← Latent Space AI
Loading summary
Transcript1 lines
- [00:00]
A
Welcome to the podcast. SpaceX has just signed a $6.3 billion compute deal with the Open Source Lab Reflection AI. Nvidia's new Ruben AI server runs at 45 degrees Celsius for the coolant, and that basically kills data center's water use. Something I am super excited about. Apple is weaving AI into iOS 27 apps and they're basically betting against the chatbot framing that mostly everyone else has decided is going to be the winner. Eco Wave is pow. A company called Eco Wave Power is going to be using Nvidia AI to run data centers on ocean waves. And AI has out persuaded human debaters in a study done by Oxford and aisi. If you want to get all of these stories into your inbox every single morning, I have a newsletter on AI chat daily.com that's my website that kind of goes along with this podcast. If you go over to the subscribe tab on that button and I'll leave a link in the show notes, you can, you can join the newsletter and get all of these stories. You get a deep dive and a link to the full article about all of these. So it's kind of like the show notes for the podcast. If you want more information, a deep dive or in your newsletter, you know, maybe you don't have time to listen to the full podcast, but you want to get a rundown of each of the top five stories of the day. Go subscribe to the newsletter and I'll send that to you. The link is aichatdaily.com subscribe if you want the newsletter. All right, let's get into this first story with SpaceX. They just signed a $6.3 billion compute deal with reflection AI and I haven't talked about reflection AI in quite a long time. This is one that I did a number of podcasts ago when it was kind of announced and launched. We haven't seen a lot coming out of it, but it's, you know, evidently they have a lot going on. If they're signing $6 billion compute deals, this one in particularly is going to give them access. They're an open source lab, but they're going to get access to Nvidia GB300 chips at the Colossus 2 data center. This is the one near Memphis, and they get that through 2029. So it is the third major company that is going to be on site. And Anthropic and Google have both signed deals as well. And I think this shows that there is a massive growing investment in a lot of these Open source AI development companies. I'm really excited about this one in particular. Reflection is going to pay about $150 million a month starting in July of this year. That's less than anthropic who's paying $1.25 billion a month. And Google is paying $920 million a month on the same, at the same site. But I mean again, this is just another one of these mega customers that SpaceX has been able to pull into, you know, paying for this data center that they've built. If you don't remember the last episode I dropped on this Reflection was founded in 2024. It was actually two former Google DeepMind researchers. One thing on these deals that were seen at this data center that I will note is that all of them have a 90 day cancellation clause after the first three months. So I mean, I'm not saying SpaceX is trying to juice the numbers, but you could imagine like let's say, you know, reflections like, hey, look, I need like three months of training compute done at, you know, at your Data Center. And SpaceX is like, sweet, why don't you sign a, you know, a three year contract with us? You can cancel after three months if you want, but like we want to make sure we got a three year contract. Okay. I don't think this is what they're doing, but I think this is what some people's criticism has been on the, on these particular deals because anthropic, Google, DeepMind and reflection all have that now. I mean, when you're talking about these kind of deals paying 1.25 billion a month or 920 million a month, those are massive sums of money. And so it kind of makes sense that, you know, if a company couldn't keep paying that they, they might need a way out. But I don't think Anthropic and Google are going to be the ones that are dropping the ball on those payments. Reflection, I'll be excited to see how they grow, but that is still a ton of money. $150 million a month. The reason why I think this matters is because SpaceX basically has accidentally become the world's biggest compute landlord for all of the top labs. They rent out all this infrastructure and originally Xai, I think built all of those to train their Groq model, but they keep just building bigger and bigger models and then they kind of had SpaceX's backing to help them roll out Colossus 2, their brand new data center. And I think they have 2.3 billion in monthly revenue flowing through Colossus 2 right now. And actually I think they have an even bigger one that they're working on now. So I think this is interesting because their model isn't really just grok, they're kind of XAI models. It's really the hardware that there's that they got for that model and then they're now renting it out to some of the other hyperscalers. Okay. Nvidia had a big conference and they unveiled a whole bunch of stuff. I have two of the big stories that I want to cover on the podcast today that I'm most excited about, but there's a ton of co Nvidia news if you care to look it up. You can go check out more of that on AI chat daily.com. i have like 10 different articles. But one of the big things I'm excited about is that Nvidia's Reuben AI servers, they're running on 45 degrees Celsius coolant. And that basically kills the data center water usage issue that we've had that a lot of people talk about. Right. There's a lot of negative PR about data centers. And maybe you're in that camp. I'm not really sure. I don't get all of the controversy about data centers. I think that they're fantastic. I mean, when I see, you know, a Farmer got offered $27 million to sell their farm for a data center. I know maybe people think that this is like, this is terrible and they'll hate me for saying it, but like, I think they probably should have sold their farm for the data center and