2026.10.01 | TB | Power | AI Data Center Electrical Infra: Meta Expert

Client-led • 17 Sep 2026

Companies: ABBN CH, ETN US, META US, LR FR, JCI US, MSFT US, SU FR, NVDA US, 2308 TW, 2301 TW

Expert Former Senior Executive, High-end Compute Infrastructure at Meta Platforms Inc

Agenda

• Comprehensive understanding of gigawatt dynamics at AI data centres, noting capacity additions, costs and pricing evolution, as well as Microsoft's plan to move from 12GW to 38GW by 2032
• Can speak to AWS (Amazon Web Services) and Google's total installed capacities, as well as data centre categories where prices could be decreasing, highlighting solid-state transformers
• Well-versed in the rationale for deploying an 800V architecture for hyperscalers, emphasising the impact on electrical providers when moving from traditional to 800V
• Knowledgeable on whether Legrand's share within Meta's CAPEX spending is increasing, as well as its ability to forecast AI electrical demand

Questions

00:05:43] Everybody in the market is talking about gigawatts of AI data centres, and their capacity. How much does it cost today to build 1GW of AI capacity, and how has this cost evolved over the last few years? What if I asked you this question in 2022? What's your opinion on the cost dynamics? Who are the big drivers of this evolution?

Expert [00:06:42]: I would say definitely compared to 2022, the demand for data centre capacity has been significantly rising. The major driver for this rise is the AI training and AI inferencing as a major driver.

Client 1 [00:08:00]: You said they will deploy 2.5GW, but what would be Meta's total capacity for 2026?

Expert [00:08:13]: Yes. You will see that Meta is certainly aggressively adding more capacity on an YoY basis.

00:08:54] Regarding costs related to capacity additions, I read some articles mentioning USD 40bn CAPEX for 1GW, but there's also an estimate at USD 50bn or USD 60bn. What's your opinion on the cost of building 1GW of AI capacity?

Expert [00:09:27]: Are you talking about specifically only for Meta or for the entire data centre?

Client 1 [00:09:35]: I'm talking about the entire data centre.

00:09:35] Can you share some specifics between Meta and other hyperscalers? What's the best way to look at it for every major company? It'd be great to get the 2026 status, and then a two- or three-year outlook. I read some articles about Microsoft planning to move from 12GW to 38GW by 2032. Did you see those as well?

Expert [00:10:53]: Yes. Let me tell you first how the hyperscalers are looking at the capacity in total for four years. In the next four years, the total hyperscaler capacity is expected to be adding around 80-90GW of new capacity. That is a new capacity that they are looking to add in the next 4-5 years alone.

Client 1 [00:11:36]: By total hyperscaler, do you mean these three – Microsoft, AWS and GCP?

Expert [00:11:50]: Only USA.

Client 1 [00:11:53]: Only the US for those three companies?

Expert [00:11:56]: No. Microsoft, Meta, AWS, Google, Oracle, OpenAI, I'm including everybody. What I'm saying is the entire hyperscaler are expecting to add around 80-90GW in the next 4-5 years.

Client 1 [00:12:17]: In the US alone?

Expert [00:12:20]: In USA, that is correct.

Client 1 [00:12:25]: Is there also CoreWeave in those numbers or neocloud?

Expert [00:12:29]: Yes. The neocloud like a CoreWeave or Crusoe or Nebius, those are also included in that 80-90GW calculation. Now your question is the breakdown for each company, right?

Client 1 [00:14:34]: Those numbers are because you said at the beginning that it's for the US only, as that's where the capacity mainly is, correct?

Expert [00:14:45]: Correct. Primarily USA, correct.

00:15:41] From what I understand, AWS's [Amazon Web Services'] total installed capacity would be 22-23GW by 2031. Is that right?

Expert [00:15:50]: Yes. 22-23GW is expected to be the total cumulative capacity for AWS, which is lower than Microsoft and Google, but higher than Meta.

Client 1 [00:16:33]: Which one has the highest total capacity in 2026 in your opinion?

Expert [00:16:41]: I believe Google has the highest capacity, followed by Microsoft, followed by AWS and then Meta.

Client 1 [00:16:52]: What's your estimate for Google?

Expert [00:17:00]: Meta probably is the lowest. The reason for that is Microsoft, AWS and Google, these are also the cloud service providers, CSP, and they have been developing a lot more capacity compared to Meta in the last many years.

00:17:45] What's your view on the total cost for 1GW? What would be the cost for the same gigawatt in 2027? What would be the evolution from 2026 to 2027?

