Anthropic, Meta and Google Release New Models, Snowflake Shares Surge, Google Avoids Break-Up

0:13Welcome, everyone, to The Information's TI TV. My name is Akash Pasricha. It is Thursday, September 3rd. Today on the show, we are taking a look at this week's new models from Anthropic, Meta, and Gemini. There is a lot to get to. We'll break down how they all stack up against each other. We'll then get into Broadcom's booming revenue forecast and understand how power could play a role in chip companies achieving the results that they forecasted. We'll also discuss the run that Snowflake is on after its big results last night and uplifting revenue forecast. We've also got PwC coming on to talk about AI M &A.

0:51And to close out the show, we'll bring on our reporter who covered the Google AdTech antitrust case last year for her reactions to the big ruling yesterday that the company won't have to break up parts of its AdTech business. It's going to be a great show, so let's get right on into it. It has been a crazy week for model releases, to say the least. Anthropic has new versions of Mythos and Fable out. Meta has a new version of Muse out. Google has released Gemini 3.8. And for a breakdown on how all of this stacks up against each other, I want to bring on Peter Gustave, AI capability lead at Arena, to help us talk through all of this.

1:32Peter, welcome to the show. It's great to have you here. Kass, great to join your show. So it's a busy, busy week for you guys. I mean, Arena is in the business of ranking all of these models. And I think it's rare that we get big releases from three of the major labs OpenAI seems to be the only company that took the week off, I guess you could say, although it had a very busy week itself from the news cycle. Which model release most surprised you? We're going to go through the one by one, but if you look at them this week, your reactions to it, what surprised you? I think Meta is the one that really stood out.

2:11And it's not necessarily because it's just an excellent model, which it seems to be getting traction and seems to be really good. But what's really interesting there is the pace of delivery. So if you remember back into Lama days, there seemed to be big gaps between model releases. And the models were, you know, debatable. Some were good, some were not so good. But when you have gaps of eight months between releases, then you just can't get enough pace to get more traction. But here we've got meta releasing models every, what, six weeks? I haven't counted, but it feels like every few weeks. It's a really impressive pace.

2:50And what we see in the data very clearly is the labs that keep releasing new models, they're the ones who are winning, not necessarily the ones who got one big model and then they wait for 10 months before the next release. Why do you think that is? Because I think there seems to be some general capability you build up, especially once we moved into this RL world where you generate more data, you create more environments, you train new models on this. And it seems to be less about one specific model, but more about this kind of process that you build up and how you keep releasing new, new, better and better models.

3:28And especially as you get more models, you've got more maybe users, you get more data, you get more of a feel how people are using it, where it falls down. And that's how you really start improving new models. So let's go back to the Muse model for a minute. 1.3 here. I mean, we know that Meta had made a big play a couple weeks ago on the coding front as well. As you look at Muse right now and where it excels against the other models, what is 1.3 great at compared to other models? Where does it fall short? Give me the overall review here. Yeah. So the interesting thing about meta models is that they really focus on your coding so the when we look at their coding performance it's been uh like the raw coding capability has been above where the kind of general agent capability has been in our data so that's been interesting to see and freezing from your reporting and more generally it does look like they're really going all out on that front.

4:35When we look at our agent leaderboard, we don't have a score yet, I think, for 1.3, but certainly looking at some of the previous models, it does look like they were a little bit behind. So in more general agenda capabilities, we wouldn't just measure the coding, which is a big part of it, but we might also look at how you do research, how you present things, how maybe you speak to the users. So there are a bunch more things. So, but in terms of the kind of raw performance versus, for example, the front-end coding capabilities has been really impressive. So it looks like they're focused on things where they could maybe get ahead a little bit or get close to the frontier and maybe pulling up the rest.

5:15That would be kind of my guess about how they're approaching it. Right. Okay. So let's move on to Gemini, which I think is also in an equally interesting moment right now they released 3.8 uh all right 3.8 flash i i guess it is uh what was your review of that model and where do you see gemini in in the model race right now broadly yeah so gemini it was kind of interesting and slightly confusing to see gemini releases because if you remember they were going with the bigger models uh a few months ago where it was Maybe 3.1 Pro was the big release, and we're kind of waiting for the next one. It never quite came.

