伯克希尔·哈撒韦公司投了100亿美元进来。

This the the big anchor on this whole sale is that they're getting $10,000,000,000 from Berkshire Hathaway.

稍后我会详细说说这件事。

More on that in a second.

现在的情况是，谷歌和它的母公司 Alphabet 在AI基础设施上的投入今年直接飙升到了800亿到1900亿美元之间。

What's going on right now is that the AI infrastructure spending of Google and, you know, Alphabet, the owner of Google, has absolutely exploded to 80 to $190,000,000,000 this year.

但问题在于，这笔支出已经超过了他们的经营性现金流。

And the problem is that that's actually outpacing their operating cash flow.

所以他们不得不转向股权和债务市场融资，因为整个超大规模云服务商群体——包括亚马逊、微软和其他所有竞争对手——都正朝着2027年合计投入一万亿美元用于AI建设的方向迈进，这简直不可思议。

So it's basically forcing them to tap into equity and debt markets as the entire hyperscaler cohort, all of the other competitors like Amazon and Microsoft and everyone else, are all collectively moving towards $1,000,000,000,000 in AI spending by 2027, which is wild.

这次融资具体包括约三千亿美元的承销发行、一千五百亿美元的强制可转换优先股，以及将在第三季度启动的四千亿美元随行就市发行计划。

This raise in particular includes about $30,000,000,000 in underwritten offerings, $15,000,000,000 in mandatory convertible preferred stock, and a $40,000,000,000 at the market program launching in q three.

伯克希尔哈撒韦这次投入了一千亿美元，但他们的持股其实一直在持续增加。

So Berkshire Hathaway's position has you know, they're putting in $10,000,000,000 now, but their position has been growing.

这对他们来说根本不是什么新鲜事。

This isn't like a new thing for them.

去年十一月他们还只有四百三十亿美元的持仓，现在已接近两千亿美元了。

They only had about $4,300,000,000 in November, and they're up to almost $20,000,000,000 now.

在苹果之后， Alphabet 是伯克希尔最大的科技持仓之一。

This is one of Berkshire's largest tech holdings after Apple.

Alphabet、微软、Meta 和亚马逊加在一起，今年的资本支出预计将达到七千亿美元，而华尔街普遍预测，到2027年这一数字将突破一万亿美元，就像我之前说的那样。

Alphabet, Microsoft, Meta, Amazon combined, they're all tracking towards the $700,000,000,000 in CapEx this year, and I think Wall Street is saying that there's gonna be about a trillion dollars, like I mentioned, by 2027.

Alphabet 等于直接承认了，华尔街早就看透了这一切。

Alphabet just admitted basically that Wall Street already knows something.

这些超大规模云服务商没法自己掏钱搞定所有这些AI基础设施建设。

The hyperscalers can't self fund all of these, you know, all of these AI build outs that they're trying to do.

伯克希尔这100亿美元的投资，我觉得意义重大，哪怕只是传递了一个信号：他们向来擅长做明智的投资。

So Berkshire's $10,000,000,000 vote of confidence, I think, is a pretty important thing if it's just if nothing other than like a signal that they're like, look, like, they make smart investments.

他们相信谷歌有能力还上这笔钱。

They believe Google's gonna be able to pay it back.

我觉得微软、Meta和亚马逊接下来也会各自发起规模庞大的融资。

I think you can expect Microsoft, Meta, and Amazon to follow in their own kind of mega raises.

这挺有意思的。

It's interesting.

我们早就习惯了看到Anthropic和OpenAI搞巨额融资，但现在连谷歌都得出去筹钱，因为实在太烧钱了。

We're all used to seeing Anthropic and OpenAI do these huge raises, but it's like Google's having to go out and get more money too because it's just so expensive.

但把这些基础设施建起来，价值是巨大的。

But it's incredibly valuable to build these out.

对吧？

Right?

我们甚至看到XAI和SpaceX，他们为训练Grok建造了庞大的数据中心，现在完全可以转手利用。

We see even x AI and SpaceX, right, with their massive data center build out that they used for training Grok, they could go and turn around.

即使Grok的使用量没那么夸张，他们也能回头把数据中心租给Anthropic，每月收15亿美元。

Even if the Grok usage wasn't insane, they turn around, and they can sell that back to Anthropic for $1,500,000,000 a month.

