SemiAnalysis's Doug O'Laughlin: Amazon's execution edge is why OpenAI chose AWS, space data centers are hype

faster. And we've been keeping Doug waiting, but let's bring him in to the TBPN Ultradome. Doug, how are you doing? Whoa, look at that.

Outer fantastic.

Galactic Doug,

I'm so glad was delivered. It looks fantastic on you.

It it dude, it's great. I've been telling we're going to make some analysis like we're going to try to do this merch. It's so good.

I got it yesterday. I actually didn't realize I was on today. I just wore it today for work. I'm like, I'm not I'm not joking. And then I was like, "Oh my god, my schedule, my calendar. I'm doing this." I was like, "Dude, this is perfect."

Yeah. I mean, we were we were talking uh and and Tyler was like, "Oh, did you book Doug specifically because you knew that Amazon was going to do this deal with OpenAI. He's the perfect person to talk to."

We booked Doug. We love Doug. I said, "No, they did the deal because they knew Doug was going on TBPN on Wednesday, December 17th. They were like, "We got to give him something to talk about. Let's do a deal." And so Sam and Andy, they got together and they were like, "This is good content for TBPN and semi- analysis." Anyway, uh take us through your reaction to the deal. And if you could actually just start with like setting the table on where are we on the Traium narrative? Is it good? Is it bad? Is it like what do we know? What how is it changing? What's the updated thinking around just Amazon's chip efforts?

So, I think it's pretty good. Uh we wrote an article about Trrenium 3. We think it's going to be a lot better than Trrenium 2, mostly because there's a lot of interaction and help with uh from the the labs themselves. And let's not forget um Oh my god, it's beautiful. Let's not forget uh one of my favorite parts of this whole story is everyone's like, "Okay, we wrote about the TPU article and people hated it because we said, "Oh, it's a it's a Reuben clickbait." Where did all the TPU engineers go? A lot of them went to Tranium. So,

this is too this brain rot's too Nvidia focused. I botched it.

We wanted to give you some brain rot. We wanted to put the brain rot directly [clears throat] in the show because we've been we've been enjoying the brain rot edits and we thought what if [laughter] we did it live but my prompting was maybe below below standard but I like the is it a cake or not.

That's pretty good. I honestly forgot about the is it a cake.

Anyway, sorry. Uh just resetting on on training. You wrote the article. Obviously there was like a bunch of back and forth but uh but but how are you thinking about it currently? So look, I think uh we actually in in the semi analysis uh premium, you know, product that I managed, we definitely we kind of made a a call that we thought there would be some kind of announcement open eye um at reinvent mostly because and like first principles here's what it is. Uh data center execution is starting to become a real issue. Oracle's delayed, fairwater's delayed, corores and core is delayed. And you know who is not delayed? Um Amazon. Amazon has like their execution is flawless when it comes to data center and power and they're I think they're going to put like 56 gigawatts online next year. It's a big number like a huge number. Wow.

Um

and so OpenAI is always trying to secure more compute at at every single point in time and who has the most power available to them and has a project uh that they you know obviously are very interested in getting more customers for. I think OpenAI's power constraint is going to like that's one of the reasons why the deal and then also you know OpenAI is always down for fresh money at this point in time. They're like uh they're hitting up Disney, they're hitting up Soft Bank, everyone who has money,

they're trying to get money investment from. So that's kind of like a you know a win-win for them.

Okay. So the question that I was kicking around was there's a there's there's this $10 billion investment into OpenAI from Amazon. Simultaneously OpenAI has around 38 37 billion of commitments with AWS over a series of years. And my my question is like how how do we see how do you think about the way the way OpenAI will be working with uh with AWS and uh and Tranium chips broadly? Is it something where they can basically make all of their models multi-platform and run them and kind of create one fluid compute uh resource, a pool of compute that that can fluidly shift across tranium and GPU from Nvidia or are they going to sort of take the Tranium chips and have a specific model like maybe they'll do their video generation over there or their image generation or they have old 40 workloads that are sticking around longer than they expected. Let's replplatform that model to Tranium and have it run really cheaply on Tranium or something like that. Like is it a per per use uh partnership or is it something that cuts across the entire org? Do you have any visibility into that?

Did we lose audio?

I'm not hearing I'm not hearing Doug.

Lost him.

Uh oh. How about now? How about now?

