← 返回

2026.09.22 | TB | Datacenter | US AI DC Construction & EPC

原标题：US AI Data Centre Construction & EPC Space – Meta Platforms & Turner Construction – Former Supply Chain Manager, Solutions at Meta Platforms Inc

#datacenter

Third Bridge · 2026-09-22

Client-led • 9 Sep 2026

Companies: META US, MSFT US, A005930 KR, ORCL US, FLR US, HOT DE, A000660 KR, GEV US

Expert Former Supply Chain Manager, Solutions at Meta Platforms Inc

Agenda

Questions

[00:05:19] We're researching Fluor and Hochtief with a focus on Turner Construction, including EPC [engineering, procurement and construction] relationships with data centres from a customer's perspective and perceptions of the different players. Turner is having incredible success in this space, while Fluor has had a few smaller projects. In theory, capacity around this industry is tight. Fluor has scope to win business. To begin, could you tell us whether the work is dispersed among different players because there’s more construction and a need to spread it around? How do you think about picking an EPC?

Expert [00:09:06]: Yes. By the way, if you go back to 5-7 years ago, the role of the EPC and the data centre contractor was highly overlapping. Now with the arrival of AI, the EPCs are becoming much more specialised in terms of the role they play, such as doing an initial design, engineering, creating a blueprint of the data centre side. Some of the top companies are Aecom, Jacobs, Black & Veatch, Olsson, WSP, Sargent & Lundy, these are some of the highly specialised EPC companies. Some of them are also involved in doing the actual construction, but primarily the role that they have been playing or they have been hired to play is to create a very site-specific design and engineering blueprint of a data centre. That is how the role of the EPC is becoming much more classified compared to how it has been interpreted or how it has been leveraged in the traditional data centre.

[00:10:36] To what extent are AI data centre sites standardised? Is every data centre its own project? Does that change with the chip generation involved?

Expert [00:10:56]: If we compare to the traditional data centre, the hyperscaler data centre sites are really, really highly customised and highly engineered according to the specific requirement they have in terms of the type of power, behind-the-meter sources, liquid cooling infrastructure, operational efficiency of the server, etc. Within the scope of hyperscaler within Meta's hyperscaler data centre site, you can find a lot of synergy and commonality across different geographical location where Meta is deploying the data centre. There are certain nuances, which could be specific to a particular geography such as climate, the weather, the soil, the type of power, etc. At a macro level, within the hyperscaler, whether building a data centre in Virginia or California or Nevada, you can see a lot of a common mechanism that use to build those locations, to build those sites.

[00:12:21] Turner’s seeing really substantial orders. Is the company designing the whole thing to own the project end-to-end? Is it mostly managing the construction? Where does the company fall on the spectrum?

Expert [00:12:51]: The reason I did not mention Turner in that context was even though, to be very fair to them, Turner does have a very strong internal design capability, engineering capability, including creating a blueprint for electrical, mechanical. Generally speaking, they are considered as a very pure-play construction company. They do have a company called SourceBlue, which is their supply chain management arm. They have a lot of capability in terms of doing the traditional electrical, mechanical-related infrastructure construction. Generally speaking, in the context of the hyperscaler, Turner is seen as a pure construction company rather than seeing them as a pure EPC company, they are more on the construction side. There is some level of or some degree of overlap between the EPC and a pure construction company. Turner does have that level of overlapping. The classification of how Meta sees a pure EPC vs a pure construction company before we talk about their level of overlap depending upon which company is chosen at a particular location.

Client 2 [00:14:20]: When you say pure construction, you're referring to the foundation and shell, and none of the electrical, HVAC [heating, ventilation and air conditioning], highly complex things and site fixtures, correct?

Expert [00:14:35]: No. Construction basically classifies and has four subcategories, civil construction, electrical, mechanical and plumbing or utility related. Turner does have all the four in their portfolio. They do that. The only reason I did not classify them as a pure EPC was that they may not be somebody who is doing the entire site design, including, for example, what transformer is required, what generator is required, what is the efficiency of the gas turbine. They may not have those technical specification creating capability on their own. They are strong on the execution side. Procurement, negotiating the price, negotiating the lead time with the supplier, but not necessarily a pure design company. Because you or your colleagues were asking me what is the difference between the hyperscaler vs the traditional. This is where major differentiation comes into play.

