Why Google Is Overhauling Its AI Coding Team, OpenAI's Jalapeno Chip, Robotics AI Rift

欢迎各位收看《The Information》的TITV节目。

Welcome everyone to The Information's TITV.

我是阿卡什·帕什里卡。

My name is Akash Pasricha.

今天是6月25日，星期四。

It is Thursday, June 25.

本期节目中，《The Information》独家披露，谷歌正在重组其新成立的专注于AI编码模型的特别小组。

Today on the show, The Information has exclusive reporting that Google is reorganizing its recently launched Strike Team working on AI coding models.

我们稍后会详细解读这些内容。

We'll dig into those details shortly.

接着我们会分析美光和Cerebras的季度财报、OpenAI新推出的Jalapeno芯片以及高通的收购案，同时深入探讨物理AI领域正在加剧的分歧。

We'll then get some analysis on Micron and Cerebras' quarterly results, OpenAI's new Jalapeno chip and Qualcomm's acquisition, plus a growing look at the a look at the growing rift currently dividing the physical AI space.

新成立的AI实验室Odyssey刚刚完成融资，估值达14.5亿美元，融资金额为3.1亿美元。

And Odyssey, a new AI lab focused on world models, has raised $310,000,000 at a $1,450,000,000 valuation.

我们会分析是什么推动了这些巨额融资。

We will look at what is driving those massive funding rounds.

今天这期节目内容非常精彩，咱们马上开始。

It's going be a great show, so let's get right on into it.

过去几天，谷歌股价大幅下跌，原因是两位资深研究员跳槽到了竞争对手的AI实验室。

Google shares have suffered a big drop over the past few days after two of its most senior researchers left the company for rival AI labs.

《The Information》独家获悉，谷歌正在重组负责AI编程模型的团队，以提升竞争力。

Now the information has exclusive reporting that the company is revamping the team that is working on its AI coding models to make it more competitive.

我的同事Aaron Wu负责报道OpenAI和谷歌，他撰写了这篇报道。

My colleague, Aaron Wu, who covers OpenAI and Google, reported that story.

为了让大家更清楚来龙去脉，我们请来我们在旧金山的首席记者Jason Dean，帮我们详细解读一下。

For more context, want to bring on Jason Dean, our San Francisco bureau chief, to help us break it all down.

Jason，欢迎再次做客节目。

Jason, welcome back to the show.

很高兴你能来。

It's great to have you here.

谢谢。

Thanks.

很高兴能来。

Good to be here.

好的。

Okay.

咱们先退后一步，理一理前因。

So let's just back up a bit.

谷歌失去了两位顶尖研究员。

Google loses two of its star researchers.

诺姆·沙齐尔去了OpenAI。

Noam Shazir goes to OpenAI.

乔恩·詹珀去了Anthropic。

Jon Jumper goes to Anthropic.

当时，至少在诺姆离开的时候，我们并不清楚他为什么做出这个决定。

We didn't have a great sense at the time well, at the time that Noam left, at least, why it was he may have made the jump.

现在我们对这些离职事件了解得更多了，对吧？

What do we know now departures?

这些事情背后还是有不少谜团，但埃琳在报道中发现了一些有意思的情况。

There's there's still quite a bit of mystery around these, but Erin has learned a couple of interesting things in her reporting.

首先，沙齐尔在离开前就告诉过人，谷歌把他的项目所用的计算资源和其他团队的合并了，结果他的计算资源被大幅削减。

First of all, Shazir told people before he left that Google had shifted away or or really kinda combined the computing resources for his project that he was working on with that of other teams, and so basically diminished his access to compute.

具体这对他产生了多大影响，现在还不太清楚，但显然这事很重要。

It's not exactly clear how that, you know, shaped his thinking, but obviously, you know, something very important.

他一直在做研究。

And he's been doing work.

沙齐尔可是那个开创性Transformer论文的共同作者之一，正是那篇论文点燃了如今这场AI热潮。

Shazir, of course, is one of the the the coauthors of the seminal AI paper on transformers that kinda gave birth to the current AI boom.

他现在正在探索超越Transformer的新型AI架构。

He's been doing work on sort of new architectures for AI beyond transformers.

计算资源的获取问题，在谷歌内部是个大问题，他离开前就明确提过这一点。

And, you know, the so the access to compute, obviously, is a is a big issue inside Google and something that he he flagged to people before he left.

另外，谷歌的首席技术官在内部宣布他离职时说，这件事来得非常突然。

The other thing is that Google's CTO in announcing his departure internally said that it was very sudden.

所以，很明显，这里发生了一次突然的决裂。

So, you know, clearly, there was some abrupt rupture here.

具体关联还不清楚，但这件事是突然发生的，而且就在计算资源问题之后。

The exact linkage isn't clear, but this was was something that happened suddenly and happened after this compute issue.

没错。

Right.

明白了。

Okay.

看来谷歌得自己收拾这两次离职留下的烂摊子了。

So Google is left to pick up the pieces, I guess, of these two departures.

我得说，他们失去了一位诺贝尔奖得主，但还有德米斯呢，他也是诺贝尔奖得主，对吧？

And I should say that, you know, they they lost one Nobel Prize winner, but they still I mean, Demis is also a Nobel Prize winner, right?

他们诺贝尔奖得主少了一半。

They're down 50% on Nobel

奖。

Prize.

是的，某种程度上是这样。

Yeah, in some sense.

我的意思是，有些情况下你没法把整个团队都堆上去。

I mean, you can't stack the whole team in some cases.