We're good. Okay, we're back. or so back um uh I I don't think we have perfect visibility first and foremost but you can look at someone who's already doing that which is anthropic uses TPUs they they use GPUs and they use trainium and so there's a certain level that you can abstract it at and I think TPUs for example is trying to do it on torch TPU right so at PyTorch and I I think um having the multiple compute like elastic pools of compute is good for negotiating power and that's probably in OpenAI's interest but It's like a pain in the ass to get this to work. So, there's like a two two brains. Um, one, if you're really focused, you just buy more of one and be the best at it. Um, two, uh, at the same time, you get to pay less gross margins if you can kind of diversify your suppliers instead of just one. Um, it's possible. Anthropic has shown you that they people can do it. They already have. And that's probably what we Yeah. So, that's kind of my brain. I I think it really comes down to the power constraint, dude.

Mhm.

That's what Amazon has that no one else has. And that's like that's the art of the deal if that makes sense. And also

is that is that a refer is that uh like does that tell me that Amazon was more excited about the AI buildout a year or two ago? They didn't pause when Sachi Nadella went through that pause. Or is it just Amazon's been building sort of linearly growing growing growing for a decade two decades and so this is just more of the same what we should have predicted from Amazon. Do you have do you have an idea of like is this them catching up? Is this them just maintaining their level of excitement?

Well, I mean, if you look at Oracle's like backlog enhancement, they were basically saying we're going to build AWS in like a couple years. Meanwhile, AWS has, you know, multi-deade head start and it's like, yeah, we're also

that like imagine they're thinking like we're going to we're going to build that capacity too and

you're welcome to like, you know, have your deal with Oracle, but uh it makes sense that Amazon can kind of like beat them to the to the punch.

Yeah, I think that that's uh I think what Jordy said is probably more down the more realistic what's happened. So like last year we had the pause from Microsoft um obviously it's a big deal but even before the pause u Microsoft at the infrastructure level was never as sophisticated from Azure versus AWS. AWS um is the OG all-in-house did everything themselves and has like a very reactive and like battle tested infrastructure and like hey they're the largest logistics company in the United States. They're one of the largest logistic companies in the entire world. before all this AI stuff, they had the biggest amount of compute and the biggest uh you know the biggest infrastructure shell and all like the whole pipeline and platform and all of this. And I think that that's the like pretty much they finally got everything turned in the right direction with a focus on more AI and then you're seeing the the results of that about a year and a half later like right but you're you're while Microsoft pulled back they pushed ahead and then now all of their long-term investments are going to start to bear fruit meaningfully in 26 and 27. So that's the I think that that's the simplistic way to think about it. Um when I work when I talk to and I work with people who have worked with let's say and this is like from the before times you know before GPUs uh the difference between AWS and Azure is like you know junior varsity varsity like they're a totally different league and I think that that's uh that's that and and you don't bet against AWS's infra uh infrastructure like execution there. We've seen almost like like Oracle is a perfect example dude. It's been delay after delay after delay. Like these timelines are starting to be pushed out and

and we think that that's going to become like you know more than just like trying to find power. It's just like converting that power into a powered shell seems to have a real execution risk and I think Amazon is more money good from that perspective.

How did you process the AWS uh like direct pipe to GCP news? It seems like they're maybe trying to get into more like cross data center training. I know Google uh has some experience there is AWS. Is that is that something that the the big labs are demanding at this point? Like because they're also multicloud like are there any special um offerings other than or is it really just like hey we just have capacity and no one else does and that's enough.

I mean I think I think the labs probably demand it. Um, and historically the the biggest rake that everyone got paid or like had to pay in these in the multi cloud world was egress and ingress. Um, I don't like this is like I don't know you remember maybe 2019 everyone was like oh these freaking egress fees are like too high and that's kind of um I think

that's like kind of how what they did is they forced all the data to stay inside their data centers and they charged you like crazy to open the door. I think um I think that's just kind of breaking because they're they're seeing that like the customers absolute size is so large and they need so much demand that you have to pay. You know, how do you win these customers? Well, you need to build a pipeline, a door to your your data center, especially if your data center isn't where they started at and they're, you know, their data lives in Azure and not in AWS. Uh right, you want to have a door to uh AWS because up until now it was everything lived in AWS. So, I think that's kind of the story there. Um yeah, I think I definitely think the multi- data center uh multi-data center training and inference probably like to a certain extent is like maybe not inference but like that's part of the story too. I also think um maybe even rolling back to the previous question in terms of like there was a paper or article where maybe it was like a conversation that Enthropic had where there are different models that have different TCOs and different values at um on different hardware. So like certain types of hardware I think is much more profitable on certain types of memory bound and or like flops bound and so I do think having all those different options is going to have um a better co cost per use per model per hardware and so that's I think that that's going to be a like strategically valuable thing going forward in terms of like lowering the cost AI.

