SpaceXAI’s Explores Texas Data Center, Oracle’s Costly Surprises, AI-Generated Movies Coming to Theaters

0:13Welcome everyone to the information's TITV. My name is Akash Pasricha. It is Wednesday, July 22nd. Today on the show, we have exclusive reporting that XAI, SpaceX AI's AI division, is working on a big new Texas data center. We'll bring on our Elon Musk reporter with more details on that effort. We also have exclusive reporting on how Oracle's push to build new data centers is hitting roadblocks facing surprise costs and heavy red tape. We'll also take a look at how NeoCloud General Compute is looking beyond NVIDIA as it seeks to build out its NeoCloud offering. And to close out the show, we'll talk to the director and producer behind what could become the very first AI-generated movie to hit theaters.

0:58It's going to be a great show, so let's get right on into it. SpaceX AI is working on a new data center in Texas. That is according to exclusive reporting from my colleague Grace Kay, who covers all things Elon Musk. I want to bring on Grace to share more about what we know. Grace, welcome back to the show. It's great to have you here. Thanks. Okay, so what do we know about this data center that Elon Musk is building in Texas? Yeah, so they've been talking about a data center expansion to Texas for a few months internally. They've been exploring a few sites and are considering potentially building multiple data centers in the state, including, you know, building it from scratch as a greenfield project and repurposing a warehouse, which is kind of what they did for their first data center in Memphis.

1:44And do we know how big this data center is supposed to be at all? Yeah, so it's very early. So there's that caveat that this could change. But based on what I've heard from sources, they want to build this on the same scale as Memphis, or potentially larger. So we've talked about the Memphis facilities a lot on this show. And I think there's what there's one that's built and two more that's being built. Is that right? it yeah they have they're all in kind of various stages of being built there's one that's fully operational was built in about 122 days which is kind of a record for the industry but yes they're spread out across memphis and south haven um nearby and i mean so this would be i guess the uh fourth facility that um spacex ai would be working on i mean let's talk about the demand for SpaceX AI's AI products.

2:42I mean, Grok, I think you've reported as well that demand has been a little iffy on that front. So talk a little bit about how they're considering building all these facilities when it's looking a little iffy for Grok right now. Yeah, I think we're seeing these compute deals with Anthropic and Google. And I think that's something that SpaceX AI has kind of realized is their secret sauce, is they're really good at building these data centers really quickly. You know, like I said, the 122 days for the first one was pretty unheard of in the industry. So I think they're seeing the capacity from other companies that are willing to pay, you know, large amounts of money to rent this compute.

3:22So, yes, you know, Grok maybe doesn't have as much demand as some of the other AI companies right now. But Elon Musk has indicated that they're going to continue to work on that. But also, you know, there's a subset of the business where they can make a lot of money on their ability to build these data centers quickly. Right. I mean, they can make a lot of money, but it's kind of funny. Like, being a neocloud, I don't think it was the vision for SpaceX AI to begin with. And so, in some ways, I wonder why keep going. I don't know. You know, I wonder if people on the ground, if they're telling you, like, is this the new vision for the company, really?

4:01Is it just something that they are pushing ahead with because Elon is telling them to? Like, what's the strategy here? Yeah. I mean, it's definitely part of, like, the Elon Musk vision for the company. He has repeatedly told staff that, like, how they're going to win the AI wars is by having the most compute. And we've seen with other companies like Anthropic, you know, compute constraints have led them to look elsewhere. So it is an advantage, and it's something that Elon Musk is pursuing. But, yes, it's very expensive to build these data centers also. So that's another factor. let's talk about the reaction to data centers so in memphis there's been some backlash to these facilities getting built are we expecting some of that in texas as well it's hard to say i mean texas is kind of the wild west for data centers right now it seems like a lot of companies are building there um so it's a little early to tell in memphis i think a lot of the backlash has been around their use of turbines gas turbines to power the data center and you know, we'll have to see if that's something they're going to pursue in Texas as well.

5:05Right. Well, it's certainly a story to watch, Grace. I want to thank you for coming on. That is Grace Kay, our Elon Musk reporter here at The Information. Oracle's data center build continues and The Information has new reporting on multi-billion dollar surprise costs that have come up through the process. I want to bring on Ann Davis Vaughn, author of The Information's AI infrastructure newsletter to walk us through what she found. Anne, welcome back to the show. It's great to have you here. Thanks, Akash. Great to be here. Okay, so let's talk about where Oracle is at with its data center build.

