Nvidia Chip Prices to Rise 17%, Starcloud CEO on Orbital Data Centers, Amazon’s Growing Importance

0:13Welcome, everyone, to The Information's TI TV. My name is Akash Pasricha. It is Monday, August 24th. Today on the show, we are diving into some new reporting, the information published over the weekend, about NVIDIA's price hikes for its chips, and also discussions about an investment in perplexity. We'll then bring on the CEO of StarCloud, the orbital data center company that raised money at a$2.3 billion valuation a couple of days ago. And finally, we are going to close out the show with our Amazon reporter, who has a new story out this morning about why the platform is becoming so important for independent merchants after they used to shy away from selling on the website in favor of DTC.

0:56It's going to be a great show, so let's get right on into it. NVIDIA is raising the prices of some of its flagship chips. My colleagues Phoebe Liu and Amir Afradi wrote a story about that over the weekend. I want to bring on Phoebe to share more about what is going into this decision. Phoebe, welcome back to the show. It's great to have you here. Great to be here. Okay, so is this all memory chip price hikes, or what's going on here? Yeah, so I would say it's mostly driven by memory, but I think a lot of things are in short supply right now. So the price hikes are about 17 % and they apply to shipments of NVIDIA's kind of last generation Grace Blackwell 300 server racks and current newest Vera Rubin 200s.

1:43And it would be four shipments that are set to arrive kind of early to mid next year. And I would say that memory prices are the biggest driver of the increase, but I think also storage and networking are starting to play bigger roles as well. Now, I mean, you said memory is the driving factor here. Can NVIDIA not just lock in prices for some of these memory chips? I mean, you know. Trust me, they do. They do everything they can. Okay. But what is it? They're just so big that they can't get all the memory they need or what? Yeah. I mean, I think literally NVIDIA is investing money into some of these companies to try to kind of get them to produce more just because they physically can't even produce as much as NVIDIA and the rest of the world needs right now.

2:33Yeah, I guess on pricing, NVIDIA renegotiates pricing agreements with the three biggest memory suppliers, so Samsung, SK Hynix, and Micron periodically. So every time they kind of renegotiate, there would be base price changes. And then there are also additional fluctuations with the various suppliers that also help procure memory and kind of put the individual components into the server racks as well. Is there any discussion from the people you're talking to, customers? I mean, my sense is people are not going to buy less NVIDIA, but I don't know. No, I don't think so. So this is what business as usual, I guess.

3:19Yeah, and it's also unclear kind of how exactly this will affect margins. I assume that most of the price increases will be passed on to the end customer. That's what the server people that Amir and I spoke to also said. But one interesting point from one of my interviews for that story is that some people are starting to expect that memory prices might ease up a little bit shortly after Q1 of next year. We will see whether that will happen. It depends on kind of whether the big suppliers are able to actually build out additional manufacturing capability by then. But if they do, things could get better.

3:57Right. Well, and I should say that it will be interesting to see then if these prices come down, because on one end, if your customers get used to paying the higher price and it doesn't affect demand, I mean, I don't know, maybe there's an argument to make saying, well, it's just extra margin for us. You can ask Apple about that one. Well, I know. Yeah, I'm curious to see what happens. But I mean, look, the iPhone demand is different than the NVIDIA demand in the current moment. So that might be a little more elastic. Phoebe, I want to ask you about another story that you published over the weekend with Amir and with our colleague Valida Pau.

4:39So you had some exclusive reporting about NVIDIA thinking or talking about investing in perplexity. Tell us more about what we know there. Yes. So right now, NVIDIA is thinking about investing kind of multiple billions of dollars into Perplexity, which has kind of AI search chatbots and AI agent products at a$30 billion valuation, which is about 50 % higher than the company's last round. I think last fall. What's really interesting about this deal is that talks were previously around kind of a license and hire deal, similar to what Nvidia did for Grok, for Infabrica, for just, I think, last week, poolside.

5:27Similar to that, but somehow, some way, it morphed into more of a traditional equity investment. And so, I mean, let's talk, the poolside deal was interesting. I think it was a$6 billion deal that was reported by newcomer initially. So this deal, I mean, why is NVIDIA going for the model companies now? What's the ambition here? Yeah, totally. So NVIDIA over the last year has really ramped up its investment in developing its own models. And NVIDIA's AI models are all pretty much open source, open way, kind of everything. They release the weights, they release the training data as much as they can.

