探索AI芯片的潜在空间

Anthropic正在组建自己的定制硅片团队来设计他们自己的AI芯片，而AMD刚刚发布了名为Helios机架系统的东西。

Anthropic is building their own custom silicone team to design their own AI chips, and AMD has just unveiled something called the Helios Rack System.

他们的目标是在AI数据中心领域正面挑战NVIDIA。

Their goal is to challenge NVIDIA head on in AI data centers.

Anthropic正在升级Claude的语音模式。

Anthropic is upgrading Claude's voice mode.

现在它将运行在Opus和Sonnet上。

It's gonna be now running on Opus and Sonnet.

谷歌云的营收飙升了82%，达到248亿美元。

And Google Cloud revenue jumped 82% to $24,800,000,000.

如果最近几周有人关注谷歌，他们公布了AI支出后股价大跌，那么显然，这就是他们宣布这一消息的原因。

If anyone has been following Google for the last few weeks when they've released how much their AI spending is and their stock took a massive hit, well, evidently, this is the reason why they announced that.

OpenAI和Anthropic都在游说华盛顿，以限制中国的开放权重AI。

OpenAI and Anthropic are both lobbying Washington to curb Chinese open weight AI.

我们在播客里已经聊过很多次了，但这是游说方面的最新消息。

We've talked about this a lot on the podcast, but this is kind of the the latest news on the lobbying front.

如果你想深入了解这些以及每天更多的新闻，我建议你去看看 aichatdaily.com。

If you wanna get a deep dive on these and a dozen more stories every single day, I'd love for you to check out aichatdaily.com.

那是我自己的姊妹网站，一个新闻网站，和这个播客是合作伙伴。

That's my own sister website, news site that is partnered with this podcast.

它基本上就像节目笔记，但内容更深入，还有我在节目里没讲到的额外新闻。

I it's essentially like a show notes, but it also goes deeper and has extra stories I don't cover in here.

每天都有几十条关于最新AI新闻的报道，网站上还有你可以订阅的新闻通讯。

There's dozens of stories every single day on all the latest AI news, and there's a newsletter that you can subscribe to on the website.

所以如果你愿意，每天早上你都会收到五条最重要的AI新闻，直接发到你的邮箱，让你了解AI领域的一切动态，而且每条新闻都有快速的事实解析，让你明白发生了什么。

So every single morning, if you like, you will get the top five AI news stories straight into your inbox to keep you up to date with everything happening in AI, and you get some really quick breakdowns of all of the facts so that you understand what's going on.

网站是 aichatdaily.com。

The website is aichatdaily.com.

我想谈的第一条新闻是，Anthropic 正在组建自己的AI芯片设计团队，目的是为 Claude 打造定制硬件。

The first story I wanted to talk about is the fact that Anthropic is building their own AI chip design team, and it's gonna be for creating custom hardware for Claude.

这和 OpenAI、谷歌、Meta 现在在做的事情一样，主要原因就是大家都想摆脱对 NVIDIA 的依赖。

It's the same thing that OpenAI, Google, and Meta, they're all doing it right now, and the big reason is everyone wants to get off of NVIDIA.

这些公司对NVIDIA的依赖程度非常高。

It's these companies are incredibly dependent on NVIDIA.

NVIDIA基本上有配额，他们分配GPU和芯片，只给不同公司一定数量，没人愿意被一个自己花大钱的供应商卡脖子。

NVIDIA basically has quotas and they dole out their GPUs and their chips, and they only give a certain amount to, you know, different companies, and nobody wants to feel bottlenecked by one player that they're obviously spending a ton of money on.

所以大家都在自己造芯片。

So everyone's building their own chips.

现在他们还在考察三星作为制造合作伙伴，这标志着Anthropic不再是最后一个没有公开芯片计划的主要AI实验室，他们觉得，好吧。

Right now, they're also scouting Samsung as a manufacturing partner, which is ending Anthropic status as the basically, the last major AI lab without a public silicone prog program, they're like, alright.

是时候认真对待这件事了。

It's it's time to get serious on that one.

Anthropic目前依赖多家供应商，包括AWS、谷歌、NVIDIA和AMD。

Anthropic currently relies on a bunch of different suppliers, so there's AWS, Google, NVIDIA, AMD.