5:40You had some reporting on how that build specifically is going in New Mexico. What did you find? So, Akash, the New Mexico project is very important for Oracle. It's one of its big five Stargate projects. and the key question here is whether it's going to get power to this site where construction is already underway and a lot of money is already being spent and it has gone through some pretty extraordinary steps to try to get a clear path to power and all of a sudden it doesn't look like what it's done is going to be enough. So how does it plan or how is it trying to fix the power issue that? So it was planning to do an off-the-grid natural gas power plant.

6:27There were concerns both about the pollution from the natural gas power as well as greenhouse gas emissions, and it switched abruptly in April to a natural gas fuel cell option with Bloom Energy. It also took some extraordinary efforts to move a pipeline. But in very recent days, the environment department in New Mexico has said, we're not sure we're ready to give an air permit. We're going to push out some hearings on that. And the land commissioner in New Mexico has for a second time rejected a pipeline route. And this means that it could take longer to come up with solutions that could get this project up and running.

7:17What does all this mean for the costs of development then? I mean, this new approach that they've embarked on with Bloom Energy and then whatever they need to do to satisfy the local governments, I presume this means that costs are going up here? What's going on? So costs are going up. And basically what we're finding with all of these AI mega campuses is that what you thought it was going to cost is, you know, it just tends to go up. And it's more than you really bargained for to begin with. Fuel cells are a more expensive solution just on their face. It's probably, you know, in at least a billion, maybe more of expense compared with the old natural gas design.

8:02And more interestingly, fuel cells need to run continuously or they degrade. So it's not as easy to switch to cheap solar power, for example, on a day when it's sunny. And that is a better way to get your power. And so both the delays and the form of energy are a more expensive thing for bloom. And so is just waiting every week, every month to get power online to their site. And it is something that investors are watching carefully. So in your column, you talked about this idea of power flexibility being something that all these data center companies are working for. And so the way I understand it here is by using fuel cells, they basically have less of that flexibility because kind of like how we've talked about it.

8:58This is an analogy, so tell me if I'm getting it correctly, but we talk about routing to different AI models and which model is cheaper, depending on what the level of sunlight is or what the energy situation is on the ground. You can also route to different sources of energy to lower your costs, but by using fuel cells, you can't actually do that. Is that right? Yeah, you can add other forms of power to any site, but there is a cost with a site with fuel cells that the equipment would degrade if you turn it on and off frequently. And these off-grid power sites do have to take on a lot of their own equipment burden.

9:44They can't rely on the grid to be there for the little extra power and to provide that stability. They're having to do all of it themselves. They're pretty expensive engineered solutions. Right. Okay, so that's what's going on in New Mexico. You also wrote about some, I mean, we'll talk about Oracle and we'll get to Meta in a second. What's going on in Wisconsin? So Wisconsin is another big area of AI development, and Oracle has another very important project there in Port Washington, Wisconsin. The public regulators, the state regulators of the grid have been insistent that the expenses of building more energy generation as well as transmission need to fall on Oracle and its partners.

10:37and Oracle has gone to court in Wisconsin, challenging some requirements about how much collateral it has to post because of its credit rating. And it is just another example of more expenses coming on these projects than perhaps the hyperscalers anticipated. So you mentioned credit rating. Let's talk about these delays. Is it impacting Oracle's credit rating right now? How is that? So S &P downgraded Oracle's credit rating to a notch above junk earlier this month and is citing a lot of ongoing, you know, upfront costs that Oracle has to pay for while the reward of these data centers and the revenue that they'll generate is years away as they continue to build.

11:28It's just a very expensive endeavor. And when you are the company that has more financial strain than perhaps the big four hyperscalers, that does mean that your capital costs could be higher. And that is S &P's concern that Oracle has quite a lot of upfront investment. So are investors fretting then? I mean, a notch above junk seems pretty close to junk. In my limited understanding of credit ratings, it doesn't seem so good. Some of the investors I've spoken with don't expect that it would go down to junk, but we're all watching it carefully. And yes, I think investors are worried about capital costs as well as delays.

12:20You know, the fact that you, you know, have a campus but can't get power to it does, you know, trigger a lot of concerns. It doesn't mean that they won't figure it out. Oracle has been, you know, doing all kinds of, you know, in the minute pivots, you know, both with public opinion and trying to make their case directly, you know, to the people of New Mexico. as well as trying to change the design so that it would work. It's very important for them to get this back on track. Right. And I mean, we should say, I'm looking at the stock chart here. I mean, the stock is down 35 % this year. Last year alone, it's 47 % down.

