Speakers discuss data center constraints, European sovereign AI demand, and agentic AI hardware shifts
Summary
Paul Walsh, Emmett Kelly, and Lee Simpson discuss constraints on data center rollouts, including rising construction costs and power grid limitations. They highlight how European data sovereignty laws are driving demand for local infrastructure, particularly from telcos entering the market. The speakers analyze the shift toward agentic AI, noting a hardware transition from GPUs to CPUs for inference workloads. Finally, they address the critical bottleneck of power availability, comparing grid connection timelines across different European nations.
Key takeaways
- 1 Data center construction costs have risen 12% year-on-year and are projected to accelerate to 15% in the following two years, driven by inflation in building inputs.
Unidentified speaker (speaker_1) · 2:03
- 2 So tech companies are very, very mobile. If they can't build in location A, they will move to location B. So where would location B? That would be other states within the US that would be more planning permission friendly. And also there could be some spill. spillover into other countries and there could be an opportunity here for European data center builders and beyond that into Asia, where we've earmarked countries such as South Korea, India, and other Asian countries where you could see greater than expected data center build.
Unidentified speaker (speaker_1) · 3:23
- 3 European data sovereignty requirements are driving demand for infrastructure that is sovereign from the software layer down to the physical data center, particularly for sensitive sectors like banking, government, pharmaceuticals, and defense.
Unidentified speaker (speaker_1) · 5:02
- 4 the layers. There's a compute layer that will carry on being done by the GPU, but the orchestration, particularly inference, is going to move to a CPU. And that CPU is going to be the new compute resource driven by this greater use of agentic AI.
Unidentified speaker (speaker_2) · 6:47
Episode sections
- 0:00 – 4:43
Data Center Rollout Constraints and US Policy Impact
Paul Walsh introduces conference themes on data center rollouts and power constraints. Emmett Kelly identifies new constraints beyond power and planning, citing political changes in the US and financing blocks, while noting that existing projects will continue but new builds may slow due to consultation with the new administration.
- 4:43 – 6:27
European Sovereign AI and Telco Infrastructure
Emmett Kelly discusses the demand for sovereign data centers in Europe to host sensitive data in banking, government, and defense. He notes that telcos are entering the data center market despite higher build costs, driven by insatiable corporate demand and strong pricing power.
- 6:27 – 9:39
Agentic AI and Hardware Architecture Shifts
Lee Simpson describes agentic AI as an 'application moment' driving a shift where inference moves from GPUs to CPUs. Emmett Kelly adds that telcos view agentic AI as a middle step between generative and physical AI, with network quality becoming a key differentiator against churn.
- 9:39 – 13:06
Power Shortfalls and Efficiency Innovations
Emmett Kelly highlights power as the primary constraint, citing Ireland's data centers using 25% of national grid capacity. Lee Simpson details future hardware innovations like 800-volt architectures and silicon carbide to improve power efficiency at the server rack and edge levels.
Claims worth verifying
Data center build costs are up 12% year-on-year and expected to accelerate to 15% next year and the year after.
Unidentified speaker (speaker_1) ·2:03
How to verify: Morgan Stanley research report on data center build costs; check for '12%' and '15%' inflation estimates in recent TMT conference summaries or equity research notes from Morgan Stanley's European Telco/Data Centers team. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
US data center capacity growth will be very strong in 2027, but may slow in 2028-2029 due to consultations with the new administration.
Unidentified speaker (speaker_1) ·2:43
How to verify: Morgan Stanley research report; verify the timeline of 'strong 2027' and 'slowdown 2026-2026' in their data center outlook. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
South Korea and India are earmarked for greater-than-expected data center build.
Unidentified speaker (speaker_1) ·3:42
How to verify: Morgan Stanley research report; check for specific mentions of South Korea and India as beneficiaries of US planning permission constraints. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Data centers in Ireland use 25% of national grid capacity.
Unidentified speaker (speaker_1) ·9:57
How to verify: Irish Grid Operator (EirGrid) annual reports or national energy statistics; verify the percentage of total national consumption attributed to data centers. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Securing a grid connection in the UK takes between five and eight years.
