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

0:13Welcome, everyone, to The Information's TITV. My name is Akash Pasricha. It is Thursday, June 25th. Today on the show, The Information has exclusive reporting that Google is reorganizing its recently launched Strike team working on AI coding models. We'll dig into those details shortly. We'll then get some analysis on Micron and Cerebrus' quarterly results, OpenAI's new jalapeno chip, and Qualcomm's acquisition, plus a look at the growing rift currently dividing the physical AI space. And Odyssey, a new AI lab focused on world models, has raised$310 million and a$1.45 billion valuation. We will look at what is driving those massive funding rounds.

0:58It's going to be a great show, so let's get right on into it. 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. 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. My colleague Aaron Wu, who covers OpenAI and Google, reported that story. And for more context, I want to bring on Jason Dean, our San Francisco Bureau Chief, to help us break it all down. Jason, welcome back to the show. Great to have you here. Thanks for being here.

1:30Okay, so let's just back up a bit. Google loses two of its star researchers. Noam Shazir goes to OpenAI. John Jumper goes to Anthropic. 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. What do we know now about these two departures? There's still quite a bit of mystery around these, but Aaron has learned a couple of interesting things in her reporting. First of all, Shazir told people before he left that Google had shifted away or really kind of 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.

2:19It's not exactly clear how that shaped his thinking, but obviously, you know, something very important. And he's been doing work. Shazir, of course, is one of the co-authors of the seminal AI paper on transformers that kind of gave birth to the current AI boom. He's been doing work on sort of new architectures for AI beyond transformers. And, you know, so the access to compute, obviously, is a big issue inside Google and something that he flagged to people before he left. The other thing is that Google's CTO, in announcing his departure internally, said that it was very sudden. So, you know, clearly there was some abrupt rupture here.

3:05The exact linkage isn't clear, but this was something that happened suddenly and happened after this compute issue. Right. Okay. So Google is left to pick up the pieces, I guess, of these two departures. And I should say that they lost one Nobel Prize winner, but they still – I mean, Demis is also a Nobel Prize winner, right? Yeah. I'm down 50 % on Nobel. Yeah, in some sense. I mean, can't stack the whole team in some cases. But okay, so Google is left to pick up the pieces. So what did we learn then about how they are reorganizing themselves to address this moment? Yeah. 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.

4:03And 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. And 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 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 OpenEye, which are kind of further ahead in this area of enterprise AI than Google is, you know, both cloud code. And Codex products are being expanded into other areas adjacent to coding, trying to develop the same success that Anthropics had with coding in other areas.

4:57So Google's change in the way 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 products where Google's been a laggard. Now, one of the things that Aaron focused on in the story is this concept of mid-training. What is mid-training and why is that so important right now? Yeah, I mean, this stuff can kind of make your head spin a little bit, all the jargon. But, you know, we tend to group the creation of these models into pre-training and post-training. Pre-training, we're sort of building the basic intelligence.

5:37You're building the model. It's learning and post-training. We're kind of adapting it to specific functions. This is a disaggregation of that. 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 how they interact with users, the model's behavior. And again, the goal seems to be to try to enable the models to be better at specific functions like coding, etc., before you get into the behavior part, just the underlying intelligence to be better before you get into the behavior aspect of it.

6:16Okay. So let's go back to the Strike team that Google has established. You know, Aaron had reported that this team, it came together a couple months ago. They've now changed the organization of it a bit. Do we know how big the team is, who's in it, how much power they have internally in the organization? Is this kind of like the equivalent to, I don't know, I'm thinking of Meta. Meta has the super intelligence group. Is this just a coding-specific group, or is this the group? I think their remit is expanding now. 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 uh veteran engineer who's been there for eight years but has also had the involvement of the deep mind cto as well as google co-founder sergey brin so this has been something that's been uh very important for the company and i think though we haven't really seen uh the tangible results yet um in terms of them closing the gap in a meaningful way with their competitors and things like coding tools.

