DeepSeq showed Wall Street that China could build cheap, powerful AI.
Now, Ji-Poo, the AI, is showing how that advantage could spread.
I've been consistently surprised by how quickly the open source has caught up.
And I think GLM 5.2, you're kind of seeing the first model where it's really competitive with like Opus 4.7
and some of these like very closed source frontier models.
What GLM 5.2 kind of shows is, you know, we might still be in this sort of three to six month territory.
Everything about the economics of AI changes depending on which outcome we're in.
We need costs to come down. That's how you drive adoption. That's how you get returns.
The AI leaderboard is starting to look incomplete.
The next fight may be over intelligence per dollar.
Models that are capable enough for real work and cheap enough to run constantly.
That is where ZAI is forcing a new conversation.
Ji-Poo's latest model, it's called GLM 5.2, comes out of China.
It has landed with a bang in Silicon Valley, blowing past all other open source models,
nearly matching the American frontier for just a fraction of the price.
Developers, they're piling in open router token traffic showing much quicker adoption for GLM 5.2
than DeepSeek's V4 launch back in April.
It was a big deal then, it's a big deal now.
This just carries the story that we've been reporting on even further.
It all matters because Ji-Poo is hitting at a different moment though.
The trillion dollar sell-off post-DeepSeek that was treated as a kind of a one-off shock,
partly because people saw it as a chat bot story.
GLM 5.2, this is different.
It's strong at agentic work and that is key.
On one agentic benchmark, it is just one percentage point away from Opus 4.8 for a fifth of the cost.
Now, Opus 4.8, that is Anthropics' most powerfully available model.
So basically, you're getting very close to the same horsepower for just a fraction of a cost.
And it really comes at a time when expensive AI is already eating into budgets.
And that is a really hard trade-off for enterprises, for companies, for Fortune 500 to ignore.
Agentic AI is only going to intensify those costs.
More complex tasks, more steps, more tokens.
So now that Ji-Poo, also known as Z.AI, is in the picture,
agentic open source, it might be the next big threat out of China.
And here's a part of the story, the AI story, that I think Wall Street is still missing here.
For the last few years, everyone has been obsessed with these AI leaderboards.
Who has the smartest model?
Who's the best at coding?
Which one is ahead on reasoning?
But that is not how companies buy software.
Companies are increasingly asking, what is good enough and what does it cost to run this thing a million times or more?
All of my employees running it a million times over the course of weeks, months, a year.
So the new metric in AI, it is intelligence per dollar.
And one reason Chinese models are pushing so hard on that metric is because of distillation,
which is basically you take a big expensive model, use it to train a smaller, cheaper model to act like it.
The American AI story, it has been built around bigger models, bigger data centers, huge spending.
But Chinese labs, they're putting these cheaper versions front and center, very, very close to the frontier performance without the cost.
That's why GPO is interesting.
Artificial analysis has this chart that looks at both sides of this.
How smart is a model and how much does it cost to run it?
Now, the most attractive quadrant, it's obvious, high performance, low cost.
That's where every company wants to be.
GPO's GLM 5.2, it's getting very close to that sweet spot.
It's not quite at the top in terms of performance.
You can see it here, very close to that green square.
So not at the very top in terms of performance, but it is close enough to the best models from OpenAI and Anthropic to make this price gap hard to ignore.
And on agentic work, that gap matters even more.
These are tasks where models don't just answer one question.
They plan, they code, test, fix mistakes, they loop, they keep going.
That is a lot more expensive.
So if you're a company, you need a great model, cheap enough to use over and over again.
And that is the setup.
I want to get into all of this with two people who are actually living this.
Aaron Levy, he's been running Box for 20 years.
He saw the move to the cloud, the move to mobile, the move to AI.
He is one of the most plugged in enterprise guys in the valley on what this stuff actually looks like when it hits a real company.
We've also got Gabe Pereira, co-founder of Harvey, which is the AI platform that half of the AM Law 100 is now running on.
Now, before that, he was a research scientist at DeepMind and Meta, so he has seen this from both sides, the model layer and the application layer.
Aaron and Gabe, it's great to have you both on.
We got Aaron too?
Okay, we're going to, I can't, let's see, do we have both their sound?
Okay, we're going to try and get this sound fixed.
So, Aaron, you're an observer now.
I'm assuming that you can hear Gabe and I, and then I'll come to you on this.
Gabe, let me start this off with what's happening with, you know, GLM.
It feels like my entire X feed, everyone I'm talking to here just cannot get enough of this new model GLM 5.2 right now.
Kind of feels like DeepSeq all over again.
People also aren't being shy and calling that maybe even bigger.
So, I guess first, how do you explain this to someone outside of tech or like a non-tech CEO who has been all in on American AI,
but sees all of this buzz around a new model that is almost as capable, but a lot cheaper?
Yeah, I would say the big question since kind of ChatGPT came out in these closed source models is just how big is the gap between the closed source and the open source?
And I think I've been consistently surprised by how quickly the open source has caught up.
And I think GLM 5.2, you're kind of seeing the first model where it's really competitive with like Opus 4.7 and some of these like very closed source frontier models.
And exactly to your point, it lets you think about the cost curves, where you need frontier intelligence, where you can use open source intelligence.
And I think this is what we're starting to figure out and a lot of companies are starting to figure out now.
Right. So explain that kind of in like just basic terms.
If you are a company CEO and you have this AI budget, but you've seen, you know, your company blowing through it, when you get something like this, what should they be thinking?
Especially when you say like it almost reaches the frontier, the very best models, but for a fraction of the price.
People you are talking to, Gabe, do you think that they're taking a second look at Chinese and these open source models or should they still be concerned about them?