they could go buy another farm down the road if they want. But that's a lot of money, especially if they're getting paid above market prices. If I were them, I'd take the deal and go buy another farm. But anyways, I'm sure that is deeply upsetting to a lot of people. And by the way, I love farms. I hope to have a farm and my own, you know, hundred acres someday. But anyways, whatever, you can't win them all. Okay. What's cool that's happening here though is one of the big negative PR things that people hate about these data centers is they're like, look, they use up so much water. I've talked about this in the past on my podcast where I said, look, all data centers don't use up water. They have closed loop systems. And I kind of explained that. And there's someone that left a comment. It's probably on this podcast if you Go look at the reviews. And it was like a one star review. And it's like, if Jaden can be so wrong about AI and, you know, data centers and water usage. I've worked in data centers for my whole career. If he can be so wrong about that, he's probably wrong about everything else. Okay, I probably didn't explain it well enough in that podcast. There are some data centers that do not use closed loop systems so the water can evaporate, but there are other data centers that do use closed loop systems and they keep the water in. And in those instances, you have a very low percentage of the water that's actually being lost to evaporation. Also, I mean, maybe if you're in a drought area like Arizona, which I come from, I maybe get that concept. But if you're somewhere where there's a ton of water and a ton of humidity and it rains a lot, like, I'm currently living in North Carolina and that's the case. I don't know. Water just, you can get water from the air, you can get water from the ocean. I lived on a boat for a whole year and we just pumped salt water and desalinated and got fresh water. So, like, I hate to feel like water is a very scarce resource, but whatever. That's probably another thing that triggers a lot of people. In any case, we're getting away from all of these triggering facts that are all of these triggering topics that I cover because we don't need it anymore. Nvidia's Ruben AI servers are going to come with their own coolant, and it brings the, the water that you need to cool these data centers down to a ground total of zero. You do not need water to cool these. A single 50 megawatt facility can save over $4 million a year on cooling energy and water costs. And they cut their water consumption to basically zero. Like it, it's, it's all the way down. So this is really exciting to me. So how does this actually work? So the coolant enters the chip at up to 45 degrees Celsius, which is hotter than a hot tub, and then it's going to leave at about 55 degrees Celsius, which basically allows Nvidia to get rid of all of their mechanical chillers and all of their fans. The conventional data center consumes about 2.6 million gallons of water per megawatt per year. And, and so with Ruben's closed loop design, that's going to cut it to zero in most climates. The other thing I'm really Excited about is that the rack density is going to jump up a ton. So the servers that used to be able to, you know, have six rack units in air cooled designs, now they can fit in two with liquid cooling systems. So they can actually get these, you know, these chips and the GPUs and like everything, the servers, they can put them all closer together. You know, before they had to give them more space to keep apart for airflow and stuff. But they can fit more inside. So what's interesting to me is we're actually gonna need to build less data centers and we'll be able to have more compute inside of them so they'll be more optimized, which is great because there's a huge percentage of all data centers today. I think it's something like 2 out of 6 are getting canceled because there's, you know, all of these big kind of these big programs pushing against data centers. And I know a lot of people have a lot of different opinions on this. I saw Kevin o', Leary, he is funding a lot of different data centers, I think in Alberta and also in Canada and also in the United States. And he was saying that based off of, you know, what, what they're seeing, there's a lot of groups tied to the Chinese Communist Party that are funding organizations that are kind of protesting against data centers in the United States. And his theory and hypothesis on why that is is that China would like to have superiority in AI if they can build more data centers and if they have more compute, they can build more powerful AI models. And potentially if we get into a big enough compute crunch, we would have to outsource our computing to China to have them do stuff for us and they kind of get our access to our data and our models and stuff like that. So anyways, it's definitely a big huge geopolitical battle and I think the United States needs to build more data centers. It's interesting. Even like recently I was on Instagram and I clicked on some clip and it was this guy and honestly, super talented musician, he wrote this whole song about the farmer lady, the viral clip where she like refused to sell her, her farm to the data center people for, you know, millions of dollars. And he wrote this whole song about how awesome she is. And all the comments on it were, I mean, you know, there's lots of people that are like supportive of it. I was like clicking around on some of the comments and it was really interesting to me because there's actually these anti data center like movements in all states and Some of them I'd click on and they'd be like a profile with no posts. The profile picture is just like, you know, a graphic of a person. And all that there's on the profile is like a link in the description to. The one I saw particularly was like, conserve Ohio. And I was like, what's this? And it's like you click on it. It's just like this anti data center in Ohio movement. It's like, sign our petition to, you know, block