
Can AI Chip Demand Outrun the Power Crunch?
Shownotes
The U.S. midterm results could lead to policy changes that worsen the power bottleneck for U.S. data centers. Meanwhile, the outlook for AI chips remains bullish. Our specialists discuss potential challenges and opportunities in the AI buildout.
Read more insights from Morgan Stanley.
----- Transcript -----
Tom Wigg: Welcome to Thoughts on the Market. I'm Tom Wigg from Morgan Stanley's TMT Specialty Sales. I'm joined by Joe Moore, U.S. Semiconductors Analyst, and Stephen Byrd, Global Head of Thematic Research.
Today, the disconnect between the AI power gigawatt shortfall, especially post-midterms, and a bullish outlook on AI semiconductors.
It's Friday, November 13th at 10am in New York.
Morgan Stanley already estimated a 30 gigawatt shortfall between now and 2028, and you revised that higher after the midterm. So maybe talk about our estimates, where they've gone now? And how you're thinking about this whole question around, you know, really Texas is kind of the ground zero for this conversation.
But it kinda happened in Ohio and other places as well…
Stephen Byrd: Absolutely. Yeah. I mean, obviously the election caused a lot of investors to focus on the political challenges, but this is not new per se. So, we've been talking about people, power, and politics being real bottlenecks on the ability to build out data centers in the United States.
Now, that creates a lot of incentive elsewhere around the world to build. But in the U.S., we have a lot of challenges in terms of the build-out. The election really underscored the importance of addressing data center concerns. A bunch of voters expressed concerns with data center development.
And to be fair, I think that's warranted in some cases because some of these data centers really are developed in ways that may not work for local communities. So, I understand the concern. Mathematically, what we essentially are looking at is that bottleneck getting larger in 2028 because of a couple of gubernatorial outcomes, Texas and Ohio being specific, where we think it's a delay.
It's not a cancellation. We actually think there's a path to getting these data centers built. It's just going to take a little bit longer because of the processes that those new governors are going to want data center developers to go through. So, we go from kind of 30 gigawatts up into the mid-30s in terms of the shortage.
But then in 2029, the deployment is much, much bigger. So, it is a relief valve, and in some ways – something Joe and I talk a lot about – helpful in some ways because this elongates that growth cycle as well.
Tom Wigg: What are the offsets that we can get? You know, [be]cause some clients will talk about, you know, Louisiana now became the new Texas that everyone…
Stephen Byrd: Yes.
Tom Wigg: … is excited about, you know; still a red state. Some clients are talking about Asia, particularly Malaysia as a region where labor's cheaper, there's no permitting. You know, maybe we can get a lot more gigawatts online.
Stephen Byrd: Mm-hmm. Yeah.
Tom Wigg: Google recently signed a deal there. Behind the meter, still an option. Grid optimization, probably not something I understand. But I know you've been a power exec, so maybe you do.
Stephen Byrd: Sure, sure.
Tom Wigg: Do you think we can kinda cobble together enough power in other ways where maybe your shortfall is too conservative?
Stephen Byrd: Yeah, I think there's a path where the shortfall is less than what we laid out, so let's go through all the options. But where there's a will and lots and lots of money and lots of innovation, there's often a solution.
So, let's go through the categories. So, off grid we like a lot, and we put in a forecast of what we think can be done in the United States. The challenge is time – in the sense that engineering those solutions, getting the labor, getting the permits, all of that takes time.
And an example, my home state of Virginia essentially outlawed on-site gas generation, in sort of anticipation that folks would go down this off-grid path. And they don't want to see those kinds of emissions.
So, we are running into challenges. When we tally that up, it's not as much power as I think many bulls would like. Just – it's not a question of demand, it's a question of all those challenges that I just mentioned. But there could be bigger solutions over time.
We have really good teams around the world that assess incremental demand and deployment. Just as you mentioned, Tom, Malaysia is one of those opportunities, though I would point to India, Thailand, a little bit in Japan, China.
On the other hand, we have issues in other parts of the world. So, Norway now has a moratorium on data centers. We're running into some problems in Australia. We're running into some problems in the United Kingdom, where a big, proposed UK project was terminated because the utility couldn't provide grid access fast enough.
So, it's a lot of puts and takes. When we add all this up, as of now, it's not enough to address the U.S. shortfall. But I would expect over time just more and more emphasis on getting solutions.
Now, one thing I do want to say where I'm not bullish is grid optimization. So, it's a cool idea. But let me just briefly explain why I think there's more hype than reality to this.
The idea is that the U.S. grid is actually only utilized on average, like, 50 percent of the time. Something like that, so there's clearly slack. The problem is that's not how a utility executive would look at the situation. Utilities manage the grid for the worst possible day or the worst possible situation. That would be a six-day winter storm, for example.
During a six-day winter storm, demand is at the peak, and they expect some of their power plants to trip offline because of the cold conditions. So, they're going to manage their system assuming every data center developer is going to want to have power and everybody else will.
So, the idea of being flexible and, and sort of filling in those gaps just doesn't work under the conservative mindset of a utility. So that part unfortunately I don't think is going to work.
Tom Wigg: Got it. And then before I switch to Joe, maybe if you can just put your political hat on for a second. I know you spend a lot of time in D.C., so if you can play pundit. There's a difference between, you know, running for office in an election and actually governing.
Stephen Byrd: Absolutely.
Tom Wigg: So, some of the commentary post-election from the bulls has been, okay, you know, this blue wave happened. But these blue governors are still going to want economic activity in their state. And if the AI companies approach this with, you know, more community-oriented solutions. In terms of, you know, giving back, building a school, building a football stadium in West Texas, paying teachers more. Whatever it is, you know, there's a better way to do this. And that should get a lot of these things through.
