00:00Let's kick it off with the Amazon story, although I think it's fair to say that we're both more interested in Alphabet today.
00:06So they have a bazillion data centers.
00:09Like, I'm actually not moved by that.
00:12I could have guessed it, and I don't think it really matters, right?
00:16How many actual places they have, does it?
00:20I mean, right now, when you compare them to a CoreWeave, which is a pure play data center play on the GPU side,
00:26CoreWeave has about 32 to 35 data centers.
00:30It's got 250 plus 100,000 GPUs, and, you know, they have a $5 billion revenue run rate.
00:38Amazon is like $130 billion plus run rate, so almost 30 times.
00:43So it doesn't surprise me they are around that 900 mark when you look at CoreWeave's footprint.
00:49But the real challenge for Amazon is they don't have the GPU capacity that some of these NeoClouds have,
00:55which is why their backlog has trailed the likes of Microsoft and Google, and now they are being more strategic about it.
01:01So this is where I'll be the one who asked the dumb question, right, for the guys in the back row here.
01:06Why does it matter if you have data centers all around the world with lightning-fast fiber?
01:12Can't you just have one that has, like, 50 gigawatts of compute?
01:16Why do you need these edge centers, like, near New York or near L.A.?
01:20So, Matt, you're serving a large-angleic model with one trillion parameters to over, let's say, a billion users.
01:28And not all the billion users are in that location.
01:30They're opening a data center.
01:32But that's where these models need so much compute.
01:35They require, you know, one trillion parameters to be loaded in that server or the GPU.
01:41But it has to be regionally accessible?
01:43I thought the idea was...
01:44That's how you minimize the latency.
01:46I mean, anyways, these models take a lot of time.
01:50That's the biggest criticism, that they are slow.
01:52So how do you serve that traffic, which could be billions of users,
01:56and that's why you need these edge locations to really cache and, you know, make sure your users have a good experience with the OOM.
02:04Okay, I thought I was more interested in Google, but now I have more questions on this, Manteep.
02:07Okay, if you are building this out and you have all these data centers, you have 900, does that also kneecap your competitors?
02:14Is it also harder for them to find the land to get the data centers if Amazon is buying up so many?
02:19Right now, they are in land grab mode because there are very few facilities with the kind of power requirements that these GPU centers have.
02:27And look, I mean, everyone is trying to make sure they get a hold of them because earlier it used to be three players.
02:33Now you have got 10 different NeoClouds that are also scrambling to get a piece of that data center pie.
02:40By the way, it's not just latency, right?
02:41The other big problem is hallucination.
02:44And you point out that Gemini 3, Google's new advancement, has a greater ability to curb AI hallucinations.
02:52And they do it using their own chips that need to be powered by less energy.
02:59So it's like a win-win-win here for Google and a loss, in a sense, for OpenAI and for NVIDIA.
03:05I mean, look, vertical integration has always proven to be a better strategy.
03:10The question was, in this case, could the TPUs from Google match up to NVIDIA GPUs in terms of performance and tokens per watt, which is the key metric?
03:20And in this case, you know, TPUs are in their seventh generation.
03:24So Google has been doing this for more than 10 years.
03:27The Tensor Processing Unit.
03:29And that's where other custom silicon efforts have still not caught up to NVIDIA because they're in their second or third generation.
03:36So, I mean, look, Google has the search index, which is where I think it's their other big differentiator.
03:42That no one, even OpenAI, doesn't have the search index for the web that Google has.
03:47So their results are grounded a lot better than any of the other LLMs out there.
03:53So when Sam Altman warns of rough vibes and warns of revenue that will take a hit because of the competition, because of Gemini edging ahead, how does that change OpenAI's behavior in the race?
04:03I think they will be more product-centric and more application-focused now.
04:08So whether it's e-commerce, which is where they can really hurt Google, because let's say OpenAI really doubles down on e-commerce with ads and, you know, integrating with a lot of marketplaces.
04:21That's where bulk of Google's search ad revenue comes from.
04:24The $200-plus billion in search ads are tied to e-commerce.
04:29So that's where I feel OpenAI is really going to either release a new model, new functionality,
04:34because e-commerce is the place where you can hurt an Amazon and a Google and all these incumbent players.
Comments