这些供应商都提供算力，但这也意味着他们无法控制定价、分配或任何路线图的时间安排。

All of them give them compute, but of course that means that they don't get to control the pricing, allocation, or any of the road map timing.

OpenAI已经推出了他们与博通设计的芯片，名为Jalapeno。

OpenAI shipped their Broadcom designed chip, which is called Jalapeno.

那是一个推理芯片，是在今年六月发布的。

It's an inference chip, that was in June of this year.

谷歌现在运行自己的TPU。

Google now runs their own TPUs.

Meta正在构建MTIA加速器，而Anthropic基本上是唯一一个没有在做这个的。

Meta is building MTIA accelerators, and Anthropic basically was the only one that wasn't working on this.

定制芯片通常需要十八到二十四个月来设计，而且每个芯片版本的成本将高达数亿美元。

Custom silicone typically takes between eighteen to twenty four months to design, and it's gonna cost hundreds of millions of dollars per chip version.

但是，他们在这方面获得的每一个效率提升，都会在Anthropic用于Claude API调用的数十亿美元中产生复合效应。

But every single efficiency gain that they can get on this compounds across billions of dollars that Anthropic and is, you know, using for these Claude API calls.

所以我认为这实际上是在打利润率的主意。

So I think this is something that is kind of a play on margin.

这不是一条捷径。

It's not a shortcut.

这需要很长时间。

This is gonna take a long time.

Anthropic可以根据Claude的具体工作负载来调整内存和数学格式，这是任何现成的GPU都无法真正匹敌的，但他们也押注了两年时间，之后这些投资才能得到回报。

Anthropic can tune memory and math formats to Claude's exact workloads in a way that no off the shelf GPU can really match, but they're also betting on a two year runway before any of that investment is gonna get paid back.

所以这不是一件很快就能见效的事情，但从长远来看，这对他们来说可能是个好策略。

So this isn't something that's very quick, but over the long run, this might be a good play for them.

在宣布与Anthropic建立战略合作伙伴关系后，AMD刚刚推出了Helios，这是一个面向AI数据中心的机架系统，他们表示这基本上是NVIDIA现有产品的直接竞争对手。

After announcing a strategic partnership with in Anthropic, AMD has just unveiled Helios, which is a Rack System for AI data centers, and they're saying that this is basically a direct competitor to NVIDIA's products that they have right now.

微软、OpenAI、Meta、甲骨文和Anthropic都已承诺以吉瓦规模部署它，并将在今年晚些时候开始发货。

Microsoft, OpenAI, meta, Oracle, and Anthropic have all committed to deploying it at a gigawatt scale, and it's gonna be starting shipments later this year.

Anthropic和AMD都宣布了一项战略合作，在Helios上部署高达2吉瓦的GPU，这与NVIDIA最大的AI训练集群规模完全相同，所以他们确实非常有竞争力。AMD还推出了Venice X，这是一款数据中心CPU，旨在与Helios搭配使用，将于2027年推出，并将与市场上的许多不同竞争对手展开全面竞争。

Anthropic and AMD both announced a strategic partnership to deploy up to two gigawatts of GPUs on Helios, which is, you know, the exact same scale of NVIDIA's largest AI training cluster, so they're, you know, they're really being quite competitive in the AMD also introduced Venice X, which is a data center CPU designed to pair with Helios, and that's gonna be launching in 2027, and it's going to compete, on kind of this full stack alternative to a lot of the different competitors that are out there.

CEO苏姿丰预测，到2030年，AI加速器市场将达到1.4万亿美元，接近当今整个半导体行业的规模。

The CEO Lisa Su projected the AI accelerator market will reach 1,400,000,000,000 by 2030, approaching the size of today's entire semiconductor industry.

所以，她说这个单一市场细分将和整个半导体行业一样大。

So, you know, this one segment of the market she's saying is gonna be as big as the entire semiconductor industry.

这是AMD首次在一个机架平台上吸引所有五大前沿AI实验室。

This is the first time that AMD has landed all five major frontier AI labs on a single rack platform.

所有客户都已经签约了。

They have everyone signed up for it.

我觉得这显示出他们确实是这些大规模AI基础设施建设项目中一个强有力的第二选择。

I think this is showing that they're really, like, a strong contender for a, quote, unquote, second place option for a lot of these massive AI infrastructure build outs that are coming up.

