The AI boom has already transformed the semiconductor industry, but the biggest question facing markets today is whether this surge in spending is nearing its peak or still has years to run. As hyperscalers continue investing billions into AI infrastructure, demand for compute, memory, and advanced chips has outpaced supply, creating bottlenecks across the semiconductor supply chain.
In this episode, semiconductor analyst CJ Muse explains why he believes AI infrastructure spending remains in its early stages. From DRAM and high-bandwidth memory to Nvidia, custom silicon, and semiconductor equipment, he discusses why today's AI buildout differs from previous technology cycles, how supply constraints are reshaping the industry, and what could determine whether the current investment wave extends through the rest of the decade.
We also examine Micron's changing earnings profile, the outlook for AI data center demand, China's growing memory ambitions, Nvidia's competitive position, and the broader economic impact of enterprise AI adoption. Whether you're following semiconductors, AI infrastructure, or the next phase of technology investment, this conversation explores the forces shaping one of the fastest-growing areas of the global economy. ----- 03:46 – Micron and the Memory Boom 08:36 – Can Memory Supply Catch Up? 11:39 – How Long Will AI Demand Last? 13:30 – Is Memory Demand Recurring? 15:13 – After 2030: Boom or Bust? 17:19 – The Next AI Bottlenecks 18:58 – CEOs Confront the AI Boom 20:43 – The Battle for AI Compute 26:10 – Why Nvidia Still Leads 28:14 – Can AI Boost the Economy? ----- Let's Stay in Contact: • Substack - https://substack.com/@themarketinsidershow • X (Twitter) - https://twitter.com/MktInsiderShow • Linkedin - https://www.linkedin.com/company/market-insider-show/ ----- 𝘛𝘩𝘦 𝘷𝘪𝘦𝘸𝘴 𝘦𝘹𝘱𝘳𝘦𝘴𝘴𝘦𝘥 𝘪𝘯 𝘵𝘩𝘪𝘴 𝘷𝘪𝘥𝘦𝘰 𝘢𝘳𝘦 𝘵𝘩𝘰𝘴𝘦 𝘰𝘧 𝘵𝘩𝘦 𝘨𝘶𝘦𝘴𝘵 𝘢𝘯𝘥 𝘥𝘰 𝘯𝘰𝘵 𝘯𝘦𝘤𝘦𝘴𝘴𝘢𝘳𝘪𝘭𝘺 𝘳𝘦𝘧𝘭𝘦𝘤𝘵 𝘵𝘩𝘦 𝘷𝘪𝘦𝘸𝘴 𝘰𝘧 𝘔𝘢𝘳𝘬𝘦𝘵 𝘐𝘯𝘴𝘪𝘥𝘦𝘳. 𝘛𝘩𝘪𝘴 𝘤𝘰𝘯𝘵𝘦𝘯𝘵 𝘪𝘴 𝘧𝘰𝘳 𝘦𝘥𝘶𝘤𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘱𝘶𝘳𝘱𝘰𝘴𝘦𝘴 𝘰𝘯𝘭𝘺 𝘢𝘯𝘥 𝘴𝘩𝘰𝘶𝘭𝘥 𝘯𝘰𝘵 𝘣𝘦 𝘤𝘰𝘯𝘴𝘪𝘥𝘦𝘳𝘦𝘥 𝘧𝘪𝘯𝘢𝘯𝘤𝘪𝘢𝘭, 𝘪𝘯𝘷𝘦𝘴𝘵𝘮𝘦𝘯𝘵, 𝘰𝘳 𝘭𝘦𝘨𝘢𝘭 𝘢𝘥𝘷𝘪𝘤𝘦. 𝘈𝘭𝘸𝘢𝘺𝘴 𝘤𝘰𝘯𝘴𝘶𝘭𝘵 𝘢 𝘲𝘶𝘢𝘭𝘪𝘧𝘪𝘦𝘥 𝘱𝘳𝘰𝘧𝘦𝘴𝘴𝘪𝘰𝘯𝘢𝘭 𝘣𝘦𝘧𝘰𝘳𝘦 𝘮𝘢𝘬𝘪𝘯𝘨 𝘧𝘪𝘯𝘢𝘯𝘤𝘪𝘢𝘭 𝘥𝘦𝘤𝘪𝘴𝘪𝘰𝘯𝘴. 𝘚𝘰𝘮𝘦 𝘭𝘪𝘯𝘬𝘴 𝘪𝘯 𝘵𝘩𝘦 𝘥𝘦𝘴𝘤𝘳𝘪𝘱𝘵𝘪𝘰𝘯 𝘮𝘢𝘺 𝘣𝘦 𝘢𝘧𝘧𝘪𝘭𝘪𝘢𝘵𝘦 𝘭𝘪𝘯𝘬𝘴.