Unidentified speaker (speaker_1) ·11:16
How to verify: UK Grid Connection Provider (e.g., National Grid ESO) published timelines or industry reports on grid connection waitlists. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
800-volt architecture is coming and will involve silicon carbide and gallium nitride.
Unidentified speaker (speaker_2) ·11:37
How to verify: Morgan Stanley research report on power electronics or semiconductor trends; verify the mention of '800-volt' and 'silicon carbide/gallium nitride' in the context of server rack power delivery. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Agentic AI is the second step of AI adoption, following generative AI and preceding physical AI.
Unidentified speaker (speaker_1) ·7:41
How to verify: Morgan Stanley research report on telco AI strategies; verify the three-step framework: generative, agentic, physical. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Europe needs sovereign data centers for sensitive data in banking, government, pharmaceuticals, aviation, and defense.
Unidentified speaker (speaker_1) ·5:02
How to verify: Morgan Stanley research report on European data sovereignty; verify the list of industries cited as requiring sovereign hosting. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Telcos are starting to build their own data center infrastructure despite higher costs.
Unidentified speaker (speaker_1) ·5:59
How to verify: Morgan Stanley research report on European telcos; verify claims of telco capex shifts into data center infrastructure. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Agentic AI inference is moving to CPUs, while compute remains with GPUs.
Unidentified speaker (speaker_2) ·6:47
How to verify: Morgan Stanley research report on hardware trends; verify the claim of 'separation of layers' with inference shifting to CPUs. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Do your own research
- Verify construction cost inflation data for data centers using industry reports from firms like JLL or CBRE.
- Check US federal and state-level planning regulations for data centers to assess potential delays in 2028-2029.
- Review European Union data sovereignty regulations (e.g., GDPR, Data Governance Act) to understand compliance requirements for banking and government sectors.
- Analyze hardware architecture trends from major chipmakers (NVIDIA, Intel, AMD) regarding CPU vs. GPU demand for agentic AI inference workloads.
- Assess telco network quality metrics and churn rates to evaluate competitive positioning in the agentic AI era.
- Compare grid connection wait times and power availability data for UK, France, and Finland from national grid operators.
- Investigate construction cost inflation trends for building inputs like steel, concrete, and electrical components.
- Look for counter-arguments regarding the timeline of agentic AI adoption versus generative AI deployment in enterprise settings.
Full transcriptexpand
0:00Unidentified speaker (speaker_0): Welcome to Thoughts on the Market. I'm Paul Walsh, Morgan Stanley's Head of European Research, and we're bringing you a special episode today from Morgan Stanley's 26th European TMT conference currently underway. And the central themes we're focused on today relate to the data center rollout story, the need for sovereign AI, the rise of AI agents, and of course the power shortfall and the potential shift in focus towards power efficiency. I'm joined by my colleagues, Emmett Kelly, head of European telco and data centres here in Europe, and Lee Simpson, head of European Technology Hardware. It's Thursday, November the 19th at 8 a.m. in Barcelona.
0:42Unidentified speaker (UNKNOWN): It's
0:42Unidentified speaker (speaker_0): fantastic to have you here. We've got over 160 corporates, over 500 investors registered from over 250 different institutions. What an opportune moment for us to reflect on the wider TMT space given some of the market moves that we've seen in recent months. And so we'd love to hear both your thoughts on maybe four key thematic areas that I think are very much in focus. Emmet, let's start with the data center rollout question. It's something you and the global team have talked about in your research. We've talked about funding constraints and how that impacts the rate of rollout and an increasing shift in focus for investors from 2027 to 2028. What have you learned on that topic?