Do you ac uh expect to see any announcements on the commerce side between open AI and Amazon? Uh we were talking earlier uh it feels like this could be a scenario where like Disney invested in OpenAI and in exchange uh as part of that Disney is giving them a one-year exclusive on all that IP which is like an insane I think underrated advantage with Amazon. You can imagine chatbt openai has been working since inception to like be able to monetize uh some of the product uh like some of the purchasing that they are driving across the internet and could be but at the same time Amazon is a $60 billion like ad business and they want to protect that and that's based on people landing on Amazon and searching for products there. My impulse is no. No way, man. That's like the, you know, that's like the unassalable. That's the unassalable. So, yeah. It's the golden goose. Why would you let Why would you let the fox into the into the [laughter] house? Like, just no way. No way.

Yeah. So, so the weaker weaker companies have the Etsies of the world have said like, "Yeah, we we'll do it because we

they more to win."

Yeah.

Your background makes it look like you're in space. Are you developing a thesis around data centers in space or is that coming? I know you probably don't want to leak too much alpha on this particular show, but how are you thinking about it? What's the timeline? Is there going to be a space denator data center model for sale anytime soon?

So, you're talking I'm from the space data center team, research team at SM analysis right now on the ISS. Yeah. Um, and and and I'm [laughter]

Santa Tracker online.

Santa Tracker online. Um, and uh I'm going to be honest with you, I'm a space data center hater. I'm a huge space data I'm a I'm a giant hater. I'm sorry. I know. I know Gavin talked about it. Gavin's really intelligent and like a very wellspoken smart individual in the space. But like guys, it is hard to train a model on Earth today. [laughter]

It's hard enough to train a model.

It's hard. It's hard to build a box and put some chips in it and plug it in and on Earth.

Yeah.

Yeah. It's It's Dude, we're telling me that we have power delays on Earth.

Yes.

What is going to happen? Like we you know,

it is hard to square those things. Um and so and so and so hater hater

I mean the push back there would obviously be like there's a lot of energy in space solar energy but but again it it uh I just felt like there was like a very organized uh effectively narrative pump

tied to this uh

I I mean do you want to know why? Come on. I can tell you the answer. I could tell you the answer. It's cuz space uh SpaceX is racing. Yeah,

you got to remember what bagger what bag is being pumped at any given time. And the fact that you heard the SpaceX round come out maybe a few weeks later is not um it's that like

it wasn't it was within the same week. It was the same week that the $800 billion round like details came out. I could see them raising more now because they've announced the like target 1 a.5 trillion. So like that that just creates space for like the preIPO round which could land at easily a trillion or north of a trillion now that all these people realize like hey it's going to go out I'm going to have like a a shorter liquidity timeline. Uh and so yeah, I would expect, you know, more

I I think that's part of it, but also uh you know, if we're talking about data centers in space, there's only one service provider, right? The whole So like, you know, one of the reasons why I'm like a hater of data centers in space is like, hey, um like, you know, a 3,200 lb or whatever a one ton GB200 on Earth costs a lot of money. Uh 10x the cost to get it into space, right? Um but if a space data center was to ever happen, right? uh there's only one company that has really lowered the cost of moving something from on Earth into space. And so if that was to even be part of the TAM, uh all of that TAM would belong to SpaceX. So that's my that's that's that's my belief at least is that like when when these like large narratives come around a very large funding round, there's no it it isn't a coincidence. Maybe that's what they were talking about in the SpaceX round and that's where you're starting to hear all this stuff. Hey, this is something in their long-term planning. And then, you know, investors who are very excited about it talk about it. That's very plausible.

And if if there are data centers in space, I promise you SpaceX will be doing it. Let's put it that way.

However you feel about it, I promise you SpaceX will be doing it. So,

it's sort of the same thing as like the Mars narrative, which is still years and years away and and you know, maybe decades away. Uh, but you still have this like call option on it because if we get to Mars, it's probably gonna be on space.

Something I've been thinking about is I mean it seems like I mean so far Paramount's uh Sky Dance effort to acquire Warner Brothers seems to be falling apart a little bit. Warner uh Warner Brothers board doesn't want to do it. Uh they're happy with Netflix. It seems like Netflix very clearly is like good for the deal and there's less certainty on the on the Paramount side or at least that's how Warner Brothers board is positioning it. And I have to imagine that there's a handful of people that are looking at that money that was like softcircled by the Gulf and they're like I want I want those tens of billions of dollars. So like if you're SpaceX or you're open AI

and you're like seeing an opportunity of like, hey, this this kind of money was softcircled for a deal that might not be going through.

Everybody should be going after that.

Yeah, you should be raising. Always be closing, bro.

Always be.

