저자: DavidCarbutt · 외국인
요약 보기
젠슨 황(엔비디아)과 일론 머스크의 협업을 다룬 대담으로, xAI 멤피스 콜로서스 데이터센터의 빠른 건설, 기가와트급 전력 수요 계산, 우주 태양광·데이터센터 구상이 주요 내용이다. 우주 데이터센터는 36개월 내 가장 저렴한 AI 배치 장소가 될 것이라는 주장과 엔비디아 칩 중심의 수익성 논리가 제시됐다. 에너지 증설이 GDP 성장과 연결된다는 주장도 포함된다.
다듬은 글
He says he wants to have 10 gigawatts by the end of 2027. That's hundreds of billions, right, spent on Vera Rubin. This is an unprecedented scale-up. You were with him again yesterday when he was also presented with a medal. Is Elon much better at building computing power today than he was then? Is this obvious to you now?
Is it a foregone conclusion that SpaceX will become the world's largest AI hyperscaler? Well, we we we used to, you know, in two parts of our chart, Elon was in the " enterprises" section. That's how we reflected it in our metrics. Now it is part of our CSP. Of course. Yes, yes. And that's why They are huge.
So I'm glad that, that, that, um, I like working with Elon. Um, he uses our technology for his own AI, for training models and stuff, and of course, for deploying models. Grokbot—it's really cool, and the demand for Grokbot, the consumption of Grokbot is really growing rapidly. And, of course, he rents it out.
Yes. And so the combination of those three things makes it pretty unique in that regard. Hmm, I guess Google does the same thing, right? This would be another, yet another example. And so I think he, um, also has the advantage of loving the intersection between computing and physical building. This is true.
You know, he loves this whole field, and no one knows energy better than him. First, tell us a little about X, their ambitions and what they have achieved. First of all, achievements should be recognized where they are deserved. From the moment of concept to the data center ready for Nvidia to put their equipment there, to the moment we turned it on, connected everything, and it ran its first training.
Yes. OK? Right. So, that first part is just building a huge liquid- cooled factory, Mhm. connecting to the grid, getting permits—in such a short time, it's, well, it's like a superhuman effort. Yes. Yeah, and as far as I know, there's only one person in the world who could do that, you know? I mean, Elon is unique in his understanding of engineering, construction, large systems, and also in resource mobilization.
Incredible. Yes, it's just unbelievable. And the amount of technology, if you saw it, just doesn't fit in your head. All this wiring and networking, and you know, Nvidia hardware networks are very different from hyperscale data center networks. Yes. Of course? The number of wires that go into one node, the back of the computer— it's solid wires.
And just integrating this mountain of technology and all the software is incredible. For comparison: 100,000 GPUs is, you know, definitely the fastest supercomputer on the planet as a single cluster. Um, the supercomputer you would build would normally take 3 years to plan, yeah. and then the equipment is delivered, and it takes another year for everything to work.
Yes. We are talking about 19 days. The feeling that they will have a very big impact on the world in the field of energy generation. They are already doing it. Um, the amount of miracles the XAI team had to pull to connect a gigawatt of power was insane. I had to put together a bunch of turbines, and then we had some permitting issues in Tennessee, and we had to cross the border into Mississippi, which is fortunately only a few miles from here.
So, we still had to stretch the high-voltage lines for several miles. And build a power plant in Mississippi. And it was very difficult to build. People don't understand how much electricity is actually needed at the generation level to power a data center. Because they look— the news looks—at the energy consumption of, say, GB300, multiply that and think that's the amount of energy needed.
What about cooling and everything else? Wake up. Yes. This is a complete noob, someone who has never dealt with iron in their life. In addition to the GB300, you need to power all network equipment. There's a lot more to it than just processors and storage. You must anticipate peak cooling needs. This means whether you can cool the system even during the worst hour of the hottest day of the year.
Isn't that right? It's damn hot in Memphis . So you would need about a 40% increase in energy just for cooling. Assuming you don't want your data center to shut down on hot days. And you want to continue working. Then you have to say: well, there's another multiplicative element on top of that, which is : do you assume that you never have any power outages?