6:11They release methodology, environments to help customers fine-tune on top of those models. They release everything essentially for free. And the idea is that it's not just charity. I'm sure NVIDIA, yeah. Anyway, it's like a very business-minded decision. The idea is that if more people outside of kind of the top few firms that buy a lot of NVIDIA chips can access affordable and like pretty good and fine-tunable open source models, they will want more compute and therefore buy more NVIDIA GPUs. And I assume they have some kind of mathematical model internally. Would love to see it on how exactly that would drive sales, but I think that's the thinking behind it.

7:00Right. So on the perplexity, remind me, perplexity, do they make their own models? My understanding was I thought they really – I mean they have a great tool and perplexity computer use I think it is. Like the agent that they have, that's been driving a lot of growth for them. Are they developing their own models right now? What's the state of that company? They're not. I think what you said is exactly right. I would say within their tools, they have software that routes particular AI tasks to different models, including models that they've post-trained on top of existing open models. So they don't develop their own models, but I think the way Perplexity and NVIDIA work together kind of involves collaboration both on kind of hardware design because Perplexity has users and shares a lot of kind of anonymized data around real-world applications for AI to kind of inform how NVIDIA both trains its own models and kind of develops its hardware.

8:10Right. I talked to someone who said the companies meet literally multiple times a week to discuss all of these things. I would love to know more of the specifics of what exactly they collaborate on, but I think the companies have gotten quite close. And what's the status of Perplexity's business right now? I mean, I remember a year ago, this was a company that seemed to have a new funding round like every other month. We were talking about the sequential funding rounds. it then kind of, I don't want to say went quiet, but we stopped hearing about the funding rounds. It seemed to focus more on product.

8:45How's the company doing in terms of revenue and growth right now? Do we know? Yeah, so I think grain of salt, because this is annualized revenue, so a revenue for a given month times 12, it seems that from conversations that we had perplexities, annualized revenue is around$750 million right now, which is a significant uptick from just a few months ago. I'm not entirely sure what's driving that demand because I agree, I feel like they went a little bit quiet in the past year or two, at least in terms of media attention. But it seems like the business is coming along. Right. Last question for you, Phoebe.

9:28I mean, And as it relates to NVIDIA's model push here, you know, developing these models, what are the open questions that you have right now with respect to how it's looking to build out that function? And we also have earnings coming up this week. I mean, I wonder if you're watching for any specific comments from Jensen along these lines of building out the Nemotron family. I mean, what should we really be watching for? Yeah, so a couple of things. i guess on the heels of the crazy poolside deal uh wondering just how good nvidia's next open model is going to be um i've heard that it's about one trillion parameters which is smaller than today's current leading edge models but bigger than anything else nvidia has released in the past and yeah curious of kind of the poolside people will help make that even better um and then if nvidia's models end up being competitive, like the best of the best, if that happens.

10:30Curious as to whether they will ever charge for them. They've insisted that they never will, but I feel like they're always searching for a kind of a new revenue stream because... It's not charity, as you say. Jensen's not running a charity. It's definitely not charity, but I think the market just expects so much out of NVIDIA's earnings. I think NPD is expected to report 90 plus percent revenue growth, but the market might still be disappointed, which is crazy. So I think Jensen and his team always has to keep looking for the next big thing because they're already so dominant in AI hardware. It's like they need to think about robotics, about model making, about so many other things.

11:18And also kind of for earnings, really interested to see what else they'll share on kind of all the financing and the backstopping and dealmaking that they're making and how that might affect how investors are perceiving NVIDIA's credit and how it's assessing its risk. Great. Well, Phoebe, I want to thank you for coming on. That is Phoebe Liu, our NVIDIA reporter here at The Information. star cloud the fast-growing startup building orbital data centers raised 250 million dollars at a 2.3 billion dollar valuation the funding round included nvidia cisco benchmark 776 and a whole host of other investors i want to bring on co-founder and ceo philip johnson for a conversation phil welcome to the show it's great to have you back thanks so much for having me Congrats on the funding round.