Yeah. Always be closing.

I wanted to ask you about storytelling, but in a very specific context. Obviously, everybody on the timeline was talking about storytelling this week. if you should hire a storyteller. Uh, but I wanted to ask you about Meta, specifically their storytelling because they came out this year with this story around personal super intelligence. And then there was some reporting recently, I forget by who that was saying like some of the execs apparently were just like, hey, we should just work on ads basically, like let's just make the core business better. And I feel like I I personally as a as a user in the meta ecosystem, I have no idea what like personal super intelligence like really means, right? Is it is it a better ver like are you trying to deliver a better version of chat GBT and get into the search business as a as a meta shareholder? I'm like I don't know. I also don't I also don't totally know what it means. And so my question is like do you do you think that Meta needs to kind of dial that in or or do we just let them cook and and we can form an opinion once they uh once they ship?

Look, so okay, I I think forming an opinion when they ship is like the most reasonable way to live your life instead of speculating. But I'm going to put my my Zuck hat on. Okay, Zuck hat on. Um let's think about how I would do this. Um, three billion people use our products every single day and it's an important part for of your day-to-day thing. I would argue uh the chat GPT universe is actually like don't even include us uh include the United States actually we are high RPO users who like pay for like max or pro or whatever and like can can advertise know how to use these things. Um the adoption in the United States is pretty high. The long tale of meta users is like pretty astounding. um you know in in some places in Southeast Asia like hey Meta or like WhatsApp which is owned by Meta is like the primary mean for people to do business on right for people to talk to other businesses. Hey um all of a sudden like I'm sure have you ever done like a random like tour in like Southeast Asia or like Japan or like wherever the hell you going um and you like you know you sign up for it and they they send all these things there's clearly like a whole business flow behind it. Um, WhatsApp actually does have a ginormous, you know, billion user moat that I think can like kind of, you know, kind of become the super app. That's like if I

Yeah, that that but the other thing they have to be feeling pretty confident with the experience with threads where they have effectively ported their user base to an entirely new app like it's I I believe they could at some point once they're confident in the product experience get the Meta AI app to half a billion users in the way that they I don't think threads is at that yet but it's at hundreds of millions of of active users and that that distribution advantage does give them the potential to like come from behind, especially in some of these uh uh sort of more international markets.

I mean, have you guys ever seen the like the really sloppy um I I think they're apploving app advertisements actually for the games. It's like the guy and he's shooting a gun and you press one. It's like the the pure brain rot video game ad that like gets you to click through. Um, Meta Meta has the ability, in my opinion, to serve that kind of crap in your feed that is so gamified it makes you press a button. Um, when I am when I am doom scrolling on reals or something like that, occasionally there will be a thread where they have like such a jabated ridiculous title where I'm like, I just have to see that. And of course, it ends like the word before the hook that you want to know the answer. And so, you click it and then you go through and then like that's how you can like juice arpoo. So, yeah, you're I think you're right. No one else has quite the on-ramp, I think, in terms of engaging active users ASAP if they wanted to.

But I think you have to have like, you know, you have to have the Galaxy brain addictive perfect use case done first. Cuz I think they have the on-ramp and they can spend enough GPUs to tell you here's how you convert a random user to try our experience for 30 seconds. And we could probably mechanically show it to enough people to get, you know, to to easily boot up 10, 20, 30 million people to to try the app very quickly. But the real execution I think it's going to be on them is to get those 10, 20, 30 million people to become addicted users who share it with other people and make that, you know, personalized super intelligence part of your day-to-day. Um, maybe

remember to the WhatsApp point, they did boot chatvt out of like the WhatsApp ecosystem like not too long ago, remember? which had like a ton of users. I think it was I I it was at least

How did they boot this out?

This was like I think a couple months ago.

That's actually a really interesting

I might have to follow up on that idea actually. That's pretty It's pretty interesting. It might kind of uh

That tells you that tells you Yeah. I mean it tells you like how how they'll integrate their own models.

Yeah.

Uh anyways, we're we're way over. I wish we had more time. But let's do it again.

Thank you for being a part of this this year. Some of our most enjoyable conversations.

Thank you so much for everything you've done to help.

Dude, thank you guys. I love I love TPB. Like seriously, I love TPBN. Like straight up. Uh we're we're on the moon, I guess, or we're in outer space. Uh you guys are in LA. I would love to do I got to do in person sometime. I'm like never out. No, I would love to. It'd be great. It'd be great.

Merry Christmas. Have a great holiday season and we'll see you in 2026.

You're the man. Yeah. Goodbye. [clears throat]

Cheers.

Public.com investing for those who take it seriously. They got multiasset