Is that your sense? Or is there something more against the data centers?
Stephen Byrd: Yes. No, I think that is my sense, at least in some states. And this is really the genesis of our big buy the drop note; was we think this is a bit of a pause in terms of the pace of data center growth. It is not a cancellation, in terms of what we see, to your point.
What I think we'll see is very smart approaches to developing data centers that work for all the constituents. Louisiana's a great example you mentioned earlier. We see projects where the developers are bankrolling new hospitals, pay raises for teachers and folks in police and fire service.
And when you do it in the right way, I think you can get support and create economic activity. Look at Louisiana. When you look at all of the activity in Louisiana. Not just data centers, but other industrial activity. That magnitude rivals large nations, in terms of the amount of economic activity in the state. And that is so good from a fundamental economic health of the state.
So, I do think there's a path. That's just going to take a little bit of time to work through.
Tom Wigg: Yep. Let's talk to Joe. So, semis have been very controversial. We had this tough summer, another sell-off in October. One of the bear cases is this power shortfall.
How is everything still so tight in semis from a supply demand standpoint when power is getting worse? And are we setting ourselves up for an oversupply situation where there's just no place to put all these chips that we're building?
Joe Moore: Yeah. So, I think for 2027, it's not just staying tight. It's getting tighter. Like, all of the component shortages are severe. You've had multiple CEOs from hyperscalers say memory is their biggest problem, not land power shell in the short term.
That doesn't mean land power shell's not becoming a bigger and bigger focus. So, we have to see what's going to happen with each individual project. I think the nice thing about selling ubiquitous compute is you can have three people that don't have a place to put it and five other people who want the thing that they didn't have a place to put it.
So, I think it's still very strong into 2028. I would kind of argue that what we're hearing is it's very challenging. If you sort of take something in Texas offline, you take something in Ohio offline that we thought we had; we have to really scramble and find another place to put it.
I think there's still the view from the people we talk to is that it's not going to slow us down. We'll find a way. And talking to Stephen, it's sort of hard to see exactly what that way may be. A lot of the international stuff may be an option. We'll have to see what the economy looks like.
In a weaker economy, I think you're much more prone to allow people to build data centers [be]cause it's jobs and things like that. So, we have… There's a lot of unknowns out to [20]28, and any semiconductor analyst talking two years out is hard to say.
But Stephen made this point. This is a duration kind of trade now, and we know it's really good. Other bottlenecks are a big risk that we need to contemplate. And I mean, the interesting thing about memory too is consumers – PC, smartphone, you know, those businesses are very weak.
And every quarter we've gone through this where the PC numbers aren't that good and there's inventory and... But the fulfillment at the data centers is really low, like 50 percent. And so that's the true north in all of this. And what that does to me for memory is it keeps the stocks very range bound. But look at the range.
I mean, we're down 60 percent, 70 percent. And we're trading at like three times earnings. So, any kind of sustainability to some of these stocks and, you know, you'll be able to buy back half the company in two years. So, I'm pretty bullish on all of it.
And I think – and none of that is discounting anything that Stephen is saying.
And I think it does reframe a little bit the competitive dynamic. Because, you know, am I taking my restricted land power shell and using it on something that may or may not work? We'll see. But I think the overall ecosystem is still very good, and I think the kinds of arguments you're hearing now is that [20]27's good.
What about '28? We heard that at the end of [20]24, [20]25. Eventually it'll be right. Eventually there will be a year where you're like, "Okay, it is actually slowing down in that year." It just doesn't feel that way to me. And given that the market is basically saying now that it does, we're being valued as it slows in [20]28. And you know, I think that visibility's going to still be good.
Tom Wigg: So, like you said, you know, the, the stocks are now essentially baking in 2028 rolling over. You think that's too bearish. When does the market appreciate that? Will we get some sort of 2028 visibility in the middle of next year. Kinda like we did this year? Do we need a new AI use case to make people think about, you know, incremental tightening in the supply chain?
Joe Moore: Yeah.
Tom Wigg: Is there more clarity on AI financing? One of the above, all of the above? Like, how do you think about that? And like, you know, when you looked at past cycles, I know this is sort of unprecedented, but like how does that play out?
Joe Moore: Yeah, I think it's very... You know, the worst case is it's just kind of a grind, and we have to prove that [20]28 is good because we have 12-month lead times, and 12 months from now we think [20]28 is good.
And quarter by quarter we continue to feel okay about it. But we are going to learn a lot about how these models are used. We're starting to see, you know, the companies that are going to come public now are the ones who actually… There's actually cost of sales for them. You know, we've seen cloud service providers doing great, but somebody's renting from them.
So, I think getting some visibility into the economics of how AI's actually getting used will be really important. You know, maybe bad, I don't know. But I think that's what will help us get that visibility. Certainly the confidence. I mean, the Silicon Valley… The dissonance between Silicon Valley optimism and Wall Street pessimism has never been this wide.
And it's not even; Wall Street pessimism is the wrong word [be]cause I feel like, you know, 60 percent of the people will be bullish, but they're very easily swayed by a couple people who pitch the -- They know things are falling apart. And so, there's like a not… The conviction isn't there that maybe you'd like to see.
I think that conviction can come back very quickly in the right sort of economic circumstance. I don't think guiding to 2030 is the answer when you're in a duration trade, you know. That just creates risk. It makes it hard for the stocks to move forward because you're already guiding for a 75 percent four-year CAGR.
I need to each and every quarter, you know, validate that. But I think in general, just put up numbers and the world will come around to if it's sustainable or not. I mean, the market's smart and the market will decide.
Tom Wigg: Got it. Stephen, Joe, thanks for speaking with me today. And thanks for listening. If you enjoyed the show, please leave us a review wherever you listen. And share Thoughts on the Market with a friend or colleague today.