他们与这五家超大规模云服务商达成的这些交易规模，我认为都是以吉瓦为单位衡量的，这表明这些超大规模云服务商都准备摆脱对NVIDIA近乎完全垄断的依赖。

The scale of all of these different deals that they've done with these these five hyperscalers, I think they're all measured in gigawatts, and I think that shows that the hyperscalers are all of them are ready to diversify away from NVIDIA's almost complete dominance.

看来市场上真的出现了第二个重要玩家，AMD将会获得可观的市场份额。

So looks like we're having a a real second player in the market, and AMD is gonna get some serious market share.

那这对NVIDIA来说是不是非常不利？

Does that, you know, is that, like, super bad for NVIDIA?

我觉得他们的市场还是会疯狂增长，但现在有了其他竞争对手。

I think their market is just gonna keep growing like crazy, but there are other competitors now.

Anthropic正在升级Claude的语音模式。

Anthropic is upgrading Claude's voice mode.

现在它将支持他们最强大的模型Opus和Sonnet。

It's now gonna work with their most powerful models, Opus and Sonnet.

所以我觉得以前，它只支持Haiku。

So I think in the past, it was only working with Haikyu.

而且我要说，我其实挺欣赏这一点的是，我一直在OpenAI和Anthropic之间来回切换，用Claude Cowork和ChatGPT Work。

And I will say one thing that I actually did appreciate about this was I've been going back and forth between OpenAI and Anthropic as far as using, like, Claude Cowork and and ChatGPT work.

我发现ChatGPT Work在升级之后，所有东西都跑在Sol 5.6 Extra High上，当我让它帮我完成一个项目，并且用语音模式的时候，它需要花一点时间来转录和写下来。

And I did find that ChatGPT work, after they did their upgrade and everything's running on sol 5.6, extra high, when I go and and ask it to complete a project for me and I'm doing the voice mode on it, it takes a while for it to kinda transcribe and write it all down.

这其实挺烦人的，那个延迟。

And it's actually kind of annoying the the leg.

就是几秒钟的延迟，我注意到不是每次都这样，但当你开启Sol 5.6 Extra High时，就会这样。

It's a it's a few second leg, and I noticed that it doesn't always happen, but when you have sol 5.6 extra high on, it does.

所以我其实挺欣赏Anthropic的，我用Claude的时候，它几乎是即时转录我正在做的事情。

And so I actually appreciated that Anthropic when I was using Claude, it was like instantaneous transcription of what I was doing.

它全都跑在Haiku上。

It was all on haiku.

现在它完美吗？

Now is it perfect?

它是最好的吗？

Is it the is it the best?

比如，它可能会出一些小差错，但它即时又超快。

Like, there's a couple things that it might mess up, but it instantaneous and super fast.

嗯，他们会升级它，因为显然他们认为需要更高一点的质量，所以现在他们把它放在Opus和Sonnet上。

Well, they're gonna upgrade it because, evidently, they think that it needs to be a little bit higher quality, so now they're putting it on Opus and Sonnet.

我希望他们不要牺牲太多速度，因为说实话，那是我最喜欢的一点。

I hope they don't sacrifice too much of the speed because, honestly, that was my favorite thing.

他们会让你直接把它接入Gmail、Google日历、Slack、Canva、Notion，所以你现在可以用语音告诉Claude处理你的邮件。

They're gonna be able to they're allowing you to plug that directly into Gmail, Google Calendar, Slack, Canva, Notion, so you can now tell Claude by voice to do your emails.

我的意思是，其他所有你可能一直在做的事情。

I mean, everything else that I you're you've probably been doing.

我甚至有一个，字面意义上的。

I I even have, like, a literal.

我有这个，来自OpenAI的设备。

I have this, device from OpenAI.

它就像一个小键盘。

It's like a little keyboard.

是的，是用WorkLouder做的，它真的就只有一个麦克风按钮，我按一下就行。

It is, think built with WorkLouder, and it literally just has a microphone button that I push.

所以，这就是我最近工作的主要方式，就是跟这些模型说话，让它们帮我去完成任务，而不是打字。

So I I this is the way I've been working a lot lately is just talking to these models and having them go and complete tasks for me, instead of typing it out.