Like, oh, well, actually, sometimes we have to turn off the generators, some of the capacity, to do maintenance. Oh, of course. Now you add another 20-25% multiplier to that. Because you have to assume that you will have to turn off the power for maintenance. So, in fact, for every approximately 110,000 GB200, including networking, CPU, storage, cooling, and power reserve, approximately 300 megawatts are required.
Sorry, please repeat that again . That's roughly, or think of it this way: to service 330,000 GB200s, including all the associated support, networks and everything else, peak cooling and some power reserve, you need about one gigawatt at the generation level. You praised the achievements of Elon and xAI in Memphis during the construction of the Colossus supercomputer.
Apparently, in record time—just 4 months. It now has 200,000 GPUs and is growing very quickly. Is there anything about his approach that you could share that would be instructive for all data center creators and what made it possible to achieve such a result? His approach to engineering, his approach to all of construction management, to everything.
First, Elon dives deep into many different topics, but he is also a great systems thinker. Mhm. Therefore, he is able to think across multiple disciplines. And he, of course, pushes things, questions everything. Is this, first of all , necessary? Secondly, is it necessary to do it this way? And, you know, thirdly, does it have to last this long?
So, he has the ability to question everything to the point that everything is reduced to the minimum necessary. You can't remove anything from there anymore. Yet the essential capabilities of the product are preserved, so it is as minimalist as it is possible to imagine, and it does so on a systemic scale.
Um, I think I also like the fact that he, um, he's represented, he, he's present at the scene. Mhm. You know, he'll just go there, and if there's a problem, he'll just go there and show me the problem. You know , when you do all of this in combination, you overcome a lot of the previous "we just do it this way" stuff .
Mhm. Um, you know, I, I'm waiting for them. Well, you know, I mean, everyone just has a bunch of excuses. And so, and then, and finally— when you personally act with such urgency, it forces everyone else to act with urgency. You know, and each supplier has many customers. Each supplier has many projects in the works.
And he does, he did, he makes it his business to be the top priority in everyone else's projects, you know, and he does it by demonstrating Speed—that's important. So, um, I'm really going to do what, um, I, I, I, I mean, I just keep taking on the limiting factor . Whatever the limiting factor of speed, I'm going to , I'm going to, I'm going to take on it.
So, um, yeah, if capital is a limiting factor, then I, I'll solve the capital issue. If that's not a limiting factor , I'll decide on something else. The reason Tesla created a chip development team, and the important thing is, well, the vast majority, Tesla—it's like a dozen startups in one. Um, and you know, they, we actually only made one big acquisition, SolarCity, and then a few very small ones.
So all of this, almost all of this is organic. So, uh, I built a chip development team from scratch and, you know, an AI team from scratch. Um, it was just because it became a limiting factor. So, um, for the hardware for the second version, we used Nvidia, but Nvidia at the time was focused on making truly server- side AI hardware, which was obviously a smart bet.
They are the most valuable, currently the most valuable company in the world. And, you know, Jensen Huang and his team have done an incredible job at Nvidia. I take my hat off to them. Um, you know, I'm a huge fan of Jensen and Nvidia. They did, they did an amazing job. But they, they didn't want to make a low-cost, energy-efficient car computer at that time, a car AI computer.
So I thought, "Well, maybe we need to create a chip development team to solve this." So, then I hired Jim Keller, and we created a chip development team. I think Jensen made 5 trillion on his own, actually. Excuse me? I think Jensen made 5 trillion single-handedly. I think it was around 3:30 or something.
Let's try, let's try to be conservative . Elon, how do we win? How do we continue to win? Well, we need a lot of power, a lot of energy. So, I believe that scaling up energy is one of the most important tasks. You have to scale energy , you have to scale chip production. I think we will win in software, in everything intelligent or digital.
But in the long term, we need to generate enough energy to compete with China, which I believe is our main competitor. And China has about three times the electricity production of the United States. And we'll figure out how to make chips. So the real major challenge, I think, in the long term is going to be how to generate enough energy and how to ensure that there is enough capacity to manufacture logic chips and memory in the United States or in a secure region.