Expert [00:18:12]: I will say even though per gigawatt cost is different for different companies. Some companies have a high AI content, some of them have a mix of AI and all. On an average, 1GW of the capacity today in 2026, the CAPEX cost around USD 35bn-36bn. Have you heard about the 800V architecture?

Client 1 [00:18:56]: I've heard about it, but not from a technical point of view.

Expert [00:19:09]: Because what's going to happen is from latter part of this year and 2027 onwards, hyperscalers are going to deploy the 800V power architecture. This architecture is expected to have higher CAPEX and also more expensive GPUs, server rack, cooling equipment, networking, etc. My belief is that 1GW of capacity in 2027, on an average, it is going to be something around USD 37bn-38bn on an average.

Client 1 [00:20:00]: That's not a huge increase compared to 2026. Why so?

Expert [00:20:04]: Because the 800V is not going to be fully deployed, there will be a slow incremental cost increase. If I were to say, if you're interested in 2030-31, in 2031, it is possible that 1GW of the installed capacity might require a CAPEX of somewhere around USD 44bn-45bn. It is going to be almost a USD 10bn increase or 33% increase from compared to 2026.

Client 1 [00:20:41]: I'm a bit surprised because there are many parts within this level of spend. If we just look at memory pricing at Micron, Hynix and Samsung, it has gone through the roof, correct?

Expert [00:21:07]: Correct.

00:21:10] If we compare 2026 to 2027, would it be a 5-7% or 8% increase in the total CAPEX cost for 1GW?

Expert [00:21:30]: I will say the price increase is going to be contributed by multiple factors. The price of HBM memory, the DDR, the NAND memory, those are expensive. Also the storage. Do you know what HDD and SSD means? The hard disk drive and solid-state drive, their prices are also going to increase because of the rise in the memory prices. In addition to that, the price of the server itself on a per megawatt compute and when I say server means you must have heard about Nvidia Vera Rubin, Rubin Ultra, right? The price of the server on a per megawatt basis is also going to rise because the demand is high and companies such as Nvidia, AMD, Google, these guys are going to charge a premium for their GPU, CPU and TPU as an example, so server is definitely also going to rise.

Now in addition to memory and server, I will say there are two more categories in which the price is going to increase. One of them is the cost of the networking. Basically, the fibre optic cables are becoming almost 30-40% more expensive than the fibre optic cable cost in the last few years. Then also hyperscalers are going to use the co-packaged optics CPO technology for the scale out and scale up. That is also going to require expensive ASICs such as co-packaged optic ASICs. The last category is the cost of the cooling, like a liquid cooling costs such as direct-to-chip cooling, immersion cooling, microfluidics. These are also going to rise from a CAPEX point of view.

00:23:45] You mentioned categories where pricing is going up. Are there any categories with decreasing prices?

Expert [00:24:04]: Have you heard about solid-state transformers? 800V data centre are going to require some of these expensive but energy-efficient solid-state transformer. There are company like DG Matrix, ABB, Eaton, Schneider, Heron Power. These guys are manufacturing, developing the solid-state transformer. Right now, the cost of the solid-state transformer is almost 4-5x of the traditional transformer. It is projected to come lower in future, but this also contributes to the higher cost from the electrical equipment point of view because of the high energy efficiency that 800V data centre require.

Client 1 [00:25:02]: Yes, I understand that. Why did you say only USD 37bn-38bn on average compared to USD 35bn-36bn in 2026? It seems quite low given what you just said.

Expert [00:25:19]: Yes. The reason is a lot of these new technology components are not available. There's a supply constraint. The transition to 800V architecture is going to be very, very slow. It is constrained by the supply of many, many parts such as electrical parts, memory, hard drive, SSD. Mainly it is slow in 2027, but gradually, it will rise because I said by 2030, the cost can go up to USD 45bn, which is almost a 30% increase.

Client 1 [00:26:11]: Let's not say in 2022 because it was maybe even before ChatGPT release.

00:26:11] If I were to ask you the same question for 2024 or 2025, what would have been your guess on the average CAPEX cost for 1GW?

Expert [00:26:39]: 2024, the average cost for a 1GW data centre is going to be something around, I will say, USD 27bn. 2025, it is approximately USD 31bn-32bn. The jump has gone been very higher compared to last two years. The reason is in the last 2-3 years, hyperscalers have deployed a lot of AI infrastructure, and they are relatively more available compared to the 800V architecture. As you can tell, in the last 2-3 years compared to today, the cost has increased by almost USD 7bn-8bn per gigawatt.

00:27:50] What's the rationale for deploying an 800V architecture for hyperscalers?