5:58And then Flash models, previously they were super cheap models, and now they became closer to the Pro models. And it looks like they're kind of slightly changing the approach to their whole model family. So we see pretty much Flash models while being cheaper maybe than other models. They're certainly cheaper than like a Fable-level model. They're certainly not as cheap as Luma model. and in terms of the performance what we're seeing on this kind of coding tasks specifically and maybe kind of data processing tasks or vision tasks it does seem to be kind of middle of the park-ish or maybe sitting on that parade to frontier which is really important but it's certainly not that very top frontier kind of model so we are not seeing it competing yet with Fable-level model, even quite with Opus or with GPT-5-6-Soul or that kind of category.

6:54But it's really very competitive versus something like a Terra or a Sornet, which is, you know, it is definitely a category where you also want to play, but it will be really interesting to see what's the big Max-Gemini model where we could really see compete. Do you have confidence that Gemini can reinsert itself into the conversation in a big way? And by that, I mean, it's very much in the conversation right now, just not for the right reasons. People seem to have lost some confidence in it. We also know that Google is shaking up their cloud business quite a bit and their deep mind. I shouldn't say cloud, more so the DeepMind business.

7:41But I mean, how much confidence do you have in Google coming back in a big way? Yeah, I actually have a lot of confidence. And I think it kind of comes back to the previous point about that you need to build up a process. And we see, if you note the last few months, I think really should change our minds a little bit. And what happened if you go back maybe, say six to eight months, It was pretty much anthropic open AI, right? Those kind of other models were kind of, okay, interesting. You can use them, certainly, but it's really about those two. And now what we see, we see Groke, we see Matter catching up.

8:17We see Kimi K3, we see QN 3.8, all amazing results. So what this tells me is that there's not any crazy fundamental reason why one lab is so much better than the other. There are obviously advantages like compute and so on. But what this tells me is that Google is serious about this. I don't think they've given up on the frontier ambitions. You can't count them out necessarily. No, no. So I think the stats... I agree. I mean, look, I agree. I mean, these models, frankly, they come out of left field sometimes, and you just never know what companies are working on. So I agree with you. I want to get to Anthropics models.

9:01So Mythos 5.1 and Fable 5.1. And first of all, explain to us and make clear to us what the differences are between these two models and what your reaction was to them. Yeah, so it is a little bit confusing with entropic models, but they've been very clear that Mythos and Fable is fundamentally the same model, but Mythos has fewer restrictions. So they only use it with select partners, which I guess makes sense. In terms of the performance, we've seen it gain by far the highest score on our web dev arena, so where we test front-end coding skills. And they made this kind of leap before with Fable 5, they were ahead.

9:44And then we saw models like QN 3.8 and Kimi K3 all kind of gaining and almost matching or even exceeding a little bit Fable 5. And now Fable 5.1 made another leap. And that really tells you something fundamentally about the model quality, but also the taste and aesthetics and just one short ability to just produce great results. How about price? How do they rank on price? You know, I was testing the model yesterday, and I read what seems like great news where they reduced the cash cost of reading the cash. And I thought, wow, that's going to make a lot of difference. And then I did a bunch of generations, which are kind of complicated prompts where it's running maybe for a few hours to complete them.

10:38And I see the prices of$40,$60,$50 per generation. This is when they compare it to 5-6 solar to something like$6. Or sometimes Kimi K3 could be$3. You know, things like that. So the thing about this is that it's kind of hard to judge, right? If your task is, I want to have the best possible output of the thing that I'm working on, it could well be worth you paying so much more. For example, they were talking about security scans. If you're doing in-depth security scans, maybe it's worth you paying, I don't know, 100K per scan. Because, okay, maybe you save 90K, but then you leave a bunch of bugs in or a bunch of security holes.

11:22So is it worth it or not? So it's a kind of time when you really start having to think hard about what kind of model you want to use. Because clearly just using Fable for everything, maybe you're rich enough to burn. not sustainable right let me ask you one last question before you go so you know open weight models have been getting a lot of traction what do this week's model releases tell us about the gap between frontier models and open weight models i mean we talked about we've talked previously about that gap getting smaller um i i wonder if this week does it is the gap getting wider again does Does it tell us anything?