这些数据中心和芯片采购里，藏着巨大的赚钱机会。

So there's a lot of money to be made in these data centers and in buying all of these chips.

明白了。

Okay.

特朗普今天签署了一项行政命令，要求AI公司在他发布前沿模型前30天自愿提交备案。

Trump is signing an executive order today where he's asking AI companies to voluntarily submit their Frontier models thirty days before they release them.

原本是90天，说实话，提前三个月，我觉得对竞争来说太早了。

Originally, it was ninety days, which I'm gonna be honest, three months ahead of time, I think, is too much if you're trying to be competitive.

我们得跟中国和其他国家抢速度。

We're trying to be competitive with China and a lot of other countries.

要是所有模型都得被政府扣住90天——我也不觉得政府真能有效监管这些事，但也许确实需要个监管机构，就像航空管理局管飞机那样。

If all of our models were gonna get held you know, ninety days for the government, which I don't even think is that great at monitoring everything, but I guess you have to have some sort of regulatory body perhaps, right, like the Federal Aviation Administration for airplanes.

也许AI领域也需要这样的机制。

Maybe we need that for AI.

我不太喜欢监管，但也许确实有必要。

I don't really like regulation, but perhaps it's needed.

不管怎样，能把九十天改成三十天，我觉得挺好。

Anyways, I'm happy to move it down from ninety to thirty days.

而且还是自愿的。

And it's also voluntary.

对吧？

Right?

所以并不是强迫他们这么做。

So it's not like they're being forced.

之所以做这些调整，是因为硅谷那边提出了反对，有人觉得这是优点，也有人觉得是缺点。

Now the reason why they've made a bunch of these adjustments, Silicon Valley pushed back, which I think some people will say this is a pro, some people will say this is a con.

他们就说，你看嘛。

They're like, oh, look.

硅谷在左右政府，告诉他们该怎么做。

Silicon Valley is pushing around the government telling them what to do.

或者说，政府只是在倾听这个领域有专业见解的人的意见。

Or you could say maybe the government is listening to the people who have expertise in this area.

事情总有两面性，你知道的。

There is, you know, two sides to the coin.

我觉得目前联邦监管本质上还是自愿的，没有强制性的许可制度，但确实已经建立了一个框架。

I think right now this is showing that the federal oversight is basically voluntary, so there's no mandatory licensing regime, but we do have some sort of framework in place.

我觉得他们现在最看重的是速度。

And I think they're just right now trying to prioritize speed.

他们也在做一些安全检查，但眼下真正要紧的是和中国竞争。

They're doing some safety checks, but we're really trying to compete with China at this point.

这份行政令里还有几处有意思的规定，明确禁止对AI模型设立强制性的联邦许可或事前审批，目前只保留自愿合规，我觉得这很好。

A couple other interesting things in this executive order, it explicitly forbids creating a mandatory federal licensing or pre clearance for AI models, so it is locking in this kind of voluntary compliance only for now, which I think is great.

特朗普在五月还推迟签署了更严格的一版。

Trump also delayed signing a much stricter version in May.

风险投资人大卫·萨克斯和业内其他一些人士一直在游说，反对延长审查周期，也就是我提到的那九十天。

Venture capitalist David Sachs and a bunch of other figures inside of the industry were lobbying against this kind of longer review window, the ninety days, like I mentioned.

这项命令要求司法部将人工智能辅助的黑客行为列为执法重点，同时给检察官更明确的授权，去追查由模型协助的入侵行为。

The order directs the Department of Justice to treat AI assisted hacking as a high priority enforcement area, and it also gives prosecutors a lot clearer mandates to pursue model insisted assisted intrusion.

这里面涉及的内容可真不少。

So there's a lot going on in here.

我很高兴政府开始关注这个问题，也很高兴他们没有像欧盟那样采取过于严苛的监管，毕竟那已经严重抑制了欧盟地区的人工智能创新。

I'm happy that the government is looking at this, and I'm happy they're not being too heavy handed on the regulation like we've seen with, you know, the EU and how that stifled a lot of AI innovation coming out of the EU.

好了，GitHub Copilot 现在遭到了大量批评。

Alright, GitHub Copilot is getting a massive backlash.

