The Exchanges

Q You think, so you think that this level of build out is healthy?

A I, I think we definitely are not nearly at the point at which there is too much AI infrastructure, given I think the number of AI workloads that are coming for the world. Um, and, um, you know, I think we're seeing on a, over the last couple of years as people use AI, they get value, they use it more, the models get better and people then use it even more for new use cases. And, and I think at this point across the industry, With AI, we're still more supply constrained than we are demand constrained. And I think, um, I, you know, I expect that to continue over the next couple of years as the technology continues to evolve and as people start to integrate AI into more and more workflows.

AI assessment note: “definitely are not nearly at the point at which there is too much AI infrastructure”

Q the AI world? I mean, like we talked about previously, GPUs were designed for gaming. They happened to do parallel processing. Actually ended up being really good for, uh, you know, large language models, the training and the, and the inference. Um, what's your perspective on, on whether this industry is going to continue to run on that type of chip, uh, and what the potential is for custom silicon?

A I think, um, a couple of things. I think one is, um, I think the increasing the, um, The, the number of tokens you can get per watt per dollar is going to be the game over the next couple years and, and maximizing, um, the ability of our cloud to deliver the best, uh, Volume of tokens for every watt of power, for every dollar that's spent, where the dollar is spent on energy, it's spent on the GPUs, it's spent on the data center infrastructure, it's spent on the network, and it's spent on everything else, is, is the thing that, again, we're laser focused on, and, uh, it is, you know, there's a bunch of steps as part of that, GPUs being a critical component of it. Um, And, you know, one of the things that our scale gives us the ability to do is to invest for, uh, kind of nonlinear improvements in that type of productivity and that type of yield. You know, if you've got, you know, A million dollars of revenue on a couple hundred GPUs. You're not going to be investing in custom silicon. Um, when you're at our scale, you will be. Um, and you're not just investing in custom skill again for GPUs for pre-training or for inferencing. You're, you're looking at what can we be doing for synthetic data generation with silicon? What can we be doing from a compression perspective with custom silicon? What can we be doing from a security perspective? And we have bets across all of those, many…

AI assessment note: “every GPU server that we're running in the fleet right now, um, is using custom silicon”

Q working life of a GPU and how long until they burn out? They are their use cases for GPUs once they are no longer top of the market, right? We hear often about, okay, well, Unlike the, the, um, the laying of the fiber, the GPU depreciates after a couple years. So I think this is a pretty important question. Can you, can you handle that, tackle that for us?

A Yeah, I think kind of going back to the comments we had earlier on balance, I think as you think about your GPU build out, one of the things that we think about is the lifetime of the GPU and how we use it. I think, you know, what you use it for in year one or two might be very different than how you use it in year three, four, and five or six. Um, and so, you know, I think that is something where, um, You know, so far we've always been able to use our GPUs, even ones that we deployed multiple years ago for different use cases and get positive ROI from it. And that's why our depreciation cycle for GPUs is what it is. Um, but I do think that's, you know, as we build out, um, our infrastructure, we are definitely consciously thinking about that because, uh, you don't want to have your entire fleet in two years suddenly have to be replaced because, um, you know, that That, that would be expensive. And so, you know, we are very thoughtful on that. Um, and again, that's, I think I talked earlier about, um, different training. I also think even as you think about training, we, we often in the past used to monolithically call training, training. There's lots and lots of different training use cases. Now there's pre-training, there's synthetic data generation that goes into training. There's post-training with RL and fine tuning and other different techniques. And, um, You know, having…

AI assessment note: “what you use it for in year one or two might be very different”

Q do you feel the, um, I don't know the right word to put it, the pressure of competing with China? Um, because from my understanding, China has a much looser regulatory approval process, and they're just stacking, you know, data centers. They have abundant electricity, uh, in the United States in particular. Um, I imagine Europe is the same way. Uh, that, that is not the case. What's it like?

A Oh, it's certainly, I think the world has a very different regulatory approval process. I mean, I think one thing that when I talk to people and they say, how can you build data centers faster? Um, you know, there's obviously things that we can do from a technology and are doing from a technology and from a manufacturing perspective, but, you know, candidly here in the U S the longest Part of building a data center is getting permitting. It's not actually the construction. It's, it's making sure that you, you know, uh, get permitting approval for all the steps that you want to take. And, you know, different states and different parts of the country have different regulatory environments. And I think even if you look at a sort of a heat map, if you will, of where data centers are being built in the U S you definitely see pockets. And I would say some of that approximates to where there is land and where there's power. And some of it really, you know, closely correlates with where it is easier or faster to kind of complete the permitting process. Um, you know, in Wisconsin, we had a, you know, a phenomenal partnership with the governor and the local county. Um, we were able to, to, uh, purchase some land and power that, um, a manufacturer was previously going to use and they, they pulled out of a project. And so, you know, I think that the, The, the local communities recognized i…

AI assessment note: “candidly here in the U S the longest Part of building a data center is getting permitting”

Q is the consensus leader in the space. They needed more infrastructure. There must have been some Calculation within your group or your company to say, uh, is it worth it for us to be the one that goes out and builds this massive, massive footprint, you know, in partnership with them or somebody else? So I definitely understand there's, there's multiple stakeholders, but what made Microsoft pause on that front?

A Well, we, we have a balanced view. And so it's, we, and we take it with a, uh, a long-term view in terms of making sure that we're building out in, All the locations that we want to build out that were being, um, uh, Thoughtful in terms of kind of the investment spend and the infrastructure that we're building and also recognize that we don't have to do it all. And so I think we're always trying to kind of take a continually balanced view of that. And, and as you've seen from our capex, and as you've seen from our earnings calls, we are investing a lot in infrastructure and building out like crazy. Um, but again, at the same time, you know, we're, we're always constantly reevaluating and watching closely, you know, which data centers in which markets. To what specifications and making sure that, that we keep, you know, good discipline as we're doing it, um, that optimizes for both the longterm, near-term and midterm, uh, horizons.

AI assessment note: “we have a balanced view... and also recognize that we don't have to do it all”