12:10I want to get all the updates on the company's operations. Well, we've had you on a couple of times with all the SpaceX madness, but last time we had you on to talk about your company was in November and you had just launched an H-100 on board a SpaceX vehicle, I believe it was. And we haven't had an update since then. So give us the update.

12:33Philip Johnston:yeah for sure so with that h100 we then trained the first model in space and we trained nano gpt from andre carpenter we then also ran the first version of gemini so working in collaboration with google deep mind we ran a version called gemmo it's like a cut down local version and then we also have been running high power inference on satellite imagery particularly on SAR data so synthetic aperture radar which is very useful for many sort of government and military applications um then yeah in march we announced the raise 170 million raise at a 1.1 billion dollar valuation led by benchmark and eqt and then yeah and then just uh friday last week we announced an extension to that so another 250 million at a 2.3 billion dollar valuation and yeah significantly we had nvidia investing in this round um as well as cisco and it was led by Manhattan West fund in LA that was early in SpaceX.

13:31So is it, let's go back to sort of the launches. So the H100 you got up there, you used it to train some models and it was that chip that you used to then work with the Gemini team as well?

13:46Philip Johnston:Yeah, exactly. So the first spacecraft, we had five GPUs on there. The most significant one was definitely this NVIDIA H100. Yeah, three of them were from video two from arm um but the h100 was the thing we used for all of those uh workloads so training inference and uh also fine tuning of a model right and so what is the timeline now from here what does the roadmap look like going ahead with these launches when are we going to see more chips and space from you guys yeah so in q1 next year we've got two launches booked and then also in q3 we've got another launch book so three launches going up next year um that's our star cloud 2 satellite so it's a um much more capable spacecraft it's about a hundred times the power generation of star cloud one we'll have by far the largest commercial deployable radiators in space um as well as you know many more h1 h100s and blackboard chips Bitcoin mining a6 and all that other good stuff right right tell me um you know I I've been thinking a lot about So, you know, we've had a couple other Orbital Data Center folks on the show.

14:54Cowboy Space Corp is the other founder that we've had. He described his approach to us in terms of he wants to leave part of the vehicle up there instead of bringing it back down, at least as far as the satellites go. So, look, everyone has their own approaches here. And look, the big debate is if and when orbital data centers will pan out the way you and founders are advertising. But I want to get into some of the more nuanced debates here on approach. You know, if we sort of compare you and all these other companies, what are the more micro debates that are playing out in terms of your approach?

15:36and then you know uh spacex's approach blue origin is trying their own um constellation of of uh launches although that one's going a bit slower compare and contrast all these for us yeah um so

15:51Philip Johnston:the the main difference in us and cowboy space is cowboy is building their own launch vehicle um um so i mean that is certainly yeah that's certainly one approach you know yeah people have built launches before so you know they should be able to do it but it is a very capital intensive and uh time consuming process to build a launch vehicle um so yeah i mean well and i'm not even i'm not even really you know i think look building that's a big difference here between you and them, but I'm talking more about the scientific debates on different approaches of getting orbital data centers. I mean, are there any other sort of hot button issues that researchers are thinking about right now?

16:37Is it pretty straightforward or are there different ways?

16:42Philip Johnston:I guess there's only one sort of outlier or there's one company doing it a bit different from most which is Sophia Space is as I understand doesn't have active cooling and all of the other companies so us SpaceX probably Cowboy well definitely Cowboy actually um uh Google's Project Suncatcher and also uh Blue Origin we all have this active cooling loop so you have a radiator which you run cooling through and then that goes back to the chips and then heat it up and back out to the radiuses. The main difference with Sophia Space is they have a passive cooled architecture. So they have this panel of solar panel and the chip is in the middle of that panel and all of the heat from the chip will radiate away on the other side of the solar panel essentially, which means they don't need cooling loops.

17:38Philip Johnston:The downside of that, of course, is you can't run very power-dense chips because you need... So you need to spread heat from those power dense chips. Otherwise, the whole thing just melts. So because you can only radiate heat at a certain amount per square meter of surface area. And so the purpose of liquid cooled architecture is that you can spread that heat across a much larger surface area, which enables you to run in much more power dense chips. Right. Now, the other issue that I've been trying to pay attention to is how you actually get the compute back down to Earth. And there's optical links as one way that people are trying to develop further.