所以谷歌只能收拾这两个人离开后的烂摊子。

Okay, so Google has left to pick up the pieces.

那么，我们从中学到了什么？他们是如何重组来应对当前局面的？

So what did we learn then about how they are reorganizing themselves to address this moment?

对。

Yeah.

其实这些变化是近几个月来一些已有举措的自然演进，早在亚伦组建谷歌的代码突击队、加速AI编程工具开发时就已经开始了。

So some of these, you know, this is sort of an evolution of things that have been in the works in recent months, going back to when Aaron scooped the creation of what they call a Code Strike Team in Google to accelerate its efforts to build AI coding tools.

他们现在做的，是把这个才成立几个月的团队重新调整、扩大职责范围，让它变成一个更长期的常设机构。

And what they're doing now is sort of that's only months old, but they're restructuring it and expanding its remit and making it a more permanent thing.

他们正在调整训练谷歌模型的方式，让提升模型能力变得更轻松，不仅限于编程，还扩展到其他白领应用场景——这正是竞争对手Anthropic和OpenAI已经走在前面的领域。谷歌的Cloud Code和Codecs产品，也正朝着类似方向拓展，试图复制Anthropic在编程领域取得的成功，应用到更多相关场景中。

Part of what they're doing is to try to make it easier for them to they're changing the way that they do training of of Google's models to make it easier for them to enhance the capabilities, not only in coding, but in other sort of white collar applications, which is something we're seeing at the rivals, Anthropic and OpenAI, which are kind of further ahead in this area of of enterprise AI, than than Google is, you know, both Cloud Code and Codecs products are being expanded into other areas adjacent to coding, trying to develop the same success that Anthropic had with coding in other areas.

所以谷歌调整了模型训练方式，还把这支突击队的职责扩大并固定下来，目的就是提升那些谷歌一直落后的产品的性能。

So Google's change in the way it's it's training these models and the expansion of the remit of this Strike Team, making of it permanent is designed to sort of enhance the ability to improve those those products where Google has been a laggard.

现在，亚伦在报道中特别提到一个概念，叫‘中间训练’。

Now one of the things that Aaron focused on in the story is this concept of mid training.

什么是中间训练？为什么现在它这么重要？

What is mid training, and and why is that so important right now?

是的。

Yeah.

说实话，这些术语真有点让人头晕。

I mean, this this stuff can kind of make your your head spin a little bit of the of the jargon.

不过我们通常把模型的创建分成预训练和后训练两个阶段。

But, you know, we we tend to group the creation of these models into pretraining and post training.

预训练阶段，我们是在构建基础智能。

Pretraining, we're sort of building the basic intelligence.

也就是在搭建模型本身。

You're building the model.

它在学习。

It's learning.

而训练后阶段，我们是在让它适应特定的功能。

And post training, we're kind of adapting it to specific functions.

这其实是对那个过程的拆分。

This is a disaggregation of that.

你把学习过程拆开，让模型在进入适应用户交互和行为的训练后阶段之前，先变得更擅长特定任务。

So you're taking that learning process and enabling the models to become more specialized in what they can do before you get into the post training process of adapting, like, how they interact with users, the model's behavior.

而且，这背后的目标显然是想让模型在涉及行为调整之前，先把基础能力提升上去，比如在编程这类具体任务上表现得更好。

And, again, this is the the goal seems to be to try to enable the models to be better at specific functions like coding, etcetera, before you get into the behavior part, just the underlying intelligence to be better before you get into the behavior aspect of it.

明白了。

Okay.

那我们再回过头说说谷歌成立的这个特别小组。

So let's go back to the Strike Team that Google has established.

Aaron 报道说，这个小组是几个月前才组建起来的。

Aaron had reported that this team came together a couple months ago.

他们现在对这个团队的结构做了一些调整。

They've now changed the organization of it a bit.

我们知道这个团队有多少人吗？

Do we know how big the team is?

都有谁在里面？

Who's in it?

他们在公司内部有多大话语权？

How much power they have internally in the organization?

这会不会有点像……我想到的是，Meta那边的团队？

Is this kind of like the equivalent to I don't know, I'm thinking it Meta.

Meta有个超级智能团队。

Meta has the super intelligence group.

这是专门搞代码的团队，还是说就是那个核心团队？

Is this just a coding specific group, or is this the group?

我觉得，他们的职责范围现在正在扩大。

I I think, you know, their their their remit is expanding now.

这个团队一直非常重要，我们之所以知道这一点，不仅是因为它由一位在DeepMind工作了八年的资深工程师领导，还因为DeepMind的首席技术官和谷歌联合创始人谢尔盖·布林都曾直接参与。

It's a pretty it's been a pretty important team in inside this, and we know that in part because not only is it it's led by a a deep mind veteran engineer who's been there for eight years, but it's also had the involvement of the DeepMind CTO as well as Google cofounder Sergei Brin.

所以这一直是公司非常重视的一项事业。

So this has been something that's been very important for the company.

虽然目前我们还没看到他们真正缩小与竞争对手差距的明显成果，尤其是在编程工具方面，但他们在巩固这个团队、让它更稳定，并扩大其研究范围，这说明他们相信这条路是对的。

And I think though we haven't really seen the tangible results yet in terms of them closing the gap in a meaningful way with their competitors and things like coding tools, the fact that they're sort of solidifying this team and and and making it more permanent and expanding its areas of focus suggests that they they think that it's working.