1:23Unidentified speaker (UNKNOWN): Yeah,
1:24Unidentified speaker (speaker_1): data centers are front and center at the TMT conference here in Barcelona this week. I think looking back over the last couple of years, there have been two main constraints: power and planning. Power is getting access to energy cheap and fast. And planning is getting proposals approved by local government and central government. Now we see new constraints emerging, and those constraints are politics. So we've had a change in the United States after the midterm elections, and then financing, getting access to capital. to build data centers is becoming an area of blockage. In our recent report, we added another constraint onto data center build, which was price. We've noticed very, very heavy inflation in building data centers. We estimate that the cost of building a data center is up twelve percent year on year, and we expect this inflation rate to accelerate to fifteen 15% percent. next year and the year after. In terms of what's happening in the United States, our view and our feedback that we've received is kind of threefold. Firstly, where there are shovels in the ground, those projects will continue uninhibited, and those data
2:43centers will get built. Therefore, we think data center capacity growth in the states will be very strong in 2027. However, there will be a consultation period between data center builders and the new administration, looking at areas such as planning permission, unionized labor, usage of water, use of renewables, etc. And this could drive a slight slowdown in data center build in 2028, and that will roll forward into 2028. 2029. In terms of the kind of second implication, our view is that data centers will build wherever they can. So tech companies are very, very mobile. If they can't build in location A, they will move to location B. So where would location B? That would be other states within the US that would be more planning permission friendly. And also there could be some spill. spillover into other countries and there could be an opportunity here for European data center builders and beyond that into Asia, where we've earmarked countries such as South Korea, India, and other Asian countries where you could see greater than expected data center build.
3:59Unidentified speaker (UNKNOWN): And
3:59Unidentified speaker (speaker_0): Lee, coming to you with the same question, what are your observations?
4:02Unidentified speaker (UNKNOWN): Yeah,
4:02Unidentified speaker (speaker_2): I think that's absolutely correct. I don't think anyone's querying the near-term demand. Really, the debate is how can we look out to 28? How can we be comfortable that we have a build-out story that still has momentum, still has growth? So everyone's querying the order book at this point. How resilient is this? How fixed are those orders? And what could evaporate if indeed the market turns a little sour, depending on what the climate the would be at that point?
4:27Unidentified speaker (speaker_0): Understood. So I want to talk about sovereign AI capability here in Europe, both in terms of compute capacity but also in terms of frontier or open source model capability. Again, Emmett, what are you hearing on that front from your meetings today?
4:43Unidentified speaker (speaker_1): So data sovereignty has been a core discussion point here again this year. This was also the case last year. because we've had huge shifts in geopolitics over the last couple of years. There is a viewpoint that Europe needs its own data centers for its own data. Europe has got huge flows of very sensitive data, in particular in industries such as banking and finance, government, pharmaceutical. pharmaceuticals, aviation, defense, military, and this information needs to be hosted on sovereign European platforms. And these platforms need to be sovereign from top to bottom, from the software layer all the way down to the physical data center. A number of telcos are now beginning to build data centers. And despite the fact that data centers are becoming more expensive to build, the feeling in Europe is the demand from corporates for data center capacity is really quite insatiable. It's growing. There's very good pricing power in the industry. So telcos, they're a little bit late to the game, but they are
5:59starting to build their own data center infrastructure.
6:03Unidentified speaker (speaker_0): I'd agree. I think this notion of there still being time to establish sovereign capabilities in Europe is a message that I think has come across in a number of the different presentation tracks that we've attended. Something that Morgan Stanley Research has written a lot about is the rise of agentic AI. Um and it's notable to me that this is something which is being talked about pretty much in every meeting. So Leah, what are your insights into the agentic piece from the meetings you've had today?
6:27Unidentified speaker (UNKNOWN): So
6:27Unidentified speaker (speaker_2): I think agentic AI is a really interesting space and just to level set everyone, the way we think about it is it's a little bit akin to an application moment on the smartphone. So where you see apps coming, we're going to see agents coming for cloud AI and maybe even at the edge eventually as well. But what's happening at the compute level is you're seeing a separation of the layers. There's a compute layer that will carry on being done by the GPU, but the orchestration, particularly inference, is going to move to a CPU. And that CPU is going to be the new compute resource driven by this greater use of agentic AI. Now, in the main, agentic AI is delivering a huge new capability set for both enterprise and other general users alike. So agentic AI is a theme that's big here this year. No surprise. really. I imagine it'll feature again next year, frankly, and something that we're seeing most of our companies lean into. Yeah,
7:21Unidentified speaker (speaker_1): I think telcos are becoming a little bit more comfortable in how they view agentic AI. So there is a transition taking place at the moment. There's three steps of AI that telcos are cognizant of. The first step is generative AI, the second step is agentic AI, and the third step is physical AI. So agentic AI is the one in the middle. Now telcos by and large seem willing to engage with agents. So some of the largest telcos have said they will deploy their own agents to deal with third-party agents, but their agents, so the telco agents, will not negotiate with the agentic AI agents. The second takeaway was telcos that have got best network or are the best value feel that agentic AI could actually present an opportunity for them. So if agentic AI does trigger churn and customers moving between telcos, the operators with the best network feel that they can win market share
8:24because one of the most common customer complaints in telcos. is my network isn't very good. So the best networks can win market share. There is a feeling that the one area you do not want to be is stuck in the middle. So you either want to have best network or offer better value, best value. But you don't want to be an operator with an average network and average prices. Because in an agentic AI world, you might get squeezed.
