Anthropic in Talks to Use Microsoft AI Chips, Biggest Reveals in SpaceX IPO Filing

0:13Welcome everyone to the information's TITV. My name is Akash Basritcha. It is Thursday, May 21st. We are kicking off the show today with some exclusive reporting that the information published around talks of a new chip deal between Anthropic and Microsoft. we'll then get to the monster nvidia results the company posted its quarter results it is expecting another big quarter ahead we'll break that down with an analyst very shortly the spacex s1 is finally here we have the numbers we are going to get to the analysis we are bringing on ross gerber very shortly and we will wrap the show with some more exclusive reporting this time around the White House's plans to review advanced AI models before they get released to the public.

0:59It's going to be a busy show, so let's get right on into it. The Information published exclusive reporting that Anthropic is looking to make a chip deal with Microsoft. My colleagues Aaron Holmes and Chan Erlou wrote that story. Aaron joins us now to share with us what he knows. Aaron, welcome back to the show. What do we know about this potential chip deal? so anthropic is in early talks with microsoft to use microsoft's own custom ai chips which are called maya 200s and the talks are early it's not clear if this is going to actually lead to a commitment but if they you know do decide to to rent the chips from microsoft this would be a huge win for microsoft's uh in-house chip effort which it's trying to to you know prop up as a meaningful competitor to NVIDIA, as well as to the in-house chip efforts of Google and Amazon and other AI players in the space.

1:54So we've talked a lot about a lot of chips on the show. We have not talked about the Maya 200 before. What is the Maya 200? So this is the latest version of Microsoft's in-house AI chip. Microsoft has been working on Maya for several years, and this version specifically was delayed a few times. And, you know, a lot of AI chips like NVIDIA's are used for both training new AI models as well as running existing AI applications. But the Maya 200 is more focused on just running AI, which is also known as inference. And Microsoft has said earlier this year that it started using Maya internally to run its Copilot tools and that it found that it was cheaper to run Copilot using Maya for some tasks than NVIDIA chips.

2:48So, you know, if it can convince customers like Anthropic that they could also run their models more cheaply on Maya than on NVIDIA chips, that could help Anthropic, you know, unlock more compute, which it desperately needs right now because of all of the surging customer demand that Anthropic is seeing. Do we have any sense around whether or not this chip is better or worse or comparable to Tranium or the TPUs that Google has put out there? It's definitely a lot more nascent than Google's TPUs and Amazon's Tranium. You know, Maya, as of now, doesn't have any big customers other than Microsoft itself, whereas, you know, Google has already been able to convince companies like OpenAI, and Anthropic and Meta to use TPUs, and Amazon has similarly said that some major AI labs are using Tranium.

3:40Generally across the board, NVIDIA is still seen as the gold standard for GPUs, but it kind of depends exactly which type of task you're trying to use them for. And I think that winning a big customer like Anthropic would go a long way in terms of Microsoft's efforts to prove that these chips are the real deal and are useful to major AI customers. Were you surprised about this when you heard this? Because I feel like we just put the Maya 200 aside for a second. I mean, chips were an area that I think we just haven't really talked about Microsoft around that much. And I think there was always a hunch that everybody's getting into the chip game.

4:22Everyone wants to be a little more vertically integrated. Microsoft probably would have got into it. But, you know, it strikes me that if they get this Anthropic deal, you know, this could mark a pretty significant turning point for its own chip ambitions in the long run. Yeah, I mean, on the one hand, it is maybe, you know, a little surprising to see a big lab potentially adopt the Maya chips, just because, to your point, we haven't seen that happen yet. But at the same time, Anthropoc has actually always had this kind of agnostic strategy where they have, you know, signaled that they're very willing to use any chips that they possibly can get their hands on.

5:00They were one of the first adopters of Google's TPUs, as well as Amazon's Tranium, in addition to the NVIDIA GPUs that they've always used. They're even in talks with newer startups about potentially using chips from those companies. So Anthropic very much has this philosophy of trying to diversify and not just rely on one architecture. And so in that sense, it makes sense that they might explore what that would look like from Microsoft. Right. And that's kind of what I was trying to get at. But it was, you know, to what extent is this a reflection of how good Microsoft's chip is? And to what extent is this a reflection of just how desperate Anthropic is or maybe how eager they are to diversify their chip supply?