Expert [00:28:10]: Let me give you a few reasons. The number one reason is the higher power efficiency. The 800V is expected to reach the power conversion efficiency of around 95%. The traditional non-800V, the efficiency is around 87-88%, so power conversion efficiency is number one reason. The second reason is the amount of copper cable required in the 800V is going to be almost 40-45% lesser compared to the traditional data centre, so copper will reduce. Thirdly, the 800V data centre will use very centralised power distribution system using battery energy storage system. The UPS will be eliminated from the data hall, but the battery energy storage system, which are considered to be more efficient, they will be deployed. The last reason is that because the rack power density is increasing from 100kW to greater than 1MW, it is practically not possible to manage the power operation as well as the thermal load using the traditional 48V. 800V is expected to be significantly much more efficient in terms of the thermal loss, power loss, etc. It is going to be much more efficient compared to the traditional voltages.

00:30:18] Are you familiar with Legrand? It's a smaller company.

Expert [00:30:44]: Oh, Legrand, yes. They are also a very strong electrical integrator for the low-voltage, medium-voltage switchgear.

Client 1 [00:31:20]: While going through earnings calls, we come across analysts asking about the possible implications of moving from the traditional architecture to 800V. What do you think about the share of Legrand's revenues within a hyperscaler, and what could be the impact of this transition for the company?

Expert [00:31:55]: First of all, let me tell you, Legrand is considered as a tier 1 vendor. Tier 1 means the vendor with the highest rating. They are considered as a competitor to ABB, Eaton, Vertiv, Schneider Electric, etc. They also compete with players like Delta Electronics, Lite-On, Flex Power. These are some of the competitors of Legrand. Johnson Controls is another one.

00:33:16] Is the share of Legrand within the CAPEX spending of Meta increasing, moving to 800V, or decreasing the share of total CAPEX?

Expert [00:34:00]: I mean, compared to the share of wallet that they have today, my answer to your question is the increase will not be very dramatic. They will face a strong competition from other players like Eaton, ABB, Schneider Electric, Johnson Controls, etc, etc. What I would say is the share of wallet for Legrand is not going to decrease dramatically or it will not decrease at all, but I do not expect the share of wallet to increase dramatically either. There will be a slightly slower increase in the share of wallet or it might remain flat because as you can tell, 800V is going to be a very competitive space and other companies are also looking to compete with Legrand. If they are able to maintain their share at the current level or are able to increase… Let's say, in 2026, they are able to increase from 10%. let's say they increase by 2-3% by 2028, that is not a small increase. That is a very significant increase for Legrand, just to think that way.

Client 1 [00:35:27]: Which players benefit from this transition?

Expert [00:35:35]: I believe ABB and Eaton are going to be the biggest beneficiary. In addition to that, have you heard about a company called Delta Electronics? Delta Electronics is, again, a US company. They are considered as the electrical equipment and cooling equipment integrator, and they compete with JCI, Legrand, Flex Power. My belief is that ABB and Eaton are expected to be growing the most aggressively. Also Legrand will see competition from JCI, Delta Electronics, Flex Power, Lite-On and these players as well.

Client 1 [00:36:28]: Is the level of competition increasing compared to traditional architecture? If so, why?

Expert [00:36:42]: I will say, compared to what we have seen in the last 3-4 years, the 800V data centre architecture is going to put a very heavy pressure on the electrical equipment provider from a demand point of view. Because of this heavy pressure, those companies who are able to meet the lead time, the price competitiveness, performance reliability and even their on-site support, those companies will see some level of winning edge compared to those companies who are not strong in pricing and the lead time point of view. I will say compared to the traditional non-800V, 800V will see the competition increasing significantly across all the players, but not necessarily 2027. As we get into 2028, 2029, 2030, that is when the competition is going to grow much faster.

00:37:47] Does the pie share for electrical providers increase when moving from traditional to 800V? Is there an addressable market increase at the hyperscaler level?

Expert [00:38:17]: I will say from a total addressable market value point of view, just to give you some reference, in 2027, the combined hyperscaler companies are going to add approximately, in my view, 15GW, cumulatively. Compared to 2026, it is around 12GW this year. It is only 25% increase. The YoY increase from 2027 to 2028, '28 to '29, it is projected that by 2030, the YoY capacity deployment will increase at the rate of 28-35%, so that means the demand for 800V electrical infrastructure, the cooling system, the GPU, the memory, etc, that will keep rising. Compared to 2023-24, the rate of increase was around, I would say, 15-18% YoY. From 15% to 18% YoY, now we are talking about 28-35% increase YoY by 2032. You can see that hyperscalers are certainly putting very heavy pressure on the electrical switchgear vendor such as Legrand, ABB, Eaton, but also very heavy pressure on the integrator such as JCI, Delta Electronics, Lite-On, and company like ABB and Eaton, they are also trying to become an integrator. I have a reason to believe that Legrand is also one of the strong integrator that data centre use for deployment of the electrical infrastructure.