18:22And then there's radio links, which is sort of the more popular version of this. Which of those two ground link connections are you trying to rely on? Yeah.

18:31Philip Johnston:So for now, almost all space to ground links are RF. Space to space is very quickly moving towards optical. And so our second space cluster has three optical terminals. the first one actually only has these RF antennas to connect to the ground but our second space cross has three including two Starlink terminals which enable us to connect directly into Starlink and then in the same way that when you use a Starlink terminal on a plane or something you have very low latency connectivity through the Starlink network will have very low latency connectivity to anywhere on earth over time it does look like the industry is moving to space to ground optical.

19:10Philip Johnston:The problem with that is obviously if they have cloud cover and you have attenuation in the atmosphere, that can significantly slow down the data rates. There are places in the world where you have much less cloud cover. So the Atacama desert in Chile, for example, there's a lot of optical ground stations being built on the tops of mountains in Hawaii and things like that. But for now, almost all data going from space to ground is RF right now. Right. What's the pitch from you guys to customers using this? One is certainly how you establish these data centers and the costs associated with that.

19:50And I think the industry is broadly pitching this idea that once we get this technology developed, it will be cheaper to set up a data center, which is the ambition. the other half of this is then the unit economics of the compute itself are you pitching that the unit economics of using compute is going to be cheaper uh so it depends on the time frame in

20:13Philip Johnston:the next two to three years it will not be cheaper because the launch cost is still quite high um so the business we're going after for the next few years at least is we are mainly selling to other space users particularly you know various government entities and uh and earth observation constellations uh for that you know that that business works very well with just the falcon nine launch cost um over time as the launch cost comes below about 500 a kilo then uh yeah to your second point it will be cheaper on a per unit of compute basis per token or per per watt um to run this in space because we have just lower energy and infrastructure costs in space including the launch post right and where where are you going to make your all of your uh equipment and satellites and everything yeah uh so right now we're in redmond washington just down the road from uh spacex starlink and also from amazon's leo constellation now sometimes people don't realize 90 of all satellites launched in the last few years were designed built tested in in redmond where we are and we're about to open a much larger facility just down the road in woodenville about 100 000 square foot for 13 at least 10 times the size of what we're in now um to set up you know more of a manufacturing line um to to start producing these so is it all is it all like former boeing people that work at these uh at these companies i mean everett is just a stone's throw away is it is that the reason that this industry has popped out over there um the initial reason I mean, ourselves, we don't have any bank people.

21:46Philip Johnston:About half our team came from SpaceX. And then the other half from various hyperscalers, you know, Microsoft and AWS. I think in the early days, the reason it was here was there was a very strong RF team at Microsoft. And also there was a strong team at T-Mobile. And T-Mobile's headquarters and R &D is all just down the road. So they were particularly, you know, things on, were working on cell towers and things like that. back in 2016 I think they started SpaceX moved up here and I think it was for that reason uh and then yeah and then the town's just growing you're still you're you're 25 people right now I read yes and what's I mean a year from now how big are you trying to be uh we'll be 50 by the end of the year um yeah you need people to make that make the satellites well yeah I mean so we're only launching I mean Adi and Ezra did the whole first base graph by themselves essentially the payload and the antennas and everything that was two people StarCloud 2 I think will essentially do with like 15 people um and then we've got three launches of StarCloud 2 next year so you know if we have 100 people that's not uh yeah until Starship is flying frequently we cannot ramp up real cadence on right on StarCloud 3.

23:06Great well uh Phil I want to thank you for coming on some exciting developments and look forward to having you on more as there are more questions to debate and discuss on the space race. That is Philip Johnson, co-founder and CEO of StarCloud here on TITV. Companies that used to rely on direct-to-consumer sales are finding that Amazon is becoming a more attractive distribution channel. Our Amazon reporter Catherine Perloff wrote a deep dive on that trend this morning in an exclusive story. I want to bring her on to share more about what she found. Catherine, welcome back to the show. It's great to have you here.