AI Generated·May contain errors
Memory stocks are on fire and a number of them have been going up significantly.
Today we sit down with CJ Mews, who's one of the top analysts in this space, to tell us what's
going on right now and how much justification is there for these valuations and will the runs
continue. This AI infrastructure build-out, you know, makes everything else pale in comparison.
Here we are in 26 and we're probably going to do about 1.6 trillion in semiconductor revenues.
Quite frankly, semiconductors are the bottleneck to building compute. It's important to continue
to remember that semiconductors are cyclical. We will probably overbuild at some point and that
means a correction of some certain magnitude that, you know, will be painful for investors.
That will definitely come. I'm Siamak and this is Market Insider.
CJ, it's great to have you on. Welcome. Yeah, thank you for having me. It's great to be here.
You're one of the top analysts in the space of semiconductor and hard drives and memory. Can you
tell us what you're seeing right now? Well, this AI infrastructure build-out is driving tremendous
demand across all of the sub-segments you talked about, you know, whether it's semiconductors,
hard disk drives, memory, including both DRAM and NAND. You know, AI is really soaking up
all of that capacity. And, you know, if you take a step back, you know, McKinsey back in 2022 said
we were going to hit maybe a trillion dollar semiconductor revenues by 2030. And, you know,
that was 22. But here we are in 26 and we're probably going to do about 1.6 trillion in
semiconductor revenues. And so, you know, the last, you know, handful of years has really led
by AI to tremendous demand for semiconductors. And quite frankly, semiconductors are the bottleneck
to building compute. How do you see the AI go from here? Yeah, I mean, you're hitting on
the most debated kind of thought in the stock market today with broad implications, not just
for semis, but, you know, really for the whole market where, you know, whether you're in energy
or industrial, you're benefiting greatly from AI as well. And so, you know, really the key
debate is the sustainability of CapEx by the hyperscalers and whether Amazon, Google, Meta,
you name it, are going to continue to invest like they are or do we see a pause? And I'd say,
you know, from our perspective, we don't think we're close yet to the 1 million megascale GPU
cluster that we know the likes of Anthropic OpenAI are pushing to. And I think that they
both think there's tremendous value in terms of reaching true AI. And so they will continue.
And so when I layer in that thought with the backdrop that semiconductors are in shortage,
which means compute is in shortage. And, you know, that is best seen at TSMC where
they can't meet demand. And I think that'll be the case 26, 27, 28. So with the bottleneck of
not being able to build compute fast enough, with this very strong demand picture, I think this AI
infrastructure build out will last through 2030. But there's definitely, you know, my competitors
or peers that think that we might peak out in 27 or 28. And really, that's probably the best way to
define the debate. You know, will spending peak in 28 or will it sustain itself into 2030 and beyond?
Let's take a look at some companies like, for example, Micron. The valuations are,
for some people, it's very hard to believe that these companies are now worth
over $1 trillion. But the earnings are going up significantly. And you've been telling people
about the earnings. Can you tell us what you're seeing there? It's important to take a step back
and understand that the DRAM industry, which is the lion's share of the revenues for Micron,
from 1995 through 2025, generated a negative free cash flow for the DRAM industry. And so,
you know, the industry was basically offering a free ride to all their downstream customers,
and the beneficiaries were the likes of, you know, Apple and other smartphone vendors,
HP, Lenovo in the PC world. And I think that that has led to the industry not wanting to
bring on too much capacity. And so, when you balance that with this spike in demand from AI,
and the data center is now driving, you know, more than 50% of DRAM and NAND demand,
that has created kind of a perfect storm. On top of that, we need HBM memory, high bandwidth memory,
for accelerators. And the great kind of issue around HBM is that for every one bit of HBM,
it's the equivalent, from a wafer perspective, of four bits or five bits of traditional DRAM.