[00:15:48] Does Meta Platforms want to work with someone who designs and constructs, or does the company actually prefer to separate those?

Expert [00:15:57]: No. As Meta, certainly prefer a company which is much more vertically integrated in terms of having the in-house design and engineering capability, including project management, cost management as part of their portfolio, their ability also to be a sourcing company who can manage the bill of material across electrical, mechanical, civil, which is what Turner does have. I did give an example of SourceBlue as one of their subsidiary, which is quite strong in their ability to source or supply chain manage the different components, products as well as the contractors like our manpower as well. There is certainly a preference for a company to be vertically integrated, but it depends upon site to site where and whether or not they have a strong footprint in that region being a vertically integrated company or a pure-play construction company.

[00:17:05] Did Meta ever use Turner, and does that perception vary among your peers?

Expert [00:17:18]: If you know Meta and Turner relationship, have been working with them very significantly in the Midwestern region like Indiana. Meta has signed a contract with Turner for a 1GW data centre site. Initially, they have spent USD 10bn to kickstart the project. This is something that you can find in the public domain. I'm not giving you anything confidential here, which I'm not allowed to. Turner is a chosen company that Meta has selected to do the entire end-to-end data centre design, construction, engineering and procurement. In that context, you can see that they do have that capability and Meta does look at those vendors who have that highly integrated role as part of their portfolio.

[00:18:20] Fluor’s less of a player in data centres. Does it have a role to play here, or become a bigger participant? Is the industry meeting the demand with the prime vendors of choice?

Expert [00:18:57]: Just to be justifying Turner's role, Meta, do not consider Turner as a tier 2 or a tier 3 player. We do consider them as a tier 1 player when it comes to building some of our critical-to-business data centre site. When I say critical to business means AI training or AI inferencing. The fact that Turner is one of the company that Meta is using for building this 1GW data centre site in Indiana, that should give you an idea that they are not considered as a lower-tier vendor, particularly when it comes to the hyperscale data centre. If you know the portfolio of Turner well enough, you will see that they actually are very strong in terms of not only the construction capability across electrical, mechanical, civil, but they also have a newer category in which they are rapidly getting into, and that is called prefabricated modular system development.

[00:20:18] Turner has many data centre orders and currently reports on them in its public reporting. Fluor is its competitor, but the company hasn’t historically competed in the same niches. Fluor has a potential opportunity to benefit from the power generation around data centres. Physical man-labour associated with such projects and EPC resources are constrained. Will Turner be your first choice for a project like Indiana’s?

Client 1 [00:20:18]: There are only so many people, presumably, at Turner who can do these jobs, and there's a fixed number of people in the industry who do this. How willing would Meta be to work with Fluor instead of Turner or a Tier 1 choice? Is that a decision point Meta faced, as a result of this demand?

Expert [00:21:36]: Fluor is a relatively smaller company. They are not obviously at the same level as a Turner.

Client 1 [00:23:02]: Everyone at the top of the list is busy with the biggest data centres. There's now some space underneath that for other smaller players to step up and do the mid-sized data centres.

Expert [00:23:26]: Yes. One thing I would like to clarify here is that given the pace at which hyperscalers are looking to build their data centre capacity, we are talking about around 18-20GW capacity per year that hyperscalers are looking to add for the next 5-7 years. What that means is that we are not necessarily in the driver seat in terms of choosing the construction company or the EPC company. It is almost like a two-way street wherein the construction company or the EPC are also being very careful who or which hyperscaler they work with. It is definitely a very competitive space compared to what we have seen 5-7 years ago or before that.

[00:24:13] You mentioned that given the supply tightness, EPCs have more leverage in negotiations than they historically had. Are you seeing that in higher prices, better margins and terms? How is that leverage showing up in the deals?

Expert [00:24:37]: Yes. As a construction company and the EPC company, they are certainly able to and are looking at some premium that they are charging from the customer. a customer, the way we try to handle the rising prices of the labour, of the material, of the schedule-based delays is by trying to see whether we can choose a particular EPC company. I'm using the term EPC, please assume that we are talking about construction company in this category.