他们昨天刚调整了定价，开始按每积分 0.01 美元收费，也就是差不多一便士一个积分。

So they switched their pricing yesterday, and they started to charge 0.01 per credit, like so basically 1p per credit.

用户反映，他们一个月的配额有时候一小时甚至一天就用光了。

And users have reported that they are using their entire monthly allotment in a single hour or a single day.

如果你订阅的是每月十美元的高级版，现在只包含一千五百个积分。

If you're on the $10 a month pro plan, that now includes only 1,500 credits.

有个用户一天就用掉了840个积分。

One user burned 840 credits in one day.

另一个用户说，他24小时内就把整个月的8000个积分用光了。

Another user said they used 8,000 monthly credits in twenty four hours.

我得说实话，现在Anthropic和其他不少公司都在大力补贴用户的token使用量。

The the thing that I'm gonna be honest here is we have Anthropic and a lot of other players heavily subsidizing their token usage.

我觉得谷歌和OpenAI也在补贴，但Anthropic在他们的Claude Max计划里补贴得最狠，你要是每月花200美元的话。

I think Google and OpenAI are also subsidizing it, but Anthropic most definitely is on their ClonMAX plan if you're gonna pay the $200 a month.

他们确实在贴钱补贴这些积分。

They are subsidizing those credits.

有意思的是，我觉得GitHub是第一个踩下刹车的。

And what's interesting is we have GitHub as kind of the first shoe to drop in my opinion.

微软 owns GitHub，按理说他们资金几乎无限，理应拼命推广大家来用他们的产品。

Microsoft owns GitHub, they should have, you know, technically unlimited money, and they should really just be trying to get everyone to use their product.

但他们反而说：兄弟们，你们得自己掏钱，而且积分很快就用完了。

But instead, they're saying, look, guys, you you gotta pay for it, and you're gonna burn through your tokens super quick.

如果你想升级，那就得全自费。

If you wanna upgrade, you just gotta pay for it all.

所以我觉得，未来很可能会出现这么一天：OpenAI 和 Anthropic 都上市了，成了公众公司，得接受更多 scrutiny。

So I think right now we're gonna see probably a future when OpenAI and Anthropic have IPO ed and they're public companies and they're under a little bit more scrutiny.

那时候，他们就没法再这么免费送资源了。

They're not gonna be able to just give away all of these free handouts.

不过，也许将来我们能降低模型的成本。

Now maybe we can make models cheaper in the future.

对吧？

Right?

比如优化模型，或者建更多数据中心，这样成本就能降下来。

Like, maybe we can optimize them and maybe data centers, we have more of them and we can make things cheaper.

但数据中心不是说建就能建的，得花好几年，这不可能是立竿见影的解决办法。

But data centers take years to build, so I don't think that's gonna be our instant solution.

至于模型优化，我还是很有信心的。

Model optimization, I have a lot of hope there.

但说实话，我预测未来一年内，这些补贴项目会大量终止。

But if I'm being honest, my prediction for the future is that we are going to see in the next year a lot of these subsidized programs end.

那我给大家的建议是什么？

So what is my recommendation to everyone?

当然，我这个判断也可能完全错了。

And by the way, could be completely wrong about this.

我的建议是，如果你现在就想做点东西，就花每月200美元订阅Claude Max。

My recommendation is to get a Claude Max subscription for $200 a month if you want to build something and build it today.

趁这些补贴还在，赶紧动手，能用就用上。

Build stuff as soon as you can while you can take advantage of, a lot of these subsidized programs.

我本人是言出必行的。

I'm putting my money where my mouth is personally.

我昨天刚买了第二个Claude Max订阅，现在手上一共有两个在用。

I, yesterday, just purchased my second Clod Max subscription, so I actually have two that I'm running now.

我知道很多人在LinkedIn上笑话我，说我居然订了两个Claude Max。

Know a lot of people, they like to roast me on LinkedIn for having two Clod Max subscriptions.

今天我干了特别多事。

I am getting so much stuff done today.

我提交了两个不同的应用到App Store等待审核。

I pushed two different apps to the App Store for review.

等它们上线了我告诉你。

I'll let you know when they are live.

我对这两个应用特别满意。

I'm really proud of them.

其中一些是已经给我赚钱好几年的生意，现在我终于能彻底升级它们，以前根本做不到。

Some of them are businesses that have been making me money for years, and I've been able to completely upgrade them in a way that I never was able to before.