So, if you're converting over traditional DRAM to HBM, you're removing significant incremental
supply. And so, that has left us to where we are today, which is significant undersupply.
Now, pricing has ripped higher, as you mentioned, and earnings power is pushing aggressively higher.
Micron, you know, we think will do $200 in 2027, which means, yes, the stock is at a trillion
dollars, but it's only trading five times. And the historical peak multiple for Micron has been,
you know, about six to eight times. So, you could argue just, you know, this time is not different,
but on relative peak, they're still upside. But then, you know, I think you take it one step
further, and it is a little more different this time, where there was no excess clean room
capacity. So, we haven't been able to add capacity like the industry would like. We are seeing that
incrementally, but our take is that the supply-demand dynamic for both DRAM and NAND only
gets worse in 2027. And so, what that means is data center players are going to have to pay up.
Consumer players, like PCs and smartphones, are either going to despec, raise pricing,
or see their unit volumes lower and decline year-on-year. And we think we're going to see all
three. You know, just last week, Apple raised their pricing across the board for their products
to reflect higher memory pricing. This is a trend we think that will last for multiple years. And so,
that is different. And so, that's the opportunity that we see near-term. And then, I think the big
news from Micron when they reported last week was that they expect to have 50% of their bids
under contract for the next five years. Now, there's a pricing ceiling in floor. And so,
the gross margins that are underneath the contract will be lower than what they're earning today,
but they get the confidence and the certainty of higher margins for a five-year duration,
which I think will drive a higher re-rating of the multiple. And so, I think those contracts
that they announced were a big deal. And moreover, one last thought, it changes the whole pricing
discovery mechanism. When you are sold out 50% of your bids already day one of your quarter,
you don't have to be aggressive. You don't have to blink during negotiations. Historically,
many of these memory companies are only 20% to 30% sold out in the first two months of any
quarter. Now, if they stay to the historical norm, there'll be 70% to 80% done. And at the last month
of the quarter, if customers want to be aggressive with them, they don't have to do the business.
They can cut their utilization a little, build inventory a little. That changes the whole
dynamic. So, I think the earnings power for the memory industry is sustainably higher. And I think
that that will enable the group to re-rate and see multiples push higher. And for that reason,
we're still very bullish on Micron, SanDisk, et cetera. Is the supply going to increase and
is it possible for other competitors and new players to come into this market if there's
that much demand? Yeah, great question. So, we're adding about 8% to supply this
year. We're going to add about 12% next year. But if you remember that 4 for 1, 5 for 1 hit as you
transition to HBM, that is the peak number on top of about 15% growth that you'll get from
just normal shrinks. So, think of 15 plus 8, 23, 15, 12, 27, but then haircut it for the HBM
transition because you're growing much faster. It means supply is going to grow for DRAM overall
about 19%, 20% this year and about 23%, 24% next year. And it's hard to get a good
vision for what the actual growth number is from a demand perspective, but I would
guesstimate that it's above 30%. And so, despite adding meaningful capacity,
we're going to remain in shortage. Now, that has opened the door for the one other DRAM player in
the market, CXMT in China. And they're about 8% of the market. They're IPO-ing this summer.
And there's a worry in the market that they will be aggressive. And I guess what I would say to
that is they don't produce the high quality bits for the data center. So, they would really be
focused on the consumer market. And we have heard most recently Apple asking Washington,
DC for approval to buy DRAM from CXMT. And we'll see if that gets approved. But I don't think it
changes the data center story here. And that's the most important part of the story. And so,
I don't really worry right now about CXMT and the added supply from them to the consumer market.
Now, a naysayer would say, well, it's a fungible market. So, if you're taking share in that market,
maybe that opens up bits for other parts of the market. And that's a fair comment. But I think
given it's only 8%, their yields are dramatically lower than the three multinational leaders.