13:09So, I mean, it seems like investors are certainly paying attention to these challenges that you've been writing about. I want to ask you one more question. We just talked to Grace about the pushback that SpaceX AI has been seeing for its data centers in Memphis. Now it's pursuing another opportunity in Texas. What does the pushback look like locally for Oracle in New Mexico and Wisconsin, where it's building its facilities right now? It's been an extraordinary kind of shift from, you know, just the focus on, you know, what is this project going to be? Did I have enough input into it to a lot of fear and concern about the impacts on land, on water, on air?

14:01And it's hard. There's misinformation out there, too. But what is happening in New Mexico, it's a water-scarce area. And people are concerned about how they're going to manage natural resources. Oracle's sought out their own water source and will treat their own water source. they're going directly to the public saying, look, you know, we're going to be responsible for paying, for putting in the infrastructure that we need. You know, tech companies are doing that around the country. They're trying to make that same case in Wisconsin, but somehow still, you know, push back on how much collateral they've got to post.

14:48So it's a delicate dance in each place and each location is different. Can I ask you one question, Anne? I mean, we're so focused on Oracle. You wrote about Meta as well in your newsletter. Are the challenges that Oracle is facing, is this an Oracle problem or is this a data center build wide problem? Are these similar problems that Meta is facing in its states, wherever it's building? You know, most of these issues go beyond Oracle. Meta has had some projects that are going pretty well. They've also got some questions that they're having to answer. We wrote in this piece about a project that Meta is doing in partnership with a local utility in El Paso, Texas, very close to the New Mexico project, but in Texas, where they've come up with what may be a less cost-intensive and less water-intensive way to bridge the power until the utility can get power to them.

15:49But at the same time, Meta is also juggling a story that has started to – I think you'll hear more about on how they handle – Is this the bacteriast? Are we getting to the bacteria story? Is that what it is? We are. Yes. So in Cheyenne, Wyoming, Meta has a data center. They were cleaning out some pipes that are supposed to be used for closed loop cooling. The local water utility has alleged that there is a bacteria that got into the water system, the wastewater system, not the drinking water system and has said that it's traced back to Meta. Meta has appealed a decision saying that it, you know, is the source, but there's some mysteries around how it all happened, but it has brought up a lot of questions about what was supposed to be, you know, a very big innovation where these AI data centers use less water, but recirculate fluid to cool their data centers.

16:59There could be more scrutiny around these closed loop cooling systems and whether the wastewater systems in the communities that they are in are prepared to handle some of the more, the buildup that comes from reusing these liquids again and again, it's something to watch. It could be an addressable problem, but as you know, it's the public perception that - Yeah, I mean, look, public perception, and as we just talked about with Grace, I mean, 122 days, I think she mentioned for one of the Colossus builds in Memphis. I mean, these projects are moving at extraordinary pace. And so... Yeah, but they're slowing down, Akash.

17:48They're slowing down. Yeah, well, that's why the reporting is so important to consider, and certainly the cost overruns. And I want to thank you for coming on. That is Anne Davis Vaughn, author of our AI Infrastructure Newsletter, here at The Information. Our next segment is with our partner, General Compute. We've covered a lot of neoclouds on this show. Most of them rely heavily on NVIDIA. General Compute is taking a different approach. I want to bring on the co-founder and CTO of the company, Jason Goodison, to talk about their strategy. Jason, welcome to the show. It's great to have you here.

18:25Great to be here. Thanks for having me. I want to start with the big news this morning. So Anthropic is buying two gigawatts of compute from AMD. AMD is investing$5 billion into Anthropic. What did you make of that deal? Yeah, honestly, I'd love to say this kind of blew the entire game out of the market. But the truth is, we've already started to see deals like this happen, right? So we have Meta that's doing a deal with AMD already. We've got OpenAI that's doing a deal with AMD and now kind of more of the same. I think the reason is because NVIDIA has such healthy margins that if you're doing AMD and you're kind of neck and neck on inference, maybe you're not as good at training because of the software ecosystem, but you're pretty much neck and neck on inference.