12:05What do you think? I would say the gap has widened with Fable 5.1. This week, yeah, that's the caveat. This week it's wider. This week it's a little bit wider. But looking at Fable 5.1, there's nothing fundamental. I don't think there's any reason to say that, okay, the next one, the next Kimmy, if they carry on on the trajectory that they were, there's no reason why they couldn't get closer in that direction on this kind of fundamental level. So it's widened, but probably it will get less wide. Come back next Thursday. Exactly. We'll recap it again. All right. Well, Peter, I want to thank you for coming on.

12:46It was a great conversation. That is Peter Gustav, AI Capability Lead at Arena here on TI TV. Broadcom made some big forecasts last night in its quarterly results. The company said AI-related semiconductor revenue is expected to double next year, and there is a, quote, line of sight, end quote, that could double it in 2028 as well. Even still, shares slumped this morning. I want to bring on Tristan Guerra, senior analyst covering semiconductors at Baird, for his thoughts on all this. Tristan, welcome back to the show. It's great to have you here. Yeah, good seeing you. Good ball, Amy. What stood out to you last night from Broadcom's results, Tristan?

13:28Yeah, so clearly these days we're talking about hundreds of billions and no longer millions in terms of AI forecasts. So if you look at the near term, what they reported in terms of AI revenue for both the quarter and what they guided, there was pretty much almost no upside. So pretty much in line with expectation. And we've seen during this fast-running season that even if you beat meaning-free, people are eager to take profit. So in this case, it was really not much new in terms of near-term momentum. And we call that they were slightly short of expectation a quarter ago where they died it because they decided not to do the wax for anthropics so that we moved some revenue.

14:22However, to your point, they provided a line of sight that was obviously meaningful with some more precision on the next year revenue outlook for AI and then extending into 28. And by 28, you know, where the forecast is inferring another double in revenue. So two consecutive years of double, which highlights that clearly they, you know, have a lot of orders in the pipeline and should ultimately kind of relieve concerns that, you know, about visibility. I think the near-term concern is some share lost at their largest customer, which is Google, giving some business to Mediatek. And we've learned more recently and something that Broadcom confirmed yesterday night that Mediatek is going to get the V8 training version and Broadcom is going to get the inferencing version, which infers that...

15:28From Google. From Google. Yeah. And so that infers that the inferencing chip is actually featuring a higher level of complexity. That's what Broadcom is getting because they're the leader in terms of capabilities, whether it's 30s and chip-to-chip interconnects, so they still get the higher value design. So it's hard to see the stuff working near term when you have such a displacement of market share. And ultimately, it shows that companies like Google and others, they own the IP so they can choose where they're going to. But at the same time, it doesn't remove the value add that Broadcom is adding and the technology, which is well ahead of some of their custom ASIC competitors.

16:18Our view is that Broadcom is actually on the right side of the equation by getting the inference and shape. Why? Because it's fair to make an assumption that by 28, In 2008, inferencing will continue to gain share versus training. And as such, pass that initial share loss to Mediatek, which by the way, also would depend on Mediatek's ability to execute. In 2008, we're hearing huge volumes of Google ships. I mean, we're looking at a 9 million units of TPUs next year going to 15, 16 in 2008. So that's a huge growth again total for Google. And we don't see how inferencing could not be more of the mix than it is in 27, which means that mathematically, Broca'm is well positioned to actually regain some share and see some rebound year over year.

17:19So without ever being concerned, I think there's probably a little bit of the air pocket near term in the stock because of the companies adjusting the supply chain in terms of different suppliers, but ultimately what can remain well entrenched as a leader in customer usage. Let me ask you a couple of questions. You got into a lot there. So the shares are dropping this morning. And so I wonder, I mean, on one hand, you could say that analysts were expecting more growth, more than double, I guess, and maybe this is a tempered expectation. I mean, that's one explanation. But did I hear you suggest that maybe that could be a reaction to the Google going with MediaTek?

18:08Is that what you were suggesting as well? Yeah, I think it's nothing new, actually. I think where there's been increased visibility over the past few months is MediaTek doubling their expected revenue coming from Google. by end of this year, they went from$1 to$2 billion. That's V8 revenue from Google that they expect as they went late this year. And I've also quantified$14 to$16 billion of revenue. But you also compare that$14 billion and$16 billion in expected revenue that MediaTek, assuming good execution again, would get from Google next year. you compare that with the type of numbers that Broadcom has provided in terms of their total AI guidance.

