How Software Firms Are Using AI Internally

Live from Morgan Stanley’s TMT conference, our panel break down where AI is already delivering real returns—and where rapid advances are raising new risks.

Read more insights from Morgan Stanley.

----- Transcript -----

Michelle Weaver: Welcome to Thoughts on the Market. I'm Michelle Weaver, U.S. Thematic and Equity Strategist here at Morgan Stanley.

Today we've got a special episode on AI adoption. And this is a first in a two-part conversation live from our Technology, Media and Telecom conference.

It's Thursday, March 5th at 11am in San Francisco.

We're really excited to be here with all of you taping live. And we've got on stage with me. Stephen Byrd, he's our Global Head of Thematic and Sustainability Research; Josh Baer, Software Analyst; and Lindsay Tyler, TMT Credit Research Analyst.

So, Stephen, I want to start with you, pretty broad, pretty high level. We recently published our fifth AI Mapping Survey that identifies how different companies are exposed to the broad AI theme. Can you just share with us some insights from that piece and how stocks are performing with this AI exposure?

Stephen Byrd: Yeah, it's interesting. I mean, we've been doing this survey now, thanks to you, Michelle, and your excellent work, for quite a while. And every six months it is pretty telling to see the progression.

I would say a few things that got my attention from our most recent mapping was the number of companies that are quantifying the adoption benefits continues to go up quite a bit. And to me that feels like that's going to be table stakes very soon as in every industry you see two or three companies that are really laying out quite specifically what they expect to be able to do with AI and lay out the math. I think that really is going to pull all the other companies to follow suit. So, we're seeing that in a big way.

We do see adopters, with real tangible benefits performing well. But a new thing that we're seeing now, of course, in the market is concerns that in some cases adoption can lead to dramatic deflation, disruption, et cetera. That's coming up as well. So, we're seeing greater concerns around disruption as well.

But broadly, I'd say a proliferation of adoption, that that universe of companies continues to grow, increases in quantification of the benefits. So, that is good. What's really surprised me though, is the narrative among investors has so quickly moved from those benefits which we've talked about into flipping that to toggle all negative, which I know some of our analysts have to deal with every day. The mapping work suggests significant benefits. But the market is fast forwarding to very powerful AI that is very disruptive in deflation. And that's been a surprise to me.

Michelle Weaver: Mm-hmm. Josh, I want to bring software into this. Your team has been arguing that AI is actually good for software. And it's really something that you need that application layer to then enable other companies to adopt AI. Can you tell us a little bit about how much GenAI could add to the broader enterprise software market? And how are you thinking about monetization these days?

Josh Baer: Of course. I think the best starting place is a reminder that AI is software, and so we see software as a TAM expander. And in many ways, even though this is extremely exciting innovation, it's following past innovation trends where first you see value accrue and market cap accrue to semiconductors, and then hardware and devices, and then eventually software and services. And we do think that that absolutely will occur just given [$]3 trillion in infrastructure investment into data centers and GPUs.

There's got to be an application layer that brings all of these productivity and efficiency gains to enterprises and advanced capabilities to consumers as well. And so we see AI more as an evolution for software than a revolution. An evolution of capabilities and expansion of capabilities. LLMs and diffusion engines absolutely unlocked all of these new features of what software can do. But incumbents will play a key role in this unlock.

And our CIO surveys really support that. Quarterly we ask chief information officers about their spending intentions, and these application vendors who we cover in the public markets are increasingly selected as vendors that companies will go to, to help deploy and apply AI and LLM technologies.

So, to answer your question, we estimate GenAI could unlock [$]400 billion in incremental TAM for software; for enterprise software by 2028. And this is based on looking at the type of work able to be automated, the labor costs associated with that work, the scope of automation, and then thinking about how much of that value is captured typically by software vendors.

Michelle Weaver: And you have a bit of a different lens on AI adoption. So, what are some of the ways you're hearing software customers using these AI tools and anything interesting that popped up at the conference?

Josh Baer: To echo what Stephen laid out, I mean, all of our software companies are using AI internally, both to drive efficiencies, but also to move faster. So thinking about product. Innovation, you know, the incumbents are able to use all of the same coding tools and, you know, …

Michelle Weaver: Mm-hmm.

Josh Bear: … products geared to developers to move faster and more efficiently on R&D. So, they're doing more. From a sales and marketing perspective, a G&A perspective, every area of OpEx, our software companies are in a great position to deploy the AI tools internally.

I think more important[ly], speaking to this TAM and expanded opportunity, is our companies have skews that they're monetizing. It might be a separate suite that incorporates advanced AI functionality. It might be a standalone offering, or it might be embedded into the core platform because the essence of software is AI and it, you know, leading to better retention rates and acceleration from here.

Michelle Weaver: Mm-hmm. And Stephen, going back to you on the state of play for AI, we had the AI labs here and we heard a lot about the developments and what's to come. So, what's your view on the trajectory for LLM advancements and what are some of the key signposts or catalysts you're watching here?

