We Need More Power - DTNS 5138  ## Metadata - Author: [[Daily Tech News Show]] - Full Title: We Need More Power - DTNS 5138 - Category: #podcasts - URL: https://share.snipd.com/episode/7ab1cecb-133f-4e12-81ef-c65a602ed071 ## Document Notes ## Highlights - **Power, Not Chips, Is The Bottleneck** - Companies racing to secure cloud compute face a bigger bottleneck: power and ready data-center shells, not chips. - Satya Nadella and Sam Altman said power availability and "warm shells" are today's primary constraint on scaling AI. Transcript: Tom Merritt Now that OpenAI and Microsoft have opened up their relationship, there's no more exclusive for cloud services between them. The race is on for companies to secure the computing resources they need to drive innovation and meet demand for all those generative model users that are coming on in. So let's run down the deals that were just made. First, OpenAI struck its first deal with AWS. Now, they had struck preliminary deals before they got the open permission from Microsoft, but they had not done one with AWS. So this seven-year deal is about 10% the size of those previous deals. It's a smaller one, relatively speaking, but it's the opening of a relationship. OpenAI has made bigger deals with Oracle and made a bigger deal with Microsoft. They're not ditching Microsoft. For the Amazon deal, OpenAI will use NVIDIA GPUs, but there is the possibility of adding Amazon's in-house Tranium CPUs to the deal as well. And the workloads will start running on AWS immediately, but of course, continue to expand as Amazon adds capacity. Microsoft has two deals to talk about. They'll spend $7.9 billion in the United Arab Emirates, that's where Abu Dhabi is, over the next four years to build out data centers there. Microsoft says it got a U.S. License to ship NVIDIA chips there for this purpose. That is a restricted area, but they got a license to ease that restriction for them. So they'll be able to sell not only all the NVIDIA advanced stuff, but the Blackwell chips, the GB300s. Microsoft said it will triple its use of NVIDIA chips in this country. This comes along with remarks from the U.S. President that the country's most advanced chips will be reserved for U.S. Companies, but Microsoft's a U.S. Company. They're just building something somewhere else. Microsoft also reached a deal with Australia's former crypto miner turned AI resource provider IREN. That's I-R They'll be able to access NVIDIA's GB300 chips as well to put in data centers in Childress, Texas. It's out in West Texas. IREN will contract with Dell to provide further equipment as it expands operations to meet the deal. So Microsoft building the stuff in and around Abu Dhabi and getting IREN to build stuff for it in Texas. This will, of course, add to the conversations about excess compute. Are companies building too much? Are we going to get all the use out of this, that all of this money and time spent building is hoping they will get? In an appearance on Brad Gerstner's YouTube podcast, both Microsoft CEO Satya Nadella and OpenAI CEO Sam Altman, who appeared together, by the way, to show that there's, you know, No bad blood between them. They said power, not compute, is the problem. They're not worried about the users. The users are going to be there. Will they be able to have enough power? Nadella said, in fact, that is my problem today, right? It's not a supply issue of chips. It's actually the fact that I don't have warm shells to plug into. Now by warm shells, he does not mean pasta. He means data centers with power that are ready to install the servers in, right? It's a shell of a data center and warm means it's got power. ([Time 0:02:46](https://share.snipd.com/snip/a1d1a7ae-1afc-4c57-b30f-a7935a9aa021)) - **AI Demand Could Accelerate Renewables** - Renewable energy deals are becoming cost-competitive and are being signed to supply data centers at scale. - Tom Merritt argues demand from AI could drive innovation and lower renewable prices further. Transcript: Tom Merritt Altman said, someday we will make an incredible consumer device that can run a GPT-5 or GPT-6 capable model completely locally at a low power draw. So it's worth noting that Meta signed three deals this week for power, a total of one gigawatts worth of solar power, two in Louisiana, one near Lubbock, Texas, also in West Texas. And South Korea will mandate that all parking lots, including existing ones, not just new construction, with more than 80 spaces must add solar canopies that will not only protect Your car, but it'll also generate energy. So Rob, let's talk about the energy use of this all. It does appear that most of the time when