5:44I mean, there is always a balance in that power dynamic. I want to dig into, though, the broader relationship between Anthropic and Microsoft. So how does this fit into, you know, the cloud deals that they have signed together and whether this could have an impact on that in the long run? Yeah, so Anthropic and Microsoft have been getting closer and closer over the past year. You know, originally OpenAI and Microsoft were joined at the hip through an exclusive deal that made Microsoft OpenAI's exclusive cloud provider and gave Microsoft the rights to OpenAI's models. But they opened up that deal earlier this year.

6:21So now, you know, OpenAI is striking deals with other clouds, and we're also seeing Microsoft get closer to Anthropic. And at the same time, Anthropic has become a big customer of Azure in the last, you know, six months or so. It has been ramping up its spending on Azure. I'm told that, you know, inside of Azure, folks have been actively setting up new capacity for Anthropic and essentially saying, we need this capacity for Anthropic so we can't give it to other customers, even some very big enterprise customers. And it makes sense. I mean, Microsoft wants to show that Azure is the place for AI labs to run their models.

6:59And at this point, Anthropic is the front runner in the AI race. So it's definitely in Microsoft's interest to keep growing and nurturing that customer relationship. Great. Well, Aaron, I want to thank you for coming on. That is Aaron Holmes, our Microsoft reporter, here at The Information. NVIDIA had a blowout quarter with revenue climbing 85%. The chip giant also projected 95 % growth in the current quarter. That would be the fourth consecutive quarter of top-line acceleration. To break it all down, I want to bring on Frank Lee, Global Head of Technology, Hardware, and Semiconductor Research at HSBC.

7:36Frank, welcome back to the show. It's great to have you here. All right. Good to be back. Okay. So what stood out to you from the results last night? Yeah. So I think going into the print, I think it was a very similar setup as last quarter. We expected we would probably see a beat and raise relative to where most consensus numbers was. But at the same time, I think we also highlighted, I think, a few days ago that NVIDIA would probably need to see more narratives emerging. I think earnings momentum and the next generation AI GP roadmap wasn't really going to be enough. I think they really need to start focusing on a couple of new narratives.

8:12And I think they've started to do that with the reclassification of their data center revenue, which we think is a potentially big event because now they're breaking down the hyperscaler and non-hyperscaler revenues. Yeah, so I want to go to that reclassification here. So it sounds like you think that that was a smart move and you support it. Yeah, I mean, I think if you think about the narrative going back to last year, you know, Jensen Huang's talked about the enterprise and the sovereign AI for quite some time now, but it hasn't really quite materialized or very obvious in their numbers because they haven't really broken that down.

8:48And so I think now that they've broken down the difference between CSP and non-CSP revenues, it will be easier for the market to really track how they're doing on a non-CSP. Because if you look at - CSP being - I mean, the hyperscodes. Yeah. Hyperscodes, okay. Cloud service providers. Yeah, because if you look at one of the issues you're seeing this year is that, it's just more obvious, is that the hyperscoded CapEx is now being divided by multiple people. So besides NVIDIA, you know, memory guys are now eating into this CapEx. You're also paying for the networking, the optical supply chain. You're also looking at, you know, the server CPU now becoming a very big theme, especially with agentic AI.

9:27So more people are now competing for that pie. But if you think about the non-hyperscalar CapEx, this is still very much NVIDIA dominated world. And that's, I think, what investors want to continue to see. So tell us a little bit about this. So the shape and structure then of these two segments, there was a question on there on the call about which one they expect to be growing faster and also where the Neo clouds would fit if they would sort of fall into the hyperscaler bucket or the non-hyperscaler revenue bucket. What was the answer to that question? And then we'll talk a little bit about how you expect these segments to grow afterwards.

10:07Yeah, so I think for the most part, everyone else outside of the four or five major hyperscalers are going to be in a non-hyperscaler bucket. So that includes, I think, the NeoCloud guys would largely be in it. In fact, if you look at the last couple of months, NVIDIA share price has had a bigger move on the back of the positive bullishness around the NeoCloud's demand rather than its own earnings beat and raise last quarter. So I think that's a sign that the market's looking for more growth opportunities to come from these non-hyperscalers. And so you're expecting, and the NVIDIA themselves, they're expecting that second bucket, the non-hyperscaler bucket, that is going to be the big growth driver for them.