00:40:19] We talked with Legrand's management about their backlog and ability to forecast AI electrical demand. What do you think is the timeframe? For example, when deploying capacity at Meta, how much time do you need to install and plug in the electrical equipment? Legrand says it only has 3-6 months of visibility into its order book for hyperscalers, which seems quite low.

Expert [00:41:24]: Yes, 3-4 months is certainly a very small backlog capacity. Generally, across the electrical OEMs, we are talking about a backlog of somewhere between 9-12 months. Now for the high-voltage transformer or like a generator or a UPS system, the lead time can be around 24-30 months. For the medium-voltage and the low-voltage switchgear, generally, we expect the lead time to be around 9-12 months. I agree with you, if Legrand is saying that their back order or their backlog is only 3-4 months, that means they have not received the order or they have not won the contract yet, and that is why their backlog is looking smaller because across all the OEM, they are talking about an average backlog of around 9-12 months. In some cases, it can be 9-15 months, but it depends upon whether we are talking about low-voltage switchgear or medium-voltage switchgear or high voltage.

00:46:03] Everybody is talking about gigawatts, but how do hyperscalers forecast short-to-medium-term demands, and how accurate have they been over time? What's the best way to assess if those levels of capacity additions are reasonable?

Expert [00:47:00]: Let me tell you this way. Let's say, Meta building a 1GW data centre in a single location. Developing a 1GW data centre takes around somewhere between 30 to 36 months. That means they are talking about approximately 5-6 years of roadmap already. When we hear Meta talking about their forecast, generally, the forecast can be somewhere between three, five years on an average. Some cases, some company might provide a forecast for 7-8 years and some will provide for 3-4 years. On an average, I will say five year is a most optimum forecast that any hyperscaler will be able to provide because in the five years, they can tell how many gigawatt they will be deploying and what is the construction status of different data centre where the deployment is going to be happening.

00:48:27] Do you have an opinion on demand? Everybody is constrained today, and token usage is increasing very rapidly and exponentially. How does it work internally at Meta? How does Meta decide to add 4GW, 5GW or 6GW? You said Meta expected to add 12-14GW in the next 4-5 years. In line with those capacity addition expectations, the company is also doing some forecasting on demand, revenue, token usage and AI inferencing compared to AI training? How do you forecast the demand level? What's your level of confidence?

Expert [00:49:56]: Yes, I do. There are two approaches. Because generally, when you think about AI training, you cannot monetise, you cannot generate the revenue from training model. You can only create the model to develop the inferencing model, agentic AI, etc. In order to serve the user base, they need to deploy AI inferencing and edge data centre so that they can monetise based on the user queries or the number of queries that the users are generating under each business operation. When I say business operation, I'm talking about Instagram or Facebook or social media, advertisement, etc, etc.

00:55:09] What's your view on the monetisation of gigawatts going forward?

Expert [00:55:16]: I mean the thing is this gigawatt deployment, it is expected to generate the revenue for the next 30-40 years and even beyond. If you remember, I told you that after 2032, they will start slowing down the deployment because the current infrastructure that they have agreed, that they have built, they don't need to build more at the same rate because they expect to continue to generate the revenue once the infrastructure is deployed. Like today, when you and I are talking, we are talking because Meta is at the stage wherein they don't have enough capacity installed, but if you and I talk in 2035, then we can think about, "Hey, Meta already has got 30GW or 40GW gigawatt. They don't need to add another 30GW in the same timeframe." Their pace of increase will slow down.

Client 1 [00:56:19]: Maybe in 2035, we'll be much more advanced in agentic AI and even physical AI. You may say that Meta still needs to add 30GW going forward. Why would it level off in 2032?

Expert [00:56:56]: Because the thing is there will be a point in time wherein they will have significantly enough infrastructure constructed that they would want to maximise the revenue or maximise the compute capacity. As the business will grow, they will add more capacity according to the business growth. Today, the capacity is added not because of the business growth only. It is being added because they are shifting their focus from non-AI to AI. That is the main difference. In the 5-7 years, mainly it is because of the strategy, but after 7-10 years, it is mainly driven by the revenue opportunity or further improvement in the compute efficiencies.