all data centers from ever being built in Ohio. And like, the graphic and logo for it is like this beautiful, like garden with these flowers and it says conserve Ohio. And it's so nice. Anyways, I just think this whole thing is quite ridiculous. When we have these closed water systems that we're building and I understand people's energy worries, like the energy costs will go up, but the Trump administration said, hey, look, if you're building a data center, you should be building your own power, you know, generation on site, whether that's solar, gas or wind. I think that's a fantastic proposal and I think that's what we should be like. If we're going to try to legislate anything, legislate towards that because the water problem is going away. If we can solve the energy problem. It's just ridiculous to me. I don't see these huge protests when someone's like, hey, let's build a Walmart. It's like, everyone's not like, no, stop the giant store that we go buy all of our stuff from. And even if people do protest a Walmart, as soon as it's built, all the same people are going there and buying stuff from it. So I don't know, it's like we want to live in a modern world of modern conveniences and technology, but like, we're fighting against the, the things that need to happen. So anyways, whatever. I know we're getting into this big, huge thing and like, I just think a lot of this is probably funded by geopolitical adversaries who don't want us to have the technology build out in our country. And even if you get like, you know, data centers, like New York, I think banned data centers or whatever. Like, even if Ohio is like, hey, look, you know, conserve Ohio, we banned all the data centers, okay, whatever, they're just going to get built in Tennessee or whatever the next state is over that doesn't have the bans and. And then that state is just going to get all the financial benefit of having billions of dollars spent there the electricians and the people that have to maintain these. So I think people are just kind of shooting themselves in the foot. But you know, that's my opinion. All right, let's get into what is happening with Apple. They're embedding AI directly into iOS 27, all of the core apps. So that's messages, mail, calendar, shortcuts. Rather than just funneling everything through Siri, they're actually adding a lot of the features into those core apps. And I think features that are shipping in the developer beta this fall is include things like bill splitting from receipt photos. Right. So you're at the restaurant, you take a picture with just your iPhone and it's going to be able to split your bill for you. They have agentic password rotation across websites. Honestly, this is pretty cool and I don't know why we didn't do this sooner. Maybe not, right? I guess passwords are kind of the bane of my existence for most websites. When I go and I have a password manager that remembers all my passwords that I've had since I was like 14 years old. But yeah, half the time I go to a website it's like, hey, you should change your password because it was found in a breach. And I'm like, oh crap. And like no one really wants to have to go and change your password for every single website you're on. I mean you can't even really remember passwords anymore. You just have to use random strings that are auto generated because it's the only thing that's strong enough. So anyways, and then you need a password manager to do it and I have to pay for my password manager. So anyways, it sounds like Apple's trying to solve this and they just have agentic password rotation. So it just always going to change your password and I guess it will tie to your face. My only concern is what if you lose your phone and then you don't know your password to anything? What if you, you know, something happens and the, the face recognition isn't working? Which, I mean, I've had a glitch out once or twice. It's not very common, but you can imagine there's some, some moments where this would not be great. But overall excited about a lot of the features Apple's added in there and the password rotation would be quite useful. Okay, let's get back to Nvidia for a second. There's a new company called Ecowave Power. They are running a data center at the port of Los Angeles and it is powered entirely by ocean waves There is no grid connection to this data center, which is pretty wild. Nvidia's AI basically predicts incoming wave patterns and then it schedules compute workloads to match peak energy generation, which basically turns what waves, like a very renewable source of energy and, and this AI into something that can be scheduled. What's cool here is I think we've all seen. In fact, I literally was just. I was on LinkedIn and you know how they have like the, the TikTok reels they're adding to LinkedIn now. I was just seeing one. I didn't click on it, but it was like this person, and I already knew what he was going to say because there's like the, the, the graphic above, it's basically like one of those giant floating docks and there's a wave that hits the floating dock and it ripples. And you know, his thing was like, this is the future of energy, whatever. I've been seeing that, that like, video in particular and that kind of concept for many years. It's funny that I actually just saw that this morning because in his, you know, basically it's a new way to generate electricity with wave power, and it's super cool. The problem is I hate all of the. The reason I didn't click on it is because I hate all of these things where they're like, look at this amazing way to create free electricity or clean electricity or this is the future of electricity. And then like, nothing ever happens. Nobody ever scales it. Nothing ever happens. And so anyways, it just makes me feel disenfranchised. I don't want to watch all these cool tech videos when it feels like nobody's doing anything about it. I want to see people actually doing something about it. So that is why I'm excited about what's going on here. Nvidia is helping support this