Yeah, I think every company will use a mix of both, but exactly to what you said, it's starting to think about which tasks do I not need to use frontier intelligence.
And so as these models get smarter, do you need the smartest model reviewing your contracts or doing simple red lines?
And so I think for every company, the same way you organize a large company by figuring out, you know, I need different seniority.
I need to pay different people for different roles.
I think you're going to start seeing this with agents where there'll be some agents that are making critical decisions.
You want frontier intelligence and there'll be some agents that are doing simpler tasks.
And there you can use open source and you can really reduce the cost of of intelligence.
Right. And are you seeing sort of the buzz that I am as well as your feed, the people you're talking to, are they very excited about GLM five, two?
And like, can you explain why this is so interesting, why this is different than deep seek?
I tried to, but I'd love to hear it from you.
Yeah, I think we're seeing the same buzz in terms of our benchmarks.
It was kind of a big open source jump.
I think it's the first model that some of my friends at the labs were like, oh, this is like a very serious open source model.
And so I think it's kind of similar.
I would say to me, the deep seek moment was much more of I think there's always this question of how much money do you need to spend to stay on the frontier?
I don't think that's changing.
I think it's still very expensive to train these frontier models.
I think the question is just how quickly can open source catch up to that frontier as the frontier keeps moving?
Right. And that gap keeps like becoming narrower and narrower.
Aaron, I think we've got you now, right?
Yes. Okay. I can hear you.
You know, it's a live stream, Aaron. I know you're used to TV.
We kind of shoot from the hip here.
I, we, we, we need AGI as soon as we can get it.
So we will have audio figured out.
Yeah. Okay. Agreed.
Okay. Give me your thoughts.
I know you've been tweeting about this.
I know you've been tweeting a lot about open source.
You've been talking about this for years, kind of as I have as well.
What is your take on this moment?
Is it different than anything else we've seen over the last few years?
It certainly feels like that from where I sit.
Yeah. I mean, I totally concur with, with Gabe.
The, the, the, you know, there's been a big question in the Valley, which is how far behind is the open weights, kind of architecture from frontier closed sourced models.
And there's like fundamentally different outcomes in the future of AI.
If we see an, sort of an exponentially kind of increasing gap or a narrowing, or at least sort of sustaining three or six month gap.
And what, what, what GLM five two kind of shows is, you know, we might still be in this sort of three to six month territory, as opposed to the kind of fast takeoff scenario of closed weights models versus open weights.
And, and everything about the economics of AI changes depending on which outcome we're in.
I think to your, to your intro, I actually do think for the, for the next, you know, year or two years, actually frontier level intelligence still is incredibly important.
I think enterprises are only going to kind of peel off sort of workloads that are sort of good enough on, on open weights models.
When they, when they really feel like we've reached a level where you can kind of, you know, bank on, on the level of intelligence we're getting.
I still think that you, you, you need fable level kind of intelligence for a lot of the types of, of use cases that we're seeing in the enterprise.
But what this shows, and you saw the founder of, of ZAI really talk about, you know, the ability to get to fable level intelligence and, you know, within the year with an open weights model.
I think that, that does fundamentally change the architecture of AI because it means that now, whether you're using Harvey or box or factory or cognition or, you know, any, any other kind of neutral, you know, AI layer.
You can kind of route the workloads between frontier intelligence and open weights models as appropriate for, for your tasks.
So it's, it's very meaningful for what the future of the kind of AI stack looks like.
Right. So the big thing right now is model routing, right?
As Erin, as you laid out, it's use the frontier models for the hardest tasks, use the, you know, open source models for the easier ones.
But you did say for the next one to two years, you think that frontier is going to be really important.
And yes, the ZAI see our founder said that it was only a year.
So what happens after that?
Is that gap becoming even shorter?
Like you said, three to six months, does that last or does that eventually go away as a whole?
I think you'd be fine with a three to six month gap.
Because that, that's not enough time for, you know, the, the big issue is, is what is the gap where if I don't use frontier on everything that my firm or my organization loses some kind of meaningful competitive advantage.
And you can totally imagine a plausible scenario where if that gap was two to three years, you would, you would never use the kind of, you know, second best model because you just simply wouldn't be able to produce as effectively as a company in, in the code you're producing or the, the legal, you know, analysis you're doing for your clients or the, the marketing assets that you're generating.
But in a world where there's a three or six month difference, that's, that's totally within the noise of, of the, the general diffusion of AI anyway, within an organization.
So, so that, that's why you kind of see this fundamental shift.
You don't need open weights to be at the exact same, you know, kind of the same week as, as frontier intelligence, but you can't have it be two or three months, two or three years behind.
And it appears that we still are remaining that really kind of important zone where, where you're going to get very, very high quality open weights models.
And then, you know, that gets to the second part, which is now you can post train these models on just your tasks.
You know, everybody kind of looks at it as a, as a pure kind of cost of the token kind of equation, but actually just as big of a deal is the fact that now I can go and post train this model just on the purpose built use cases that my, you know, particular domain requires.
And that has a different set of performance access, axes that matter, which is like, what are the capabilities I need the model to be able to do that actually now can be like built into the model in a way that, that you don't always get from kind of general purpose intelligence.
Right. And that changes the dynamics, like once again, right, Gabe, I was going to come to you because you're actually operating in this app layer and like at the beginning, you're dependent on the frontier labs, open AI, Anthropic.
And that's what we're starting to do for kind of both costs and performance reasons.
And I think to the kind of longer term question of what is going to happen with frontier intelligence, I think the shift we're starting to see from models to agents.
I think when you just use models, it was easy to kind of train a model and then that was separate from the rest of your business.
And you could swap these out. If a model got better, that was kind of easy to replace.
