The Information's TITV
00:16:31 · Akash Pasricha
[00:16:31] S0: But I guess my question is, all of the investments that Amazon is making into these data centers and this backend AI infrastructure that Meta is saying is making it easier for them to target customers, etcetera. I mean, I just feel like we should be seeing a little more from Amazon in advertising acceleration because of that. They didn't even really talk about it at all on the call.
00:17:00 · Ken Gawrelski
[00:17:00] S2: Well, I do think that they alluded to some of the advertiser innovations they have forthcoming. I do believe that they... So that there's a real difference here. You think about the innovation that Meta is seeing and their investments in GPUs and in in their internal models driving better performance. It's really two pieces. One is better targeting. Right? And that's better ad relevance and better targeting. That clearly Amazon is right in Amazon's wheelhouse. Amazon will do more of that. They do some of that now. We should expect they will be able to see those same kind of efficiencies and innovation. On the other hand, the other side of this is the audience side, which is meta is driving better... Driving better context or content recommendations. The idea is when you're on Instagram, instead of watching nine reels, you're watching the tenth. How how do they make that tenth reel so compelling to you that they increase your usage by 11% during that session? That... That's the part that Amazon does not play in. Right? Because it does have Prime Video, but that's a that's a different ballgame. What we're really talking about is advertising on the retail platform. So they can't really do much to increase their supply the same way Meta is on its content recommendations. What Amazon can do is really improve the ad relevance the
00:18:37 · Akash Pasricha
[00:18:37] S0: ad performance. Right. Right. I should correct myself. It's not that they didn't talk about the advertising. They did mention during the prepared remarks, Moor, it just wasn't as much of a focus as I thought. The focus though of much of the call was on AWS and on the capacity that they're building there. There's been a lot of focus this week on how much revenue TPUs are helping to generate for Google Cloud. Andy Jassy last night talked a lot about Trainium, how they're rolling it out. Do you have a sense for maybe how much that is contributing revenue wise to AWS, or if it is as important to AWS as TPUs are to Google Cloud?
00:19:27 · Ken Gawrelski
[00:19:27] S2: Sure. So, Trainium is important to Amazon and to AWS. However, it's much earlier in its development. You have Trainium three out there now with Trainium four that is likely forthcoming. TPUs are on version seven on... And and eight is forthcoming, I believe. So we... TPUs... Google TPUs are much further along in their development. There's a broader ecosystem, and Google has become much more aggressive in selling its TPUs externally. Right? And this is, you know, has just started really, we think in the the first sales were really came in in a small amount in 2Q, more so in 3Q, and it will grow quite substantially from here. What Amazon is talking about is by the end of twenty seven, they would expect to start to see some delivery of Trainium chips. That's external sales, and that will be more of a factor in '28. But that doesn't... That's not the whole story on Trainium. The story on Trainium is that much of the inference that happens over AWS today is served over Trainium. It is... There's a inefficiency to that for Amazon. It's a lower cost for them. And so it is very important to the AWS growth. It is one of the pieces of capacity for AWS that's critical. Again, we're in a world where today where demand exceeds supply for compute capacity, and Trainium is a critical piece of of Amazon's compute capacity in addition to partners with NVIDIA GPUs and and CPUs. Right.
00:21:14 · Akash Pasricha
[00:21:14] S0: I wanna ask you lastly about CapEx and how they're building for the long term here. I wondered on your end, when you're doing your modeling, how are you doing the math on whether or not the CapEx is reasonable, whether it's too much, too little? Just give us a view into your own modeling exercises and what inputs you're putting into that.
00:21:38 · Ken Gawrelski
[00:21:38] S2: Yeah, I think the way we look at this is the payback periods and the return profile. Right? And this is not a static thing. Right? Because we can, in a moment in time, we'll say, okay, here's what we think the return profile is for the cohort of investments of CapEx in '26, in 2026. But then as market conditions change, both the revenue that they can attain for that capacity, but also the cost. And this has been a big issue. Right? You've seen supply chain inflation, memory, all along the supply chain. And so the key is to continue to monitor this. This is an ongoing thing. What we see here is this. We've done this work. We we see high teens IRRs on on... And on compute capacity investment. We're talking about a three to four year payback period. That is longer than you would might hear out of a SpaceX, for example, where they're talking about one to two two year paybacks, maybe even shorter than that. But I think there's a difference here, which is Amazon and AWS is doing longer dated contracts, longer term agreements that are not right at the, you know, spot market price. So we... What we do is we track to make sure that over time, these return profiles are steady to improving. And what we see so far is, if anything, would say a bit of an improvement in return profile. Some of that I think is because you have... We talked about the training chips at Amazon. You talk about the TPUs at Google. I think these alternative sources of supply, proprietary sources of supply, are allowing some of the return profiles to improve for these hyperscalers.