19:04while you can pencil some of these deals that are just more advantageous for Anthropic because they can save money, their total cost of ownership would be lower. So it seems obvious, like, why wouldn't they do a deal like this? And this is the new rack system that AMD has launched this week. But what's your sense on how that rack system, Helios, is going to compete with NVIDIA? Yeah, I mean, so we're doing alternatives to GPUs. So we're not going to probably buy either of them. But the truth is that these GPU systems are really good at TCO. And TCO is total cost of ownership. Essentially, how fast and how cheaply am I running my AI systems?

19:45And so the Verirubin and the Helio system are really good at running AI very slow. And so doing it at a cheaper price if you want to run tokens really slow. What we're focusing on as a company is buying alternatives to GPUs that optimize on speed. So what we're looking to do is target 1 ,000 to 2 ,000 tokens per second. And Helios and Verirubin are both fantastic machines, and they're going to do really well at helping people compete on price when they want to generate 40 to 50, maybe 150 tokens a second. But fundamentally, that's just not really what we're focused on in this business. So you're looking at alternatives to GPUs.

20:21So who are you working with and which chips are you going to put in your facilities? Yes, I mean, we've spoken to most of the, they call these ASICs, right? We've spoken to most of these ASIC providers. And the ones that we're the most excited about are Sambanova chips. So we already have a cluster live with Sambanova's last generation hardware. We already serve models like Minimax about five times faster than competitors. We're doing it about 500 tokens per second. But what we're really excited about is the SN50 chip that's coming out at the beginning of next year. We have a large cluster Q1 that's going to come out.

20:56And these chips are going to be 20 kilowatt rack draw. So we can basically put them into any existing data center across America. They're going to be able to operate at 1 to 2 ,000 tokens per second on large frontier models. We're going to be able to put a 10 trillion parameter model on a single rack. So really, when it comes to efficiency and speed, these things really just blow the rest of the competition out of the water. And we're really excited about it. Why did you pick Samba Nova? I mean, we've had Rodrigo on our show a couple times before. I know that their chip traditionally has been – it's been very focused on running open source models from what I recall.

21:34And open source, open weight models have obviously taken off, at least in the conversation, the last couple weeks. Was that a reason that you pursued the SN50 and Salmonova specifically? Were there other reasons that you picked them over some of the other chip companies? I mean, there's, you know, let's talk D matrix, there's X, there's a cerebris. I can never pronounce it properly. Why did you go with Sominova? Yeah, I think the answer is because we think that there's a weird combination here of once in a lifetime opportunity to do frontier level intelligence extremely fast. So when you think about, for example, the difference between dial up internet and broadband, I might be dating myself a little bit here.

22:19But, you know, sitting and waiting 30 seconds for a page to load would be considered unacceptable now. But, you know, sitting and waiting for 10, 20, maybe even an hour for a coding agent to run would seem par for the course. And that just doesn't seem like something that will happen in the future. So Cerebris, I think, is a really interesting machine. The engineers there are exceptional. In fact, what you'll find is every ASIC company has an exceptional engineering team because what they're doing is fundamentally just one of the hardest things in the world. And so Cerebris can run really fast, but we found that there's a bit of a limitation on the model size.

22:54And as you scale up to more and more and more of these wafers, the cost just goes through the roof. So if you did want to run a really large model, you'd be serving at just like a cost that you just could not afford to. And so there's a unique combination with Sambanova where you can run large models extremely fast and actually at a reasonable and competitive price. And so I think that's the sweet spot that we've been waiting for, for this breakout success of fast inference. And I think we're on the cusp of our broadband moment. People are talking about open-weight models as a way to lower their costs significantly.

23:29And I just want to make sure I got my understanding of this. Do open-weight models, do they use less compute? And also, does that inherently mean that maybe the market size for the chip business and cloud services is lower than we thought? No, not at all. There's a Jevons paradox to this, and I'll get back to the open weight model question, but what we found is that the cheaper the tokens have gotten over time, the more usage has actually happened. And so we're seeing as tokens are getting faster and cheaper to use, you're just using more of them. And you could consider, imagine using a coding agent where you're building a website.

24:06Why build one website at a time? Why not build 10 variants of it and pick the one you like the most? These are things that you're just never seeing right now, but I expect to see in the future. And so open weight models, they're really flexible, because you can kind of get in there and get dirty and understand how they work and just tune them for your own use cases. And so you're seeing people use them to get even better evals, evaluations, you know, on their their private data set. So let's say you're a company that's doing law AI, you want to be really good at that. So you make a test suite that's optimized on that.