7:30The fact that they're sort of solidifying this team and making it more permanent and expanding its areas of focus suggests that they think that it's working. Yeah. Why do you think that we haven't seen results in coding? I mean, Google has excelled at seemingly every other part of AI, including chips now. Coding seems like, I don't know, your software, like, how did you miss the boat on coding? Why do we think they haven't been able to correct that? Well, I think, you know, part of it is they were late to it, and kind of everybody was late to it except for Anthropic, right? Anthropic figured this out before others, whether that was through some great visionary foresight or sort of they stumbled into it uh the you know there there could be a debate but they were first to it um that's been the killer app uh in the enterprise in terms of generating revenue um open ai is trying to be a fast follow there google is trying to be a uh a fast follow also perhaps not not as fast i think also um what aaron's reporting shows is that there was an initial faith that if you build the underlying model, the coding capabilities will come.

8:46And it's clearly not that simple. You need to do more than just make a really smart model that can do everything. You really need to focus on enabling those capabilities. And that's part of what they're trying to do now. Right. Last question for you, Jason. The latest Gemini model, we are expecting it. I think Aaron had reported we were expecting it in june it hasn't been released yet do we have any updates on that uh yeah i mean you know there's still a few days left in june which is what they said this is this is a gemini 3.5 pro they they came out with flash at google io last month and said that the pro version would be the sort of flagship model um would be available in june not here yet um that's looking much more like it's going to be uh july i mean the exact timing you know if it slips into into july perhaps not the biggest dealing the bigger question is will it you know what will it be and will it be yeah much better yeah right and i mean early in the i guess uh based on aaron's reporting it seems like that still is not going to rise to the level of of fable yet right right which you know maybe that's a good thing maybe that means it won't be sort of uh banned or restricted by the administration.

10:05Yeah, I mean, Anthropic is still quite loud about it. You go on Claude, I mean, they still have the bar there. You know, as long as they have that rule in place, they are very loud about the fact that, look, it's not our fault. They're getting marketing mileage out of it, if nothing else. Yeah, yeah. All right. Well, Jason, I want to thank you for coming on. That is Jason Dean, our San Francisco Bureau Chief, here at The Information. Shares of Micron are jumping after the chipmaker more than quadrupled its revenue in the latest quarter. The results offer some relief for the semiconductor sector after Cerebris flagged shrinking margins earlier this week in its first report since its IPO.

10:44I want to bring on Matt Bryson, Managing Director of Equity Research at Wedbush, and Mustafa Nimuchwala, partner at New Enterprise Associates, to unpack both of those results. Welcome to the both of you. It's great to have you here. Matt, I want to start with you. What did you make of Micron's very strong report last night. I think the most interesting piece beside the beaten race again was the granularity they were able to give around these longer term contracts that they've set up where you're looking at it feels like roughly half the annual value of their output to these customers as being guaranteed or cash payment that is forfeit if the customer breaks the contract.

11:34And that sets up a level of certainty for Micron and I think the industry that we haven't seen before. I think that largely is what the stock's reacting to. The idea that instead of trying to pick the end of the cycle, now we've got a longer term guaranteed cash flow, if you will. and how do you value that within this industry that used to be completely cyclical? 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? Oh, I think it's certainly how tight memory has become. They just have more leverage than they ever had before.

12:15Right. Mustafa, what stood out to you? 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? I think the cyclicality that we have 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. 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.

12:50And so far, they're sold out through 27. margins are double what they were at prior peaks. And so even if this locality 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? And if you look at, you know, Micron is probably relatively more expensive, but you look at Hynix, you look at Keoxia in Japan, a lot of people's math will tell them by the end of 28, you get the entire enterprise value back in just the cash flow these companies generate. 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.