00:57:43] Is there a way to measure the utilisation level of the infrastructure currently being built? You're essentially saying that there's a massive build-up of capacity and therefore, capacity utilisation will be relatively low in 2031-32, and we'll be able to support an increase in the utilisation of that infrastructure. Would you have indicators? What indicators should one follow over time to measure it and the need for additional capacity?

Expert [00:58:36]: Generally, the capacity utilisation is qualified under the IT and non-IT category. In the IT category, we talk about whether the GPUs or the CPUs, they are operating at 90% utilisation or 85% utilisation. Are they serving the workload without any buffering or without any shortage of the memory, etc. That means because of the infrastructure, because of the software algorithm, they can make improvement in the utilisation efficiency. That is one parameter. It is related to availability of memory, availability of the networking, etc, etc. In the non-IT category, the utilisation is defined in terms of…

Have you heard something called PUE, power utilisation effectiveness? PUE means, let's say, you have a 100MW of compute, how much power you really require to support the 100MW? In some extreme cases, it can even come down to 1.07. That means the amount of power required for each megawatt can be much more efficient if you are able to maximise the utilisation. When I was giving you the example of 800V data centre, this is one of the big goal to reduce the power utilisation effectiveness by improving the power delivery, reducing the wastage because of heat, etc, etc. Now the other ways in which you can measure the performance or the utilisation is water utilisation. Hyperscalers are very heavy consumer of the water. They are also trying to significantly reduce the water utilisation for cooling down their server GPUs, etc, or even the rack, etc.

The other areas which are critical on the utilisation efficiency is your network utilisation, how you maximise your network bandwidth at the scale out, scale up, scale across level. That is another one. Also the storage capacity, when you deploy the HDD and SSD, how effectively you are able to use the storage, dynamically allocate the storage capacity rather than building very large storage capacity, you can maximise the utility by dynamically allocating the read and write between the hard drive and SSDs at different levels. There are various ways in which through improvement in the hardware and improvement in the software, such as AI orchestration, the hyperscalers are looking to improve the operational efficiency and utilisation of their infrastructure.

01:02:43] What's your view on GPU utilisation vs GPU depreciation, as it's a debate in the market? How long do GPUs last before being replaced in data centres?

Expert [01:03:10]: That's a very good question. The goal is to maximise the utilisation, but also extend the life of the GPU in a data centre. Rather than replacing every 3-4 years, we will try to use the GPU for like 5-6 years and longer. Even then those GPUs which are depreciated rather than recycling them, we can repurpose them into other data centre, which can operate with less powerful GPUs. There is a goal to maximise or extend the life of the GPU to the maximum extent possible.

Client 1 [01:04:05]: Do you achieve those goals?

Expert [01:04:07]: I mean it is right now still too early because GPU installed base is only 3-4-year old. We feel that we are going to be very successful. Have you heard something called AI orchestration? AI orchestration is one of the most powerful mechanism that we want to maximise the utilisation of the GPUs or the other AI infrastructure to the maximum extent possible. This is still a growing category. I don't think any company can claim that they have a mature AI orchestration capability available. We are still too early to know how we will be able to maximise the use of GPU, but the goal is to not depreciate them very fast.

01:05:24] Are all large hyperscalers on the same level in executing this massive infrastructure build-out? Which players have advantages vs disadvantages?

Expert [01:05:50]: I would say among the CAPEX given that they have partnership with Nvidia, AMD, Intel, Broadcom, Lumentum, etc, hyperscalers certainly are looking to operate at the same level of architecture, design, infrastructure availability deployment. Other players like neoclouds or co-locator or other enterprises, they do not have the same advantage or same dollar power or purchasing power. From a technology point of view, I would say hyperscalers are like marching steps. They are marching in a very similar steps, but other players or other (audio distorts) are quite far behind.

01:06:48] When we asked, you said the evaluation of demand was the responsibility of the front end, and you said you do back end, so you're not exposed. Is there a particular type of job or someone responsible for that that we could try to speak to?

Expert [01:07:20]: Mostly, I would say you have to talk to somebody from the business leadership point of view or product management point of view. Mostly the business leadership point of view or like a financial operation point of view, those guys are going to be more… Because they are at the front end, they are developing the revenue model. They understand the different market segment from which the revenue will be generated. I gave you example of some of the metropolitan region like a Tokyo, Seoul, Boston, Berlin, Paris, because for Meta, there's a strong user base, particularly in the urban region and Meta is looking to deploy it there. Now how that is going to grow, that is something you need to find out from somebody at the front end.