Loading summary
A
In case you missed it today, we're bringing you a special encore release of a recent episode. We'll be back tomorrow with a brand new episode.
B
Welcome to Thoughts on the Market. I'm Tom Wig, head of Specialty Sales in the Americas at Morgan Stanley and a sector specialist in technology, media and telecom. We wake up every day to new AI product releases. So it's easy to lose sight of the unprecedented nonlinear improvement in AI capabilities. But things are about to get weird. It's Tuesday, April 28th at 8am in New York. The market has been thinking about AI in linear terms, but we need to reframe that assumption of only incremental improvement and think about exponential improvement. That was my takeaway from a conversation with Steven Byrd, Global Head of Thematic and Sustainability Research at Morgan Stanley. In our conversation, we zeroed in on Steven's bull case for broader AI model improvements. First, I want to talk about one obsession that you've been writing about for the last several months. Is this idea that we're going to see nonlinear improvements in the frontier models coming out this spring.
A
Yes.
B
There's been, you know, some big headlines around new models, benchmarks coming out publicly. Is this, you know, your bull case playing out on these models and what are the implications?
A
Absolutely, Tom. So we have. To your point, we are obsessed, and I know, I'm not shy about that, with the nonlinear rate of AI improvement, it is the most important impact to so many stocks that I can think of in the sense that it can impact all industries, all business models. So what we've been saying for some time is if you look back over the last couple of years at the relationship between the amount of compute used to train these LLMs and the capabilities, we have a very clear scaling law and approximately the law is if you increase the training compute by 10x, the capabilities of the models go up by 2x. Now, as you've now talked about this a lot, just meditate on that for a moment. I think things are about to get weird in the sense that on the positive side, we're going to see all kinds of underappreciated capabilities across many industries. So this disruption discussion I think is going to spread, but it's also going to require investors going to be more thoughtful about what they do with that concept. Meaning you can't sell everything in the sense that AI will disrupt some businesses. I actually think this is healthy in some ways because now it forces investors to really look at each business model and assess which is going to get disrupted, which can get supported and enabled by AI which are immune because there are some business models actually are immune. But essentially from here, Tom, I'd say we are expecting through the spring and summer to see multiple models that are able to perform a much greater percentage of the economy at better levels of accuracy at incredibly low cost, which I know you and I have talked a lot about the cost of actually doing this work from the LLMs. This is massive. This is going to impact so many industries. I think this is all to the good for the AI infrastructure plays because it shows the importance of getting more intelligence out into the world.
B
So you mentioned the constraints we're seeing across compute memory and power. It seems like Most of the CEOs of the labs and hyperscalers are talking about this. Investors are bullish in terms of the ownership in memory, optical, semi cap, et cetera. But the question I'm getting more recently is around what's the ROI on all the spending? And does the market action in these hyperscalers, which has been pretty bearish year to date, force a cut on capex? So maybe if you can marry that with what you're picking up on the ground in terms of compute spend and whether the frenzy still continues versus the ROI and what could happen.
A
Yeah, the short answer I'm going to go through detail is I think the bullishness is going to get more bullish over the coming months. Let me walk you through a couple of the mathematics and then just what I'm seeing on the ground to your point, Tom. So the mathematics. We have a token economics model that looks from the perspective of a hyperscaler or an LLM developer in terms of if they sell their token at a certain price and you fully load the cost of a data center and all associated costs financing, you name it, in what are the returns. And the bottom line is the returns are excellent. The other element we spend a lot of work on and you and I talk a lot about is the demand for compute in this world where the LLMs are increasing in capability and the token usage goes way up with agentic AI, video world models, all that stuff. We think that there is a massive shortage of compute. So if you're lucky enough to be a hyperscaler with the compute, with the power, we think that they will have a lot of pricing power on the tokens. Let me explain why we see pricing power on the tokens. Now I'm going to flip to the perspective of an adopter. Let me give you just rough mathematics. There was a study last year from one of the big labs showing that on average, an enterprise user using an LLM might be able to replace work that would take about one and a half hours. From a human, that would save about $55 of cost. A million tokens depends on whether you're looking at input or output, but let's just call it $5 for a million tokens. The average usage case today for a fairly complex agentic task in an enterprise setting is in the tens of thousands of tokens. Okay, let's just do that math again. $55 of savings. Okay, a million tokens cost $5. And a typical agentic usage is far less than the million tokens today, though that will accelerate. The economics are a home run for adopters. We're in a situation where compute is very scarce. I see pricing power all over the place for those who have the compute and have the power.
B
When you put it like that, Steven, it seems so inevitable and obvious. But I wonder why the hyperscalers are trading the way they are and when do they see the revenue inflection you're talking about. Is this like a stay tuned kind of 2026 event? Is this something we have to wait for for 2027, 2028? Like, how do you think this flows through to the extent that the market will get more comfortable, that all this free cash flow pressure is worth it on the other side?