对。

Yeah.

你觉得为什么我们在编程领域还没看到成果呢？

Why why do you think that we haven't seen results in coding?

谷歌在人工智能的其他几乎所有领域都做得非常出色，连芯片都搞定了。

I mean, Google has excelled at seemingly every other part of AI, including chips now.

编程不应该是你们这种软件公司的强项吗？

Coding seems like, I don't know, your software company.

你们怎么会在编程上掉队了？

How did you miss the boat on coding?

那为什么我们认为他们一直没能突破这一点呢？

Why do we think they haven't been able to crack that?

我觉得部分原因是他们起步太晚了，除了Anthropic，其他公司都慢了一拍。

Well, I think part of it is they were late to it, and kind of everybody was late to it except for Anthropic.

Anthropic比别人更早搞明白了这件事，不管是靠远见卓识，还是偶然撞上了门道。

Anthropic figured this out before others, whether that was through some great visionary foresight or sort of they stumbled into it.

这中间当然有争议，但他们是第一个做到的。

The you know, there there could be a debate, but they were first to it.

这已经成为企业市场里最赚钱的杀手级应用。

That's been the killer app, in the enterprise in terms of of generating revenue.

OpenAI正努力紧追其后。

OpenAI is trying to be a fast follower there.

谷歌也在追赶，不过可能没那么快。

Google is trying to be a a fast follower also, perhaps not not as fast.

Aaron的报道还指出，当初大家普遍相信，只要把底层模型做出来，编程能力自然就会跟上。

I think also what Aaron's reporting shows is that there was an initial faith that if you build the underlying model, the coding capabilities will come.

这显然没那么简单。

And it's clearly not that simple.

光做一个什么都能做的聪明模型是不够的。

You you need to do more than just make a really smart model that can do everything.

你得真正专注于把这些能力落地。

You really need to focus on enabling those capabilities.

这正是他们现在正在努力的方向。

And that's part of what they're trying to do now.

没错。

Right.

最后一个问题，杰森。

Last question for you, Jason.

我们都在等最新的Gemini模型。

The latest Gemini model, we are expecting it.

我记得亚伦说过，我们原本指望六月份发布。

I think Aaron had reported we were expecting it in June.

还没发布呢。

It hasn't been released yet.

关于这个，有什么新消息吗？

Do we have any updates on that?

是的。

Yeah.

毕竟六月还没过完，他们之前说的就是这个时间。

I mean, you know, there's still a few days left to June, which is what they said.

这是 Gemini 3.5 Pro。

This is this is Gemini 3.5 Pro.

上个月在 Google I/O 上，他们发布了 Flash，还说 Pro 版本会在六月推出，作为主打模型。

They they came out with Flash at Google IO last month and said that the Pro version would be the sort of flagship model would be available in June.

到现在还没出来。

Not here yet.

现在看，很可能要到七月了。

It's looking much more like it's gonna be July.

我的意思是，具体时间要是拖到七月，其实也没那么严重。

I mean, the exact timing, you know, if it slips into into July, perhaps not the biggest deal.

我觉得更重要的问题是，它到底会有什么改进，会不会有明显提升？

I think the bigger question is will it know, what will be, and will it be that much better?

是的。

Yeah.

对。

Right.

而且根据亚伦的报道，看起来目前还远没到那种夸张的程度。

And and, I mean, early in the I guess, based on Aaron's reporting, it seems like that still is not going to rise to the level of fable yet.

对。

Right.

对。

Right.

这说不定反而是件好事。

Which, you know, maybe that's a good thing.

也许这意味着政府不会禁止或限制它。

Maybe that means it won't be sort of banned or restricted by the administration.

对。

Yeah.

我是说，Anthropic 还在大力宣传这件事。

I mean, Anthropic is still quite loud about it.

你去用一下 Claude 就知道了。

You go on Claude.

他们的底线还在那儿呢。

Mean, they still have the bar there.

只要这个规则还在，他们就不断强调：这不是我们的错。

As long as they have that rule in place, they are very loud about the fact that, look, it's not our fault.

至少他们能借此做营销。

They're getting marketing mileage out of it, if nothing else.

对。

Yeah.

是的。

Yeah.

好的。

All right.

杰森，非常感谢你今天来做客。

Well, Jason, I want to thank you for coming on.

这位是杰森·迪恩，我们《信息报》旧金山分社的社长。

That is Jason Dean, our San Francisco bureau chief, here at the information.

内存芯片制造商美光科技上季度营收增长超过四倍，股价应声上涨。

Shares of Micron are jumping after the chipmaker more than quadrupled its revenue in the latest quarter.

这一结果为半导体行业带来了一些缓解，就在本周早些时候，Cerebras在上市后的首份财报中曾指出利润率下滑。

The results offer some relief for the semiconductor sector after Cerebras flagged shrinking margins earlier this week in its first report since its IPO.

我们请来Wedbush公司股权研究主管马特·布莱森，以及新企业联合公司合伙人穆斯塔法·尼穆奇瓦拉，一起来分析这两份财报。

I want to bring on Matt Bryson, Managing Director of Equity Research at Wedbush, and Mustafa Neemuchwala, partner at New Enterprise Associates, to unpack both of those results.

欢迎两位的到来。

Welcome to the both of you.

很高兴你能来。

It's great to have you here.

马特，我想先听听你的看法。

Matt, I want to start with you.

你对美光昨晚那份强劲的财报怎么看？

What did you make of Micron's very strong report last night?

