Our Semiconductors and Software analysts Joe Moore and Keith
Weiss dive into the biggest market debate around AI and why it’s
likely to shape conversations at Morgan Stanley’s Technology,
Media and Telecom (TMT) Conference in San Francisco.
----- Transcript -----
Joe Moore: Welcome to Thoughts on the
Market. I'm Joe Moore, Morgan Stanley's Head of U.S.
Semiconductors.
Keith Weiss: And I'm Keith Weiss, Head of
U.S. Software.
Joe Moore: Today on the show, one of the
biggest market debates in the tech sector has been around AI and
the Return On Investment, or ROI. In fact, we think this will be
the number one topic of conversation at Morgan Stanley's annual
Technology, Media and Telecom (TMT) conference in San Francisco.
And that's precisely where we're bringing you this episode from.
It's Monday, March 3rd, 7am in San Francisco.
So, let's get right into it. ChatGPT was released November 2022.
Since then, the biggest tech players have gained more than $9
trillion in combined market capitalization. They're up more than
double the amount of the S&P 500 index. And there's a lot of
investor expectation for a new technology cycle centered around
AI. And that's what's driving a lot of this momentum.
You know, that said, there's also a significant investor concern
around this topic of ROI, especially given the unprecedented
level of investment that we've seen and sparse data points still
on the returns.
So where are we now? Is 2025 going to be a year when the ROI and
GenAI finally turns positive?
Keith Weiss: If we take a step back and
think about the staging of how innovation cycles tend to play
out, I think it's a helpful context.
And it starts with research. I would say the period up until When
ChatGPT was released – up until that November 2022 – was a period
of where the fundamental research was being done on the
transformer models; utilizing, machine learning. And what
fundamental research is, is trying to figure out if these
fundamental capabilities are realistic. If we can do this in
software, if you will.
And with the release of ChatGPT, it was a very strong,
uh, stamp of approval of ‘Yes, like these transformer models
can work.’
Then you start stage two. And I think that's basically November
22 through where are today of, where you have two tracks going
on. One is development. So these large language models, they can
do natural language processing well.
They can contextually understand unstructured and semi structured
data. They can generate content. They could create text;
they could create images and videos.
So, there's these fundamental capabilities. But you have to
develop a product to get work done. How are we going to utilize
those capabilities? So, we've been working on development of
product over the past two years. And at the same time, we've been
scaling out the infrastructure for that product development.
And now, heading into 2025, I think we're ready to go into the
next stage of the innovation cycle, which will be market uptake.
And that's when revenue starts to flow to the software companies
that are trying to automate business processes. We definitely
think that monetization starts to ramp in 2025, which should
prove out a better ROI or start to prove out the ROI of all this
investment that we've been making.
Joe Moore: Morgan Stanley Research projects
that GenAI can potentially drive a $1.1 trillion dollar revenue
opportunity in 2028, up from $45 billion in 2024. Can you break
this down for our listeners?
Keith Weiss: We recently put out a report
where we tried to size kind of what the revenue generation
capability is from GenerativeAI, because that's an important part
of this ROI equation. You have the return on the top of where you
could actually monetize this. On the bottom, obviously,
investment. And we took a look at all the investment needed to
serve this type of functionality.
The [$]1.1 trillion, if you will, it breaks down into two big
components. Um, One side of the equation is in my
backyard, and that's the enterprise software side of the
equation. It's about a third of that number. And what we see
occurring is the automation of more and more of the work being
done by information workers; for people in overall.
And what we see is about 25 percent, of overall labor being
impacted today. And we see that growing to over 45 percent over
the next three years.
So, what that's going to look like from a software perspective is
a[n] opportunity ramping up to about, just about $400 billion of
software opportunity by 2028. At that point, GenerativeAI will
represent about 22 percent of overall software spending. At that
point, the overall software market we expect to be about a $1.8
trillion market.
The other side of the equation, the bigger side of the equation,
is actually the consumer platforms. And that kind of makes sense
if you think about the broader economy, it's basically one-third
B2B, two-thirds B2C. The automation is relatively equivalent on
both sides of the equation.
Joe Moore: So, let's drill further into
your outlook for software. What are the biggest catalysts you
expect to see this year, and then over the coming three years?
Keith Weiss: The key catalyst for this year
is proving out the efficacy of these solutions, right?
Proving out that they're going to drive productivity gains and
yield real hard dollar ROI for the end customer. And I think
where we'll see that is from labor savings.
Once that occurs, and I think it's going to be over the next 12
to 18 months, then we go into the period of mainstream adoption.
You need to start utilizing these technologies to drive the
efficiencies within your businesses to be able to keep up with
your competitors. So, that's the main thing that we're
looking for in the near term.
Over the next three years, what you're looking for is the
breakthrough technologies. Where can we find opportunities not
just to create efficiencies within existing processes, but to
completely rewrite the business process.
That's where you see new big companies emerge within the software
opportunity – is the people that really fundamentally change the
equation around some of these processes.
So, Joe, turning it over to you, hardware remains a bottleneck
for AI innovation. Why is that the case? And what are the biggest
hurdles in the semiconductor space right now?
Joe Moore: Well, this has proven to be an
extremely computationally intensive application, and I think it
started with training – where you started seeing tens of
thousands of GPUs or XPUS clustered together to train these big
models, these Large Language Models. And you started hearing
comments two years ago around the development of ChatGPT that,
you know, the scaling laws are tricky.
You might need five times as much hardware to make a model that's
10 percent smarter. But the challenge of making a model that's 10
percent smarter, the table stakes of that are very significant.
And so, you see, you know, those investments continuing to scale
up. And that's been a big debate for the market.