19:00And then, you know, easy to tell that Broadcom remains the undisputed leader in custom ASIC. And Broadcom also mentioned that they expect Entropic and OpenAI to become their two largest customers in terms of the XPU. Yeah. So I think, yeah, you know, some of the diversification of custom ASIC procurement that Google has initiated, you know, is and has been already putting pressure on Broadcom stock. I think the upside is, you know, the one that we're going to see at the Frontier AI nabs, Mediatek potentially shipping a little bit less. It depends on execution, but also... What did you think of the NVIDIA MediaTek investment?

19:50Was that... Is that something that you... That endorsement from NVIDIA, is that a factor that you think might be putting pressure on the stock at all? I don't know that it's putting pressure on... Not today. Yeah, not today. I just... I mean, this was announced a couple of days ago. I just mean, what does that tell us about the competition between Broadcom and MediaTek? Very, very fierce and specifically, you know, on one end, you have Broadcom encouraging the hyperscars and Frontier AI Labs to use XPUs, you know, custom ASICs instead of GPUs. And then you have NVIDIA, you know, with a ongoing partnership with MediaTek because it's not new.

20:36It's something that started to be announced at GTC in 25. that is basically helping MediaTek with a range of IP licensing and that ranges from custom ASIC, but also what MediaTek is doing in agentic AI-based PCs where clearly there is a ramp of local agentic AI nodes that you're going to see emerging in smartphone in 28 and in PC, and that's something where... Let me ask you maybe a broader question here about these results. So, I mean, Broadcom is booming. Nvidia, they're also saying, well, we can expect 70 % growth next year as well. So the chip businesses are all doing very well. At the same time, you have these power constraints, which came up on the Broadcom call last night.

21:39And the question that we've been thinking about here in our newsroom is the extent to which these forecasts from the chip companies can remain intact given that there's not enough power to supply to these chips and get these chips connected. What's your view on that? I mean, is that going to cause any kind of a reckoning here for these chip sales? And I'm not suggesting they're going to fall off a cliff. I'm just saying, with the power constraints, can they really achieve their utmost ambitions here? Yeah, so definitely very topical. And there is some elements of uncertainty. As you've mentioned, both NVIDIA and Broadcom have said that there is meaningful upside to the forecast that they're providing.

22:30based on availability, so electricity, as you've mentioned, but it's also all types of shortages, right? Starting with memory, whether it's HVM, whether it's DOM, and also substrate, which, by the way, on the memory side, you're going to see some relief really starting mid-20s. There is really a ramping capacity, which is supportive of some re-acceleration. So I think, you know, when you hear some of those companies guiding, they basically claim they took into account, you know, the world out of, you know, new grid connections to data centers. But I think it's very difficult to assess. Are you applying any other adjustments to your models to factor in power constraints?

23:19Do you think that their considerations are enough on the constraints for power? Yeah, good question. So as much as we've been above the street in terms of our AI estimate for 27, we're actually slightly below Broadcom's forecast for 28 to account to some of those risks, even though Broadcom has already accounted for some of those risks. it brings a theme that you I know you didn't ask but basically which is really how to address those power shortages before you actually get more power delivery directly from the grid and that's really the migration to 400 and ultimately 800 volt native architectures which is going to be really a big investment theme for a number of analog power companies to basically significantly increase the efficiency of data center before you actually get, you know, more power delivery directly delivered to those new data center.

24:19Right. Great. Well, Tristan, I want to thank you for coming on. That is Tristan Guerra, a senior analyst at Baird here on TI-TV. Snowflake also reported roaring results last night. The company accelerated its top-line growth for the fourth consecutive quarter. However, it is also now expecting to grow product revenue 36 % up from the previous 31 % expectation it last gave for the year. I want to bring on Mike Paulus, founder of PCM Encore, for his thoughts not just on the Snowflake results, but also on the software market at large and what it means for late-stage private companies right now. Mike, welcome to the show.

25:00It's great to have you here. It's great to be on. Thanks for having me. so uh the snowflake results i mean i don't know how closely you follow the the the company i trust you you watch everything uh given given your role and your uh pedigree here what did you think of the results last night i think you can't help but uh but love them and it was a full circle moment for me many years ago as a uh as as a young deal partner at andreessen horowitz i i passed on a promising snowflake. It's a bit of deja vu harkening back, but it was a remarkable set of growth to have an accelerating growth rate. What was really interesting, I think to see Coco, their agentic tool, have remarkable adoption, really maybe changes the narrative.