And then you could swap these out. If a model got better, that was kind of easy to replace.
And then you could swap these out. If a model got better, that was kind of easy to replace.
As you start moving to the to these agentic systems, this AI, these agents are getting so embedded in your business that creating the next version of frontier intelligence.
I think a lot of that training data is going to come from the companies themselves.
And so as a simple example, when a law firm works for like a large private equity firm and they help them do M&A, like that training data and creating frontier models can do that type of work like that training data can't go into the general models.
And so I think there's going to be this really interesting question in the future of where is all of that training data eval coming from as these models start doing more and more economically valuable work.
And part of the motivation for us of training these open source models is how do we allow our customers and each of our law firms or kind of in-house teams to build these systems?
Right. And keep them for yourselves, right? You don't want to train someone else's model so that they can go and compete with you.
What I'm hearing from both of you does not sound good for Frontier Labs or a company that has built its entire business model on staying ahead. Am I reading that right, Aaron?
I feel like you're not going to say that, though, because you work with all of them. But like, what is the case for continuing to buy AI?
I feel like I'm trying to ask that. What's that? Sorry?
You know, what's the case for continuing to buy AI from Anthropic or OpenAI when that gap you essentially said is like, we're kind of there. It doesn't matter anymore. We're there.
Well, I actually didn't say that. And this is where it's very nuanced. I think what you're going to see is sort of this barbell dynamic where Frontier Intelligence, again, you know, kind of might be always at two or three, five, 10% better than whatever the open weights equivalent is at any given time.
Like that's the kind of scenario that as just a mental model is sort of, you know, you might want to just sort of, you know, kind of establish. And so what you'll have is that you'll still actually have an incredible demand for Frontier Intelligence, but often as things like the orchestrator or the planner or the reviewer of the work that's happening.
And so interestingly, you might spend the same amount of money on both Frontier Intelligence and Open weights intelligence, but the token volume might be entirely different.
You might have, you might have, you know, a third of or a fifth or a tenth of the tokens going into the Frontier Intelligence. Those will actually cost five or 10 times as much.
But the bulk of sort of the heavy document processing, heavy analysis and combing through every sort of, you know, piece of text in the repository that you're trying to analyze.
All of that might be happening on, again, a cheaper, faster model that is sort of more tuned for that workflow.
So I actually don't know that this spells anything negative for even the Frontier Labs right now, assuming you are bullish on AI.
If you're bullish on AI, what happens is we're going to be using, you know, a thousand times more tokens in 10 years from now than we are today.
And you're still going to want Frontier Intelligence for a lot of your most important sort of work that an agent is doing.
But there's just so many tokens that you now need for these agentic systems that actually you do have to then have something that is sort of cheaper and faster for a large amount of the work.
The other nuance that I would just throw out there is none of this means that a Frontier Lab won't have cheaper and faster models.
And so right now we're seeing that come from these open weights models as an example.
But, you know, OpenAI released an open weights model a year ago as an example.
And there's sort of no reason why if they became a great inference cloud as an example, that you wouldn't just run that model in their environment.
So it's not obvious that this kind of tells you really anything about the market structure at the Frontier Labs other than the market structure of the applied layer is now the very logical place to have a lot of your routing go because it is sort of independent and model agnostic.
And so it's going to just route the workload depending on what the use cases.
Right. So that, you know, three to six month gap that is going to exist is going to be worth paying for, you're saying, for some of the hardest tasks.
But, you know, Aaron, you bring up a good point. Yes.
OpenAI, Google, Gemini have trained open source models, but they don't show up in terms of the adoption data and really on the benchmarks.
Why are they letting the Chinese models run away with this?
And I mean, right now it feels like it's maybe looking for something from reflection.
But where are the Frontier Labs here? Do they turn their attention to it?
It's hard to make money from open source. Is that part of it?
Gabe, maybe you can go first.
Yeah, I think it's kind of depends on where you are in the stack.
So if you're kind of the chip providers, the cloud providers, I think you want there to be a lot of good open source models.
I think if you're a closed source, you'd rather this gap be a bit bigger.
But I think to Aaron's point, there's room for both of these.
And I think like the US versus China have just kind of taken different strategies of I think we are pushing the frontier with closed and their strategy is kind of doing more open.
But I think you've seen companies flip flop. Right.
You've seen companies catch up and then go close source or, you know, release open source.
And so I think it's not clear what the right strategy is yet.
But I think China's like open source models are getting like very good.
And that's something to like pay attention to.
Right. And I wanted to sort of make this connection to I know, Gabe, like Harvey is looking at post training.
I don't know what kind of model you're going to use.
Is it a Chinese open source model, do you think?
We need to support kind of a range of these.
So we think of it like the same way we're agnostic to closed source models.
We need to be agnostic to open source or we have some customers that can't use Chinese models for certain sensitive legal work.
We have European customers that want to use Mistral.
We have customers that want us models.
So I think you'll want kind of this healthy ecosystem of open source models by region, by country.
Yeah, that's right.
Yeah. And I would I would say that that if I if I could kind of like design the perfect strategy of, you know, kind of one of the top few labs and Google is sort of the closest to this with with Gemma.
I actually think you would want to consistently have a just close to frontier open weights model because now you kind of you get the ecosystem effects on both sides.
And and that barbell that I was talking about of you want frontier intelligence from GPD, you know, 5.5 or Fable or, you know, the rumored GPD 5.6.
But there's no reason that the the cheaper, faster version isn't in the same kind of model family from the same from the same sort of underlying weights and provider.
So I actually think from the game theory standpoint, you'd probably want to be heavily supporting open weights.
Now, there's a probably a philosophical reason why Anthropic in particular might not do that strategy.