But we've heard from most of the big spenders in the market that
we are continuing to scale up training. And then after that
happened, we started seeing inference suddenly as a big user of
advanced processors, GPUs, in a way that they hadn't before. And
that was sort of simple conversational types of AI.
Now as you start migrating into more of a reasoning AI, a multi
pass approach, you're looking at a really dramatic scaling in the
amount of hardware, that's required from both GPUs and XPUs.
And at the same time the hardware companies are focused a
lot on how do we deliver that – so that it doesn't become
prohibitively expensive; which it is very expensive. But
there's a lot of improvement. And that's where you're sort of
seeing this tug of war in the stocks; that when you see something
that's deflationary, uh, it becomes a big negative. But
the reality is the hardware is designed to be deflationary
because the workloads themselves are inflationary.
And so I think there's a lot of growth still ahead of us. A lot
of investment, and a lot of rich debate in the market about this.
Keith Weiss: Let's pull on that thread a
little bit. You talked initially about the scaling of the GPU
clusters to support training. Over the past year, we've gotten a
little bit more pushback on the ideas or the efficacy of those
scaling laws.
They've come more under question. And at the same time, we've
seen the availability of some lower cost, but still very
high-performance models. Is this going to reshape the investments
from the large semiconductor players in terms of how they're
looking to address the market?
Joe Moore: I think we have to assess that
over time. Right now, there are very clear comments from
everybody who's in charge of scaling large models that they
intend to continue to scale.
I think there is a benefit to doing so from the standpoint of
creating a richer model, but is the ROI there? You know, and
that's where I think, you know, your numbers do a very good job
of justifying our model for our core companies – where we can
say, okay, this is not a bubble. This is investment that's driven
by these areas of economic benefit that our software and internet
teams are seeing.
And I think there is a bit of an arms race at the high end of the
market where people just want to have the biggest cluster. And
that's, we think that's about 30 percent of the revenue right now
in hardware – is supporting those really big models. But we're
also seeing, to your point, a very rich hardware configuration on
the inference side post training model customization. Nvidia
said on their on their earnings call recently that they see
several orders of magnitude more compute required for those
applications than for that pre-training. So, I think over time
that's where the growth is going to come from.
But you know, right now we're seeing growth really from all
aspects of the market.
Keith Weiss: Got it. So, a lot of really
big opportunities out there utilizing these GPUs and ASICs, but
also a lot of unknowns and potential risks. So, what are the key
catalysts that you're looking for in the semiconductor space over
the course of this year and maybe over the next three years?
Joe Moore: Well, 2025 is, is a year that is
really mostly about supply.
You know, we're ramping up, new hardware But also, several
companies doing custom silicon. We have to ramp all that hardware
up and it's very complicated.
It uses every kind of trick and technique that semiconductors use
to do advanced packaging and things like that. And so, it's a
very challenging supply chain and it has been for two years. And
fortunately, it's happened in a time when there's plenty of
semiconductor capacity out there.
But I think, you know, we're ramping very quickly. And I think
what you're seeing is the things that matter this year are gonna
be more about how quickly we can get that supply, what are the
gross margins on hardware, things like that.
I think beyond that, we have to really get a sense of, you know,
these ROI questions are really important beyond 2025. Because
again, this is not a bubble. But hardware is cyclical and
there; it doesn't slow gracefully. So, there will be periods
where investment may fall off and it'll be a difficult time to
own the stocks. And that's, you know, we do think that over time,
the value sort of transitions from hardware to software.
But we model for 2026 to be a year where it starts to slow down a
little bit. We start to see some consolidation in these
investments.
Now, 12 months ago, I thought that about 2025. So, the timeframe
keeps getting pushed out. It remains very robust. But I think at
some point it will plateau a little bit and we'll start to see
some fragmentation; and we'll start to see markets like, you
know, reasoning models, inference models becoming more and more
critical. But that's where when I hear you and Brian Nowak
talking about sort of the early stage that we are of actually
implementing this stuff, that inference has a long way to go in
terms of growth.
So, we're optimistic around the whole AI space for
semiconductors. Obviously, the market is as well. So, there's
expectations, challenges there. But there's still a lot of growth
ahead of us.
So Keith, looking towards the future, as AI expands the
functionality of software, how will that transform the business
models of your companies?
Keith Weiss: We're also fundamentally
optimistic about software and what GenerativeAI means for the
overall software industry.
If we look at software companies today, particularly application
companies, a lot of what you're trying to do is make information
workers more productive. So, it made a lot of sense to price
based upon the number of people who are using your software. Or
you've got a lot of seat-based models.
Now we're talking about completely automating some of those
processes, taking people out of the loop altogether. You have to
price differently. You have to price based upon the number of
transactions you're running, or some type of consumptive element
of the amount of work that you're getting done. I think the other
thing that we're going to see is the market opportunity expanding
well beyond information workers.
So, the way that we count the value, the way that we accrue the
value might change a little bit. But the underlying value
proposition remains the same. It's about automating, creating
productivity in those business processes, and then the software
companies pricing for their fair share of that productivity.
Joe Moore: Great. Well, let me just say
this has been a really useful process for me. The collaboration
between our teams is really helpful because as a semiconductor
analyst, you can see the data points, you can see the hardware
being built. And I know the enthusiasm that people have on a
tactical level. But understanding where the returns are going to
come from and what milestones we need to watch to see any
potential course correction is very valuable.
So on that note, it's time for us to get to the exciting panels
at the Morgan Stanley TMT conference. Uh, And we'll
have more from the conference on the show later this week. Keith,
thanks for taking the time to talk.
Keith Weiss: Great speaking with you, Joe.
Joe Moore: And thanks for listening. If you
enjoy Thoughts on the Market, please leave us a review wherever
you listen and share the podcast with a friend or colleague
today.
Kommentare (0)
Melde dich an, um einen Kommentar zu schreiben.