所以我现在在搞一堆特别有意思的东西。

So there's a lot of exciting stuff that I'm doing.

我建了网站，也做了SaaS产品。

I've been able to build websites and SaaS.

如果你去aichatdaily.com看看，那是跟这个频道相关的资讯网站，整个网站我都是用Claude代码写出来的。

If you go to aichatdaily.com, which is kind of the news website that's associated with this, I vibe coded that entire thing using Claude code.

真的特别 impressed。

Really impressed with it.

顺便说一下，这些数据指标我可能下期再细聊，但用户量和各种数据都在疯狂增长。

And so and by the way, the metrics on it, I'll probably cover them in the next episode, but they are exploding as far as users and all sorts of cool stuff there.

所以我看到的都是价值，但我觉得用不了多久，这些东西的成本就会暂时上涨，因为那些补贴项目都快结束了，就像 GitHub 已经停掉那样。

So I'm seeing so much value, but I think that we're gonna come to a point in the near future where all this stuff actually gets more expensive for a while as these subsidized programs are ending just like GitHub has ended theirs.

明白了。

Okay.

Opal 在转型。

Opal is pivoting.

顺带提一句，如果你还不知道他们的话，这是一家做网页摄像头的创业公司。

It's a web webcam startup, by way, if you didn't know them.

他们现在改名叫 Opal Electronics，转去做 AI 硬件了。

But they're rebranding as Opal Electronics, and they're pivoting to AI hardware.

去年第一季度，他们刚从 OpenAI 融了四千万美元，估值直接干到了两亿七千五百万。

They just raised $40,000,000 from OpenAI, in q one last year, which basically valued them at $275,000,000.

我觉得OpenAI是在押注未来会出现一些实体的AI设备。

OpenAI, I think, is betting that there's gonna be some physical AI devices.

我们已经看到不少失败的例子，比如Humane针、Rabbit R1，还有那个Friend.com吊坠。

Now we've seen a lot of flops with the Humane pin, with the Rabbit r one, with, I think, even the friend.com pendant.

这些产品都挺惨的，根本没真正起来。

We we saw we've seen a lot of flops that don't seem like they've really made it.

那这会是AI硬件的未来吗？

And is is that kind of the future of AI hardware?

在我看来，完全不是。

Absolutely not, in my opinion.

因为我们其实也看到过不少成功的案例。

Because we've also seen some big wins.

Meta的Ray-Ban智能眼镜特别火。

The Meta Ray Bans are incredibly popular.

我经常看到有人在用它们。

I see so many people using those.

所以我觉得，有些工具不适合用硬件来实现AI，而有些硬件则非常适合。

And so I think we're gonna come to a place where some some tools are not some hardware isn't a good fit for AI while other hardware is.

我们知道，OpenAI正在和苹果前设计师乔尼·艾维合作开发一些尚未公开的硬件产品。

And we know that OpenAI is working on a lot of hardware that is unannounced with Johnny Ive, the former designer of Apple, the iPhone.

接下来肯定会有一些令人期待的成果。

And so there's gonna be some interesting things.

目前OpenAI是Opal的最大股东，但他们对这家公司的知识产权和产品设计没有任何控制权。

Now I think OpenAI is Opal's largest shareholder, but they hold zero rights to the company's intellectual property or product design.

在2022年与萨姆·阿尔特曼团队讨论过将ChatGPT整合进硬件后，Opal就已经暂停了所有产品的发货，持续了三年。

And Opal has also stopped any sort of product shipments for three years after they had a 2022 meeting with Sam Altman's team about integrating ChatGPT into hardware.

我觉得OpenAI这步棋下得很聪明。

I think this is a smart play from OpenAI.

在我看来，这其实是他们和乔尼·艾维一起推进硬件计划的一种风险对冲策略。

Basically, it's a hedge, in my opinion, on their hardware goals that they're that they're doing with Johnny Ive.

通过Opal和乔尼·艾维的合作，他们能看清哪些硬件设备真正能站稳脚跟。

Between Opal and Johnny Ive, they can see which of these kind of hardware devices stick.

这也很有意思，因为这并不是一次直接的收购，但说白了，感觉就跟收购没什么两样。

It's also interesting because it's not a a straight up buyout, but, I mean, basically, it feels like a buyout.