我觉得很多人都在这么做。

I think that's what a lot of people are doing.

现在有很多像WhisperFlow这样的软件，就是干这个的，所以选择很多。

There's a lot of software like WhisperFlow that's out right now that does that kind of stuff, so, a lot of different options.

但Claude显然决定要在这里做一次大升级。

But Claude obviously decided they wanted to to put a big upgrade here.

语音模式现在会默认使用你上次在文本聊天中使用的Claude模型，所以如果你用Sonnet工作流，音频也会自动用Sonnet，这有点像OpenAI现在做的，说实话，这让我有点烦。

Voice mode is now gonna default to whatever Claude model you last used in text chat, so a Sonnet workflow is automatically gonna have that for the audio, which is kinda what OpenAI is doing right now, and I'm not gonna lie, that's actually annoying to me.

我希望你能自己选。

I wish you could choose.

说实话，关于音频，我发现当它做转录并打字出来的时候，我其实更想要一个更快但差一点的模型，因为很多时候我用的就是苹果的语音转文字，那个真的很简陋，不太好用。

The audio, honestly, what I find is when it does the transcription and when it types it out, I would actually rather a faster, worse model because you can like, many times I'm literally using Apple's, like, voice to text, which is really ghetto, not very good.

而且，当我用的时候，会有很多错别字，但Claude能理解，而且真的会修正，按我想要的方式做好一切。

And, when I when I do it, there's, a bunch of typos, but Claude understands, and it literally fixes it and does everything the way I want.

所以我发现Claude很擅长提取意思。

So I find that Claude is very good at getting the meaning out.

我宁愿底层模型更好，语音转录更快，因为我特别讨厌说话之后还要等五秒钟它才打出来，然后才能按发送。

I'd rather just have the underlying model better and having the voice transcription faster because I absolutely hate talking and then waiting five seconds for it to type out before I can hit send.

不过，这只是我太没耐心了。

But, I mean, that's just me being impatient.

所以，我想质量提升可能是好的，但我希望他们能在设置里让你选择哪个模型来驱动你的语音。

So, I guess the quality improvement is probably good, but I wish they would let you choose in the settings what model was powering your voice.

免费用户会继续用Haiku。

Free users are gonna keep getting haikus.

也许我得去转成免费用户了。

Maybe I gotta go switch to being just a free user.

开玩笑的。

Just kidding.

而且会有付费用户使用Opus和Sonnet以及其他模型。

And there's going to be paid users that get the Opus and Sonnet and all the rest of it.

顺便说一下，这个支持很多不同的语言。

They have a lot of different languages supported on this, by the way.

英语、法语、德语、印地语、日语、韩语、西班牙语。

English, French, German, Hindi, Japanese, Korean, Spanish.

而且它当然会自动检测，所以这太棒了。

And it, you know, of course, auto detects it, so that's fantastic.

谷歌云上一季度营收增长了82%，达到248亿美元。

Google Cloud revenue is up 82% to 24,800,000,000 last quarter.

这完全超出了华尔街的预期。

That absolutely crushed the expectations that Wall Street had.

我认为这证明了该公司在人工智能上的巨额投入确实得到了回报。

I think it's proving that the company's huge AI spending is actually paying off.

5140亿美元的已签约客户合同积压，我觉得这正说明对AI基础设施的需求非常巨大，而且这些服务是实实在在的，并且已经锁定了好几年的合作。

The $514,000,000,000 backlog of signed customer contracts, I think is showing that there is a huge demand for AI infrastructure, and service is real and it's locked in for years.

所以谷歌云签的很多合同，不只是今年或者这个月的，而是多年的长期合同。

So a lot of these deals that Google Cloud's doing, it's not just for this year or for this month, but these are multi year deals.

云业务的增长真的在飞速加快。

Cloud growth is accelerating really fast.

从上个季度的63%涨到了这个季度的82%。

It's up from 63 last quarter to 82% this quarter.

这主要靠很多公司租用他们的算力，还有购买他们现成的AI产品来推动。

It's mostly driven by a lot of different companies renting compute power and buying prebuilt AI products from them.

Gemini的月活跃用户达到了9.5亿。

Gemini reached 950,000,000 monthly active users.