Stephen Byrd: Yeah, this is for me, maybe the most important takeaway of the conference – is this continued non-linear improvement of LLMs, which we've been writing about for quite some time. And just to give you an example, we think many of the labs have achieved a step change up in terms of the compute that they have, in some cases 10 x the amount of compute to train their LLMs. And that [if] the scaling laws hold – and we see every sign that they will – a 10x increase in compute used to train the models results in about a doubling of the model capabilities.

Now just let that sink in for a moment. Let's just think about that. A doubling from here in a relatively short period of time is difficult to predict. It's obviously very significant and I think several of the LLM execs at our event sounded to me extremely bullish on what that will be. A lot of that I think will be evident in greater agentic capabilities.

But also, I'd say greater creativity. It was about three weeks ago, three of the best physics minds in the world worked with an LLM to achieve a true breakthrough in physics – solving a problem that had never been solved before. A couple of days ago, a math team did the same thing. And so, what we're seeing is sort of these breakthrough capabilities in creativity. This morning I thought Sam speaking to, you know, incredible increases in what these models can do – which also brings risk. You know, I think it was interesting he spoke to, you know, the risk of misalignment, the risk of what these models are doing.

But for me, that's the single biggest thing that I'm thinking about, and that's going to be evident in the next several months.

Michelle Weaver: Mm-hmm.

Stephen Byrd: So, you know, on the positive side, it leads to greater benefits from AI adoption. And to Josh's point that, you know – more and more the economy can be addressed by AI, I do get concerned about the risk that that kind of step change will create greater concerns about disruption and deflation.

That causes me to think a lot about that dynamic. Interestingly, we think the Chinese labs will not be able to keep pace just for one reason, which is compute. We think the Chinese labs have everything else they need. They have the talent, the infrastructure. They certainly have the energy, power. But they don't have the chips.

If what we laid out with the American models turns out to be true, I could see a chain reaction where the Chinese government pushes the Trump administration for full transfer of the best technology to China. And China could use their rare earth trade position to ensure that. So, that's sort of the chain reaction I've been thinking about.

Michelle Weaver: Mm-hmm. So, let's think about then bottlenecks in the U.S. Power is still one of the main bottlenecks. We had several of the solutions providers here at the conference. So, what are you thinking in terms of the size of the power bottleneck in the U.S. and how are we going to fix that?

Stephen Byrd: Yeah, absolutely. I am bullish on the companies that can de-bottleneck power, not just in the U.S., a few other places. Let's go through the math in terms of the problem we face and then the solution.

So, we have this very cool – it is cool if you're a nerd – power model that starts in the chip level up, from our semiconductor teams. And from that, we build a global power demand model for data centers. We then apply that to the U.S.

Through 2028 we need about 74 gigawatts of data centers, both AI and non-AI to be built in the United States. I don't think we'll be able to achieve that for lots of reasons. But starting from that 74, we have sort of 10 gigs that have been recently built or are under construction. We have 15 gigs of incremental grid access, but after those two, we have to go to unconventional solutions, meaning typically off-grid solutions, over 40 gigawatts of unconventional solutions.

So that will be repurposing Bitcoin sites, which could be sort of 10 to 15 gigawatts. That'll be big. Renewable energy, fuel cells will be part of the solution. Gas turbines will be a big part of the solution. Co-locating at a few nuclear plants. I'm less bullish than I used to be on that. But when we net all that out, we think the U.S. is likely to be 10 to 20 percent short of the data center capacity that will need to be in.

It's not just a power grid access issue, though, that's a big one. Labor is now showing up as a huge issue. Many of the companies I speak to trying to develop data centers struggle with availability of labor. Electricians being one very tangible example. In the U.S. we need hundreds of thousands of additional electricians.

So, for any of your children, like mine, thinking about careers, you know, you'd be surprised [at] the amount of money that people are making in the infrastructure business that does feel like it's a labor shift that's going to have to happen, but it's going to take years. So, in that context, we had a number of the Bitcoin companies at our event here. And the economics of turning a Bitcoin site into hosting a data center are extremely attractive. I mean, extremely attractive.

To give you a sense of that. Before this opportunity presented itself to these Bitcoin players, those stocks tended to trade at an enterprise value per watt of about $1 to $2 a watt. Then we started to see these deals in which the Bitcoin players build a data center and lease them to hyperscalers. Those deals – depends a lot on the deal but – have created between $10 and $18 a watt of value. Let me repeat that. 10 to 18 – relative to where these stocks were at 1 to 2.

Now many of these stocks have rerated, but not all of them. And there's still quite a bit of upside. And what we've noticed is the economics that the hyperscalers are paying are trending up and up and up. Because of this power shortage that we're dealing with. So, a lot of exciting opportunities are still in the power space.

Michelle Weaver: Great. Well, I think that's a good place to wrap this first part of our conversation around AI adoption and the state of play. We'll be back again tomorrow with Part Two, looking at financing and risks.

To our panelists, thank you for talking with me. And to our audience, 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.