我觉得除了再次超出预期并上调指引外，最值得关注的是他们对长期合同的详细说明——这些合同确保了大约一半的年产量有明确的付款保障，如果客户违约，这些款项将被没收。

I think the most interesting piece beside the the the beat and raise again was the granularity they were able to give around the these longer term contracts that they've set up where you're looking at it feels like roughly half the annual value of their their output to these customers as being guaranteed or or cash payment that is forfeit if the customer breaks the contract.

这为美光、也为整个行业带来了前所未有的确定性。

And that sets up a level of certainty for Micron and I I I think the industry that that we haven't seen before.

我认为这正是股价上涨的主要原因。

I I think that largely is is what the stock's reacting to.

现在我们不再需要猜测周期何时见底，而是有了长期稳定的现金流保障。

The idea that instead of trying to pick the the end of the cycle, now we've got a longer term guaranteed cash flow, if you will.

在这样一个曾经完全受周期驱动的行业里，你该怎么评估这种稳定性的价值呢？

And how do you how do you value that within the this this industry that used to be completely cyclical?

这一切都是因为内存芯片短缺造成的吗？还是说他们能提供如此详细信息的根本原因是什么？

And is that all a result of the memory chip shortage, or what's the root cause that they were able to give that granularity?

我觉得这跟内存行业格局的变化有很大关系。

Oh, I I I think it's certainly how tape memory has become.

他们现在的议价能力比以往任何时候都强。

They they just have more leverage than they ever had before.

明白了。

Right.

穆斯塔法，你注意到什么特别的地方了吗？

Mustafa, what stood out to you?

我觉得过去几个季度，内存行业最让人关心的问题就是：放缓什么时候才会显现？

I would say, I think the biggest question in memory for the last couple of quarters has been, when does the deceleration start to show up?

我们一直谈论的周期性，其实隐含着一个前提：总会有那么一个季度，市场预期中迟早会迎来放缓。

I think the cyclicality that we've all been talking about implicitly implies that there'll come a quarter, or there'll come a quarter sometime far out because markets look forward, where we'll have the deceleration start to show up.

但现在，由于像马特刚才提到的战略客户协议，我们看到的能见度甚至能延伸到2028年，所以现在真正的问题是——如果周期性真的还会回来，它到底什么时候才会再次显现？

And right now, because of the SEAs, the strategic customer agreements that Matt was just referencing, we know that visibility is not until perhaps '28, and that's really what we're talking about right now is when does that cyclicality, if it ever comes back, start to show up?

到目前为止，他们的产能已经卖到2027年了，利润率是过去峰值的两倍。

And so far, they're sold out through '27, margins are double what they were at prior peaks.

所以，即使周期性确实存在，我们一直在讨论的另一个问题是，从盈利角度看，你能收回多少现金？

So And even if the cyclicality concept is there, I think one of the other things we've all been talking about is how much cash do you get back from an earnings perspective?

如果你看的话，美光的估值相对要高一些。

And if you look at you know, Micron's probably relatively more expensive.

但如果你看看海力士，还有日本的铠侠，很多人算下来会发现，到2028年底，仅靠这些公司产生的现金流，就能把整个企业价值赚回来。

But you look at Hynix, you look at Kioxia in Japan, a lot of people's math will tell them by the end of 'twenty eight, you get the entire enterprise value back in just the cash flow these companies generate.

所以，马特，我很好奇的一点是，当你看美光和整个内存芯片公司的市盈率时，跟其他芯片公司相比，它仍然很低——我今天早上查的时候，市销率大概只有七倍左右。

So Matt, one thing that I was curious about is when you look at the multiple for Micron and memory chip companies, broadly speaking, compared to some of the other chip companies, it's still I mean, I think it was trading at something like seven times last I checked this morning, and that was forward sales.

这比GPU公司的估值要低。

That is lower than the GPU companies.

所以我想问的是，你觉得这些估值差距以后会收窄吗？

And so my question is, do you see those multiples converging at some point?

这正是投资者们正在讨论的问题。

I mean, that's exactly what investors are talking about.

对吧？

Right?

如果我愿意为一些加速器股票支付三十、四十甚至五十倍的估值，那为什么我不愿意为美光支付十倍或十一倍呢？毕竟它的盈利是稳定的。

If I'm willing to pay thirty, forty, 50 times for some of these accelerator names, you know, then why am I not willing to pay 10 times, 11 times for a Micron if I believe that there's there's earning stability out there.

但我真觉得美光会涨到五十倍吗？

Now do do I ever think Micron's gonna get a a 50 x multiple?

我觉得不会。

I don't.

我始终认为，随着晶圆厂不断扩建，而客户需求又不够透明，迟早会出现供过于求的情况，那时候议价能力就会下降。

I I think at the end of the day that eventually because you bring on these fabs and big chunks because customers are somewhat opaque, end up with with oversupply at some point and so you lose leverage.

但与此同时，这个问题确实值得思考。

But at the same time, I I think it's it's a question worth asking.

再说一遍，如果你愿意为AMD支付四十倍，那为什么不愿意为美光支付十一二倍呢？

Again, why wouldn't you be willing to pay 11 times, 12 times for Micron if you're willing to pay 40 times for AMD?

所以，穆斯塔法，马特的意思是，存储芯片的估值倍数可能会有所上升。

So Mustafa Matt is saying that the memory multiples are likely to come up a bit.

换种思路，可能是GPU的估值倍数会下降。

The other way of thinking about it is maybe the GPU multiple come down.