I just don't see them being terribly disruptive. And the last point would be they need to get in
line for equipment to build out their capacity. ASML is the clear bottleneck. And so, we'll see
how much tooling they're able to get. And then the second point is China is restricted from EUV
lithography. And all three of the three other players have it. And so, if you're thinking about
high quality in the data center, next generational offerings, again, not having access to EUV will
be a real challenge for CXMT to compete with, to deal with. You mentioned AI demands are not
very clear. And there has been a lot of money that has been spent and they're going to get
spent on data centers. Where do you think this demand go from here? Is it going to slow down?
And if it slows down, what does it look like for this space? No, I mean, I'd say that everything,
every indication we're getting from the hyperscalers is that their CapEx is going to surprise
meaningfully to the upside in 26 and 27. I think that new financial markets are emerging.
I was talking to a gentleman in the private equity, private credit world,
and he said that the GPU, TPU, XPU will be a multi-trillion asset class. And so, directionally,
I think that's where we're going. And I think two other thoughts, right? One is computers sold out.
So, that's elongating the cycle, making it more durable for the investments. And then the other
thought is when you look at Anthropic and OpenAI and the tremendous revenues that they're generating
as agentic AI has taken off, you are seeing on the true demand side downstream,
magnificent growth. We've never seen the type of growth that we're seeing, for example,
at Anthropic in the history of time. And so, I think the demand is there. I think we've got
tremendous engineers that are adding value, both hardware and software. And then that will drive
cost per token per watt to decline 95 plus percent per year. And I think Jevon's paradox
is alive and well, and that will absolutely support strong demand over multiple years
for DRAM and for AI, sir. And there's a refresh cycle as well. So, is this demand becoming more
of a recurring demand for these RAM players, memory players? Yes and no. There is a specific
lifetime for memory. It does wear down. And for DRAM, you really care about speeds and feeds.
For NAND, you really care about the quality and the read-write cycles. So, the lifetime is not
forever, but I don't normally think about the market as a replacement-driven cycle. It's really
just overall bit demand. And so, I wouldn't rely on replacement. If that's what you're relying on,
then that would probably make me nervous on the cycle. Are you seeing us being in the early stages
based on what you're saying, or you think we're in the middle of this growth?
We're talking about numbers that were unimaginable just two or three years ago.
So, in that sense, I think the market is worried that we might be in the fifth,
sixth, or seventh inning, if I use baseball parlance. But my view is that if we're really
going to drive the cost per token per watt down by 95 plus percent per year, that that will drive
more and more demand for compute. And in that regard, here we are in 26, we're two plus years
into the real build-out. I think this lasts until 2030, 2031. That's my base case assumption.
And that would suggest we're in the fourth inning.
So, where do you see the revenue for these companies going at that stage? Do you think
the growth will stop, or do you think the revenues will go down from there?
It's important to continue to remember that semiconductors are cyclical. We will
probably overbuild at some point. And that means a correction of some certain magnitude that
will be painful for investors. Our working assumption is that is beyond 2030. So,
that will definitely come. It's not if, it's when. On the other hand, I don't want to underestimate
the intelligence of engineers and entrepreneurs in how they harness AI. And I do think you're
going to see tremendous new opportunities that we can't even contemplate today that will continue
to deliver tremendous GDP growth and productivity growth that will be fantastic for the global
economy. And if that's right, then the world's going to continue to need more and more compute.
They're going to want faster and faster compute. They're going to want lower power, lower latency.
And those are drivers for better silicon. And so, I do think we will have a pause at some point,
but the growth will absolutely return. So, from what you're seeing,
this is kind of the beginning stages of this build out.
I think that agentic AI and anthropic through enterprise was your first meaningful,
powerful signal of what AI could bring to the table. And we haven't seen
kind of other areas emerge. And so, from that perspective, I do think we're early.
Now, in terms of these memory companies, going back to them,
are there any bottlenecks for them in terms of materials, raw materials, and other areas that
people should pay attention to or other opportunities that could come up? Because
as they are growing their volume, they might benefit some other spaces. Are there any places
people should watch out for? Yeah. I mean, I think, you know,
substrates, advanced packaging, semi-equipment, these are all areas that either are in shortage
today or will be in shortage soon. So, those are the three major areas that I would focus on if
I'm an investor looking beyond kind of memory and looking at memory derivatives.
