MU & SNDK Show Memory Shortages, Energy Next Key Bottleneck for AI Trade

Alex (0:00)

Welcome back to Next Gen Investing. It's time to take a closer look at the tech sector. And more in particular, I want to zoom in on this hardware space. We're joined now by Brendan Burke. Brendan is the Research Director for Semiconductor Supply Chain in Emerging Tech at Futurum. Can't think of a better guy to talk to you with all that's going on in the last couple of weeks. So we're just talking meta, Brendan, and it seems this Muse AI is really spurring some optimism on that front. You go back to last weekend, not this one we just got through, but the one before you had Dario and, and Sam and Elon sort of sounding alarms with how much this technology has advanced.

But it seems under the surface, the realities are these companies just can't make these chips fast enough. So I'll leave it open-ended, Brendan. Where do you assess the state of this industry as we speak today?

Brendan Burke (0:55)

Alex, it's great to be speaking with you. Thanks for having me on.

The industry is in a position of meeting demand that's rapidly evolving and changing in scope. And you mentioned the new agent products that are coming out onto the market.

The chip industry is just now reckoning with the fact that agentic AI can produce a 100-fold increase in the amount of compute needed to carry out AI tasks. And that comes along with the need for custom accelerators and AI-specific CPUs that fit that demand profile. This is something that chip makers are having to keep up with in their capacity planning and find new resources for. And this latest move up today shows that there's still additional blue sky in the market.

Alex (1:49)

When you look at this space and we sort of think about how it's evolved over the last several years, it really started with the hyper specialized high powered chips that companies like mostly just Nvidia initially rolled into Broadcom. AMD finally joined the party, even Intel started to join.

TSM has been a big part of the story as it was the main manufacturer of these products. But what really changed in the last 18 months was the realization that the memory side of things was really where the bottleneck was starting to show up. You saw SanDisk as a spinoff, go bonkers to the upside. You saw names like Micron go berserk as well. When you look at where the bottlenecks are primarily at now, has the story evolved or changed, or has it only just gotten bigger?

Brendan Burke (2:46)

Everything has gotten bigger, but there are new constraints coming into the picture that require additional supply on top of what we've gotten so far. Memory was supplied by HBM over the past two years, and that mix is shifting. But going forward, the constraint is on power. We need to be able to power up chips and be able to actually build the data centers to put these chips in.

And the next leg to accomplish that will be in the intraconect space, and it will be lowering the amount of power needed to connect chips to each other. And so that's quickly turning into a new field of photonic chips that will be supplemented with memory to get the memory bits from those very expensive HPM wafers to the chips efficiently. And that's a big part of the AI infrastructure build out. What I was hearing at AI Infrasummit last week is that we've got to be able to not just store data, but move it quickly and we're working on the Internet interconnect technologies to make that possible.

Alex (4:05)

Brendan, let's take a step back. So, I always have said this about the healthcare space. I think it's one of the highest barriers to entry from an understanding standpoint. You have to know a lot of things. But I'd argue, as much as things are advancing, the jargon, even just listening to you around this semiconductor and sort of computer chip world, there's a pretty high skill cap here to understanding things. And it's changing dramatically as this technology evolves.

Outside of, of course, reading research from you and your team and others across the industry, what can that next generation of investor who's involved, maybe self-directed, in their research and decisions kind of monitor? Is there, are there new metrics and terms that are, that are essential to know right now?

Brendan Burke (4:58)

The key building blocks of the chip trade right now are gigawatts and wafers. That's what we're connecting together at Futurum in our end-to-end living data center semiconductor model, because you can't treat any one component as independent from the others. And so the data center buildout is an area that we think kind of progressed towards 80 gigawatts of capacity added in 2030, based on power that's coming online and data center projects that are being started now. And then on the wafer side, we're looking at new fab capacity across logic, memory, as well as on the photonic side that can support large incremental markets and get to a data center capex opportunity that could amount to nearly 4 trillion or around 3.7 in our forecast, based on the pricing of all the components that will be needed in the agentic data center of the future. And so, focusing on those big building blocks, and then learning from what is an increasing, increasingly transparent sector, through events like AIM for Summit that we had in San Clara last week, or even today's coherent photon link launch, where they go into the way that customers are deploying optics technologies.