你觉得哪种情况更有可能发生？

What do you think is more likely here?

我觉得大多数人现在都会说，英伟达股价便宜，而且一直都很便宜。

I would say, I think most people would tell you right now that NVIDIA is cheap and NVIDIA has been cheap.

根据不同的数据，市盈率差距大概在十倍、十二倍，甚至十五倍之间。

You know, we're talking about, depending on your numbers, ten, twelve, 15 times gap PE.

所以很难想象英伟达的估值倍数会下调，除非它的盈利数据真的下滑。

So it's tough to see, you know, necessarily the NVIDIA multiple coming down, unless, of course, the metric comes down.

我觉得更值得问的是，现在超大规模云服务商的资本开支里，有四到五成都花在了内存上。

I think the the I the way I would pose this question is, right now, we're talking about 50 40 to 50% of CapEx of hyperscaler CapEx going to memory.

我们得想清楚，这种情况的上限到底能到哪里？

Think I the question we have to ask ourselves is what do we think the art of the possible on that can be?

然后你自然会考虑到行业的周期性、现金流的稳定性，还有利润率，再反过来推算。

And then I think you you obviously talk a lot about the cyclicality, about the cash certainty, about the margin profile, and then you do the math backwards.

换一个角度看，还可以从市值的角度来分析。

Another way of thinking about it also is from a mark from a market cap perspective.

如果我们把美光、海力士、三星、铠侠，还有西部数据这些做存储芯片的公司都算上，它们都受同样的宏观趋势驱动，那它们的盈利能力和对应的市值加起来会有多大？

If we were to get Micron and Hynix and Samsung and Kioxia and, you know, WDC, Western Digital, the drives, the memory folks that are underlying the same macro trend, how much implied earnings power, and then as a result, market cap are we talking about?

你很快就会发现，这个数字比所有客户加起来还要大得多。

I mean, you pretty quickly get to a number that's much larger than all of their customers combined.

这就引出了一个关键问题：价值到底在哪里积累？

And I think that's where you get into this whole question of, you know, where does value accrue?

当然，现在内存是瓶颈，但从长远来看，你觉得这些内存公司会不会比所有超大规模云服务商加起来还大？

And obviously this is the bottleneck today, but in the arc of time, do we think the memory companies are gonna be larger than all the hyperscalers?

这其实引出另一个有趣的问题：这个瓶颈会持续多久？

Well, I think that would lead to an interesting question of how long that constraint lasts.

就像马特提到的，大家都说晶圆产能会陆续上线。

And as Matt mentioned, we all talk about, fab capacity will come online.

关键在于什么时候能到位，以及到时候供需失衡会是什么样子。

The question is when, and the question is what does the supply demand imbalance at that point in time look like?

对。

Right.

马特，我想聊聊Cerebras。

Matt, I want to go to Cerebras.

他们提到，市场对他们的利润率存在误解。

So they pointed to margins that they said were misunderstood.

安德鲁·费尔德曼在CNBC上就是这么说的。

That's what Andrew Feldman said on CNBC.

你对全年预测中更低的利润率怎么看？

What was your read on the margins, which were narrower for the full year in the forecast?

是的。

Yeah.

Cerebras在IPO期间就明确说过，随着OpenAI的加入，会对毛利率造成压力。

Cerebras, think during the IPO proceedings was pretty open about the fact that as OpenAI comes on, it it it's a drag on on gross margins.

我觉得那里头有两方面原因。

Just I think there's two things going on there.

第一，这是一个非常大的客户，所以他们会拿到比较优惠的条件。

One, very large customer, so they're gonna get somewhat favorable terms.

第二，因为要花时间增加数据中心的容量来满足这份协议的首批需求，他们不得不把之前卖给其他客户的容量重新收回。

But then two, because it takes time to bring data center capacity up for them to satisfy the the first tranche of that agreement, they have to go and write back capacity that they've they've sold to another customer.

而这比直接用自有容量为OpenAI提供服务要贵得多。

And that's just more expensive than servicing OpenAI out of their out of their own capacity.

所以我觉得，你看到的就是这个情况。

So I I think that's what you're what you're seeing.

说实话，我觉得投资者真正感到意外的并不是这个。

I honestly don't think that's necessarily what what surprised investors.

相反，实际的毛利率下滑程度，可能还没大家之前预测的那么严重。

And and if anything, the the actual gross margin deterioration isn't isn't quite as bad as I I I think people had modeled.

真正拖累股价的，是这种特殊的限售安排——在第一个财报期结束后，股份就会解禁。

Rather, I think what is really weighing on stock is it it has this strange lockup configuration where shares get unlocked after that that first earnings period.

所以在我看来，市场上可流通的股票变多了。

So, you know, the way I look at it, there's just there's just more stock available.

这可能是股价下跌的主要原因。

And that's probably the the the greater reason the stock's down.

说实话，今年发生的事对Cerebras来说其实没那么重要。

I I mean, honestly, what happens this year doesn't matter that much for Cerebras.

你投资Cerebras，是因为你相信它未来会成为AI算力市场的主要玩家之一，占据5%、10%、15%的份额，不管具体数字是多少。

You own Cerebras because you think they are going to be one of the major players in in serving the AI compute market and that they'll have 5%, 10%, 15% share, whatever the number is.

如果这个情景真的发生，这家公司的价值会远高于现在。

If that scenario plays out, the company is a whole lot more valuable than it is right now.

没错。

Right.

