00:00Meta planning to deploy millions of NVIDIA processors in the next few years, a deal that could add billions of
00:05dollars to NVIDIA's balance sheet.
00:07Bloomberg Intelligence Global Head of Tech Research Mandeep Singh joins us now.
00:11Mandeep, first of all, I mean, does NVIDIA need this? It's up pretty sharply off the back of this.
00:15I thought there was already like pretty robust demand for what NVIDIA is selling.
00:19There is. And look, this is probably Meta's way of locking in that NVIDIA supply.
00:24So not only are they using the GPUs, but they are also talking about using NVIDIA's CPUs.
00:30So that's new. I mean, traditionally, all these companies used x86 architecture and Intel CPUs along with the GPU cluster
00:40that NVIDIA had.
00:41But now, you know, with this deal, they're talking about using NVIDIA CPUs.
00:47So that's one big change.
00:49The other thing for NVIDIA here is Enthropic has had a very steep rise in the last few months.
00:56Enthropic models aren't trained on NVIDIA GPUs.
01:00They're trained on Google TPUs, Amazon.
01:03So from that perspective, you have a company that's doing really well on the model side that is not trained
01:10on NVIDIA GPUs.
01:12And so from that perspective, I think for both of them, it's important that, you know, Meta comes up with
01:17a model that is really good in terms of being a frontier model like Anthropics that is trained on NVIDIA's
01:26latest cluster.
01:27And obviously, Meta needs to lock in the supply.
01:29So it's, I think, symbiotic.
01:31Yeah, that's a good point, too, because also Gemini, I mean, of course, they're using Google's.
01:36I mean, in all of this, can you, will other companies look at the advances of Anthropic, of Gemini, and
01:44start to rethink their supply chain because of that, that they don't want to be too reliant on NVIDIA?
01:48There is that aspect, but what NVIDIA is saying is our latest Blackwell chips will give you 30x more token
01:57output than our prior architecture from Hopper.
02:01So for the same unit of power, you can have 30 times more output.
02:06And so that throughput aspect of NVIDIA chips is always there.
02:10And that ties into how good the model will be that is trained on the latest chip.
02:15So really what NVIDIA needs is companies like OpenAI, like Meta, training their models on NVIDIA's chips and having somewhat
02:25of an advantage compared to the ones that are not trained on NVIDIA.
02:28Well, what's been the usage of, like, AMD?
02:30Where do some of these other competitors stand?
02:32I think AMD gets hit even with this tie-up between Meta and NVIDIA, where NVIDIA is now supplying the
02:40CPUs.
02:42Previously, AMD or Intel was the supplier of CPUs to the data centers for Meta.
02:47Now, with NVIDIA selling CPUs, that could have an impact on AMD.
02:53Are you surprising that a company like Intel that has a government backing, there haven't been more political forces?
02:59Because this is what was talked about, right?
03:00Yeah.
03:01The U.S. takes a stake and pushes companies to start using Intel for things like CPUs.
03:05Are you surprised that we haven't really seen that?
03:08Well, Intel still has the foundry capacity.
03:11So the difference between an Intel and AMD is Intel owns its foundries where it makes its chips.
03:17AMD is going to TSMC.
03:19So from that perspective, you're right.
03:21I mean, Intel has a local foundry that all these chip makers could use.
03:26But obviously, the government role comes into play in terms of pushing these companies to start using Intel foundry more
03:33and more.
03:33So just the one other concern that we have to bring up, as we do with any of these conversations,
03:38is the circularity of deals.
03:39The fact that Meta already makes up a nice, chunky portion of NVIDIA's revenue.
03:44Presumably, that just reinforces that.
03:46You say it's a good thing for NVIDIA.
03:48But is there also kind of like a dark side to the silver lining?
03:52I mean, look, all these companies have an ecosystem.
03:56So that circularity also exists in, you know, Google investing in Enthropic and Enthropic using Google's chips.
04:03Same thing with the Amazon Enthropic partnership.
04:05And at the end of the day, it is an ecosystem because you're optimizing on a certain chip vendor for
04:12your model training and your model inferencing.
04:14And you can't have a generic architecture.
04:17And that's why, you know, they are very keen to partner with the companies of Metascale.
04:22You want to make sure they optimize on NVIDIA chips as opposed to using someone else's.
04:27Mandeep, thank you so much.
04:28Never a boring day in your world.
04:30Mandeep Singh of Bloomberg Intelligence.
04:32You probably wish for a boring day.
04:33That might be nice.
Comments