Do you have certain companies that you're looking at in those areas?
On the substrate side, it's 100%, you know, Asian companies that I don't cover. So, I would
pass that on for investors to speak to people on the ground covering the companies there.
In the equipment space, front-end equipment, you know, the best levered names to memory
are Lam Research, LRCX, Applied Materials, AMAT, and ASM Lithography, ASML. And then in
advanced packaging, you know, it's smaller for AMAT and Lam. It's very meaningful for Onto. So,
O-N-T-O is a name that my colleague Matt Prisco really likes. So, that's where we'd be focused
on those two areas. And CJ, you've been talking to these CEOs, you've been covering these companies.
Are they surprised by what's happening to them right now?
Unequivocally, yes. I think that very good CEOs, you know, saw the chat GPT moment and began to
for stronger growth. But to level set, you know, as an example, semi-equipment,
you know, pre-COVID, we were a $65 billion business. You know, we saw a huge demand for
PCs, smartphones, and data center, you know, as we were living in the work-from-home kind of world.
And WFE went to $100 billion. You know, we're now talking about by 2030,
a stretch goal of perhaps as much as $350 billion WFE. So, $65, $350, I don't think anyone, you know,
two, three years ago would have thought that was even a possibility. But that gives you the best
kind of understanding of how, you know, important this AI infrastructure build-out is. I mean,
this is the once-in-a-lifetime kind of impact that I will see. I won't see anything like this.
And, you know, I was here during the kind of internet and internet bubble. You know,
we had the advent of the smartphone, advent of the data center, you know, automotive becoming
more important for semiconductors. But, you know, this AI infrastructure build-out, you know,
makes everything else pale in comparison. I know you put the price target of $1,500
on Micron in the next 12 months. NVIDIA, you're seeing a good amount of growth for them. Your
price target is at $300. Can you tell us more about this space? You know, NVIDIA, in our view,
is the unequivocal leader in the accelerator space, which is, you know, the most important
product to deliver AI or to deliver the lowest cost per token per watt. What they're doing
is kind of industry altering in the sense that they've moved to an annual technology cadence.
And so, they are upgrading every 12 months. And that is having a tremendous impact on the entire
supply chain as everyone tries to keep up with their innovations. And whether it's going to
whether it's building their own CPUs, DPUs, building their own GROK kind of decode inference
accelerators to the adoption of CPO as opposed to copper, they're driving, you know, much of
the great innovation. And so, you know, one of our big takeaways for NVIDIA is that they're
increasing their competitive mode every year. On top of that, they have a great mode on the
software side. And so, from a technology perspective, we think they are by far the
leader. On the other hand, you know, the hyperscalers who are spending, you know,
in aggregate, a trillion dollars this year in CapEx, they'd love to put pressure, you know,
on their costs and drive down lower costs. And so, what have we seen? We've seen a desire for
custom silicon. And Google is on their ninth generation TPU. And that is a great chip. You
know, if Jensen were here, he'd say, you know, there's three great chips and there's the two,
you know, Hopper Blackwell from NVIDIA and the TPU from Google. And so, they're going merchant
and selling to Anthropic, to OpenAI, potentially others. So, that's a competitive threat.
You've got Amazon building their own Tranium. They're working with Marvell,
as well as others. And you also have AMD, right, coming in. And, you know,
they have signed agreements with OpenAI, Anthropic, Meta. And so, you know, I think
they're being a fast follower. And so, you know, I think within this world, we're going to be
probably on a revenue basis, 70% GPUs, 30% custom silicon. And as of right now, I think the key
takeaway is that compute is sold out. It's sold out through 28, 29. And that this is a rising
tide that should lift all boats. And so, within that world, you know, we've been talking about
NVIDIA, Broadcom, AMD as our three favorite compute names. And then I'd say what we've seen
in the last six to 12 months is the impact of agentic AI on server CPUs. And so, this has been
a big game changer where the traditional or historical makeup for a server was you'd have
about one CPU for every four or eight GPUs. But with agentic AI, more and more of the workloads
are optimized for the CPU. And the goal, given how expensive it is, is to run your GPU at 100%
utilization or as close to that as possible to best utilize that more expensive chip. And so,
what does that mean? It means you need more CPUs. And so, we're actually seeing the ratio go from,
you know, 1 to 8, 1 to 4, now 1 to 1. And that is driving tremendous demand for CPUs, particularly
x86, which is Intel and AMD, where customers really are relying on reliability, serviceability,
and security. And that is, I think, differentiating x86 from ARM. And the winner there is both AMD and
Intel. But I think particularly AMD, they're coming to the market with multi-thread solution.