13:39That is lower than the GPU companies. And so my question is do you see those multiples uh converging at some point i i mean that's that's exactly what investors are talking about right if i'm willing to pay 30 40 50 times uh 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 now do i ever think microns you got to get a 50x multiple i don't 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 you end up with with oversupply at some point and so you lose leverage um but at the same time i i think it's it's a question worth asking again why wouldn't you be willing to pay 11 times 12 times for for micron if you're willing to pay you know 40 times for AMD so Mustafa Matt is saying that that the memory multiples are likely to come up a bit uh the other way of thinking about it is maybe maybe the GP multiples come down uh what do you think is more likely here I would say I think most people tell you right now that Nvidia is cheap and Nvidia has been cheap you know we're talking about depending on your numbers, 10, 12, 15 times gap PE.

15:02So it's tough to see necessarily the NVIDIA multiple coming down, unless of course the metric comes down. I think the way I would pose this question is right now we're talking about 40 to 50 % of CapEx, of hyperscaler CapEx going to memory. And then the question we have to ask ourselves is what do we think the art of the possible and that can be? And then I think you obviously talk a lot about the cyclicality, about the cash certainty about the margin profile, and then you do the math backwards. Another way of thinking about it also is from a market cap perspective. If we were to get Micron and Hynyx and Samsung and Keoxia and 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 implied market cap are we talking about?

15:50I mean, you pretty quickly get to a number that's much larger than all of their customers combined. And I think that's where you get into this whole question of where does value accrue? And obviously, this is the bottleneck today. But in the arc of time, do we think the memory companies are going to be larger than all the hyperscalers? Well, I think that would lead to an interesting question of how long that constraint lasts. And as Matt mentioned, as we all talk about, fab capacity will come online. The question is when, and the question is what does supply-demand imbalance at that point in time look like?

16:17Right. Matt, I want to go to Cerebris. So they pointed to margins that they said were misunderstood. That's what Andrew Feldman said on CNBC. What was your read on the margins, which were narrower for the full year in the forecast? Yeah, Cerebris, I think during the IPO proceedings, was pretty open about the fact that as OpenAI comes on, it's a drag on gross margins. And just I think there's two things going on there. One, very large customers, so they're going to get somewhat favorable terms. But then two, because it takes time to bring data center capacity up for them to satisfy the first launch of that agreement, they have to go and rent back capacity that they've sold to another customer.

17:10And that's just more expensive than servicing open AI out of their own capacity. So I think that's what you're seeing. I honestly don't think that's necessarily what surprised investors. And if anything, the actual gross margin deterioration isn't quite as bad as I think people have modeled. Rather, I think what is really weighing on the stock is it has this strange lockup configuration where shares get unlocked after that first earnings period. um so you know the way i look at it there's just there's just more stock available um and that's probably the the greater reason the stock's down i mean honestly what happens this year doesn't matter that much um for cerebrus you own cerebrus 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 and if that that scenario plays out the company is a whole lot more valuable than it is right now right mustafa i'm curious you know you invest in early stage companies and chip companies are certainly part of the pool of companies that you consider 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 out.

18:40It's not really a question of can you take market share away because the market is growing so fast. So I guess my question is, how do you even decipher these? How do you distinguish between these chip companies when does it really matter which one is better if the market is growing so fast? I mean, let's be super real. I think the biggest question for any customer is when can you get me those chips? 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, et cetera. But that, of course, is impacted by architecture.

19:12And 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. You don't have to rely on COOS. You might not have to even rely on the leading edge KSMC node. Those three things, specifically leading edge node, COOS, and HBM are constraints or bottlenecks. So the less bottlenecks you have, even if that reduces performance, the faster you can get your chips out. And the flip side, everyone's obviously ramping capacity and thinking about architectures and doing a lot of the constraint optimization around the same time. And obviously we'll talk about jalapeno in a second, but if we're talking about 27, 28, we have a lot of chip capacity coming online.

19:47That's where I think it becomes more of a question of what kind of what you said, which is where do you slot in? And I think all of us have been talking about pre-fill decode, the disaggregation of inference. 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, where HBM can matter. Now, obviously, it's interesting because if you take out the bottlenecks, people would probably, a lot of people would still like to have all of those nice to have, advanced packaging, COOS, of course, HBM and fast memory, because you can handle a lot more scale in these model sizes as well.