But for open AI where they already have precedent for doing so for Google, which clearly is doing that with Gemma, which is actually pretty actively used.
It's just not not as obvious that it's at the at the frontier of a GLM or or a Kimmy.
I think actually, you know, from a game theory standpoint, they probably should be pursuing this strategy and historical. Right.
I mean, they sort of led the way on open source and only ecosystem for Android.
Why wouldn't they do that again? It surprises me, too, that we don't hear or see more.
I do think about cursor as well. Right. That was sort of laid the playbook.
Gabe, maybe trained a custom model built on Kimmy's architecture. That's Moonshot, a Chinese company.
It's now one of the hottest AI companies right now. And the whole secret sauce runs on a Chinese base model.
And it feels like nobody actually remembers that or cares that much, especially in the corporate world.
So you can see how you can sort of use these Chinese open source models to get ahead just in terms of especially if you're on the app layer.
Aaron, though, I wonder like because I love that you bring up game theory.
It feels like there's a lot of that going on right now.
And so when I think about Fable and Carson, one of our viewers wrote in with a question on this, too,
it kind of seems like a disaster for Anthropic.
I mean, right after they cut off access to Fable Mythos, their most powerful model,
you had all of these sort of policymakers around the world say, you know, how can you build on a model when your access could be cut off at any time?
And it felt like it just gave so much momentum to open source.
Do you think that that matters?
Do you think that, you know, that was misplayed by Anthropic?
What are the what's the impact of it?
Well, it's it's not obvious that well, it kind of depends on, I guess, what what, you know,
you know, either who you kind of like, you know, maybe blame for the situation we're in or kind of what transpired.
But because it might have been misplayed by the US government for the same for the same kind of economic and national security reasons long term.
If if if if you're kind of like encouraging sovereign AI in these other countries by sort of, you know, being the first to show that will will block access to an AI model through export controls.
So so, you know, I actually think from a game theory standpoint, it might be more on the US government side to to have to kind of consider the implications of.
And but absolutely, like like if you are another country right now, this was a sort of shock to the system that says, oh, actually, like, like, AI is not going to be like most other technologies where you sort of, you know, just kind of assume that software is going to work globally.
There's a few nuances where, you know, there might be some privacy requirements in different regions, but generally it's considered just like a an export that the US is always just going to put out there and and to our kind of, you know, kind of economic benefit.
AI, because the export itself is coming with with, you know, possible risks in at least in the table case, you know, the perceived risk of cybersecurity or other factors out now it's it's showing up as maybe it's something that the government wants a bit more control over and who should have access to it, at least now the precedent has been set for that.
So then, you know, by definition, what another country should at least be doing is saying, oh, actually, I need to be able to protect my my ability to have access to intelligence.
And if you were, you know, the EU and if you were, I mean, you know, China's fine because they're building their own models.
You probably would sort of say, hey, we should actually be, you know, fueling, you know, more of either open source investment or post trained, you know, models and be able to get really, really good at making these things better and better.
At a minimum is a hedge, you know, maybe you still do as much work as you can with US labs.
But now you do need some sort of, you know, backstop.
Because if for some reason, we treated access to AI as another kind of geopolitical or economic kind of weapon, you could see, you know, incredibly interesting consequences as a result of that, you know, imagine, imagine AI being used as a lever in in trade negotiations or trade deals in five years.
You would you would sort of that would that could very much threaten your your kind of economic and national security if that was the case.
So so we're, you know, it's really interesting because what's happening is we have these sort of blunt instruments, you know, export controls is a really interesting blunt instrument that I, you know, again, like very smart people probably predicted this.
But like a year ago, I would not have put that in the in the board of possible chess moves of blocking access to an AI model.
Well, now that that's that now that's a possible move, you sort of open up.
Well, well, wait a second, maybe this is like going to be treated as like nuclear arms or not.
I mean, I mean, not nuclear arms, but, you know, any kind of like weapon, in which case, you know, the set of controls that the government has of who has access to it is is way greater than what we had sort of considered before, which again, then means downstream as another country.
You have to be preparing for that. What you just laid out, though, is kind of like exactly what we saw with chips.
Right. And I'm with you. I never thought that could apply to models, but you saw that.
So this is just telling sort of every government around the world, you need sovereign AI, you need to own something.
And this process of distillation, which chips were actually not even as blunt of an instrument, though chips were even more sort of fine grained.
This was this was this was very, very blunt is like non U.S. citizens is like, wow.
OK, there's like no no real way to kind of execute that at scale.
It's even more blunt, yet the solution is like sitting right there for anyone to take.
Right. And it's just distillation. That is what the Chinese models are doing.
It's still inside baseball, but it's this idea that you take a big, expensive, powerful model and use its outputs to train a smaller, cheaper model to behave like it.
And Gabe, I wonder, like, this is how Chinese AI is spreading around the world.
And what Aaron just laid out is the case for governments.
But, you know, Harvey operates in a very sensitive industry.
Do you see the same thing happening? Have you seen the same blowback happen after what happened with Mythos and Fable?
Yeah, I would say distillation is one way.
But I I think these Chinese labs are actually just training their own models.
Like I think it's a bit like extreme for us to just say the only reason they're catching up is because they're distilling this.
They're U.S. models. I actually think they're doing incredible research and like their open source, even independent of distillation, I think is competitive or ahead of ours.
And then I think what you're seeing the labs do is making it harder to distill their models like that's what you saw with Fable.
And it's like I think that strategy will work as these models get smarter.
You can just tell the model, don't let users distill me when they use them.
And it's as they get smarter, they'll get better at preventing that.
So I think that is defensible.
But then I think the kind of the blowback we saw is kind of exactly what Aaron mentioned of even within the U.S., as these systems start becoming mission critical, like you just need to think about business continuity.