you hear about a deal for power for data centers, it is a renewable resource. And the word on the street is that buying wind and solar is actually cost competitive when you're talking about the other existing power resources out there. Robb Dunewood Yeah, so power apparently is a big deal here, a really, really big deal here. And I'm thinking to, you know, the deal between Microsoft and OpenAI. You know, one of the things that is kind of interesting is that, yeah, we're not necessarily cool, but it's okay. You go do your thing and get, you know, compute from other places. And it's looking like, well, because Microsoft may not be able to supply everything that OpenAI needs. It looks like OpenAI is growing. It just they're growing immensely fast. They're getting bigger and bigger and bigger. Microsoft is, too, but I don't know at the same clip. So does it make sense for OpenAI to go and use these other things when Microsoft is still getting, you know, clearly close to 30% of what OpenAI makes, it's probably not a bad deal for both Companies. Yeah, exactly. Tom Merritt It diversifies the resources for OpenAI. They're not reliant on one company. Not that Microsoft's going anywhere anytime soon, right? But, you know, it's just always good to have lots of different resources. That's why I think that AWS deal is even though it's smaller, is that it's something they can build on. And if they start to find out, wait, AWS works really well for us, they can expand that deal fast. But again, I think there's a lot of concern. In fact, I saw it on our subreddit where people are saying, how are you going to power all this? There's the assumption that it's going to be polluting resources. It doesn't have to be I know it's a lot of the times easier for it to be, but we shouldn't discount the scenario that what this does is drives down the price of renewable resources. One of the problems with the price of renewable resources was you just didn't have enough people wanting to get into buying it. The demand wasn't there. It was one of those chicken and egg problems. Well, now you got the demand and you're seeing companies like Meta saying, yeah, we'll sign three deals in a week for a gigawatt of solar. ([Time 0:06:07](https://share.snipd.com/snip/09c36481-9fc7-4b6c-87c7-dd40a9d8347d)) - **Data Centers Follow Renewables, Not Old Grids** - Building data centers near wind and solar assets avoids an aging grid and secures reliable power. - Robb Dunewood explains hyperscale sites favor wide open areas where renewables are abundant. Transcript: Robb Dunewood Makes sense. And one of the reasons why these big data centers are interested in renewables outside of the fact that it's better for the planet and they're playing the long game is like we would like To actually have power where water won't be eventually because of global warming. There's that part of it. But the other part of it is that we have a very, very antiquated and old grid in this country. So it's oftentimes just easier to build the data center near where the renewable power is. So like wind, solar, they tend to be in big wide spaces, a lot of open land. You can build ginormous buildings and put computers in there, not necessarily next to the coal plant that has been there for 70, 80, 100 years. Tom Merritt The Louisiana deals, the power from the solar grids is going to go straight to the data centers. But in the Texas deal, it goes into the grid and then the data centers are just pulling it out of the grid. It's one of those situations. So they're putting in as much as they're pulling out in that case, at least theoretically, according to the deal. Right. But they're not drawing it directly from the solar panels when you think that that would be the place they would. All that said, you're right. ([Time 0:09:13](https://share.snipd.com/snip/57c46e4d-68ab-4e44-9a7b-bc4674507e68)) - **Choose The Right Power Sourcing Model** - Match your power procurement model to how you locate data centers: direct PPA or grid-backed sourcing each has trade-offs. - Consider whether power will feed a data center directly or enter the grid first when planning capacity. Transcript: Tom Merritt The Louisiana deals, the power from the solar grids is going to go straight to the data centers. But in the Texas deal, it goes into the grid and then the data centers are just pulling it out of the grid. It's one of those situations. So they're putting in as much as they're pulling out in that case, at least theoretically, according to the deal. Right. But they're not drawing it directly from the solar panels when you think that that would be the place they would. All that said, you're right. These