The Information's TITV
00:37:43 · Stephanie Palazzolo, Amanda Peterson Corio
[00:37:43] S0: me. So, you know, looking back over the past two years, it feels like, you know, originally just getting the chips themselves was the biggest issue. Now to add on top of that, we have all these other bottlenecks like memory, energy, obviously, which you're very focused on. I mean, how much computing capacity today is being held up by power bottlenecks for Google specifically? It's [00:38:05] S7: a really interesting question. Think, today most people think about the hold up to AI being either capital or access to chips, as you mentioned. But there is an inherent disconnect between, we are now updating software in seconds. We are now updating our chips in six months to eighteen month cycles. But to build out large power infrastructure takes years. And so, really, when we work in our field and infrastructure, we're not thinking about next year or the year after that. That's already been written. That's all the work that we've done the years prior. We think in terms of '28, '29, '30, or even beyond. And so we're always building that infrastructure three years ahead or four years ahead to be able to meet, future demands or needs. There's plenty of power. It's just being able to thoughtfully design and structure in the right way to be able to get it built in time to meet the needs.
00:38:05 · Amanda Peterson Corio, Stephanie Palazzolo
[00:38:05] S7: a really interesting question. Think, today most people think about the hold up to AI being either capital or access to chips, as you mentioned. But there is an inherent disconnect between, we are now updating software in seconds. We are now updating our chips in six months to eighteen month cycles. But to build out large power infrastructure takes years. And so, really, when we work in our field and infrastructure, we're not thinking about next year or the year after that. That's already been written. That's all the work that we've done the years prior. We think in terms of '28, '29, '30, or even beyond. And so we're always building that infrastructure three years ahead or four years ahead to be able to meet, future demands or needs. There's plenty of power. It's just being able to thoughtfully design and structure in the right way to be able to get it built in time to meet the needs. [00:39:00] S0: And, know, as you're saying, the power that we have access to today, maybe that was decided two or three years ago, where I'm sure you, you know, Google out of anyone probably knew that the big AI boom was coming, but I'm sure even you didn't realize how incredibly huge this moment would, would, would, would be. I mean, when you think about how much power you have access to now and maybe the limits that places on compute capacity, are we talking... You know, is there more than a gigawatt that's being held up because you just don't have enough power? Is there a way to kind of quantify how much capacity is being held up at at the moment because of power?
00:39:33 · Amanda Peterson Corio, Stephanie Palazzolo
[00:39:33] S7: I... At at this moment, we're continuing to build out and, as you mentioned, having to be creative in how do we access more more capacity to meet ever growing demand. And that looks like building on new structures. So we think not only about how do we build and interconnect to the grid. You see a lot of what's called colocation of power, so building power faster, so you don't actually have to wait on transmission being built to to build it next to the data center. But there's also a lot of solutions that we're opening up to meet demand. We recently announced a partnership with Voltus where we're going to do over a gigawatt of what's called, virtual power plants or demand response, where the problem on on meeting power requirements isn't necessarily around the clock. It's just in a few hours of the year when everybody turns on their air conditioner or everybody's using the grid at the same time. So actually getting creative and working with partners to figure out how can we reduce loads on the grids just during a few hours of the year can actually open up that capacity so it's no longer a bottleneck for us on the grid. [00:40:38] S0: Mhmm. And for that deal specifically, that means paying these homes and businesses Right. To stop using electricity for some of these hours of super high demand.
00:40:38 · Stephanie Palazzolo, Amanda Peterson Corio
[00:40:38] S0: Mhmm. And for that deal specifically, that means paying these homes and businesses Right. To stop using electricity for some of these hours of super high demand. [00:40:47] S7: Yeah. So, it's us actually paying as another customer on the grid to say, hey. How can we have flexibility? It could look like demand response. It could look like switching to a battery at a house or a battery somewhere on the system, that makes more sense. So it's the... In aggregate, there's a lot of efficiencies that we can still realize on our grid to meet the capacity needs. [00:41:08] S0: Mhmm. And what are some other, you know, interesting new types of deal structures that you guys could con... You know, could consider moving forward, like the the Voltus deal that we were just talking about?
00:41:18 · Amanda Peterson Corio
[00:41:18] S7: Yeah. I think we're looking at, a lot of interesting deal structures. We think about, it's a... I've often said it's an all of the above solution. There's not one solution that makes sense or doesn't make sense. So some of those structures may look like, again, building a microgrid closer to the data center. So having, solar, wind, storage, grid connected assets, or geothermal, that has a more around the clock profile, that's that's one asset. And then thinking over time, how do we better integrate the full stack through our data centers to the power supply to optimize that more firmly? Other structures we're looking at, do look more like these virtual power plants where it's like, how can we actually still connect to the grid but actually give distributed batteries to the communities around us or put solar on roofs to help them reduce their energy burden to continue to, have affordable power, but also is a better system and grid resource on the grid. We've also built a lot of structures with our utility partners. So in a lot of states in The US, you can't just go build a power plant next to your data center. You actually have to work, through your regulated utility to buy power. And one of the challenges that that they're having is just the scale and the speed and the and the amount of capital required to meet our demands. So we're innovating with them to say, how can we, commit to today to pay, upfront and commit to paying over fifteen, twenty years, whatever the term is, no matter what. We may not draw an electron from the grid, but we will commit to paying the capital today to make sure the infrastructure is there in the system to be able to service. And that may sound obvious now, but it's a structure that wasn't in place before. And now through our work, working with these utility and the grids have figured out how to get that approved in over nine states in The US, and that's continuing to grow.
00:43:11 · Stephanie Palazzolo
[00:43:11] S0: Mhmm. And, you know, you mentioned geothermal, solar. I know that you recently, you know, announced this major nuclear agreement with Constellation. Talk to me about how Google's thinking about that, you know, source of energy moving moving forward.