穆斯塔法，我很好奇，你投资早期公司，芯片企业当然也在你的关注范围内。

Mustafa, I'm curious, you invest in early stage companies and chip companies are certainly part of the pool of companies that you consider.

我的意思是，现在如果是一家芯片公司，需求这么旺盛，你看有几十家、上百家芯片初创公司冒出来。

I mean, is the strategy right now that if I'm a chip company, there's so much demand, and look, there are dozens, hundreds of these chip companies that are starting starting out.

现在的问题根本不是能不能抢到市场份额，因为市场本身增长得太快了。

It's not really a question of can you take market share away because the market is growing so fast.

所以我想问的是，你该怎么去分辨这些呢？

So I guess my question is, how do you even decipher these?

当市场增长这么快的时候，你怎么判断哪家芯片公司更优？真的有那么重要吗？

How do you distinguish between these chip companies when does it really matter which one is better if the market is growing so fast?

说实话，咱们得直面现实。

I mean, let's be super real.

对客户来说，最关键的问题其实是：你什么时候能给我供货？

I think the biggest question for any customer is when can you get me those chips?

只要你现在能给我芯片，管它能效比怎么样、封装条件如何，我都买。

If you can get me those chips today, I'll buy whatever that chip is, whatever the perf per watt is, whatever the packaging constraints are, etcetera.

当然，这背后其实跟架构密切相关。

But that of course is impacted by architecture.

如果你走晶圆级设计，就不必依赖HBM，也不用担心先进封装，不用操心COWAS，甚至可能都不必用最先进的台积电制程。

And so if you're doing the wafer scale approach, you don't have to rely on HBM, you don't have to rely on advanced packaging, don't to rely on COWAS, you might not have to even rely on the leading edge KSMC node.

这三样东西——先进制程、COWAS和HBM——才是真正卡脖子的瓶颈。

Those three things, specifically leading edge node, COOS, and HBM are constraints or bottlenecks.

所以，你遇到的瓶颈越少，哪怕性能稍微降一点，也能更快把芯片推出来。

So the less bottlenecks you have, even if that reduces performance, the faster you can get your chips out.

另一方面，大家都在同时扩大产能，研究架构，优化各种限制条件。

On the flip side, everyone's obviously ramping capacity and thinking about architectures and doing a lot of the constraint optimization around the same time.

当然，我们一会儿再聊Jalapeno。

And obviously, we'll talk about Jalapeno in a second.

但要是看27年、28年，会有大量芯片产能陆续上线。

But if we're talking about 27, 28, we have a lot of chip capacity coming online.

我觉得那时候更关键的问题是，像你刚才说的，你到底能卡在哪个位置上。

That's where I think it becomes more of a question of what kind of what you said, which is where do you slot in?

而且我觉得，我们一直都在讨论预填充解码、推理任务的解耦。

And I think, you know, all of us have been talking about prefilled decode, the disaggregation of inference.

你可以设计专门做推理的芯片，专门提升响应速度，满足模型推理时对即时交互的需求，同时把KV缓存的处理、HBM的使用这些部分单独拆开来看。

You can have these inference focused chips that can handle more of the speed and the kind of fast essentially, interactivity that you need to have with these models on inference time, but you can separate how KB caches are dealt with or HBM can matter.

不过有意思的是，一旦你去掉了这些瓶颈，很多人还是想要那些高端配置——先进封装、COWAS、HBM和高速内存，因为这些能让你更好地支撑更大规模的模型。

Now, obviously, it's interesting because because if you take take out the bottlenecks, people would probably a lot people would still like to have all of those nice to haves, advanced packaging, COWAS, of course, HBM and fast memory, because you can handle a lot more scale in these model sizes as well.

所以我觉得关键还是时机，还有每瓦性能，当然，这是在美国的情况。

And so I think really it's about timing and then performance per watt, of course, The US.

在外部，每瓦性能就没那么重要了，能源不是那么大的限制因素。

Externally performance per watt matters less, energy is less of a constraint.

所以这也是另一个考量维度。

So that's another dimension as well.

说到底，这全都是在各种约束条件下的优化问题。

It's it's really all constrained optimization based

跟什么相比呢？

compared with what

就是你刚才说的，关于整个市场本身的情况。

what you said around the kind of market itself.

马特，你对这个问题怎么看？

Matt, what what do you think of this question?

我觉得芯片的质量依然很重要。

I think that that chip quality still matters.

先说一下，我学的是历史。

And to preface it's I I'm a history major.

所以我并不设计芯片。

So I I I don't design chips.

我也不选芯片。

I don't choose chips.

但我一直在关注微软。

But I I I look at Microsoft.

我也在关注Meta。

I I look at Meta.

他们已经为定制ASIC投入了两三年，但你没看到多少量产的迹象。

They've been working on custom ASICs for two, three years now, and you're you're not seeing a lot of volume.

所以，只要有个芯片就够了。

So just having a a a chip was good enough.

我觉得微软和Meta早就该加大投入了，可实际上并没有。

I I I think you would have seen both Microsoft and Meta ramp their efforts, and and we haven't.

说到Cerebras，让我印象最深的是，他们明显想和OpenAI合作。

With Cerebras, the the the thing that stands out to me is, look, they they they wanna deal with OpenAI.

他们也想和亚马逊合作。

They wanna deal with Amazon.

但你没怎么看到Tenstorrent有类似的动静。

You don't see the same news for Tenstorant.

Grok那边也没传出过类似的消息。

You didn't see the same news for Grok.