And if one of the labs that you're using runs out of compute, the government prevents you from accessing their models.
And so I think that's the way that you're using the data to be able to do that.
And so I think that's the way that you're using the data to be able to do that.
And so I think that's the way that you're using the data to be able to do that.
And so I think that's the way that you're using the data to be able to do that.
And so I think that's the way that you're using the data to be able to do that.
And so I think that's the way that you're using the data to be able to do that.
And so I think that's the way that you're using the data to be able to do that.
And so I think that's the way that you're using the data to be able to do that.
And so I think that's the way that you're using the data to be able to do that.
That's so fascinating. So you see it among sovereigns and then you also see it among corporate America.
And that's why I guess I started by saying it felt like a miscalculation.
You're right, Aaron, probably by the government, less so anthropic.
But the result is sort of the same.
Aaron, I want to give you a word on distillation, too, because, you know, I feel like the smartest people I talk to say that distillation should not be a dirty word.
Right. When you see that anthropic is suing Alibaba, that is non sanctioned distillation, but can actually be a really positive force.
Right. Turning expensive AI into cheaper products that can actually scale.
And as Gabe was saying, like, do you agree that the Chinese are really doing some stuff on the leading edge, innovative edge?
And it's too simplistic to say they're just distilling.
Yeah, I would say I'm not as close to the latest views on what the new method, what the latest methods are.
And I would probably bias toward Gabe's view of that.
And like it also is just like first principles like China has incredible, you know, smart scientists.
Some of them are the ones that came over here and are, you know, leaders at many of these labs.
So it's sort of like a it's a very tractable problem to get very good at at AI and AI training.
And really, it's just been classically like the chip problem of do you have enough compute?
And that's also tractable for the most part.
And then from a kind of philosophical standpoint on distillation, I don't know, I don't I'm not bothered by distillation because, you know, if I were bothered on distillation, I'd probably have to be bothered on the fact that these AI models are just trained on then the public Internet.
It's like we're all benefiting from the collective sort of creation of information and knowledge and everything is sort of riding on on some other set of information that's out there.
Like, you know, I would be bothered by, you know, a model being trained on Wikipedia if I if I was bothered by bothered by distillation.
So so to me, it's sort of all within the realm of, you know, you need access to data.
Obviously, the more intelligent data you can get, the better your model becomes.
And and at the same time, I think that there should be kind of a cat and mouse game because for competitive reasons, you probably want to block that if you're a frontier lab.
But I don't think of it as a as as some kind of, you know, unethical or or, you know, overly kind of, you know, complicated problem.
Yeah, it's a great point. It kind of goes back to the beginning of this, like modern AI era.
All of these models are trained on, you know, the Internet's all of the Internet's knowledge.
That's a form of distillation.
They were they were trained on my my trolling tweets from 15 years ago.
I mean, how mad am I about that? I don't know.
That's a scary, scary thought.
Aaron Levy tweets and all of Reddit.
I don't know, Gabe, I haven't gone too deep into your X feed, but it's trained on that, too, I'm sure.
Last one I wanted to ask you guys about because this is a little bit of a turn.
I want to ask you about Claude Tagg because producer Jasmine, she's very excited about it.
She's trying to get me there.
I don't fully, fully understand the appeal, but she is always right about these things.
I wonder, Aaron, first, I know you've been tweeting about this.
Like, can you break it down?
It's something that always lives in your slack and builds context over time, learns your company.
It feels like lock in to me, but it also seems really important.
Aaron, is this like going to be as big as Co-Work or some of the other kind of land shifting things that Anthropoc has done this year?
I heard to kind of put it in the in the overall kind of graph on that front.
But the reason why it's a pretty big deal is we're very used to AI in kind of single player mode where I have access to Claude or Codex or ChatGPT or Co-Work.
And this is sort of my relationship with AI.
And that AI is effectively acting as me in a variety of systems.
And so it's incredibly powerful because I can then spin up tasks that I would otherwise do a bunch of times across these agents.
But it's again, very single player mode.
And what Claude tag kind of tries to do, and this is, you know, I think building on the recent Zeitgeist of OpenClaw and Hermes and even kind of you're seeing this in agentic coding systems like Factory and Devon, is no, like what if it's a co-worker?
It's not rep, it's not Aaron or Gabe or anybody in the in the organization, it's its own sort of entity.
And it has access to a set of resources that any other kind of entity would have had access to.
And you kind of interacted with it like another user in your system.
And the reason why that's meaningful is that it has shared context across whatever that group that has access to it is doing.
And then you just sort of punt tasks to it and it comes back.
But those are again, kind of group enabled tasks just as again, a colleague would do inside of a Slack channel in this case.
So, so it's, it's, it's probably a very big deal in, in sort of the philosophical direction that it's moving toward.
And, and it has just like a different vector of use cases that, that it enables.
It's less about my personal productivity and more about sort of a shared intelligence for some, from, you know, group of users.
So you're in a Slack channel called the sales team Slack channel.
And, and you're, you're punting off tasks of, Hey, can you generate a deck for this sales presentation?
We're, we're about to go into.
And it needs access to your underlying, you know, marketing assets that are stored in box and your other data, you know, stored in Salesforce.
Just as you would a colleague that was about to go in that sales presentation with you.
So, so that's the kind of benefit of it.
And I think it's another kind of, you know, point in the timeline of like, what is the future of how we're going to work with these agentic systems?
What is the new user experience?
What is the, the, the new interface?
And, you know, from a kind of a lock in point, I think there's an interesting point.
Like on one hand, yes, you're, you're sort of, you know, you're, you're, you're, you're using the cloud particular paradigm.