Intel is only single thread in 2027. And then probably more importantly,
is that the relationship between AMD and TSMC has only improved, you know, greatly in the last kind
of 6 to 12 months. And I think today, TSMC really views, you know, NVIDIA, Apple, and AMD as their
tier one plus customers. And that's the rank order and revenues that TSMC sees out of its top
customers. And so, they will prioritize wafers to AMD for agentic servers. And so, we think they're
going to surprise the market in terms of their growth, 26, 27, and beyond.
Now, CJ, you also met with the NVIDIA leadership team, and you were more confident about the
company. What did you see that kind of made you be certain about NVIDIA?
I've had the good fortune of spending some time with Jensen, Colette, the CFO, and the rest of the
team. And, you know, Jensen is one of the most competitive people I've met on Earth.
He is not going to sit still. He wants to be the leader. He has deep memories of when he almost,
NVIDIA, almost went bankrupt in the 1998 kind of timeframe. And so, he will be relentless
in driving technological advancements to make sure that NVIDIA keeps its lead. And so, you know,
when we look at that stock, you know, we think they can earn 20 plus dollars in 28. We think
they can do 15 in 27 or more. And so, right now, the stock is trading less than 10 times
2028 earnings. And if the S&P at 23 times, trading at half the multiple of the market is,
you know, in my view, utterly insane. I get it, though. The market looks at NVIDIA and says it's
$5 trillion market cap. Everyone already owns it. You know, who's the incremental buyer?
We're worried that maybe Google TPU is a real competitive threat. Same thing, AMD.
Same thing, Amazon, Tranium. But again, I think when you think about how aggressive they are in
both hardware and software, it's hard to want to fight against, you know, their technological lead.
And so, for that reason, you know, I do continue to think that they will maintain their number one
position in the market. They'll maintain their top gross margins in the market. And they will
continue to deliver year in, year out, you know, excellent earnings power. So, you know, that's
definitely a name we have high conviction in. The only question is really more of what's the
catalyst. And I'm not sure, because, you know, year to date, the stock is only up 3%, you know,
the stocks is up 92%. And so, hard to call. But I would say, you know, for patient investors,
we will eventually see a catch up trade here. Now, do you think this movement in AI would
trickle down to the rest of the economy? Do you have any thoughts on that?
I mean, I think the answer is 100%. I would expect GDP in 27, 28 and beyond to start to,
you know, show upside. And I think that's where you'll see it. I think from an enterprise
perspective, you know, maybe one in eight globally have fully embraced AI. So, you know, what that
tells me is we're really early in adopting AI meaningfully, and really changing business
practices. So, I think there's optimization and adoption that is still very much on the come.
But make no mistake about it, that AI is life altering, I think, for every enterprise. Because
of that, if you're not adopting AI, your competitor is, and that is going to put you at a competitive
disadvantage. And, you know, you then you'll have new winners emerge. And, you know, I think that
will drive, you know, excellent incremental growth for the industry. You have seen the technology
advancement in the last few decades. People are worried about employment. Do you think
this will happen that we're going to have significant amount of unemployment? Or do
you think things will change? There'll be more innovation in newer types of jobs?
Personally, as I think about the impact of my business, over time, you know, my team of five,
maybe we could do our business just four. So, you know, 20% less kind of employment,
because we're just that much more productive. On the other hand, new industries are going to emerge.
When the internet first came to fruition, we could never have imagined the same thing
will happen here. So, if we're looking at this on a 10 plus year period,
I don't think we'll have too many things to complain about.
CJ Muse, it was great to have you on Market Insider.
Yeah, great to spend time with you.
What do you think about the memory stocks and the run in AI space? Do you think it will continue?
Make sure to tell us in the comments. I'm Siamak. This is Market Insider. We'll see you next time.