A
Yeah, this is. In short, I think this is a 2026 event. But let me dive into that because what you just asked is so important for so many stocks. So let's talk through this. The capabilities of the models are advancing so fast that the average corporate user is not yet keeping up. There is this gap, but that will happen quickly. And we're seeing signs from these labs of revenue at the lab level that is accelerating. So that's a good sign. What we're seeing, though, among fast adopters is those adopters who really understand the capabilities are quickly realizing just how economically beneficial. There is an example. One of my best friends founded a software company many years ago. Last month was. That was the last month in which his programmers wrote code. They're. They're done with writing code. The efficiency benefits for his business are absolutely massive. But he feels like he's just scratching the surface. And he's about as technically capable as anyone I know. He has two PhDs in the subject. He's. He's very, very good. So long way to say that. We're living in almost two worlds where the fast adopters will show what's possible. The average utilization for enterprises will still take Some time. But I do think that the market will react to what see from the fast adopters in the sense of the tangible economic benefits are so big now on the ground. What I'm seeing on the infrastructure side, my friends in power tell me that a couple months ago is when they saw this sense of urgency from the AI community go up a couple of notches for them to get the infrastructure they need. So they saw this explosion in compute coming in the last two months. The weekly usage of tokens according to open routers up a couple hundred percent a couple months. So I do think we're seeing this. So this is, it's happening quickly. What I would say is the market will have these signposts in every industry of early adopters showing this benefit. I think that's enough for us to start to get bullish. We also I just think when you look at the demand for compute, the compute numbers need to go up and with that, you know, everything in the AI value chain, infrastructure value chain, the volumes need to go up.
B
One bear case that I wanted to interrogate was there's one view that yes, there's a token explosion right now, but it's because the first use case is coding which is inherently, you know, very developer friendly and token intensive relative to other knowledge work. Can you talk about, you know, whether you subscribe to that or whether the token intensity will be as high or lower as this expands to other areas of knowledge work in the next several years?
A
Yeah, it's a great question. The short version is that yes, it's true that software usage is more token intensive. However, what we're going to be seeing, we're starting to see it is in almost every knowledge based job we're going to move to agentic AI and when we do that you tend to see an explosion in compute. Let me walk you through the numbers. There are a couple of studies that show essentially when you go from a query based usage of LLMs to an agentic use for any occupation, you see about a 10x increase in token usage per use of those models. And you can see why. I've anecdotes of some of my friends who are newer to this who set their agents loose overnight to do non coding work and in the morning they get some pretty amazing results. But they also used a lot more tokens.
B
Like a five grand credit card bill.
A
Exactly. It's like maybe next time you put a few parameters around that. But long way to say it's agentic across every workflow that I can think of that will still result in an explosion in token demand.
B
It's definitely a good idea to put some parameters around your agentic workflow. My thanks to Steven for that conversation, and thank you for listening. Let us know what you think of the show by leaving us a review where you listen and if you find thoughts on the market worthwhile, tell a friend or a colleague about us today.
A
The preceding content is informational only and based on information available when created. It is not an offer or solicitation, nor is it tax or legal advice. It does not consider your financial circumstances and objectives and may not be suitable for you.
Host: Tom Wig, Head of Specialty Sales, Morgan Stanley
Guest: Steven Byrd, Global Head of Thematic and Sustainability Research, Morgan Stanley
Date: May 7, 2026
Length: ~10 minutes
This special encore episode dives into the current state and profound future implications of artificial intelligence (AI), focusing on the rapidly accelerating, exponential improvements in AI models—particularly large language models (LLMs). Host Tom Wig discusses with Steven Byrd how these advances will disrupt industries, shift investment strategies, and present both massive opportunities and urgent challenges in terms of infrastructure and compute resources. The discussion probes into return-on-investment (ROI) for AI spending, the timeline for revenue inflection, and the evolving intensity of AI workloads.
[00:07–03:05]
[03:05–05:48]
[05:48–08:19]
[08:19–09:51]
On the exponential nature of AI:
“Things are about to get weird in the sense that … we’re going to see all kinds of underappreciated capabilities across many industries.”
— Steven Byrd, 01:41
On AI economics:
“The economics are a home run for adopters. … compute is very scarce. I see pricing power all over the place for those who have the compute.”
— Steven Byrd, 05:28
On industry impact:
“You can’t sell everything … investors have to really look at each business model and assess which is going to get disrupted, which can get supported and enabled by AI, which are immune.”
— Steven Byrd, 01:54
On the agentic AI surge:
“When you go from a query-based usage … to an agentic use for any occupation, you see about a 10x increase in token usage per use of those models.”
— Steven Byrd, 09:06
On real-world consequences:
“Like a five grand credit card bill.”
— Tom Wig, 09:37
“Exactly. Maybe next time you put a few parameters around that.”
— Steven Byrd, 09:39