8:46Unidentified speaker (speaker_0): And And the other thing, maybe just as an extension of that, is the migration of AI to edge applications. as well. What have you heard around that?
8:53Unidentified speaker (speaker_2): Yeah, I think at some point we will be transferring much of this back out to the edge. What we've done in the cloud will come to the edge. And here it will have to be a lot more power efficient. It'll have to be secure if you're going to run agentic AI at the edge. So the possibility again that some of the computer compute architectures have to change to allow for that security. It is not something that's front and center. yet, but you can see that it's lurking maybe 12-18 months off as a theme.
9:20Unidentified speaker (speaker_0): And finally, Emmett, from my side, the power shortfall is one of the limitations that you and the global team have written about extensively in terms of data center rollout. But one of the things that does do is shift the focus towards power efficiency and efficiency more broadly as we're adding more and more compute. Would love to hear your thoughts on what you're hearing.
9:39Unidentified speaker (speaker_1): So power has been the greatest constraint in the data center world for the last five years. And to give you an indication of where data center usage of power can go, you just need to look at the Irish data center market. Ireland now sees 25% of its national grid capacity used by data centers. It's the most advanced data center. market in the world on a per-pop basis, but it just shows how power-hungry data centers are. The feedback I've had from the TMT conference is that data center owners would, in an ideal world, like to have access to nuclear energy. Nuclear comes with a very big baseload, Base load. it comes with potential behind the fence pricing. It comes with water and security as well. However, in say the real world, my feeling is that data center owners would like to have a smorgasbord of power solutions. So they'd like to have a mix of grid, backup with batteries, maybe some access to nuclear, renewable energy, wind, solar. It's a combination of different power solutions. But power is definitely going to remain front and center of
10:57the data center debate in Europe. To get a grid connection in the UK can take between five and eight years. Some countries are very much at the forefront of offering power solutions to data center operators, and I would highlight France and Finland as markets where the power offering. is ahead of other European countries.
11:17Unidentified speaker (speaker_2): I mean, power is it's a huge currency in AI, be it in the supply of power to the data center, be it in the efficient use across the data center, particularly in the server rack, or even at the edge device. All of this is going to continue as a theme, and I think this will roll for many years to come. The sort of things that we're hearing at this point is in the server. ract will be new powered architectures. Of course, everybody knows there's an 800-volt architecture coming at some point. This will involve new compound semis like silicon carbide and gallium nitride, but also new device architectures closer to the processor. So you can deliver power or even voltage potential much Mu closer to that processor as efficiently as you can. As you mentioned, we've got to think about the actual delivery of power from the grid to the data center as well. And perhaps also if indeed AI moves out to the edge, particularly on phones or other edge devices, that compute has to be done much more efficiently because you're gonna be limited by a lithium-ion battery. And on top of that, you have to be done securely. So a a lot of evolution around power efficiency to come. Certainly a theme that's not going away.
12:26Unidentified speaker (speaker_0): Yeah, certainly an area of innovation for sure. Emmett, Lee, thanks very much for joining me as always.
12:33Unidentified speaker (speaker_2): Thank you.
12:33Unidentified speaker (speaker_1): Thanks very much, Paul.
12:35Unidentified speaker (speaker_0): And thank you for listening. Let us know what you think about this and other episodes by leaving us a review wherever you get your podcasts. And if you enjoy thoughts on the market, please tell a friend or colleague to tune in today.
12:49Unidentified speaker (UNKNOWN): The
12:49Unidentified speaker (speaker_3): 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
13:03Unidentified speaker (UNKNOWN): you.