On the other hand, you can choose to decide how much or how little you want to sort of use the cloud sort of intelligence layer.
Like you can plug in an MCP server of an entirely other agentic system.
So cloud could just be a router to a bunch of other, other tools.
I'm making up the use case, but I'm sure it could be a router to Harvey.
So, so I, I think there's actually like it, you can just design how you want to implement these systems to whatever level of lock in or, or not lock in you want to, you want to drive.
Right.
Feels like it's certainly like the whole industry is heading that way, more of like a buffet.
You can choose what you want.
Gabe, last word to you.
I guess, what should we be looking out for in the next week or so with these live streams?
We really try to be on like what Silicon Valley is talking about this week.
It's very much ZAI's GLM 5.2.
What do you, what should we be watching for the next week ahead?
I feel like it'll be interesting to see what happens kind of with the fable five kind of in the next week.
When that comes back, I think you're going to see more about kind of what Aaron was talking about with Claude Tagg.
Like I think that philosophical shift of just thinking about agents more and more as just employees in your company.
Like you saw Rippling kind of announce they're building kind of a unified like data layer of employees.
I think more and more we're thinking about our product as how do we give agents to a law firm or an in-house department.
And instead of kind of one player like Aaron mentioned, how do you start organizing these kind of like teams of agents and humans to complete these very complex tasks?
So I think you'll kind of keep seeing that trend.
And that's kind of like the trend we're most excited about.
Right. We'll call it multiplayer mode since Aaron's been talking about single player mode.
Thank you both so much, Aaron and Gabe.
Lovely to get you both here.
Thanks for all your insights and hope to talk to you again soon.
Thanks a lot.
All of what we talked.
Distillation open source.
It runs through the same question for investors.
What does it do for the chip trade?
That has been super relevant to Wall Street, of course.
Yesterday we got a big data point.
OpenAI and Broadcom unveiled Jalapeno.
That's OpenAI's first custom inference chip built from scratch in just nine months.
OpenAI used its own AI models to help design it.
Broadcom CEO Hawk Tan says that it cuts inference costs by roughly 50% versus current Nvidia GPUs.
The person I want to break all of this down with is Stacy Rascon.
He has been covering semiconductors at Bernstein for over a decade called the Nvidia trade early.
Stacy, you don't sugarcoat either.
Two decades.
Almost two decades.
Almost 16 years. Is that right?
I should have rounded up.
Oh, then I definitely needed to round it up.
Well, I know that's why we always love talking to you.
You also have all of the technical side, but you can break it down.
Give us your thoughts on Jalapeno.
I mean, to me, it just struck me as a huge, like, develop, like maybe a new era for chip development.
Right? These are some of the most complicated objects ever built by humans.
I guess, like, do you buy it?
Is it this, has it improved this much?
Well, so look, so chips are the most complicated things that humanity has ever built.
And as you know, we've talked before.
I love this space.
And that's one of the reasons why.
So, look, this was not unexpected.
I want to say that.
So we've known for a long time that Broadcom is working on a custom chip with OpenAI.
And in fact, the two companies have a 10 gigawatt deal over the next five years.
And we'll see how much of it actually ships.
But they've got a 10 gigawatt deal for multiple generations of these chips in play already.
On Broadcom's last earnings call they talked about, I think, in 27, they're supposed to ship, I can't remember, 1.2 or 1.3 gigawatts to OpenAI.
So that is presumably this chip.
So the fact that they announced it was not a surprise.
And in fact, one would hope that they would have been announcing it soon.
Because it's supposed to start shipping next year.
They've been talking about it.
Yeah.
Now that being said, look, it's great, right?
And I don't, you know, they put out a lot of stuff.
I guess he had said that, you're right, he said it would lower the cost by 50.
I don't know if that was the cost of inference or the cost of the chip itself relative to the current state of the art, which would probably be an NVIDIA GPU.
You have to be a little careful with those kind of comparisons as well, because you don't exactly know what they're talking about.
Like, people tend to cherry pick stuff.
And at the end of the day, it's not really the cost of the chip all by itself that necessarily matters.
It's the performance per watt, you know, performance per dollar total cost of ownership.
What the press release was talking about was sort of substantial improvements in performance per watt.
I think it said it was still an early test, early phase testing.
They're not in volume production with this yet, but it sounds like the early tests are very, very encouraging along those fronts.
And so, look, I'm sure it'll be a good chip.
And it sounds like they're getting ready to ship a lot of them.
Don't open AI, like I said, well over a gigawatt next year.
And like over the next five years, potentially up to 10 gigawatts.
I'm sure it's not going to be a good chip.
Okay, so I'm glad you stepped back and you said, okay, it comes as a surprise to no one that they have a chip.
They needed to.
If you're like a hyperscaler or an AI lab right now and you're not working on your own chip in some way, like you are doing something wrong here.
Fine.
They're all buying a lot of GPUs too.
So again, to go back to open AI.
So they have this 10 gigawatt deal with Braggam.
They also have a 10 gigawatt deal with Nvidia.
And they have a six gigawatt deal with AMD.
Everyone is just trying to get compute.
They're trying to get their costs down.
They're trying to be less dependent on Nvidia.
So let me tell you the thing that I thought was most interesting about this announcement.
And I don't know, like, that's why I'm curious if you thought it was interesting.
Nine months built in nine months from scratch.
Like that to me is just, that's like mind boggling.
These are the most complicated things to make in humanity, as you said.
So do we believe it?
So that is fast, but Braggam again has talked about exactly that.
And that's part of their differentiation on, on this custom chip.
There's a number of players that can do custom chips.
Braggam has literally said in the past, we can design a chip in nine months.
Like that, that's not something new that they've said.
This is the first time I think that they've actually shown a real example.
Okay.
