- Publisher show notes · Excerpt 1
The biggest AI stocks have had a remarkable run – but questions still remain. Our Head of Americas Specialty Sales, Thomas Wigg, speaks with Global Head of Thematic and Sustainability Research Stephen Byrd and Global Head of Public Policy Research Ariana Salvatore about the competition and durability of the investment cycle.
- Publisher show notes · Excerpt 2
Read more insights from Morgan Stanley.
- Publisher show notes · Excerpt 3
----- Transcript -----
- Publisher show notes · Excerpt 4
Thomas Wigg: Welcome to Thoughts on the Market. I'm Tom Wigg, Morgan Stanley's Head of Americas Specialty Sales.
- Publisher show notes · Excerpt 5
Stephen Byrd: I'm Stephen Byrd, Morgan Stanley's Global Head of Thematic and Sustainability Research.
- Publisher show notes · Excerpt 6
Ariana Salvatore: And I'm Ariana Salvatore, Morgan Stanley's Head of Public Policy Research.
- Publisher show notes · Excerpt 7
Thomas Wigg: Today, the rally in AI CapEx beneficiaries has taken a breather in recent weeks on concerns of competition from open-source models, backlash to token-maxxing, and growing political opposition to data center builds.
- Publisher show notes · Excerpt 8
It's Tuesday, July 7th at 10am in New York.
- Publisher show notes · Excerpt 9
Let's start with you, Stephen. There's a lot of discussion recently around a backlash at token-maxxing. Essentially, enterprises trying to curtail their high spending on AI tokens from the frontier labs, and, in many cases, shifting to cheaper open-source China models.
- Publisher show notes · Excerpt 10
Can you first offer some perspective here on the value of tokens for enterprises? I know you have a popular token factory model that walks through the economics of agents.
- Publisher show notes · Excerpt 11
Stephen Byrd: Yeah, Tom, we do have this model that really walks through token economics, both from the adopter side as well as the hyperscaler side. So, let's do the adopter side.
- Publisher show notes · Excerpt 12
So, there's a study out that shows a whole range of enterprise use cases of AI, and the average single use case that they identify would save a company about $55 or provide that much benefit. And while we don't know exactly how many tokens it will require, we can make some educated guesses as to a typical token usage to achieve that $55 outcome.
- Publisher show notes · Excerpt 13
And we know that a typical American model, though this varies a lot, you can think of as the cost per million tokens being in the range of $5 per million. Some will be lower, some will be higher. So, for a few dollars of token cost, an enterprise can generate benefit of $55.
- Publisher show notes · Excerpt 14
So that doesn't make me overly concerned about token spend and concerns about token-maxxing. I know we're going to get into that, but the foundation here is really good in the sense that enterprise use cases are very much in the money.
- Publisher show notes · Excerpt 15
Thomas Wigg: How do you think market share ultimately shakes out on tokens? Do the cheaper models overtake the frontier AI labs? Do tokens bifurcate based on the complexity of workloads? How do you think this plays out?
- Publisher show notes · Excerpt 16
Stephen Byrd: What we continue to see is this relentless pace of innovation and cost reduction. So, the frontier keeps going out – meaning model capabilities continue to increase, and, with that, we see enterprise adoption growing quite a bit.
- Publisher show notes · Excerpt 17
Long way to say there is a role for both the frontier as well as these open-source models, and we'll continue to see both flourish. What I see is a lot of tokens will be spent on open-source models. A lot of the value will be in the higher end models because that's where enterprises are going to go. Let me give you an example.
- Publisher show notes · Excerpt 18
I was speaking with one of our programmers about a recent project, and he used a very high-end coding tool, an American coding tool. And for him, that incremental cost of the tokens was very much worth it. And here's a very practical example as to why it makes sense for many enterprises to use the higher end models.
- Publisher show notes · Excerpt 19
If a coding tool gets one of the thousands of lines of code wrong, the cost to remediate is very, very high. In other words, that incremental cost – in this example I'm thinking of, it's a few dollars incremental cost – is so worth it because if the quality is not there, the cost to any enterprise to go back and remediate is so high.
- Publisher show notes · Excerpt 20
And that's true in a lot of enterprise use cases, but not in every use case.