20:22So I think really it's about timing and then performance per watt, of course in the u.s and externally performance per watt matters less energy is less of a constraint that's another dimension as well it's really all constrained optimization based paired with what you said around the kind of market itself matt what do you think of this question um i think that the chip quality still matters um and the preface is i i'm a history major so i i I don't design chips. I don't choose chips. But I look at Microsoft. I look at Meta. They've been working on custom ASICs for two, three years now. And you're not seeing a lot of volume.

21:02So just having a chip was good enough. I think you would have seen both Microsoft and Meta ramp their efforts. And we haven't. Um, with Cerebris, the, the, the thing that stands out to me is, look, they, they, they want to deal with open AI. They want to deal with Amazon. You don't see the same news for 10 store. And you didn't see the same news for Grok. Um, haven't heard anything of similar from Sabanova. These are all companies that are out there with, with solutions. And so I think that, that effectively tells you that Cerebris has a, has a good chip. And I think that the quality of the solution certainly matters at this point.

21:45Now, having said that, I think you look at NVIDIA, one of the reasons I think that Wall Street is underestimating NVIDIA is not only do they have a good solution set, but they also have locked up so much of the supply chain that you just can't get competition because no one can get HBM, Substrate, Coase, go down the chain. So, okay, so now let's talk about the quality then of Jalapeno, Mustafa. I mean, if that matters so much, what are early reactions that you're hearing to how good Jalapeno might be? This is, of course, the new chip that OpenAI and Broadcom unveiled. Yeah, I would say this is a classic example of buy the rumor, sell the news.

22:28I think while it's interesting 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 and the tooling and the models themselves, OpenAI said, to no surprise, that they use the OpenAI models to help design the chip. And of course, there's a lot of companies in the private markets, including recursive that are helping people from the Google DeepMind alpha chip design solution.

23:00How do you design chips with AI? How do you make that velocity faster? So I think for me, the biggest real takeaway was around the velocity side of things. And then, of course, the timeline that they talked about as well. So we're talking about a chip that, like I mentioned earlier, has HBM in it, has eight stacks of HBM. And I'm sure over time we'll have more and more. So they're not avoiding the HBM constraint. Given they're working with Broadcom, they'll probably have leading edge TSMC capacity and they're doing co-loss and advanced packaging. So this is not a bottleneck optimized chip. This is a performance optimized chip.

23:26This is a chip that's tailored to the workloads that they have. And I think a lot of what we expected, 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 filmy conductor manufacturing process. Right, right. And Matt, last question for you. You cover Qualcomm. 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. It was a big deal. I think it was reported to be$4 billion, thereabouts.

23:59It was an all-stock transaction. 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. Why do you think Qualcomm wants to get into this business, and do you think they have a chance here? um so just to preface this i i don't cover qualcomm okay it falls it falls in my universe because it's it's it's hardware semiconductor i i read it in your note this morning so i thought i thought maybe it was in your coverage um no but i i mean i i do i talk about everything semiconductor and hardware so what what i would say is it almost seems to me that that that software burden um or or mode that could have created um has dissipated a bit at least in terms of inference um right to some extent it feels like a a token is a token um 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 now just to speculate a bit though it does seem like software is very important in targeting some of these world model type applications, if you will.

25:17So the automotive market or physical AI. And maybe that's what Qualcomm is looking at. I mean, certainly Qualcomm has put together an impressive set of assets in terms of all the acquisitions that they've made. And I see modular as part of that. that um what i'm still waiting for though is for qualcomm to uh generate uh substantial revenues from those acquisitions um right it seems like today uh qualcomm's efforts outside of handsets have been a little bit uh less less successful uh than than they've been in the handset side great well uh matt and mustafa want to thank you for coming on that is matt bryson from wedbush and Mustafa Nimuchwala from NEA here on TITV.