And called it out that it was developed in nine months.
But again, they have said in the past that they could do that.
And so it's nice to actually see that they can.
It's still that, but okay, here's what I mean.
Does this change the whole chip space?
We've talked about, you know, the cycle getting shorter and Nvidia putting out new models now on a yearly basis.
If you can create it, I know you, I feel like I'm getting skepticism from you.
Is that because?
I think that, no, no, no, no, I'm not skeptical at all.
No, no, no, it looks right.
But I mean the chip space is changing anyways.
Right?
Okay.
So is this an aspect of that?
Absolutely.
Are we seeing a tremendous amount of new advanced silicon designs?
Because now we actually have a reason to do them and someone that's actually willing to pay for them.
Yes.
I'll give you an example.
You know, it's not even this, like we're seeing like chip startups now for a change for, for example.
Yeah.
I remember over 10 years ago, I had started to do a piece or to write a piece on venture capital investments in semiconductors.
I shelved it because there wasn't any, like all of the VC investments back then, it was a corporate VC.
That was it.
There wasn't a lot of, of, of, of actual, like, like a standard VC investment because it costs a lot of money and costs a lot of time.
And you know, the exits, uh, opportunities weren't certain.
And back then they would have just like invested in SAS.
It was a lot easier.
Yeah.
Um, we're actually starting to see a lot of VC investments in, in real meaningful startups that are getting valued both in the private as well as sometimes even in the public markets now, like in the billions or even tens of billions of dollars.
So there's an actual reason to do this now.
Like there, there, there's a, there's a use case and an application where like dollars are flowing in.
And so I think the space is, is absolutely changing.
And I, this is one aspect of it clearly.
Right.
Yeah.
And okay.
So as we've seen more competition, the story has always been like, what does it mean for Nvidia only game in town, but like the whole pie is expanding.
Right.
So that's what.
Yeah.
Tell me.
So as, as, as, as an equity analyst, we get paid to overcomplicate things sometimes.
And there's always.
Worries about competition.
And I would say, first of all, you need to step back.
This is semiconductors.
There was always competition.
There was always somebody like looking to eat your lunch if you will let them.
Right.
So it is imperative on the leaders in this space to make sure that they stay that way to, and to invest in the roadmap and to hire the best people and so on and so forth.
And they're all doing that.
I would also say the opportunity is getting so big that right now it almost doesn't matter.
My general view, for example, you know, you talk about Broadcom versus Nvidia, they talk about GPUs versus TPUs.
And my general view is that I get that, but it's, it's probably the wrong question.
Like the right question probably right now is more, is the opportunity in front of us still bigger?
Is it not?
Because if it's big, you know, everybody will thrive.
And if it's not, everybody's screwed.
Yeah.
And so far it's really big.
The assumption is that it's bigger, right?
The assumption is straight up and to the right.
Well, it's been, it's been straight up to the right so, so far, right?
Is there any reason to think it might be different?
Well, I mean, you're seeing it in G, it's not even just in this space, you're seeing it here.
And you're also seeing it in networking, you see copper versus optical and all kinds of
where there's all, there's always going to be alternatives to this market is, you know, hundreds
of billions or potentially even trillions of dollars.
Somebody is always going to have alternatives, but you can look at the numbers.
So Nvidia, I don't know, they'll do what $500 billion next year, right?
Broadcom got you to a hundred billion, right?
Nope.
And they'll probably do more than that, right?
And I don't know, Qualcomm had an analyst day yesterday and they said they're going to
do $5 billion from zero basically, but still billions.
And you got MediaTek and you got Marvell and AMD and all that.
They're all doing fine.
They're all growing, growing tons.
I mean, so in, in, in the, if you were to calculate the percentages, like our shares
moving up probably a little bit, but is it a really a problem at this point?
I don't think it is.
Right.
And you're saying, even though we have so many of these different players, this is still
a supply problem.
Like we see that especially in memory.
And so I wanted to ask you, I wanted to kind of connect this to the earlier conversation.
We were talking about Chinese open source, right?
Being a solution because it's more abundant.
It's a lot more efficient.
Yeah.
And it's almost at frontier level.
Does it, I feel like as an onlooker and you tell me if this is right or not, the same,
a similar thing is happening in hardware.
Like no China is nowhere close to the leading edge in terms of chips, but in terms of memory,
in terms of, you know, some of the other equipment that you need to build these data centers or
build electronics.
They're getting there.
What does that mean to you?
How is that shift?
How is that shifting?
Well, the Chinese are constrained in some sense, right?
So, you know, I did a piece while back.
It was called the U.S. has chips, but no power.
China has power, but no chips.
Like he's bringing more capacity online.
And the answer was the U.S. is bringing more online.
It's not even close.
The Chinese have been constrained by some of the export controls and the sanctions, particularly
on things like semiconductor manufacturing equipment that forces them to make their local
chips on effectively substandard process technology that impacts their yields and ability to really
produce like local chips at high volume, which is one reason that I think that they have been
forced to innovate along other vectors like like model efficiency and things like that.
They are constrained in terms of the resources that they can deploy.
And so they're forced to do as best as they can with those.
And they're very, very, they've been very, very good at it.
Right.
And this is one thing, you know, engineers are smart.
Right.
You know, if you give them constraints, they'll they'll find their way to make them fit.
And the Chinese are clearly doing that.
There are Chinese memory players.
But I mean, from what we're seeing in memory right now, I mean, you know, I don't cover Micron.
It's a colleague of mine, but I mean, Micron reported last night.
And I mean, it's it's looking like things are going to be tight for a long time, probably.
And it's both a supply issue as well as a demand issue.
And supply will come online over time.
And they have to actually build the buildings first before they have somewhere to put the tools.