- Publisher show notes · Excerpt 21
And what we are seeing is these open-source models that are cheaper will be very good for a variety of more mundane use cases that are still very valuable.
- Publisher show notes · Excerpt 22
That said, what we've seen in data from places like OpenRouter is dollar-weighted, meaning valued by enterprise spend, the vast majority is still the proprietary models.
- Publisher show notes · Excerpt 23
But even within proprietary models, we could have more expensive and less expensive models.
- Publisher show notes · Excerpt 24
You do not need to go to the frontier.
- Publisher show notes · Excerpt 25
Where I come out on all this is that I'm very confident that the demand for compute is going to exceed the supply.
- Publisher show notes · Excerpt 26
What is difficult to exactly know is who are the winners, what is the exact mix.
- Publisher show notes · Excerpt 27
But the fundamentals of the demand for compute look extremely strong.
- Publisher show notes · Excerpt 28
Thomas Wigg: So, I think you just gave me the answer, but I do want to bring this all back to AI CapEx. Now, last year, when the market sold off on Deep Seek concerns, the concept of Jevons paradox ultimately prevailed, where the cheaper pricing led to even greater demand and CapEx went higher.
- Publisher show notes · Excerpt 29
Do you think the same plays out here?
- Publisher show notes · Excerpt 30
Stephen Byrd: It does look that way very much. And the Jevons paradox dynamic is what we still see today in the sense that as the models get better, what we can do with the models increase, the cost of tokens will keep dropping, the cost of compute will keep dropping.
- Publisher show notes · Excerpt 31
But let's talk about what might derail that, just to make sure we're thinking about all the risks.
- Publisher show notes · Excerpt 32
If somehow commoditized models could perform at the same level as proprietary models in all situations, then I would feel differently.
- Publisher show notes · Excerpt 33
What I see is that these newer models really do have capabilities that are fairly breathtaking and that are worth that extra money.
- Publisher show notes · Excerpt 34
But if somehow, we hit a wall where these models aren't getting better and therefore the sort of the open models are going to catch up, then I'd feel differently about that.
- Publisher show notes · Excerpt 35
This is where Ariana will, will come in in terms of policy and, you know, this comes up a lot when we think about U.S. versus China.
- Publisher show notes · Excerpt 36
How do we think about, you know, access to different models?
- Publisher show notes · Excerpt 37
How do we think about the cost of different models?
- Publisher show notes · Excerpt 38
What about the risk of appropriation of capabilities by the Chinese firms, for example? That comes up a lot in policy circles. But the base case that I have is this just looks more like Jevons paradox, and there's going to be continued innovation, continued reduction in the cost of producing these services from these models. That looks like more of the same.
- Publisher show notes · Excerpt 39
Thomas Wigg: Let's shift to Ariana to talk about the political angle here. The cover of Barron's over the weekend was a guy wearing a no data centers T-shirt. And this does seem to be one of the few bipartisan issues of agreement heading into the midterms.
- Publisher show notes · Excerpt 40
The stat that the article gave was that 75 data center projects worth $130 billion were blocked or delayed in 1Q26, which is equal to the total number for 2025. This is according to Data Center Watch.
- Publisher show notes · Excerpt 41
Now, most of this is in blue states like New York, Michigan, Illinois, Minnesota considering a statewide moratorium, but you're also seeing Pennsylvania, Arizona, Ohio, parts of Texas restricting tax incentives here.
- Publisher show notes · Excerpt 42
So as this gets louder into the midterms, how do you think this plays out?
- Publisher show notes · Excerpt 43
Ariana Salvatore: So, this is definitely one of the big wedge issues, not just for the midterm elections, but for 2028. And to your point, it's expanding into something that's got bipartisan momentum behind it.
- Publisher show notes · Excerpt 44
Our view is that as long as the Trump administration is in power, something like a federal ban is unlikely to come to fruition.
- Publisher show notes · Excerpt 45
That's because we think the administration is still broadly supportive of the AI data center build-out.
- Publisher show notes · Excerpt 46
And I think even if you were to see a Democrat in office further down the road, that position is the same.
- Publisher show notes · Excerpt 47
And the reason is, it's just too difficult to imagine the U.S. giving up that strategic imperative relative to China.
- Publisher show notes · Excerpt 48
So, while it is true that voters are against AI, while it is true that you ar