10:50Well, maybe not growth driver because hyperscalers are so big, but that will be a faster growing segment is the point. Yes, I think that's what they're being alluded to. And I think that is probably going to be a more meaningful narrative to help re-rate the stock than its own earnings momentum by itself. So then I want to go back to what you were saying at the start is you said the company was looking for its new narrative or its next narrative. And this reclassification being part of that, I mean, what would be that narrative? I mean, you know, companies outside of the hyperscalers, I mean, I get that, look, you could say that enterprises broadly are turning to us, the neoclouds are turning to us.

11:33Is that enough to move the needle here? Or, you know, what do you think it actually takes for the stock to move even higher? Because it didn't really move on the results last night. Yeah, because, excuse me, I think if you look at the split right now, the data center revenue split is now almost 50-50 between the hyperscalers and the non-hyperscalers. So I think if we see the non-hyperscanner portion continue to rise and exceed 50%, that would be a pretty positive indicator. The other thing I think as a second narrative is this focus on the server CPU, especially with the attention being paid this year to GenTech AI.

12:09This has been basically an area that NVIDIA is now pushing more aggressively. They talked about potentially$20 billion in revenues from the server CPU itself. Now, I guess the pushback, though, would be$20 billion for NVIDIA is not a huge number. I mean, it's crazy to think, but given how big the revenue already is. So I think there's still some debate about how meaningful the server CPU could be to its overall business. But it is an area that I think the management understands that the market is quite focused on right now. And so I just want to come back to making sure people understand how the results are reported.

12:44CPU revenue, server CPU revenue, where would that fall in that reclassification now under the current structure? Well, yeah, it's a good question. I mean, I think it still falls under compute. Now, how it breaks down, I think you'll see basically CPU revenues both in the hyperscaler as well as the non-hyperscaler, especially if they sell standalone. So that, I mean, they pulled out a number of, excuse me, of 20 billion, but they haven't really, in terms of where it classifies between that, I think you're going to see CPU revenues in both types of customers. Do we lose any level of granularity with this new reclassification?

13:23You know, it's always kind of a tug of war, right? Like you get a little bit more clarity in one sense, and then you lose some detail in the other way. Is that a concern at all for you? I think from a, I mean, NVIDIA itself is also getting much more complicated in terms of reporting, right? Because it is moving beyond just a GPU revenue company. It's becoming more system-oriented. So I think the way they have been reporting wasn't going to be fully reflected anyway. I think the end customer base is what ultimately is becoming a more important driver. So I do believe this is the right direction for them.

13:59The other thing that came up on the call last night is CFO Colette Kress. she talked about rental prices for older NVIDIA chips, and she mentioned they had actually come up. What do you think is driving the increase in those rental prices? Is it just the GPU crunch broadly? Is it customers saying that, hey, maybe we don't need the latest and greatest we can make do with the Hopper family or the A100s? What do you think was behind that? I think, excuse me, I think ultimately what you're saying is that, is that you're still seeing demand is just still outstripping whatever supply is out there. So the compute demand is still exceeding most people's expectations.

14:43So I do think that is why the rental costs are still going up. We've seen that already from the neoclouds. I've been talking about that in the last couple of months as well. So it still paints an overall picture that there's just not enough compute demand out there. But I do think one other thing that's interesting is, you know, the supply side bottlenecks continue to diffuse out to more and more places. So one thing I think that another narrative that NVIDIA has talked about is all these recent deals on the optical side. If you look at these deals with Coherent and Lumentum and Corning, these are all meant to secure the supply side on the optical side, which we think will be a very big driver.

15:22So I think NVIDIA agrees with that, which is why they want to make sure that they have enough supply there. Can you explain that a little bit more for people who don't follow optical as much? Just unpack that a little bit more in layman's terms. Sure. So as you see these data centers build out, one of the bottlenecks is really the bandwidth, the speed has to get faster and faster. And one of the things that's driving that is the optical demand. This is going to be, we're going to move from copper to optical, and this is going to create a significant amount of demand to help improve performance.

15:54So it's not purely just based on the actual computing power of the chip. It's a much more system integrated, and that includes the networking piece. And so NVIDIA has been moving there already with its Mellanox. But in addition to that, they got to make sure the rest of the supply chain is there with them. For example, if you look at memory, this has been a big bottleneck, and they have done a lot of things to make sure they have secured enough memory capacity, given the tightness we're seeing. They're just extending that to another part of the supply chain. Got it. So when you say optical, you're talking about the networking products that NVIDIA has moved deeper into.