25:46And I think from an anthropic sort of having won the market to potentially a world where a lot of different companies that have different strengths and areas of leverage will have coding agents. I thought up there, a time to market was really interesting. You know, as an old enterprise guy from my Adapar days, going from three-year deployments to eight-month deployments, the time they saw it at 80 % usage of their software, a really exciting kind of newer AI story. And I think it changes the narrative along with maybe what Salesforce did, you know, just a couple of weeks ago with Anthropic, where with the right strategy, some of these companies that we thought might be legacy, might be exposed to AI, can actually pivot and be both significant providers and have it really help their core businesses.

26:39So, I mean, what's your view here then on the SaaSpocalypse here, broadly speaking? I mean, Snowflake, I was sort of trying to figure out why its results seem so impenetrable. I mean, this is the fourth consecutive year of revenue growth acceleration. You know, on one hand, and we've talked on the show about the backend software in AI, anything to do with data, I mean, they're getting a big boost right now. It's the application layer software that we thought would be most under threat. Does that explanation still hold, or did the Salesforce results basically make that explanation moot as well?

27:22I think the size of the AI boom and the spending behind it just cannot be understated. And I think we're constantly surprised. Whenever we think we size this market, it's sort of like out of zero. So I think, first of all, it just turns out that Snowflake can grow, and the HyperStalers can grow, and Databricks can grow, and all the AI folks on top. So I think right now, there's a story of just grabbing market share in a rapidly growing market. but I do think with the SaaS apocalypse what you've seen is if you have the verified data layer it's critically important and in a way what you have is you both have the ability to build on top of that with your own agents but you also might go from a human querying a stove database 20 or 40 times to an AI agent querying it a thousand times so I think it turns out that having really accurate source of the truth data is quite valuable.

28:21And it's a fundamentally different problem than having a frontier model. And I think what we've seen, I'd say, with Salesforce and now with Snowflake is that's actually becoming quite valuable. And then there is leverage. You mentioned Databricks. Actually, before you move on, I do want to ask you about Databricks because, I mean, it's certainly booming in the private markets. I think the last valuation was$190 billion. They raised$5 billion, I think, earlier this summer it was. So here's a company that we talked about going public at one point. Now it's doing just, you know, probably enjoying its sanity in the private markets in some ways.

29:03I mean, what do you see for the IPO roadmap for a company like that? Does it ever need to go public? I'm not asking for an expectation here, but what do you think the calculus is here for that company? I think they need to go public, and I think they probably need to do it in the next 24 months. I think if you look at their early investors, the battery ventures that are now, what are they, eight, ten years in, wrapping up that first round, I just think the core economics support it. There's not a gun to their head. And I think what they're looking for is, if you squint one way, you say, man, there's a big valuation gap between late stage private and public, maybe three to four times, you know, but by certain sets of metrics.

29:52And in that sense, it's very scary. And then you look at what a SpaceX has done, and, you know, maybe a little bit of the recent read-down, and it gives you hope that the right AI-oriented public company, you know, can trade really well. Palantir is an example of that. I think all eyes are on Anthropic, to be honest. And I think if that IPO is successful in price as well, that will give Databricks a lot of confidence. I actually thought of Databricks as really the canary. I think that will be the most interesting one once they go public. I think there's probably 200 companies watching them, and they all sort of have a similar narrative around maybe being in another market, pivoting to AI, having had some success in that pivot.

30:39Right. What do you think we learned from the SpaceX IPO with respect to the market's ability to absorb these size of offerings? I mean, we could talk about the share price in one way. It's now back up above its IPO price a little bit. It's sort of dipped down below a little bit. How do you think about that issue with Deanthropic OpenAI IPOs coming down the pipe? SpaceX is so hard because Elon is one of one. You look at our Tesla trades and pittings. I'm always nervous. But maybe just the size of the offer, like the ability for the markets to absorb it. I mean, whether it be the bond markets, the multi-trillion dollars that we've seen issued, the ability for global and American capital markets to step up to play around AI in just a monumental way continues to surprise me.

31:32and I think Anthropic almost has a blank check, assuming that they have the right set of metrics. And I think what we're seeing is the proof has been under the pudding, right? You had NVIDIA knocked out earnings. We saw what Microsoft did in their cloud business. So, so far, kind of so good on these investments, I'd say. But what do they sell to get into Anthropic?