Well, you We need the Wright Brothers to get to work. Yes. Well, it turns out the Wright brothers are already working on it. Elon, if you deploy 10 gigawatts, which SpaceX has publicly stated, that would be about 25% of all the energy added in America. And I know you have real rocket engineers working on energy supply.
Yes. Help me create turbines. Can you tell me about it? Yeah, I mean, just to outline this: the average energy consumption in the United States is about 500 gigawatts. So every 5 gigawatts of additional sustainable generation is a 1% increase in energy use . And I think that actually materializes in about 1% GDP growth.
Because intelligence per watt is constantly increasing. So if you just plot a graph, how does useful intelligence per watt increase? As on the hardware level, Jensen GPUs are getting better in terms of compute per watt. Also, algorithms are getting better at getting more intelligence per watt. So I think that's a pretty interesting observation.
I'm willing to bet anyone that a 1% increase in energy consumption corresponds to about a 1% increase in GDP. And so, 10 10 gigawatts would mean a 2% increase in GDP. Um, I think with rubrics. Em. Is this like, is this the plan to send this into space after we cover Nevada with solar panels? I think it would be pretty hard to cover all of Nevada with solar panels.
You need to get permissions there, like permissions for Try to get permission for that. So space is really daring. This is actually a regulatory game. It's kind of harder to build on earth than in space. This is harder to scale on the ground than it is to scale in space. But you also get about five times the efficiency of solar panels in space compared to on the ground.
And you don't need batteries . Um, I almost wore my other shirt that says it's always sunny in space. Which is true. So, um, because you don't have day and night or seasonality or clouds or atmosphere in space. Because we lose about 30% of energy through the atmosphere alone. So any solar panel will produce about five times more energy in space than on Earth.
And you avoid spending on batteries to get through the night. Um, so it's actually a lot cheaper to do it in space. And I predict that the cheapest place to put AI will be space in 36 months or less. Maybe in 30 months. 36 months? Less than 36 months. Um, and from a SpaceX perspective, we're looking at the long term : how do we access significantly more energy than we can get on the ground?
And this is what space provides. So orbital calculations will be a very important thing. Um, actually, Google agrees with that too . They just sent or are about to send some TPUs into orbit. I think there was an H100 there, which was launched into orbit last year, I think. Mhm. Um, so it works in space.
And SpaceX, together with Tesla, aims to produce 200 gigawatts of solar energy per year. And if you transfer this to space , you get the nominal power or even more. Whereas on earth you'd only get a fifth to an eighth of the solar power generation, and you'd need huge batteries. So it's essentially always sunny in space.
So, uh ... And now is a great time to do it, because the economic returns on these computing systems are so high that the launch cost, while still quite significant, becomes quite acceptable in the context of the overall economic benefit . Certainly. This is the perfect moment for this. Using Nvidia computing systems in space is the best way to do this.
Yes, climbing the Kardashov scale using American energy and American chips on American rockets. I like it. Yes. Are you increasingly confident that we will have significant computing power in space, say, with Starship by 2028-2029? I mean, Elon put out big numbers, but do you believe we can achieve serious amounts of data output in space in this timeframe ?
Only with Nvidia calculations. And let me explain why. It's not just about technology . Nvidia's performance per unit of energy is the best in the world. So we call it performance per watt. But the most important thing is this . Nvidia's computing performance is so high that it means an excellent return on invested capital.
Yes. So you can put them in space and the launch cost would be only a small percentage of that. OK. The startup cost is quite significant. So you won't be bringing cheap computing systems there. You will be bringing high-value computing systems there. And we are the most expensive computing system imaginable, the most powerful system in the history of computing.
So we are ideally suited to launch into space right now, regardless of the cost of the launch, it will still be a great level of return on investment. So, one of the new markets that we are going to explore together is the placement of data centers in space. So our two engineering teams are working very closely to create a new type of server, a new type of architecture, that can be extremely efficient both in space and on Earth.
Comments 0