16:29Yeah, like they're basically the transceivers that goes into the backbone in terms of providing the networking bandwidth, both their scale out and scale up. You're going to need a lot of networking for that. Right. A couple other questions for you. So Vera Rubin, the new family of chips, when you talk to channel partners or people in the orbit of customers, etc., what are they hoping for out of this new family of chips? And, you know, how confident are you that NVIDIA can deliver that? Well, there's always potentially a concern that whenever you have a new ramp, there could be some potential issues in supply chain.

17:11there has been some potential concerns about memory not catching up, or at least not quick enough. But we think there's going to be relatively short term. At the end of the day, I think NVIDIA is still pushing performance-based. Rubin does increase performance. But that also brings up another question is that we're seeing some competition from the ASIC side, especially Google's TPU. We've seen a lot of very strong momentum there, which is why I think the market is looking for them to extend their customer base. Because I think if you look at NeoClouds, right? They're not going to be moving into the ASIC side.

17:44So they're going to be very much within a GPU camp. So I think a lot of that Rubin demand will also go into these non-hyperscaler customers as well. Great. Well, Frank, I want to thank you for coming on. That is Frank Lee from HSBC here on TI TV. SpaceX's IPO filing is officially out. It was a document that Silicon Valley markets, the whole world was waiting for. In some cases, there was a lot there. And so to break it all down, I want to bring on Ross Gerber, president and CEO of Gerber Kawasaki Wealth and Investment Management. Ross, welcome back to the show. It's great to have you here. Great to be back.

18:23So SpaceX's IPO find, there's nice photos in them. You could make a little coffee table book out of it, maybe. Yeah, you know, for an S1, it was the first S1 I read with pictures, you know. Yeah, yeah. Yeah. So, I mean, look, we'll get into it, but high level, better or worse than you thought? You know, I kind of knew most of the information, to be honest, because Reuters had gotten a hold of this somehow earlier. So I kind of knew. And I should say, we also had a number of the files. So not several media outlets. But the information also reported a lot of it, yeah. Yeah. So because of my relationships, I kind of had some of the information already.

19:01What I thought was really interesting was the way they laid out their business plan because it's really a three-pronged business. And I think a lot of people are really focused on financial numbers and all the stuff that came out, like how much money XAI is spending on servers versus the fact that Starlink, for example, is a wildly profitable, successful business. Space launching is also not per se a bad business either. And so when you see the vision laid out, the three prongs of the vision, which is really about, you know, first, you know, connectivity through Starlink. And then secondly, this, you know, space data center things, which would bring ideally down the cost of compute substantially, but also be a connection to further life outside of Earth.

19:49You know, it's kind of this crazy vision. But, you know, somebody's got to do it. I guess it's Elon. So there was nothing in there. That surprised you, really? Not surprised me. I think the one thing that I didn't really, I guess, understand as clearly, and I recently met with a lot of SpaceX employees because we're working with a lot of SpaceX employees at my firm. And, you know, you see different parts of the vision, but the way the S1 is lined out, you see this, like, grand, grandiose vision of the future. And really this idea that humans could be, like, extinct. And so he wants to create this, like, colony somewhere else so that we'll never be extinct.

20:31You know, I'm more in the camp of, like, why don't we just, like, save the Earth we're on now versus trying to live somewhere else. But certainly Elon and SpaceX's vision is massive. so that's the part that surprised me right it's just like wow you know like a billion people living on mars come on you know i can't get a million people to move out of la to somewhere you know else you know it'd be great so let me ask you this do you do you uh recall when the tesla s1 came out i think it was 2010 was it you remember reading i mean i i kind of remember that i remember of the IPO because we watched it.

21:07Um, and you know, but at that time I wasn't like that into Tesla, you know, like somebody brought it up to me and I knew of it. And so I didn't actually care that much, but once it went public, it was actually the model S which rest in peace, you know, died yesterday. Uh, they had the funeral at, at, uh, Tesla yesterday online and which I couldn't watch cause I was so sad because that was the vehicle when I saw it, I knew Tesla had something very special, you know? And so, you know, it's a very different experience because Tesla was like a three or$4 billion company. It was a startup and we were betting on something new.