26:05Speaking of world models, there is an intensifying debate between world models and vision language action models, or VLA's. That debate is creating a sharp divide among tech leaders and venture capitalists over the future of robotics and physical AI. Our AI and robotics reporter, Rocket Drew, wrote about that exact topic in this morning's AI Agenda newsletter. I want to invite Rocket on to talk about it and help us understand what these terms are all about. Rocket, welcome to the show. It's great to have you back. Hi, Akash. Great to be here. Okay. World models and vision language action models, right?

26:43Do I have that right? VLAs? What are each of those? Define them. Yeah, definitely. So a VLA is a little bit more straightforward. Let's start there. You start with a language model, like an actual language model, like the same kind that powers your favorite coding agent or a chatbot. And then you customize it. You train it on further data so that it's able to control a robot. And the hope, the idea here is that you start from a base that's already kind of intelligent. It knows how to reason. It 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.

27:23That's kind of the vision there, okay? World models are kind of a separate approach. When we talk about a world model, really we just mean an AI model that has a really good understanding of physics. I think sometimes the lines get a little blurred here. People could say, well, ideally we'd have a VLA that has a really good understanding of physics. But in practice, people mean something kind of different when they talk about world models. They're often talking about something that was trained on a lot of video, meaning it saw a lot of videos, say like 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.

28:06So why would world models have anything to do with robotics, you might ask? Well, there's sort of two different approaches. One is that if you had this model that was really good at understanding how the physical environment is going to evolve, maybe you could just use that as the brain of the robot. You just put it in the robot's brain, you have it send, you know, predict what the robot should do next, which actions it shouldn't take, how hard does it need to pick up different objects around it to move them around, that sort of thing. But short of that, even if it can't be the full brain of the robot, maybe it's too big or slow or it only works in certain environments and not others, there's a hope that the world model could still create simulated environments.

28:47The way right now, roboticists use computer simulations to train their robots, but you could have even better simulations because you're using AI to create them. So that's kind of the hope for world models right now. Does that get at the difference? Does that make sense? I think so, yeah. So, I mean, are these approaches, you talk in your newsletter about there being a debate as to which one is, I guess, better for physical AI and for robotics. Why is it that they are so, I mean, it seems like they're trying to get the same issue. So why is it that this Rift even exists? I think the Rift exists because neither of them work very well yet.

29:26I mean, everyone is hoping that robotics eventually is able to scale the way like language models have been able to scale. And we can just get bigger and bigger models that are more and more capable. But robotics doesn't have access to the same kind of data that language models use, right? There's no internet worth of training data, just waiting there for robotics to take and to train on. So roboticists are like painstakingly collecting more physical data to train their models with. And so far, there just hasn't been enough data for these models to work reliably. Also, they have to work really reliably, right?

30:00Because like if If your chatbot wasn't trained on enough data and it makes a mistake or it hallucinates, I don't know. What's the worst that happens? You hallucinate a case as a lawyer and then you get in trouble with the judge. But as a robot, if you hallucinate something that isn't there, you trip, you fall over, you break, your parts break, maybe you hurt someone nearby, and the stakes are just much higher. So roboticists are still trying to figure out what is the right way forward. If we're going to take advantage of AI for robotics, what is the right model and what are sort of the right mixes of data that we need in order to make that possible?

30:36So am I, just to go back to the definitions here, because I think it's worth actually drilling in on this. Am I right that vision, language, action models, VLAs are more about collecting data from the outside world and using that data to then train the models that robots would use? whereas world models is a little bit more about using simulations as a way of sort of predicting the physics of the real world? Is that the idea? In both cases, you're going to want some data from the real world to make sure that the outputs of the model are as realistic and grounded as possible. The difference really comes down to in the VLA case, you're starting from a language model and you're hoping that your robotic model, your VLA, sort of inherits some of the intelligence and language understanding from that language model that you started with.