这比两个季度前的7.5亿又涨了不少。

That's up from seven fifty two quarters ago.

所以，他们自家产品的使用量在加速增长，同时，越来越多的公司也在跟他们签约，用他们的基础设施来支撑自己的AI，这个数量也在增加。

So the usage of their own products is accelerating, but also the amount of companies signing deals to power their AI on their infrastructure is also increasing.

而且Alphabet公布的季度利润达到了1121亿美元。

And Alphabet posted a 112,100,000,000 in quarterly profit.

这几乎是去年同期的281亿美元的四倍。

That's almost four times the 28,100,000,000 from the same quarter last year.

这太疯狂了。

That is wild.

1120亿美元，一年后就变成了210亿美元。

112,000,000,000 up to 21,000,000,000 just one year later.

这简直难以置信。

That's absolutely insane.

谷歌每年在数据中心和芯片上的支出高达1800亿到1900亿美元，你知道，这在前几周公布时导致股价大跌，大家都觉得谷歌花钱太多了。

Google's 180,000,000,000 to $190,000,000,000 in annual spending on data centers and chips is now, you know, it's something that kind of crashed their stock a couple weeks ago when it came out and everyone was like, Oh my gosh, Google's overspending.

但我认为这恰恰说明需求是真实存在的。

But I think this is showing there is serious demand.

需求正在疯狂增长。

The demand is accelerating like crazy.

他们知道自己能把这笔钱赚回来。

They know they're gonna get that money back.

所以我觉得他们预计到2027年，这个大规模建设就会开始真正产生现金流。

And so I think they expect that 2027 is gonna be when this massive build out starts turning into some real cash generation.

但为了拿到那些钱，他们得先花钱才能赚钱，而且他们能看到需求很大。

But in order to get that money, like, they have to spend they gotta spend the money to make the money, but they can see, a lot of demand.

所以很明显，他们正在全力投入。

So, evidently, they're they're going all in.

OpenAI和Anthropic都在游说美国监管机构，限制中国的开放权重AI模型，比如Moonshot的Kimi，而且你知道，K3出来有一阵子了，当时特别火。

OpenAI and Anthropic are both lobbying US regulators to restrict Chinese open weight AI models like Moonshot's Kimi, and, you know, k three came out a while ago, and it was all the rage.

然后还有最新的千问模型，几天前刚发布。

And then we have the latest Quen model came out a few days ago.

他们在这里说的是存在国家安全方面的担忧。

What they're saying here is that there are national security concerns.

当然，他们这么说的时候，时机正好很合适。

They, of course, are saying this at a time when it would be quite convenient.

这会把这些公司推向，你知道，成为这个领域的赢家。

It would push these companies towards, you know, being the the the winners in this space.

没人能真正取代他们。

No one could really dethrone them.

这些开源模型没有一个能对他们造成毁灭性打击。

None of these open source models could go scorched earth on them.

Dean Ball 提出，美国应该对开源权重模型制造监管上的恐惧、不确定性和怀疑，以削弱它们的竞争地位，然后他又有点收回这个说法，但那是他最初的观点。

Opening eyes Dean Ball argued The US should create regulatory fear, uncertainty, and doubt around open weight models to weaken their competitive position, and then he kinda walked that argument back, but that was like his first thing.

我们在加密货币领域经常听到这个，就是 FUD，他公开说应该制造一些 FUD，然后他又说，不，不，实际上这是严肃的。

We've heard this in crypto a lot back in the day, the FUD, that he's openly saying that they should create some FUD, and he's like, No, no, actually it's serious.

这不仅仅是无谓的 FUD。

It's not just FUD for FUD's sake.

Kimi 在上个月底发布时，我想很多人都在恐慌。

Kimi's launch at in the end of last month, I think a lot of people were panicking.

有一个病毒式传播的演示，展示了它在三十分钟内生成一个类似 Mac OS 的 UI 原型。

There was a viral demo showing it generating a Mac OS like UI mock up in thirty minutes.

我觉得在那之后，很多热度就消退了。

I think a lot of the hype faded after that.

很多演示都是那种‘天哪，看这个疯狂又不可思议的东西！’

A lot of demos are like, Oh my gosh, look at this crazy incredible thing!