companies like the press hit of saying, hey, we're doing a renewable thing, but they are doing it more often because it is cheaper and it is flexible. ([Time 0:10:05](https://share.snipd.com/snip/517ac0ad-607f-4b34-822b-c372ac8caf57)) - **Apple Using Gemini But Running It Locally** - Apple plans to use Google's Gemini to power a new Siri, running a custom Gemini model on Apple servers. - Tom Merritt highlights the marketing challenge of using Google tech while assuring user privacy. Transcript: Robb Dunewood Every Sunday, Mark Gurman posts his Power On newsletter on Bloomberg, generating that week's reliable Apple leaks. A lot of it this week is about Apple celebrating its 50th anniversary on April 1st. However, the remark getting the most attention is Apple is betting heavily on the new Siri, which will lean on Google's Gemini model and introduce features like AI-powered web search. Gurman has reported on this before, but repeated that while Anthropic offered a better model, Apple chose Google because of the price. He also repeated that Gemini-based custom model will run on Apple's private cloud server. So I'm wondering here, is this deal between Google and Apple, is this the Google search deal 2.0 for those two companies yeah it does feel like that right uh makes sense for google to say Tom Merritt Like hey you know that that search deal is going to get less and and less important for both of us uh yeah and we don't want you to ditch us for somebody else uh so let's provide gemini that Can power your search eventually and uh and power siri as well. Uh, I have to say Gemini is very good. Uh, it has improved tremendously. I, I lean on it, uh, quite a bit and I, I find it to, to be on par anyway. None of these models are perfect. They all, they all have their quirks and their, their weaknesses, but it it's decent. And I think this is where my head was at Rob is that, uh is going to need to buy while it builds, which it does all the time. It used to buy cloud services. It used to buy maps. It's been a buyer of services a lot. It's always bought search, even though it is toyed with the idea of creating search. And it can do both well. It can build things like chips and eventually use those or it can just buy things like search and integrate them in a way that's seamless the other interesting thing is that this will Robb Dunewood Run in apple data centers not it's not just running on google stuff they're actually going to put this model inside of apple's data center and run on their hardware which is probably Going to mean a lot for security for Apple. I know they're very, very security conscious, probably the most out of the two big phone manufacturing operating systems. But that's a that's a good thing for Apple. Tom Merritt If this continues to be the case with German stuff, I always believe that he's right now. It doesn't mean that they will continue to do that. They could change their mind between now and next year. But assuming this continues to be the case, it'll be an interesting marketing challenge for Apple to acknowledge that they have a very good model powering Siri without sounding like Google is going to get all your information, right? I mean, from what it says, it's going to be on Apple servers. It's going to be a customized model. It's going to be able to look into your device for you, but keep it private. But if people hear Google in this, if they hear Gemini and think Google, it doesn't matter what Apple says, they're going to think, oh, you're sharing my data with Google. ([Time 0:14:10](https://share.snipd.com/snip/b76618e2-981a-4c7f-bb7d-77e352595323)) - **On-Device/On-Prem Hosting Reduces Perception Risk** - Apple will run the customized Gemini model inside its data centers to address security and privacy concerns. - Robb Dunewood says hosting on Apple hardware helps reassure users despite Google's involvement. Transcript: Tom Merritt If this continues to be the case with German stuff, I always believe that he's right now. It doesn't mean that they will continue to do that. They could change their mind between now and next year. But assuming this continues to be the case, it'll be an interesting marketing challenge for Apple to acknowledge that they have a very good model powering Siri without sounding like Google is going to get all your information, right? I mean, from what it says, it's going to be on Apple servers. It's going to be a customized model. It's going to be able to look into your device for you, but keep it private. But if people hear Google in this, if they hear Gemini and think Google, it doesn't matter what Apple says, they're going to think, oh, you're sharing my data