21:44SpaceX is a very well-established multi-trillion dollar company that now in its maturity is giving investors an opportunity to get involved with. Now, I was lucky enough to get into the SpaceX investment through my investment in Twitter many years ago. So I'm happy to, full disclosure, I do own SpaceX, so I'm excited about this stage. But what I'm most excited about actually is somebody actually has a grandiose vision in America. You know, like you can't get somewhere if somebody doesn't try. And I don't necessarily believe we're going to be living on Mars or not. That we can debate that for the next several decades.

22:23You know, I ask my kids, my kids are like, why would anybody want to live on Mars? Like, you know, humans don't do good with no gravity, it turns out. You know, things like that. So but I think space as an actual place, it's like a new sector. And this sector, in what I've learned, has enormous opportunities. And so you can look at this grandiose stuff or you can focus on the stuff that's real, which is how much weight they can bring into space on an annual basis to build up satellites and put up actual infrastructure in space. And that number is growing exponentially. We have a launch this afternoon, Starship launch.

22:59you know hopefully hopefully it goes well it'll go if the question is like how far you know yeah yeah so i i do want to talk though about you know you mentioned what's real and so there were a couple numbers in the filing that that were uh interesting the specifically the numbers i'm thinking about are the arpu numbers the average revenue per user numbers for the starlink business and you know we saw that those numbers uh have started to come down even despite the number of subscribers having gone up and look i think but that's also because of the starling mini like i like i have two starlings one i pay 120 the other i only pay 50 because it's the mini right and so i think there's a good understanding out there as to why that's the case my question is i mean do you expect those to come up do you expect those to continue to come down how are you thinking about that figure um you know because it's kind of like a new business and i find their price point to be very competitive compared to like normal internet and starlings are a lifesaver i you know it saved my life in the palisades fire in a lot of ways like we were stuck for three days without any communication and without starling it's it's winning the war in ukraine so we can't underestimate the value of starling and what will eventually be charged to customers it seems like they're bringing down prices to really just be competitive because their margins are huge at Starlink.

24:25So, you know, once they get it up there in the sky, you know, it's a very profitable system to run. So Starlink is clearly the, you know, the profit monster of SpaceX. And so what I think the real question is, can they like double, triple, you know, the amount of subscribers and maybe rates go down a little bit, but you've still got a rapidly growing business that's highly profitable. So that's definitely the queen of the crop of businesses inside SpaceX as far as from profitability perspective. Now, the other thing that was interesting was the lockup period for the shares here. It's shorter in some cases for some employees than the 180 days that everyone has been traditionally used to.

25:08And, you know, my sense is this is becoming a little bit more common for tech companies. It wasn't always the case, but what did you make of the language around the lockup period? And then we'll talk broadly about whether you think this is going to change in the long run. I think it's a byproduct of the just lack of supply of IPOs in America where companies can like do these things and investors will take it. I mean, the first page of the S1 is just absurd. It's basically like this is Elon Musk company. Texas has a rule like you have to have independent directors. That's all thrown out the door.

25:41It's called a controlled company. it's like flat out Elon will do whatever he wants you are investing in this on whatever terms he wants if not stay away so these lock up periods and all this he doesn't care what anybody says he's going to do whatever he wants to do and if you don't want to be a part of it don't be a part of it and that's the attitude of SpaceX so so they're going to do what they want to do and they're going to let their employees uh have liquidity sooner and and I don't per se have a big problem with that as an investor as long as it's disclosed. People know this and a lot of liquidities, we're working with these employees.

26:15I'm telling you, a lot of them have been there for 10, 15 years and waiting to get a little money to guarantee their future. And we're working with these people and there's a lot of them. So it's hard to predict the supply and demand dynamics of IPO, but I know the demand is also really high for shares. So this is going to be a fun time. It's going to be a very, I think, record-breaking IPO in almost every way. Sadly, it reminds me a little bit of the AOL-Time Warner merger with all that euphoria in the late 90s that I lived through. And we're starting to see a little bit of that with SpaceX. So part of that's good and part of that's bad.

26:54Why do you draw that? I mean, go deeper on that connection. Why? Because we want markets to function without too much emotion. OK, like markets can go pretty out of control, bearish or bullish, depending on sentiment, you know, in human behavior. And we're getting into a stage in the cycle where there are many things that are concerning, like higher rates and inflation, which could actually really hurt the economy. But yet our economy is just like booming right now, despite these things. But it doesn't last forever. And so what happens is you get a lot of euphoria about a big IPO like SpaceX. And then you get Anthropic and you get OpenAI.