24:39and then you use an open weight model. Instead of using one of Anthropic's models, you know, off the shelf, you get to tune something specifically for your use case. And the last thing to say on that too is, let's say you have proprietary data. You know, if you're going to use that with Anthropic, you're sending the data to Anthropic and, you know, you're kind of praying and hoping that Anthropic doesn't just, you know, kind of bake it into their next model because that's all of your IP gone. And so these open weight models are something people are using to keep their data private and also get better unit economics and just better AI in general that's optimized on their use cases.

25:14And that's how we're seeing a big shift to it. So I want to get back to your decision to not just use Samanova, but not use NVIDIA, because we've done some reporting here at The Information, that if you look at NVIDIA's share of the inference market, it's actually been going up according to our reporter's estimates and data sources that they've consulted. And so in a world where customers still want more NVIDIA for inference, why then do you still believe that it's wise to not go with an NVIDIA for your offer? I mean, it's a fantastic question. I think the answer is that we're in a compute dry market and NVIDIA has really robust supply chains and they've got basically a head start on everyone.

25:58So people need tokens and they produce tokens. And so it's pretty simple, you know, give me more tokens. but what I think is going to happen in the future is what we're doing is we're optimizing for where the puck is going right which is we want fast tokens and this market of fast tokens is just kind of it's nascent it's just starting to take off you've seen the cerebris IPO is market indication that people care about this stuff obviously open ai is running some of their latest models on cerebris so that market is just getting started and we really want to own that market so that's where we're going that's where the puck is going but AMD and Nvidia you know they're fighting it out here you're also seeing nvidia is theoretically you could say subsidizing demand to some degree by by doing some of these circular uh financing deals and even amd now that we're we're discussing what happened today you know this morning i i they invested five billion dollars into it dropping so you're seeing like you're seeing subsidization um yeah so that's part of the reason that's accelerating it and these asic providers none of them have the balance sheet to do that and why would they right they're still getting started working capital requirements for building ASIC is unbelievable.

27:03And so they're just getting started. And so what we're doing as a business is we're partnering with debt financers to help fund these ASIC chips. And we do not expect that the ASIC is going to underwrite the investments because they don't have the balance sheet for it. So that's particularly what's difficult about doing what we're doing. That's why we're so early in it, but that's why it's going to grow to where we're going, where the puck is going, essentially. And you don't think that the Verorubin will materially change that game or make people want NVIDIA even more based on what you're seeing?

27:37Well, I think, so obviously the Grok acquisition happened, right? So there's only a handful of pieces of silicon in the world that can do fast imprints. You've got Cerebris, you've got Grok, you've got Sambanova. The Grok acquisition was an interesting one. And we're starting to see, you know, there was some report from Semi-Analysis about how there's some delayed timelines on the Verirubin already for the Grok rack, the Kyber. And, you know, that was rejected by the company. So yet to be seen what actually happens there. But the truth is, and not to get too technical here, but the Grok chip is not really the optimal chip for how big models are getting these days.

28:13So they put every single thing on the chip. And the difficulty with what they're doing is they're doing what's called attention FFN disaggregation. I don't want to get too into the weeds, but the TLDR on it is that every time they produce a token on a Grok chip, they have to go back and forth to the Verirubin rack and back to the Grok chip. So think about it. If you have an 80 layer neural network, you're going to go back and forth 80 times, which is just insane. Whereas on the SN50 chip, what we're going to do is just kind of run the entire decode on this silicon all at the same time. Again, not to get too technical, but the point is that the chip is just not optimized for the use case.

28:50I think it's going to have a really hard time running large models in particular. and even if it was going to be optimal at it, we're not going to see it for a while. And so these SN50 racks we're getting are air-cooled. They're a 20-kilowatt draw. The Kyber racks that we're talking about are up to 600 kilowatts, which is insane. They also use... So basically what you're saying is the power that is going to be demanded by the Rubin family of chips, and the Grok chips, from what I hear you're saying, that's going to be significant. I figure that's going to be a problem that are challenged that NVIDIA will have to deal with.

29:26And that's one of the reasons why you've started to diversify away from it. Exactly. I mean, look, we've been talking about it on the show, and I think people are certainly looking for alternatives. And so I think it's certainly a smart strategy. Jason, I want to thank you for coming on and speaking with us. That is Jason Goodison, the co-founder and CTO of General Compute here on TI TV. Amazon-backed AI studio Fable is producing and distributing the first AI-generated visual feature film to hit U.S. theaters. The movie is called Gods Don't Give Gifts, and it is by YouTube creator Zach London.