with Google. Absolutely. Robb Dunewood I think perception is they're just trying to make sure, hey, this is on our stuff. Settle down. Yeah. ([Time 0:16:20](https://share.snipd.com/snip/f9389a5f-a1a1-413c-a713-60c690fb59f4)) - **South Korea Builds AI Film Production Stack** - South Korea is building AI-focused production infrastructure and power to create film and TV with generative models. - The joint venture pledges fully licensed models and datasets to address IP concerns. Transcript: Tom Merritt Now, this one I think is just notable. Companies in the U.S. Are going slow on generative models in filmmaking, not because the company's making the generative models want them to, but because filmmakers are very cautious. There's a lot of upset about copyright provenance, intellectual property, replacement of actors, replacement of production staff. So a joint venture in South Korea is creating infrastructure, both the data centers and the power supply to our earlier conversation, for producing movies and TV shows with those models. So, you know, while the feet drag in the United States, you're seeing South Korea jump to the front of the line, at least they'd like to. This joint venture is between a production company called Utopai Studios, spelled U-T and an investment firm called Stock Farm Road that happens to have the grandson of the founder Of LG on it. The venture will co-produce film and TV projects and expand access to Korean intellectual property. The first production is expected to be released next year, and they say every model and every data set is fully licensed and contractually approved. ([Time 0:20:39](https://share.snipd.com/snip/6b8172c6-0312-46ae-89be-2f547142f264)) - **Global Adoption Could Outpace Hollywood** - If U.S. studios hesitate, international players will adopt model-driven production and gain competitive advantage. - Robb Dunewood warns adoption abroad could pressure Hollywood to follow or lose ground. Transcript: Robb Dunewood Going to be interesting to see what the industry from a U.S. Standpoint looks like, because if you look at what SOAR 2 is doing, if you look at what Google is doing with Gemini and the video that you can create, the realism that you get out of this, There's definitely cause for concern if you are an actor or around that industry. So just because we're not necessarily moving full steam ahead here in the United States doesn't mean Earth is going to adhere to that. So this is going to be very interesting to me to see how we actually go forward with AI in this space. Tom Merritt If there is a way to do this cheaper, at quality or better, it will be done. And I think that's why this story is interesting, because South Korea is a growing content powerhouse, winning Oscars, winning Grammys. Well, they haven't won a Grammy yet, but there's a lot of momentum behind South Korean TV shows and movies. And that is happening not just in Asia, but in the United States as well. And so a company that is willing to work with its domestic industry and take advantage of the momentum behind that is going to push Hollywood. It's going to say like, hey, if you're not going to do it, we will. ([Time 0:21:55](https://share.snipd.com/snip/3c27581c-33a0-4ff1-ad31-4430109b972b)) - **Listener Railroad Analogy Sparks GPU Lifecycle Debate** - Listener Andrew compared data centers to railroads but noted GPUs become obsolete fast. - Tom Merritt and Robb Dunewood discussed how warm shells remain useful while GPUs cycle rapidly. Transcript: Tom Merritt Yeah, Andrew writes, long time listener, love your work. So what are you preparing me for, Andrew? But thank you. That's a nice way to start the email. Very quick note on your railroad analogy for data center infrastructure. It works, except that railroads remain useful for many decades with minimal maintenance, while GPUs will be obsolete in five. Data center investors don't have much time to find ROI. Keep up the great work. Robb Dunewood That's actually a really good point. Data centers, the technology that goes into today will not be there 20 years from today, let alone 100 years from today. Tom Merritt Yeah, yeah. No, that is a very good point. It is different. But let's let's walk through the implications of that. The warm shells that they build continue to be useful, right? As long as you got power and a place to put the racks, you can swap out those GPUs. So if you need more warm shells, building more warm shells is still good. And it's not a bad thing for the GPU makers that you have to swap out the GPUs regularly, right? That's gonna be a good thing. ([Time 0:26:52](https://share.snipd.com/snip/de5c0828-ac72-4435-9279-f6577dd0dbb3))