company, Eco Wave Power, that is actually doing something about this. They're actually taking the power of these waves and using them to power data centers, which is awesome. Apparently wave energy is 800 times denser than air, which is letting floaters generate comparable power to wind turbines. And they're using a fraction of the footprint. And by the way, like, I don't know if you guys have looked into wind turbines. Those things are insane pieces of infrastructure. The amount of concrete and rebar for the base of those alone, that gets put in, it's like, it's like insanity. I mean, I've even heard some people say, and like, not to be a hater, but like, the amount of emissions off of like the concrete and the steel and everything that goes into building a wind turbine and the amount of energy to create one, it like doesn't actually really offset it. So it's more like a tax thing. I don't know if that's true. It's pretty insane what goes into those. And I think waves are perhaps a much better solution. The footprint is much smaller and it is 800 times denser the energy that comes out of it than air. Just waves are so powerful. If you've ever been hit by a wave, you know how powerful they are. If you've ever been hit by the wind, I mean, in a hurricane it's powerful, but that's about it. Okay, so the U.S. energy Information Administration estimates that wave energy could supply over 60% of annual U.S. electricity consumption. But it is basically the most untapped. So wave energy could actually be a huge solution for us. So how is Nvidia helping with all of this? Well, they have something called the Nvidia Omniverse Digital Twin. And basically this is like a simulation of the real world that they've created. Like this real world simulation, which is wild. And what they're doing in this particular case is they're simulating wave conditions and floater stress before they're actually deployed. So while live AI models can forecast sea conditions, they can then cue energy intensive training jobs into the high output windows. Right? So they're like, look, there's a huge storm coming or a bunch of really big waves coming. Let's get all of our floaters like queued up and ready. And we have this, you know, this like high output training job we're going to do. We're going to run it exactly when the waves are coming. So they like predict when the waves are coming and they kind of forecast and schedule everything to happen and to be trained at that time. That is pretty, pretty cool if I'm being honest. Okay. There is a new study that came out from Oxford and asi and in it it says that frontier AI models now persuade better than expert human debaters and professional fundraisers. They did debaters and fundraisers, which is interesting. And they did this across about 19,000 real conversations. So what their findings showed was that AI raised fundamentals three times more donations for charity and it shifted policy opinions more effectively than elite debaters. But what's interesting about this particular study is they said it wasn't just because of like its rhetorical brilliance. Like these models are so good at debating. They said the biggest edge actually comes from speed and throughput so they use Claude Opus 4.1 and 4.6. Those were the two best, according to this study, followed by GPT4O and GPT 5.44. They said Gemini 2.5 Pro and Grok 4.2 were the stragglers on the study. And by the way, they had 6,900 participants in this. When it was forced to match human writing speed and message length, then the AI's advantage over kind of these kind of coached elite debaters dropped from plus 4.1 percentage points to statistically insignificant to zero. But I mean, I guess what that does show is like if they're forcing to. If they're forced to talk at the same speed, speed and the same message length as a regular person, it's basically the same as a regular person, it's just as persuasive, it's just as good at getting those donations. But if it can go at its own speed and its own length, then it is 4.1% better at closing the deals, getting donations. The AI exceeded professional canvassers by 10.8 percentage points in actual donation size. So both of the donation rates and the average gift amount for Save the Children, they did a whole thing where they had it doing, you know, being a canvasser for Save the Children, and it was able to raise 10 percentage points in actual donation size. So more donations, which is actually pretty crazy. The study basically, I think, is flipping a lot of the narratives we hear. AI doesn't win debates because it's like it has like, superior reasoning, it's like way smarter than humans or just we can't even wrap our minds around its arguments. It's really winning because it has information density and because it can, it can give that information very quickly. So if you throttle its bandwidth, then a human could match it. But it's, you know, showing that this, the, the speed of thinking is a big advantage, which is so fascinating. Guys, thank you so much for listening to the podcast. If you want to get access to the top 80 different AI models all in one place for 8.99amonth, and if you want to get an MCP to put all of those AI models into ChatGPT, Claude or Gemini. So, like, if you use Claude and you're like, gee, I wish it had image generation or audio or video generation, I'd love for you to check out my startup AI box. AI, we have an mcp. So basically you give this MCP link into the connectors in Claude and it can generate image, audio, video, or any other tools that you've created on AI Box you can call them and it can port them all straight into Claude and generate all of that context and content that Claude can't normally generate inside of it. So go check it out. It is AI Box AI mcp. There's a little page I've built there that describes how to do this. It is super fascinating and super useful. I hope this really helps your workflows and everything you're building with AI. Go check it out. There's a link in the description and I will catch you in the next episode.