Constellation's Wang on Google-Nvidia Chips Rivalry

Bloomberg

Watch Constellation's Wang on Google-Nvidia Chips Rivalry - Bloomberg on Dailymotion

Transcript

00:00Ray, it's good to have you. First of all, maybe start with the whole idea of TPUs versus GPUs.

00:07What are some of the differences and why do you think TPUs are getting a bit more traction or at

00:12least more talk these days? You know, you're absolutely right on the analysis. The reason

00:17TPUs are so important, these are tensor processing units. They're purpose-built for AI, for deep

00:22learning. This means training and inference. This means lower total costs. This means they're more

00:27power efficient than GPUs. And of course, these are things that Google has been working on for quite

00:32some time. And what we've seen over the course of the seventh generation of Google's TPUs is their

00:38ability to make them super efficient. Now, think you're meta and you've got a supply chain

00:44diversification problem. You're only getting everything from NVIDIA and you need the ability

00:49to actually crank out training and learning and all the other things you need for reasoning for AI.

00:53And suddenly you're like, hey, where else can I go find chips? You can't do the regular chips.

00:58But Google's shown that they can do TPUs. And so the logical thing is to rent those TPUs from Google

01:03and Google Cloud. And, you know, up until recently, people thought, you know, Alphabet was lagging

01:10behind in this whole AI race. But the fact that, you know, there has this whole, it's one of the few

01:14companies that have this so-called full stack when it comes to computing then. So, you know, what are the

01:20prospects, you know, is this sleeping giant finally making a big comeback now?

01:24Well, we've been talking about how Google is ahead in AI for the last 18 months. And part of it is really

01:30because when Microsoft fired that shot saying they were ahead in AI, Google finally got its act together

01:35about 18 months ago. But they've been working on these chips for the last five years. So it's not something

01:40they just came up with over time, but they are fully vertically integrated, like you said, with a complete

01:45stack. We call it chip to app. And what that does, it gives them massive efficiencies of scale. If you

01:50think about having a fully vertically integrated iPhone, you know the advantages as opposed to

01:55getting components from different places, just like with a Mac. Now, as I said, over time, like you'll see

02:00Google is leading the AI battle because it's not only diversifying income streams, it's taking all parts

02:06of the value chain and bringing it together back into search. Now, one of the things that we spend a lot of

02:12time thinking about is AI chip demand. And the question really is, is NVIDIA going to take a

02:17dent from TPUs from Google or is it going to make a difference from AMD? And the short answer is

02:23there's $7 trillion market demand in AI chips by 23. I mean, that's where we see the market in 2030.

02:31So we think there's so much demand to go around that this may be a time to buy the dip for NVIDIA

02:36or buy the dip in other places where you start to see some weakness.

02:39Interesting. So there's plenty of demand out there, I guess you say. It's not really kind

02:45of a zero-sum game, right, Ray, as you highlighted, right? It's not like I choose

02:48between TPU or GPU. There's plenty to kind of, you can do both for a lot of these hyper-series.

02:54You're definitely going to do both. You're definitely going to do both. And part of the

02:58reason is you want to diversify your chip base. And part of that, you know, part of that

03:02diversification strategy is if something works with CUDA and NVIDIA well, but something works better

03:07with TPUs. It's just the same reason that you diversify whether you put your cloud in AWS

03:12or in Google or in Oracle or in Microsoft. It's the same reason. And people are also diversifying

03:17their chip sets because each of these chips are going to do something differently. And for example,

03:21you might use deep learning chips for training and inference with, you know, TPUs more than you're

03:27going to do with the GPUs that you're going to do for just training. And so each one is going to

03:31play a different role, especially in a mix for some of the advanced companies using AI.

03:35Yeah. I mean, beyond just reports of meta looking into TPUs, who else do you think would be,

03:43you know, adopting and seeing more of a wider adoption of these TPUs? Who else would be using

03:48them, you think, Ray?

03:49I think every hyperscaler that doesn't feel like they're competing with Google will be looking at

03:54this as an example. So for example, I mean, Oracle's probably going to look at these TPUs. I think,

04:00you know, you might even see Microsoft and Amazon take a stab at it. Amazon has its own

04:04training chip as well. So they're probably using that. But when we get down to the companies that

04:09are using it, pharmaceutical giants, energy companies doing exploration, governments using

04:14weather, people building sovereign AI, they want to be able to make sure that they're going to

04:18consistently get chips on time. They're going to want to be able to make sure they have a second

04:22choice or third choice in source. So you'll see AMD, you'll see Google play that role in terms of

04:27providing the alternative to the NVIDIA dominance.

04:30What about when it comes to, of course, our large language model, the Gemini 3?

04:36How does it stack up with the likes of ChatGPT? Because, you know, there's been talk about,

04:41you know, it is one of the leading top tier sort of language, you know, large language models out

04:45there. How do we see consumers? You know, are they getting more attractive to Gemini?

04:53There's definitely a lot of attraction to Gemini because you get the full stack of Google,

04:57back to your point earlier. And we're seeing Gemini 3 beat out ChatGPT in a lot of instances

05:02and use cases. If you go to some of those LM leaderboards, you're starting to see that

05:06Gemini stacking up against perplexity, Gemini stacking up against Claude, and you're seeing

05:12these head to head battles. And I think you're seeing it's going to come down to really your

05:16use case. General purpose, there'll still be Gemini, still be ChatGPT. You'll see Claude there,

05:21especially for software development. It's one of everyone's favorites. And then, of course,

05:25if you're looking for open source Chinese, open source LLMs, Nguyen and DeepSeq are still leading

05:31the pack. We're seeing some great advantages there as well as people are looking at smaller

05:35language models and use cases. What does this mean for some of the Asia sort of chip makers out

05:40there, right? I think Samsung is one of the key suppliers of Google's TPUs with their HBMs.

05:45Could they really start to catch up with the likes of SK Hynix in this space?

05:49Well, it depends on the volumes, but you're definitely going to see that diversification.

05:55Some spreading the AI chip wealth around to some of the other vendors is going to happen.

06:00But the NVIDIA lead is definitely hard to beat. If you look at what's going on,

06:05even with TSMC making some chips in the U.S., they only get half the chips and then the rest

06:09are assembled in the U.S. I mean, there's something special about what's coming out of TSMC.

06:13And of course, the HBM high bandwidth memory that's coming out of SK Hynix and some of the

06:22other places, because what we are seeing is a massive demand for speed and massive demand

06:27for compute power.

06:32Can you tell us more about what this, I mean, you mentioned that this doesn't really make

06:36a dent when it comes to NVIDIA, but the stock really has been taking a hit recently, right?

06:41Whether it's concerns about, you know, this AI bubble, I think Michael Burry has led to

06:46a little bit of that, you know, scrutinizing of these sort of deals and the like.

06:51I'm just wondering overall, you know, how much more upside do you think NVIDIA does have

06:55moving forward?

06:57Well, if we actually do some models and some of our models, we basically see there's another

07:01trillion dollars in sovereign AI in market cap, and there's another trillion dollars in

07:06physical AI in market cap.

07:07So somewhere about six and a half trillion, seven trillion market cap is probably where

07:12we think NVIDIA is going to peak.

07:14And that's coming from basically the demand as we start seeing that shift on sovereign

07:18AI.

07:19That's probably the thing you're going to be hearing about all year from Davos all the

07:23way down to next year by CES.

07:26That's the sovereign AI, sovereign AI deals, companies that are building data centers, physical

07:30AI.

07:31Those two things are going to drive the market and drive the headlines in 2026.

Category

Transcript

Be the first to comment

Recommended