The mechanism that we devise to get a better price is by looking at the EPC to see whether they can offer more vertically integrated services and that vertically integrated service approach allows to negotiate based on the volume of the business that we are giving to them rather than trying to save on by bringing in multiple vendors, because generally, for a particular data centre site, it is always and always preferred to use a single vendor for the entire data centre site. It is not always possible, but that is the preferred direction that hyperscaler will look at to choose a single vendor or a single EPC company for the entire data centre site. By doing that, potentially able to negotiate a better price by offering them a volume service role or volume construction role within their portfolio that they are capable of capturing or bidding for.

[00:26:35] Construction contracts have a structuring element that's probably not pure price. If like-for-like were compared before the AI data centre boom, how much has pricing on data centre deals held up against customers such as Meta? How much more risk are you aware of? How would you frame it?

Expert [00:27:02]: Compared to the traditional data centre, certainly have seen a significant price difference. For example, if you think of a 100MW data centre, for a moment, let's not include the cost of the server networking story. Let's just focus on the cost of the site. The site means the civil construction, electrical, mechanical, everything. Traditional data centre, generally, you are looking to spend around for 100MW, approximately USD 1bn-1.2bn for 100MW. Compared to that, the hyperscaler data centre today are looking at spending around USD 1.5bn-2.0bn for creating the same 100MW infrastructure.

You can see that there is a significant, almost 35-40%, increase in the construction cost of a hyperscaler data centre. Out of the 35-40%, I will say approximately 70% is attributed to the actual material, 70% of the increase and 30% of the increase is attributed to the design, engineering, specialised services that these hyperscaler data centre require. Certainly, the cost has increased from an upfront CAPEX point of view, there are ways in which we are going to amortise or bring a cost parity, but that is not something that you can achieve within a short period of time. As an upfront cost, we do spend 30-35% more in a hyperscaler data centre compared to a traditional data centre.

[00:29:02] You mentioned that an enormous amount of capacity is still being brought online every year. It seems the pace of additions will slow down. Will costs continue to go up? Does the infrastructure exist to deliver on the 18-20GW of new hyperscaler capacity per year? Will it strain the system even more in the next five years? Should we expect costs to continue to rise because of that?

Expert [00:29:39]: The pressure from the demand side certainly is expected to remain for the next seven-year minimum given that hyperscalers are still at a low 10-20% of the construction capacity that they are looking at compared to where they want to be. Certainly, you will see the demand side pressure there. In terms of the cost, I will say the material level, it is unlikely that the cost will come down dramatically anytime soon because of the timeline, the lead time, the urgency, etc.

From a services point of view, when you think about the EPC role as a designing, servicing, consultation, etc, that is something we expect to leverage a lot of existing knowledge to build the new data centre wherein, for example, we have chosen Turner as an EPC partner in Indiana and assume that we want to build another data centre site somewhere in California. If Turner is able to win the bidding for California, one of the reasons could be that we are leveraging Turner's servicing portfolio into the California site as well, and that is allowing us to negotiate a much better saving on the servicing part.

a customer, are certainly going to find every possible avenue to reduce our per-megawatt cost, but the opportunity is more going to be in the services area, which is the EPC in the engineering procurement design area. Material-wise, it is going to be very much driven by the demand-supply pressure, including some of the dependency on the tariff-related material such as electrical or thermal, etc, that we are going to be dependent on outside of the USA. Material is going to be lesser, but services is going to be more friendlier in terms of the cost coming down as we move forward.

[00:32:04] The number of gigawatts added per year flattens out. Much of the work being done today is creating IP leverageable in the future. Turner doing 1GW of data centre for Meta in Indiana will have IP leverageable for the next data centre. In theory, Turner’s contract is smaller on the next data centre, even if it's 1GW. Is that a fair way to understand the go-forward here? Meta should be benefiting. There should be an EPC deflation, setting aside material costs.

Expert [00:32:49]: Yes. From a customer point of view, certainly are trying to save. From a Turner point of view, they're also winning, compared to their competitor, they have a much better opportunity to win the contract because they are being paid not only for the design services, they are being paid for the actual construction services as well. They are able to expand their volume of the business by using their existing IP. In some ways, you can say it is a win-win situation.

The vendor will also would like to benefit because they then can leverage their existing IP, but they are able to win some other contracts or outbid the other bidders because of their existing relationship. call it something called vendor stickiness. You probably have heard this term before. Vendor stickiness certainly plays a very important role, wherein if you check, okay, Meta might be using Turner, while Google might be using Holder Construction or AWS will be working with Mortenson, for example. There is a level of relationship that gets developed between the customer and the vendor that might be carried over to multiple data centre locations and over the multiple years of partnership.