Right.
To make the chip.
So it takes time.
But, you know, the question will be once that capacity comes online, like, does the demand rise rise to meet it?
We've actually seen this, by the way, like broadly in semis.
It's really interesting.
We've had this sort of rolling wave of bottlenecks.
Like, AI has gotten so big, it's kind of dragged everything along with it.
And one at a time, the different parts of the industry have sort of been hitting their limits and the stocks have been ripping.
And, you know, we went from the accelerators themselves being the bottleneck a year or two years or whatever.
And then it went to memory.
And then it went to semi-cap.
And then it went to optical and networking and power semis and more recently CPUs.
And you could have almost owned anything in the space.
I think that the SOX index is, you know, it's an index of semi-cap.
It's up 100% year to date.
It's been incredible.
You could have owned anything.
You would have been fine.
Like almost anything you would have been just fine to greater or lesser degree.
In the hardware space.
And it's all being ruined.
What's that?
Yes.
And it's basically you're owning the underlying things, but the app layer or the model layer, it's been a little bit more complicated.
Okay.
So like Stacey, you've been covering this almost two decades.
We're going to say two decades.
This stuff is cyclical, but right now it doesn't feel like cyclical.
And everyone's talking about a super cycle.
Is that still the case right now?
Do you see anything to throw that off?
Well, for now, yes.
Right.
I mean, look, I've been hearing the word super cycle as long as I've ever been doing this job, but this is probably the first real super cycle we've seen.
You know, there's a few different types of cycles, right?
You have inventory cycles, like semis are the back of the supply chain and fluctuations in any demand can propagate backwards.
And they tend to be shorter term typically.
You can get supply cycles.
Supply is tight and pricing goes up.
We're having a big one in memory right now.
You can have product cycles or socket cycles.
I was like, you know, I, I, I win or lose a chip that goes into an iPhone and that can drive my revenue up and down by big amounts.
And then you've got what we have today, which is, I mean, it's a true demand cycle.
And it's not that, you know, we didn't, you know, clearly we didn't have enough supply.
The reason is demand has gotten gotten so big.
It just overpowered and any of the wildest forecasts that were there, like not that long ago.
And if you want to define that as maybe, you know, indicative of what one might call a super cycle, maybe, maybe I would, because it is purely demand driven and it's driving everything.
And so the big question is how long does the demand last?
And I think that's the trillion, maybe it's the quadrillion dollar question.
Like, I don't know.
Right now it's still up and to the right.
And at some point, maybe that won't be the case, but I think all, all signs right now, the only thing we're hearing from anybody is they can't get enough compute.
Right. It's more just sort of the makeup of that, which we discussed before.
Is it going to be the labs?
Is it going to be the open source models, but infrastructure is sort of the system.
For my goals, I don't really care.
I mean, it's computed like Nvidia benefits clearly from both, you know, closed and open source.
Yeah.
Broadcom is doing A6, you know, for the vendors.
I mean, it'll depend on, on building for their models.
Right. But at the end of the day, I think compute is, is demand for compute is good.
It's, it's my guys in some sense, like they'll, by and large, they'll probably do fine as long as compute demand is going up.
Right.
Well, Stacey, it's always great to get your insights.
What's that?
They're selling the picks and the shovels and the gold rush.
So they'll, they'll be fine as long as the gold rush is going on.
Yeah. But have you seen that cartoon?
I feel like I've seen it a lot over the last few months, especially where, you know, someone's like, yeah.
What are you here to do?
No, one's actually mining the gold.
They're just bringing more picks and shovels for like a fraction.
They are, they are mining.
I think we can have genuine questions on, on return in our ROI, because I think that's really where the debate is.
Like what determines whether or not demand continues is, is there a return on the spending or is there not?
And it's still early, but I already think we're seeing evidence.
I mean, we've got you on the rental side.
Yeah.
They're sold out.
There's clearly a return.
I do wonder though, Stacey, how is it that the Chinese are able to do so much on a fraction of the CapEx?
Well, again, you know, it's, they're, they're being forced to be innovative, right?
Yeah.
And by the way, I do not view that as a bad thing.
Like this gets back to the whole deep seek scare from a year and a half ago.
You remember that?
Oh, I do.
I think we might be.
Everybody freaked out.
I know.
Okay.
Last thing I'll say.
Yes.
That was not a blip though.
I mean, we, we may be having another deep seek like moment.
What happened?
So people were worried.
It was like, Oh my God, these guys are so much more efficient.
We won't need as much computers.
We're building.
What happened?
Only thing we've seen since then is skyrocket.
We need costs to come down.
That's how you drive adoption.
That's how you get return.
And everybody throw this company, Devon's paradox, which it's been thrown around to death.
And the idea is when things get cheaper, people use more, but you have to remember I'm a semi
guy.
It's made out like that.
But yeah.
And look, I was, of course I believe in Jevon's paradox and semis cost got cut in half every
two years for six decades.
Was that a bad thing for semi?
No, it was a fantastic thing for semiconductors and for everybody else.
So I think lower, lower cost computers.
Good.
I started by calling you a semi guy.
Now you're calling yourself a semi guy.
I love it.
It was perfect.
You are our semi guy, Stacy, lots of energy, right?
Right place for the last few years.
Thank you so much for coming on the live stream.
We'll talk to you again soon.
I'm sure.
Thanks, Stacy.
Thank you guys for joining another live stream.
Thank you to Jasmine and Janice and Divya, Robert and Evan, and the behind here and Bud
in the control room.
We'll be back next week.
Thanks for watching guys and keep giving in those comments and questions.

Thank you.
Thank you.

Thank you.
Thank you.

Thank you.
Thank you.

Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.
Thank you.