AI Chip Race Heats Up With Amazon's Trainium3

Bloomberg

Watch AI Chip Race Heats Up With Amazon's Trainium3 - Bloomberg on Dailymotion

Transcript

00:00So did Amazon manage to solidify its prowess when it comes to vertical integration, cloud and chip

00:06offering? Well, they seem to be doing a really excellent job of that. And what's interesting

00:11about what Matt just said is they're going to let customer demand drive, you know, how they build out

00:17these different capabilities for AI applications, which tells us that there are going to be

00:24different models purposed to different types of workloads. So everybody's thinking, oh,

00:29NVIDIA is the market dominant player right now. They have massive market share and they're treating

00:35it like it's one sort of big thing that's homogenous. And I think what people are going to

00:40realize over time is that there are going to be different types of models for different workloads

00:43and different chips will be useful for developing such different models and for running those models

00:49for doing the inference. So I think there's a lot of room for different chip players here. And we

00:54should expect more of the in-house chips like what Amazon does in collaboration with Marvell

01:00or what Google does in collaboration with Broadcom to continue to rise in prominence as we've been

01:05seeing in the recent news coverage. I mean, yeah, let's talk about Marvell out with earnings, out with

01:09also a key purchase of Celestial AI that's all about lasers and can talk in lasers a little bit more.

01:15But are we seeing ASICs really starting to put a concern towards NVIDIA? Should they be threatened

01:21by this custom built model? Well, no company likes any kind of competition, right? So I'm sure NVIDIA

01:27would just be happier if they didn't exist. But clearly, look, ASICs are very powerful. That stands

01:33for application-specific integrated circuit, right? So these, I believe, these kind of XPUs or TPUs

01:40are going to be designed and going to be put to use for the type of applications in AI for which

01:46they're best designed. So it's clearly the case that NVIDIA is going to lose share over time

01:54in the big world of AI compute. But nobody should be surprised by that. And the pie is still growing

02:00so massively. They can't meet demand. They're unlikely to be able to meet demand for some years

02:05to come. And so there's a lot of room for ASIC approaches to play roles in various types of models.

02:11So we should expect that. Before we talk more lasers, I want to go back to where we sort of

02:16started. The fact that you're getting specific chips for specific applications, but also specific

02:20models. And I think there's some threat in the market today around Microsoft's other part of

02:25the flywheel. Because everyone thinks Microsoft's got a very strong cloud offering. And clearly,

02:29we've seen Azure not be able to fill demand with its supply. But we're now worried just basically

02:34about the applications, about companies being willing to pay more for these AI agents and actually

02:39the use of AI. What are you hearing in the market? How much does that impact your sector

02:44of chips?

02:46There is a lot of sporadic, I should say, information about how much AI agents and AI is being used

02:53in different settings. So, right, there was a story today about how Microsoft has reduced

02:59its sales targets for getting AI agents to be adopted by their customers. And so that has

03:05people concerned. Salesforce, similarly, is trying to roll out their AI agent capabilities. And some

03:12people are wondering whether that's going to happen. But there are so many other applications

03:15in addition to those kind of smart helper AI agents. And I think the earlier in the week announcement

03:22with NVIDIA and Synopsys is a good example of that, right? The industrial uses, whether it's these

03:28digital twins, basically being able to model out a whole factory or a whole building or any other

03:34physical presence in simulation, you need this kind of computational power, you need AI.

03:39And so that kind of collaboration, I think, really points to lots of other applications

03:44for AI. And then you go to biotech and pharma and, you know, computation, we've just been held back

03:51for so many years by its limitations from the CPU side. And now that we have these GPU-based

03:56accelerators and innovations in fabric, like what Celestial is bringing to Marvell and what Broadcom

04:02already does, is speeding up all of these computations, making far more applications possible

04:07across all industries.

04:09And some would say limitations have been in the fabrication in the United States,

04:13energy just making sure that we've got more solid and trustworthy supply chains. I think about just

04:20the news also earlier this week, $150 million stake in a chip startup by Pat Gelsinger. Of course,

04:25we all know him from Intel. But the U.S. is making that investment in X-Lite, a startup. Joanne, you have some

04:31real insight as to how this technology is being put forward. It's all about lithography, taking on ASML.

04:38Yeah, so X-Lite is developing a new laser technology. And that is a fundamental ingredient to doing

04:45lithography, right? ASML being the leader in extreme ultraviolet lithography, which depends on these very high-end

04:51lasers. So X-Lite's taking a different approach to laser development. And what's interesting is, and also

04:55reminds me of my past, is that this $150 million investment by the federal government, by the Chips and

05:02Science Act, the work is going to be done at a place called Albany Nanotech, which is where I got my start

05:08in the semiconductor industry, you know, 20-plus years ago. And it's a location for collaborative R&D,

05:15for development of new manufacturing techniques, development of new materials,

05:19development of new recipes for designing and building chips. And it's all the major players

05:25in the industry are there, and it's pre-competitive. And so for X-Lite and Pat Gelsinger to be doing it

05:30there with the support of the federal government, sort of enhances this model of get some public

05:36support for a technology that can be spread and help the industry, broadly speaking, and does go

05:45towards this effort to enable more chip equipment design and manufacturing to eventually occur in the

05:51United States. It's going to take a long time. These programs, you know, as I was involved with back in

05:55the early 2000s, they take many years. But it's a great way for the U.S. to subsidize a general

06:02technology, which could ultimately help the U.S. position in equipment manufacturing.

Category

Transcript

Be the first to comment

Recommended