- 2 months ago
Category
🗞
NewsTranscript
00:00Caroline, I think you can fall into the camp of people who've been like, hey, I've known this for a while.
00:04I've known that TPUs are so efficient.
00:06Does it really make a difference that you have Meta now buying into this?
00:10What does it mean to have a market awaken to the possibilities of what Google has developed?
00:14I think what's different here is before Alphabet had been selling it as part of the cloud winning formula.
00:20Come to us, you can use our cloud and use our TPUs, which are on their seventh generation now.
00:24They've been developing these for 10 years.
00:25They've been offering them to other clients since 2018.
00:28We haven't seen them really sell them specifically to others to use within their data centers, not just Meta, but maybe banks who need to trade at high frequency.
00:37That's a clear area of edge for them.
00:39So they're starting to get this inbound.
00:41In fact, a lot of analysts, Stacey Rackson over there, we've heard also from Bernstein.
00:45You've also heard from Gil Luria saying this is kind of an obvious dimension of the business model.
00:50The question, though, is what will NVIDIA do about it?
00:53Now, we've heard before reports from the information that OpenAI was getting very close to potentially taking TPUs from Google Alphabet, maybe using their cloud more ever since that relationship when Microsoft got disrupted a bit.
01:04And then NVIDIA went and put $100 billion into OpenAI and said, oh, please do use our GPUs if you wouldn't like to.
01:11What is a meta going to suddenly get offered from Jensen Huang?
01:14That's my key question, because you've seen big deals with Anthropic, but Anthropic suddenly has got some more investment from NVIDIA as well.
01:20There's going to be a tussle at the top here, but be sure, Jensen Huang's been eyeing TPUs for a while, and Alphabet, this has been a winning formula.
01:27That's why Gemini 3 has been so efficient, and efficiency is everything.
01:31It's all about energy and what more they can do with it, because that's going to be the bottleneck.
01:35Yeah, we have actually updated our circularity chart, which continues to grow and boggles the mind, no matter which form you look at it in.
01:46Michael Ball, is this a problem for the whole market?
01:49I mean, so much is built on NVIDIA reigning supreme.
01:53Here's the new version, right?
01:55And you can see all of the deals that they do with other big players, like AMD, like Oracle, like Microsoft, like Intel, and that those players do with them.
02:05So if NVIDIA is brought down, is it a problem for everybody?
02:08I don't think so, and this has been the question that the market's trying to figure out, because capital so far is staying in the market.
02:15So although we've had NVIDIA now break below that $180 level, which was a key battleground and had a lot of sort of positive, negative connotations if it was below or above,
02:22the fact that Google has taken on a trillion dollars and that the market is just finding a more efficient answer to the AI CapEx question.
02:29And that's what we want to see.
02:30We want to see competition.
02:31We want to see these guys have choices.
02:33We don't want the hyperscalers all bottlenecked into one choice that has backlogs growing and energy efficiency questions.
02:39We want optionality.
02:40We want capitalism to work.
02:41And that's sort of what we're seeing.
02:42And although it's violent in some ways, given the price action of the winners and losers, it is a positive development in the long run.
02:48Okay, so if we do get, and we have, something more efficient, something more energy efficient,
02:53Caroline, is that enough?
02:54Or is this an ecosystem that's incredibly sticky?
02:56So even if there are hyperscalers and others who want to use Alphabet's chips, if they're already with NVIDIA, can they just so easily change over?
03:05What's going to be interesting is how the development of the software works out here.
03:08Because we understand that Alphabet is doing a lot to show the prowess not only of the TPU, but also the software that goes with it.
03:14And some of the ability to use software that's adjacent, not just their jacks, which is their own programming language that they've developed to go alongside TPUs, but you can use others that use different ecosystems.
03:25Look, the thing with NVIDIA is that they have CUDA.
03:28And everyone is locked in not only to the hardware that NVIDIA offers, but the entire ecosystem that NVIDIA offers.
03:35If Alphabet can start to chip away at that and show, look, with us, you can have prowess the entire way through, I think there is going to be a real recognition that this is a clear alternative.
03:44But remember, the whole vertical integration story, what Google has shown so well is that it can develop incredible models because it's using its own tech stack, and then it's putting it out into the ecosystem where we're all using it.
03:57Look what Alibaba is just announcing.
03:59They, too, want to be a full vertically integrated company that makes their own chips, that develops the models, that offers it out on their cloud, and then offers the models as well.
04:07But also remember, this is what Amazon does.
04:09Yes, we're talking about TPUs, but Tranium is on their third iteration.
04:13We understand that many people are using not only for training, but are we going to use it not just for inference, but the training of models as well.
04:19How much can Amazon start to show its competitive edge as well with its chips, not just its AWS offering?
