Sovereign Vs Hyperscaler Demand

Bloomberg

Watch Sovereign Vs Hyperscaler Demand - Bloomberg on Dailymotion

Transcript

00:00Daniel Pilling is with us, Sands Capital Portfolio Manager. So why is the stock down?

00:06Well, I think there's a few reasons. One, stock had a good run into the quarter. And two,

00:11it probably slightly missed sort of the buy side expectations. And it was mainly around,

00:16I think, China. And then secondly, Black Hole is still a little bit supply constrained. So

00:21they're ramping it as fast as they can. But I guess the ramp wasn't fast enough. But

00:25I would say, though, to us, those were pretty strong results. And maybe one interesting fact

00:30to look at, if you take out the China revenues completely out of the data center, then the

00:36rest of the world for them actually accelerated from about 60-ish percent growth in Q1 this year

00:41to a guided 80 percent growth in Q3 this year. So actually, NVIDIA x China has been accelerating,

00:47which I think is really strong. I think that's a really interesting take. We had Vladimir

00:51Galibov from Omdia yesterday talking about how Europe is stop-start in certain countries.

00:58You've got Denmark really going hard on sovereign AI, other countries less so. We got told $20

01:03billion for sovereign AI. Is that strong enough for you, Daniel?

01:07Yeah, I think that's a great number. But the truth of the game is still within hyperscalers

01:11and AI startups. I think that's where the real action is for now. And it's a much,

01:16much, much larger number. And I think the interesting aspects there are one,

01:20as Jensen noted, that reasoning AI is so much more compute intensive. He quoted 100 to 1,000 times

01:26more on the one inside. And then on the other inside, though, it's also much better. So it

01:30reduces hallucinations. And he also mentioned that the return on invested capital you receive

01:35from buying these chips is phenomenal, right? He mentioned if you invest 3 million today,

01:39you can make 30 millions in token revenues in short order. So I think that's where the real action

01:44is as of today.

01:44And is that why the real action is talking about where Rubin takes us, the cadence of innovation

01:49over at NVIDIA?

01:50I think so. I think Rubin will, again, the most important thing that NVIDIA is doing

01:56is they're increasing the performance per watt. And he noted that it costs about $50 billion for each

02:02one gigawatt data center. Rubin will, again, improve that performance per watt, which means that NVIDIA's

02:08competitive situation improves on the one inside. But also on the other hand side, you can do more

02:13compute, you can build better models with Rubin, which means it's a self-fulfilling prophecy a little

02:17bit because whatever comes out of these models will, again, have higher IQ and effectively bigger and better apps.

02:24You're not worried in any way about Amazon's Tranium, about Google's TPU, the fact that the customers are

02:31innovating themselves.

02:33Yeah, I know. I think it's our job to worry sometimes. But we've looked at this in detail. And I think the idea here is,

02:39again, to say, well, so here's NVIDIA. They're building a giant system of a company. They have both networking, CPUs, GPUs,

02:45hundreds of different suppliers, and they put the software on top. They say, and we believe in this,

02:52that they are the best performance per watt supplier in the industry, bar nobody else. And as long as that

02:58is true, something like Tranium cannot compete simply because they don't have the scale, they don't have

03:03the system, and they don't have the software. So NVIDIA, to us, remains the winner in this sense.

03:09And that's what's interesting about Jensen Huang trying to say, look, we are more than just hardware in

03:14many ways. We are the one-stop shop, and we're leaning more into that. Did he lean more into the

03:18future opportunities enough for you, the robotic side of things? Of course, automotive remains a bit

03:23of an overhang when it comes to China. But thinking also about quantum and real future here.

03:28I think NVIDIA is probably the company in the world that runs the fastest that I've ever seen.

03:35I mean, what they're trying to do is they're trying to align hundreds of different suppliers.

03:38They try to design a GPU on an annual cadence, and they're trying to update their systems on an annual

03:43to 18-month cadence, which is unheard of. At the same time, they're focusing on quantum robotics,

03:49et cetera, et cetera. So I think NVIDIA is running as fast as anything else I have ever seen. And if

03:55I may mention, too, I think quantum is probably going to slot into the NVIDIA ecosystem by saying

04:01they have CPUs, they have GPUs, and at some point, they'll have quantum somewhere in there, too. And you

04:06can just switch between those various sort of compute layers.

04:08And that was what was interesting with IBM and AMD's partnership, the idea that almost

04:13you reference how these reasoning models work. Well, I'm thinking of GPT-5, how it just pushes

04:17you towards the best possible model for your question. We're going to have the best possible

04:21computing standard for your general task.

04:25Yes, I fully agree with that. The best possible, as well as the cheapest and the fastest.

04:30And again, I mean, as we think through the next few years, the question is, like, how much more

04:34compute can NVIDIA sell? And it's truthfully continues to be a question of the scaling laws,

04:39right? And we've seen with Grok 4 that compute still provides better and bigger models. We're

04:44going to have Blackboard clusters scaling in the next 12 months or 6 to 12 months. And hopefully,

04:49we'll be seeing good results coming out of whatever comes after Grok 4 and GPT-5. And that's

04:54what ultimately truly going to drive the bus for NVIDIA in terms of demand and in terms of the

04:58revenue outlook.

04:59So you think we should think more about Colossus? We should think more about what Meta's doing in

05:04terms of, well, it's scaling in Louisiana. Should we figure China out entirely, considering the

05:09government wants to push its own homegrown domestic talent anyhow?

05:14Yes, I think we should. And I would actually argue what China is doing seems to be relatively

05:20rational. What China is basically saying is that you're trying to sell as a chip that's based off

05:25of an architecture that's three and a half years old. And secondly, we've also grown a local AI

05:30architecture or a stack that has caught up somewhat to that old chip. So unless you start

05:36selling us a newer derivative of Blackwell, we don't really maybe need you as much anymore

05:42as we needed you three years ago. So that all seems to make sense. And I don't know whether

05:46the licensing will work out or not. But the interesting piece is if it does, that's a big, big upset for

05:52time, for Nvidia, because Jensen noted is a $50 billion market. And if they're allowed to ship a

05:59derivative of the newest chip, their market share there should be enormous, because there's nothing

06:03in China that even can remotely compete with what Nvidia has.

Category

Transcript

Be the first to comment

Recommended