对吧？

Right?

他们投了四千万美元进去。

They put $40,000,000 in.

他们是这家公司最大的股东。

They're the biggest, you know, stakeholder in this company.

但有意思的是，如果Opal的产品像Friend.com、Humane和Rabbit那样失败了，也不会影响OpenAI的品牌。

But what's interesting is if this product flops from Opal, like we've seen with friend.com, Humane, and Rabbit, it's not gonna tarnish OpenAI's brand.

对他们来说，这根本算不上损失。

It's not like it was a loss for them.

他们就是说，哦，好吧。

They're just like, oh, yeah.

我们只是找了个合作伙伴，本来指望他们帮我们推出来。

There's just someone we partnered with, and they were gonna roll us out.

但如果它成功了，对它们来说就是个证明：AI硬件真的能留住用户，而他们正在推出的、由乔尼·艾维参与的新产品，也可能大获成功。

But if it does work, it's proof for them that, look, AI hardware can be sticky, and this new product we're coming out with Johnny Ive is it could be a success.

OpenAI 这步棋挺有意思。

So an interesting play from OpenAI.

Uber 目前把每个 AI 工具的员工开支上限定在 1500 美元，因为他们四个月就把全年预算花光了。

Uber is currently capping their AI employee spending at $1,500 per tool because they burnt through their entire annual budget in four months.

这其实跟我之前说 GitHub 遇到的问题是一回事。

I mean, this is kind of the same problem that I was talking about with GitHub.

如果你直接给公司所有员工开放 Claude 的 API 访问权限，而不是让他们用每月 200 美元的 Claude Max 订阅，那钱真的会花得飞快。

There is an issue where if you just give all of your employees at a company API access to Claude, for example, and not the Claude Max subscription that's $200 a month, it's very easy to burn through an insane amount of money.

我上周就遇到过这种情况，我的 Claude Max 订阅额度用完了。

Know, I run out of my Claude Max subscription, which happened to me like a week ago.

我当时想，行吧，那就买点超额积分吧，毕竟可以单独购买额外的使用额度。

Was I like, okay, fine.

你可以直接买一些超量积分。

I'll just buy some extra tokens because you can can buy like overage tokens.

所以我心想，干脆花200美元试试，看能撑多久。

So I was like, I'll just spend $200, see how long it lasts me.

结果我一天就把这200美元花光了。

I've spent the $200 in one day.

我其实挺乐意每个月花200美元的，因为按我每天的用量，一个月下来光token就得花六千美元。

So I spent I'm happy to spend $200 a month because the token cost for me would be that much every single day times thirty days, which is $6,000.

不过这事儿其实应该是几周前发生的。

And that was also I I this was actually maybe this was a few weeks ago.

那时候他们还没宣布因为和SpaceX的合作，把使用量和token限额翻了一倍。

This was before they said they doubled the usage and and the token limits because they they made the SpaceX deal.

所以那时候我的消耗大概在六千到一万两千美元之间。

So somewhere between 6 and $12,000.

现在每个月花200美元，我真觉得太划算了。

I am really happy for the $200 a month.

我觉得这补贴力度大得离谱，我之前就说过，大家都该去用一用。

I think that's insanely subsidized, and I recommend, like I already mentioned, everybody get access to that.

好了，这期播客就到这里。

And that's it for the podcast.

非常感谢你们收听。

Thank you so much for tuning in.

如果你们想试试我的创业项目，花每月8.99美元就能在一个平台访问80多个AI模型，去aibox.ai看看吧。

Guys, if you wanna try my startup and get access to over 80 different AI models in one place for $8.99 a month, go check out aibox.ai.

你可以和80多个AI模型聊天，支持视频、音频、图像，所有功能都集中在一个平台上。

You can chat with 80 different AI models, video, audio, image, and you get access to everything in one platform.

不用到处登录不同网站，不用把信用卡信息分散给多个平台，也不用记一堆密码。

You don't have to log in to a bunch of different places, share your credit cards with a bunch of different platforms, remember your passwords.

一个平台，所有数据和信息都集中在一起，就是aibox.ai。

One place, all of your data, all of your info in one platform, aibox.ai.

我在简介里放了链接。

I'll leave a link in the description.