04:24By the way, maybe you can clear something up.
04:26I don't really understand Jensen Wang's argument that the GPUs they sell, what are we in now, Blackwell or Hopper?
04:33Rubin, Vero Rubin for the next one.
04:35Right, so those are capable of many different things, right?
04:39The graphics processing units are flexible, as you may have read last night.
04:43Whereas the tensor processing units are supposed to be ASICs, so they're application-specific.
04:51However, if that application is AI, I mean, I don't care if GPUs can show me good graphics.
04:59I want AI, right?
05:00And I think this is where the battleground is going to be.
05:02How much efficiency can you wring out from this hardware with the software within it?
05:06And if Jensen Wang can say, look, my GPUs have been able to accelerate your AI models, your machine learning that much faster, and by the way, I've got a chunk of change that I can continue to invest in you as a company and continue to invest in your entire ecosystem, maybe there's the winning formula.
05:20But this is why Google has made everyone sit up straight and why analysts have become noisier and noisier as to how much this has been a winning formula for them, is that DeepMind were able to go directly to their semiconductor team and say, look, we're learning this.
05:36Let's work in it together.
05:37Let's understand where the efficiencies can be brought.
05:39And I really do think that the infrastructure issue is not going to be whether you eventually can lay your hands on the hardware.
05:45It's going to be whether you can do more with the grid, with the electricity.
05:49And if they can show energy efficiency at this point, that's going to be what's music to the government's ears and therefore to users' ears.
05:56But I just think this turf war is just starting.
05:58We're all going to start talking about Inferentia and Tranium and what Amazon's doing and what everyone else is doing in ASICs.
06:04Because remember, that's what Broadcom has been doing so well on, all these coming, Meta's designing.
06:07Broadcom was up 12% yesterday.
06:09OpenAI is developing its own chips.
06:11Eventually, maybe that's going to become commoditized, not just large language models.
06:15So it's a world where there are different changes, Michael.
06:17I mean, it's no longer sort of like the tide is lifting all boats.
06:21You have to be more discerning with what these companies are doing.
06:24Right.
06:24And I would add that the depreciation question now is even more forefront, right?
06:28Because we don't know to everything you just said.
06:30I do this, I watch this, and my head is blowing.
06:32Like, it's just so much to keep track of.
06:34And as an investor, if someone is trying to look at a balance sheet, whether it's a hyperscaler or whether it's the provider of these chips, looking at the backlogs, looking at the inventory builds, how much is this all going to be worth?
06:43If there's all these new options of energy efficiency versus training inference kind of duration, how long something stays as a practical and efficient class?
06:50Well, this is why Zuck is supposedly thinking about renting them, right?
06:54When I look at a new EV out there in the car world, that's where my mind goes, I think, why don't I lease it instead of buy it?
07:01Because there's going to be something better in two years anyway, and who knows about the battery?
07:05I want to ask, though, more about Danny's question, which is that you have less sector correlation if two people in the same sector are at odds, right?
07:14Well, funny enough, tech's still been under pressure, so looking at what I like to look at at the option markets and seeing where sort of bullish skew slash bearish skew is, people are still very negative on tech.
07:24People have obviously rotated into more value-focused sectors like health care, like defensives.
07:29And that makes sense in a lot of ways because the valuation thing over the last three weeks has been what spurred this.
07:33But to your point, that, like, internal battle of is this good for the overall AI CapEx story is still being thought out, and we're not seeing it with people trying to regain exposure to it.
07:43Google calls are expensive, the vol is a little bit high, but it's not necessarily being reached for like we've seen in NVIDIA in the past.
07:49Can I ask whether this whole short open AI suppliers, long alphabet suppliers is still working?
07:56Because yesterday Oracle pushed higher, but today Oracle's back trading in negative territory before the market opens.
08:02Is that long in the tooth?
08:03Are we still thinking about that trade?
08:04You know, that's a great question.
08:06I don't understand, quite frankly, if everyone's sort of in this together in a lot of ways, why there's been such differentiated to this scale.
08:12Now, I get the weaker links were being picked off, but we're looking at CoreWeave and Oracle's down almost 50%, 40% in their equities, and then obviously we know what's going on with their CDSs.
08:20So it looks like we've had that emotional pulse come through the markets, as we often do, and now valuations have sort of reset in a lot of ways, and it's suddenly going to be a point where people might try to pick bottoms.
08:29Now, I'm not calling for that, but I guess to your point, it can't all just go back into Google now.
08:34Like, we've just got to have this ecosystem going on.
08:38Apple's next.
08:38Like, is that the moment it's NVIDIA, Apple's at what, 4.2 trillion, there or thereabouts?
08:45I think the Apple questions are incredibly fascinating.
08:48I think the Apple questions are really interesting.
Be the first to comment