然后等它真正发布，人们真正上手玩的时候，就会觉得‘好吧，其实没那么夸张。’

And then once it actually comes out and people actually play with it, they're like, Okay, it's not as crazy as it seems.

所以我才有点怀疑——不过，算了。

Which is why I'm actually a little skepticalI mean, whatever.

你可以说我傻，但这就是为什么我对Anthropic有点怀疑，他们总是把模型吹得天花乱坠，只给企业用，不给普通人，然后说‘看，它能做这些疯狂的事。’

You can say I'm dumb, but it's why I'm a little skeptical of, like, Anthropic always hyping up their models to an insane degree, only giving them to enterprises, not to the regular people, and then saying, look, it can do all these crazy things.

我不知道。

I don't know.

他们不给你用，这让我很怀疑，但也许它真的很危险。

It makes me skeptical when they won't give it to you, but maybe it's truly dangerous.

我不知道。

I don't know.

这已经是我们过去不到一年里第二次遇到这种情况了，就是有一个超便宜又超强的模型出来。

This is the second time that we've had, in the last under a year, basically, where we've had this kind of cycle where there's a huge cheaper model that comes out and it's super capable.

DeepSeek之前也有类似的情况，然后它出来之后，就有很多人呼吁要出台政策限制。

DeepSeek had something that that kind of, was along these lines, and then afterwards, there's a bunch of policy restrictions that people were calling for after it came out.

所以我觉得我们基本上是在反复看到同样的模式。

So I think we're seeing basically the same pattern play out over and over again.

中国的实验室会推出一个超级便宜的开源模型。

Chinese labs are gonna ship like a super cheap open model.

它在基准测试上会非常有竞争力。

It's gonna be really competitive in the benchmarks.

很多美国高管会开始警告安全和监管问题，说我们需要限制。

A lot of the American executives are gonna start warning about security and regulation, and we need to have restrictions.

当然，真正赢家是那些大型闭源模型公司，比如OpenAI、Anthropic等等。

And, of course, the people that are really winning are those big closed model companies, Open AI, Anthropic, etcetera.

所以我觉得，不管这些安全担忧是否合理，事实是，所有要求政策变化的人恰好都是会从中受益的人。

So I think whether those security concerns are legitimate or not, it doesn't change the fact that all of the people asking for these policy changes happen to be the people that are going to be winning from it.

说实话，我对此有点怀疑。

So I am a little bit skeptical of it, if I'm being honest.

而且我确实认为开放权重模型对创新非常棒，比如对我自己的公司来说，能够使用一个比闭源模型便宜90%，甚至便宜90倍的模型，如果它能完成我们反复需要的一些任务的话。

And I actually think open weight models are fantastic for innovation and like for my own company being able to use a model that's 90, you know, 90% cheaper, 90 times cheaper than an open or than a closed source model if it can do some of the same tasks that we need it to get done on a repeated basis.

对我来说，这很有意义。

To me, that makes a lot of sense.

但关于这个话题肯定有很多争论。

But there's a lot of, there's a lot of debate on this topic for sure.

各位，非常感谢你们收听本期播客。

Guys, thank you so much for tuning in to the podcast.

如果你喜欢这一集，记得在你收听播客的地方留下评分或评论。

If you enjoyed this episode, make sure to leave a rating or review wherever you get your podcast.

如果是在Spotify上，需要点击“关于”标签。

If that's over on Spotify, gotta hit the about tab.

如果是在Apple上，你可以留下评论，打几颗星。

If it's on Apple, you can leave a little comment, drop some stars.

我知道有人抱怨苹果系统上留评论不太方便，但你可以去播客主页，往下滑，点几颗星，留个评论。

I know people have complained that Apple's not intuitive for leaving reviews, but go to the main podcast page, scroll down, hit some stars, leave a review.

这对我们帮助很大。

It helps us show out so much.

我会非常感激的。

I would be, incredibly grateful.

如果你想看这些新闻的文章版，可以去 aichatdaily.com 看看。

Go check out aichatdaily.com if you wanna get all of these news stories in article format.

如果你想每天收到新闻简报，可以订阅我们的通讯。

And if you wanna get a news brief sent to you every single day, you can subscribe to our newsletter.

我们下期节目再见。

I will catch you guys all in the next episode.