23:45Hey, Akash. So what's going on with these independent merchants in Amazon right now? Yeah. So Amazon is becoming a bigger source of their sales. You know, 10 or so years ago, there was an idea that the Internet could kind of democratize retail and cut out the middleman of the big box store and also cut out the middleman of Amazon. Amazon has been around for 30 years. But that is becoming harder to do for some reasons we got into. And yeah, Amazon, you know, for a lot of merchants is becoming sort of their biggest, fastest growing retail channel. And I mean, you spoke to a number of independent merchants for this story.

24:31And I just want to make clear, I mean, we're talking basically about these brands that, you know, I used to buy from tons of websites on their website, brands that used to rely on direct to consumer. The websites, I guess, have become less popular now is the idea, right? That's right. Yeah. I mean, you know, some it's still a viable sales channel, but it's not something you can necessarily rely exclusively on. You know, some people used to see Amazon as sort of optional. Maybe you put a couple products there, but you want to drive all, you know, sales traffic to your own website because you have higher margin on your own website.

25:11You don't have to give up the fees to Amazon or other retailers. But the economics that sort of made that make sense are more challenging now. Okay. Why is that? What's so great about Amazon or what's so bad about DTC? Yeah. Yeah, I mean, and listen, you know, I don't know if there's still a lot of challenges on selling on Amazon and there's still a lot of fees. Which I want to get into that, too. I mean, I want to get into sort of the double-edged sword here. But I mean, why? What's so attractive about the channel right now? So, DTC became harder for two, maybe three main reasons. First, ad costs got more expensive.

25:50And, you know, I think partly just as the model became more popular, that means more people are trying to do it. more people are paying for ads to drive, you know, shoppers to their site. Between 2021 and 2026, ad costs on Google and Meta went up more than 20 % and actually have not gone up as much on Amazon. So that's one thing. Another thing is it is harder to find websites now than it used to be. Google has sort of changed the UI of its search engines. So now you see an AI overview. There might be a few websites in there, but certainly not as much as when, you know, it was just 10 blue links.

26:28And that has made a lot of merchants who rely on their sites get less, what they call in the business, organic traffic. So traffic that doesn't come from ads. So it's just harder to find their sites. So, you know, it used to be that, yeah, maybe less people would come to your site than Amazon because, you know, Amazon is a big clearinghouse for everyone knows about. But it was cheaper to get those people to go there. And now it's more expensive to get someone to go to your site. And I guess the last thing that, and we can talk about this more later, but like the trust factor of Amazon has maybe increased a bit.

27:04But people have told me that people who find brands on TikTok are more likely to buy things on Amazon. And I think it's partly people sort of have been trusting reviews more. I think TikTok sort of encourages this sort of, you know, familiar style, not influencers of just like, you know, this is a product I liked. And I think people kind of trust the reviews on Amazon that you can see and don't always trust reviews on brands on websites. I thought half of these reviews were like AI generated. Am I dreaming? I mean, I don't know. I don't have any data here. I just, it was, you know, there's, I don't know, there's discussions.

27:46No, I mean, I think it's always great to think about like, are the reviews legit? There's lots of review manipulation going on everywhere. But I think that the idea is that reviews on Amazon are more legit because Amazon has less of an interest to promote any one product than obviously a brand on their own website. And yeah, I mean, these reviews are so effective that one brand I talked to, Flex, which is a menstrual product, started printing the Amazon reviews on their Subway ads because they were so effective. And another sort of sign of this Amazon trust boost is that an e-commerce consultant I talked to found that if someone buys a brand on Amazon, they're more likely to then maybe try it and buy it on the D2C site.

28:34But they kind of need to know that they can return it on Amazon, which Amazon has a good returns policy. They don't necessarily trust the D2C site as much to let them return it, that it's good, et cetera. Right. Okay. So what are the costs here? I mean, is it all great selling on Amazon? I mean, some of the brands you talked to in the story said that Amazon is now, in some ways, their biggest customer, which is significant. Are there costs that come with that? Yeah. I mean, Amazon is, you know, there's a lot of problems. You give up a ton of fees. You're in their ecosystem. You have to buy ads to stand out.