Sabanova那边我根本没听过任何类似的消息。

Haven't heard anything similar from Sabanova.

这些公司其实都在推出自己的解决方案。

These are all companies that are out there with with solutions.

所以我觉得，这本身就说明Cerebras的芯片确实不错。

And so I I think that that effectively tells you that Cerebras has a has a good chip.

我觉得，现在解决方案的质量真的特别关键。

And I think that the the quality of solution certainly certainly matters at this point.

话虽如此，我觉得你看英伟达，华尔街低估它的原因之一，不只是因为他们方案出色，更关键的是他们牢牢掌控了整个供应链——别人根本抢不到HBM基板、COWOS这些关键部件。

Now having said that, I I think you look at NVIDIA, one of the reasons I think that Wall Street is is underestimating NVIDIA is not only do they have a good solution set, but they also have locked up much of the supply chain that you just can't get competition because no one can get HBM substrate, COWOS go down the chain.

明白了。

So, okay.

那我们现在来聊聊Jalapeno芯片的质量吧，穆斯塔法。

So now let's talk about the quality then of Jalapeno, Mustafa.

既然质量这么重要，你听到的早期反馈是，Jalapeno到底有多强？

I mean, if that matters so much, what are early reactions that you're hearing to how good Jalapeno might be?

这可是OpenAI和博通刚发布的全新芯片。

This is, of course, the new chip that OpenAI and Broadcom unveiled.

是的。

Yeah.

我觉得这正是典型的‘买预期，卖事实’。

I would say this is a classic example of, you know, buy the rumor, sell the news.

我觉得，特别是听到他们九个月就完成流片的时间，确实让人兴奋——芯片迭代速度现在快得惊人。再想想电子设计自动化，怎么设计芯片，到底能产出多少芯片，这些都因为速度提升而变得完全不同了。

I think, you know, while it's interesting to to particularly hear some of the numbers of the nine months to tape out time, which I think continues to make the kind of chip velocity really exciting for people, and as you think about electronic design automation, and how do you design chips in the first place, and how many chips can we have, because that velocity now is so much faster.

至于工具和模型本身，OpenAI也毫不意外地表示，他们用自家的模型来辅助芯片设计。

And the tooling and the models themselves, OpenAI said the and to no surprise that they use the OpenAI models to help design the chip.

当然，私营领域还有很多公司，比如Recursive，正在帮像谷歌、DeepMind这样的机构开发芯片设计解决方案。

And of course, there's lot of companies in the private markets, including Recursive, that are helping people, know, from Google, DeepMind, Alpha Chip design solution.

你到底该怎么用人工智能来设计芯片呢？

How do you design chips with AI?

又该怎么让这个设计速度更快呢？

How do you make that velocity faster?

对我而言，最关键的收获其实是关于设计速度这一点。

So I think for me, the biggest real takeaway was around the velocity side of things.

当然，他们提到的时间线也很重要。

And And then of course the timeline that they talked about as well.

我们说的这款芯片，就像我之前提到的，内置了HBM，而且有八层HBM堆栈。

So we're talking about a chip that, like I mentioned earlier, has HBM in it, has eight stacks of HBM.

我相信随着时间推移，这种堆栈还会越来越多。

And I'm sure over time we'll have more and more.

所以他们并没有回避HBM的限制。

So they're not avoiding the HBM constraint.

既然他们和博通合作，大概能拿到台积电最先进的产能，而且还在用COLOS和先进封装技术。

Given they're working with Broadcom, they'll probably have leading edge TSMC capacity, and they're doing COLOS and advanced packaging.

这并不是一个针对瓶颈优化的芯片。

So this is not a bottleneck optimized chip.

这是一个追求性能的芯片。

This is a performance optimized chip.

这是专门为他们自身工作负载量身打造的芯片。

This is a chip that's tailored to the workloads that they have.

我原本以为功耗规模会更大一些，但也许他们是分阶段推进的，毕竟还得走传统的半导体制造流程。

And I think in a lot of what we expected maybe maybe I expected the gigawatt size and the scale to be a little bit larger, but perhaps they're just sequencing that given they have to kind of go through the traditional semiconductor manufacturing process.

对，没错。

Right, right.

马特，最后一个问题。

Matt, last question for you.

你负责报道高通。

You cover Qualcomm.

高通这周举办了投资者日，宣布收购了一家叫Modular的AI初创公司，我们《信息》这边之前就一直在关注他们。

Qualcomm had their Investor Day this week, and they announced the acquisition of Modular, an AI startup that we've been tracking here at The Information for a little while.

这事儿挺重要的。

It was a big deal.

我听说收购价大概在40亿美元左右。

I think it was reported to be $4,000,000,000 thereabouts.

这笔交易是全股票支付的。

It was an all stock transaction.

我真正想问你的问题是，Matt，Modular的专长我记得是想挑战CUDA，也就是围绕芯片的软件生态。

My question really for you, Matt, is Modular was sort of I mean, their specialty here was they were trying to take on CUDA from what I recall in terms of the software around the chips.

你觉得高通为什么想进入这个领域？他们有希望成功吗？

Why do you think Qualcomm wants to get into this business, and do you think they have a chance here?

先说一句，别去报道高通。

So just to preface this, don't cover Qualcomm.

好的。

Okay.

这属于我的领域，因为是硬件和半导体相关。

It falls in my universe because it's hardware semiconductor.

我今天早上在你的报告里看到过，还以为你在覆盖这家公司。

I read it in your note this morning, so I thought maybe it was in your coverage.

没有。

No.

但我一直关注所有半导体和硬件领域的动态。

But I talk talk about everything semiconductor and hardware.

我觉得，Kunu当初打造的那套软件负担或模式，至少在推理层面，已经没那么重要了。

So what what I would say is it it almost seems to me that that that software burden or or mode that that Kunu created has dissipated a bit, at least in terms of inference.

对吧？

Right?

某种程度上，现在一个令牌跟另一个令牌也没太大区别。

To some extent, it feels like a token is a token.

所以我对这在多大程度上改变了高通瞄准令牌市场的叙事，有点拿不准。

And so I'm a little bit less certain around how much it it changes the narrative for for Qualcomm in the sense of just going after that that token market.

不过，稍微推测一下，软件在瞄准这些世界模型类应用时确实很重要，你说是吧？

Now just to to speculate a bit though, it it does seem like software is very important in targeting some of these world model type applications, if you will.

比如汽车市场或者物理AI，也许这正是高通在关注的方向。

So the automotive market or physical AI, and maybe that's that's what Qualcomm's looking at.

毕竟，他们在收购方面已经整合了一整套相当出色的资产。

I I mean, certainly, has put together an impressive set of assets in terms of all the acquisitions that that they've made.

而且我也注意到模块化是其中的一部分。

And and I see modular modular is part of that.

但我还在等，等高通能从这些收购中真正带来可观的收入。

What I'm still waiting for, though, is for Qualcomm to generate substantial revenues from those acquisitions.

对吧？

Right?

目前看来，高通在手机以外领域的表现，似乎还不如他们在手机业务上那么成功。

It seems like today, Qualcomm's efforts outside of handsets have been a little bit less successful than than they've been in the handset side.

很好。

Great.

感谢Matt和Mustafa今天做客。

Well, Matt and Mustafa, want to thank you for coming on.

这位是来自Wedbush的Matt Bryson，还有来自NEA的Mustafa Neemuchwala，欢迎来到TITV。

That is Matt Bryson from Wedbush and Mustafa Neemuchwala from NEA here on TITV.

说到世界模型，现在业内对世界模型和视觉-语言-动作模型（VLA）的争论越来越激烈。

Speaking of world models, there is an intensifying debate between world models and vision language action models, or VLAs.

这场争论正在科技领袖和风险投资人之间造成明显的分歧，焦点是机器人和物理AI的未来走向。

That debate is creating a sharp divide among tech leaders and venture capitalists over the future of robotics and physical AI.

我们的AI和机器人记者Drew，今天早上在《AI简报》里专门写了这个话题。

Our AI and robotics reporter, Drew, wrote about that exact topic in this morning's AI agenda newsletter.

我想请Rocket上来聊聊，帮我们理清这些概念到底是什么意思。

I want to invite Rocket on to talk about it and help us understand what these terms are all about.

Rocket，欢迎来到节目。

Rocket, welcome to the show.

很高兴你再次来做客。

It's great to have you back.

嗨，阿卡什。

Hi, Akash.

能来这儿真的很棒。

Great to be here.

好的。

Okay.

世界模型和视觉语言动作模型。

World models and vision language action models.

对吧？

Right?

我说得对吗？

Do I have that right?

就是VLAs，对吧？

VLAs?

这些分别是什么？

What are each of those?

解释一下它们。

Define them.

是的，没错。

Yeah, definitely.

VLA 这个概念相对简单一点。

So a VLA is a little bit more straightforward.

我们先从这个说起。

Let's start there.

你先拿一个语言模型，就是那种能驱动你最爱的编程助手或聊天机器人的模型。

You start with a language model, like an actual language model, like the same kind that powers your favorite coding agent or a chatbot.

然后你再对它进行微调，用更多数据训练，让它能控制机器人。

And then you customize it, you train it on further data so that it's able to control a robot.

核心想法是，你从一个本身就具备一定智能的基础开始——它懂得推理，能理解世界并制定计划，再把这种智能应用到现实世界中，教会它如何操控机器人。

And the hope, the idea here, is that you start from a base that's already kind of intelligent, it knows how to reason, knows how to make sense of the world and come up with plans, and you take that intelligence and then you apply it to the physical world by teaching it how to control a robot.

这大概就是他们的愿景了，明白吗？

That's kind of the vision there, okay?

世界模型则是另一种思路。

World models are kind of a separate approach.

当我们说世界模型时，其实指的是那种对物理规律有深刻理解的AI模型。

When we talk about a world model, really we just mean an AI model that has a really good understanding of physics.

有时候这里的界限确实有点模糊。

I think sometimes the lines get a little blurred here.

有人可能会说，理想情况下，VLA也应该具备很强的物理理解能力。

People could say, Well ideally we'd have a VLA that has a really good understanding of physics.

但现实中，人们提到世界模型时，意思其实不太一样。

But in practice, people mean something kind of different when they talk about world models.

他们通常指的是那些在大量视频数据上训练过的模型。

They're often talking about something that was trained on a lot of video.

也就是说，模型看过很多来自互联网的视频，比如人们与周围环境互动的场景，或者环境随时间变化的画面，它通过预测下一帧画面，来学习一些关于物理规律和世界运行方式的知识。

Meaning it saw a lot of videos, say, just videos collected from the internet, that show people interacting with their surroundings, that show environments changing over time, and the model learns to predict the next frame of those videos, and in doing so, hopefully learns a thing or two about physics and how the world works.