[00:34:42] Are contracts typically fixed-price or lump-sum? For example, is it like: this is a 2GW site. This is how many dollars you agree to pay for a scope of work. Is the contractor responsible for executing it under that and get paid a fixed sum?

Expert [00:35:16]: Yes. Generally, whenever sign a master service agreement or a long-term agreement, one thing is that even though, site to site, there could be a lot of similarities in terms of the type of infrastructure being deployed, the type of machines, components being constructed, etc, etc.

There is definitely a fixed cost that we are calculating that potentially we will be expecting the company like Turner to provide a quotation. Generally, that fixed cost is determined by how much compute capacity we are looking to deploy at that location. The compute capacity is one of the yardstick or one of the measures that we use. In this calculation, this is more a high-level calculation that, let's say, 100MW data centre, assuming that Turner is one of the leading vendor, maybe their contract might be, let's say, USD 500m, because the remaining USD 1bn might be because of material. Let's assume that Turner cost is USD 500m.

That agreement, USD 500m or whatever that number is, is first and the foremost, calculated or negotiated based on the compute capacity are looking to deploy. That is the number one factor. Now there are certain variabilities such as the labour cost in that area or any specific kind of building codes or a specific type of electrical equipment that requires special handling or special construction. In a particular location, there might be more generators rather than the power from the grid. Depending upon how the equipment is going to be deployed, the associated labour cost, there could be certain variations in there. In a summary way, my answer is yes. Fixed cost is a mechanism that use depending upon the site, compute capacity that are looking to deploy there.

Client 2 [00:38:07]: Are there issues with the weather or delays with materials? Who is responsible for it?

Expert [00:38:21]: If there are delays related to the material delivery such as generators or a rigging machine, those equipment. In case there are delays in the delivery of the material or the construction or some of the machines that the GC is responsible for. Meta is going to hold the GC responsible for that or there will be certain penalties or provision in the contract in which GC has to expedite those. The construction company like Turner, their primary responsibility is to make sure that the site is being constructed according to the schedule and the milestone that the contract has initially outlined.

Let's say, it's a 500MW data centre site, which is expected to be completed over the course of four years and X percentage or Y percentage of the site will become alive or will become operational. Let's say, 25% capacity will be operational by the end of year two, 60% will be operational by year three and 100% will be operational by year four, then Turner is going to be responsible for making sure that those milestones which were outlined in the original contract, they are being met with. For any delays related to the labour, construction material, the project management, those are the responsibility of Turner, and they have to make sure that they are meeting those timelines and the milestones.

Client 2 [00:40:46]: What if there are permits or power equipment-related delays?

Expert [00:40:56]: Those are the responsibility of Meta or Meta's general contractor because it is conceivable that Turner is not responsible. They are not purchasing the generator or transformer and Meta has directly contracted the delivery schedule. In case of delays, then Meta is going to bear the cost or bear whatever that might entail. That is not going to be considered as a Turner responsibility because Turner, in order for them to meet each of those milestones, they do have a dependency on availability of the material under different categories such as electrical, mechanical. In this case, we are assuming that Turner is a wholesale company in which they are providing the entire civil, mechanical, electrical in case they are doing only civil construction, then that question may not be even part of the contract because they are not doing the electrical deployment.

[00:42:09] A project had issues because it hadn't really been properly thought through, such as provisioning power at the site. It was moving forward, but all of a sudden, it came crashing to a halt. Hyperscalers are running way ahead of themselves and not thinking about whether they could access a natural gas BTM [behind-the-meter] solution deployed somewhere, only to realise they aren't allowed to plug into the pipeline, cannot get the air permits or for other reasons. How well put together and thoughtful are the build plans across hyperscalers?

Expert [00:43:13]: The fact that you have heard about Oracle's or there are some other issues also, that gives an indication that these are more of an exceptions rather than a rule because generally, when you think about company like Meta, Google, Microsoft, AWS and even Oracle for that matter, they are certainly looking at heavy-duty outsourcing these services to the third-party expert companies, whether it's a pure-play EPC or a construction company or a general contractor because these companies certainly do not have not only the technical capability, they don't have the required resources, qualification, certification or licensed engineers or a technician to manage the progress.

Certainly, a very heavy reliance on a third-party company. Obviously, we will hear about certain exceptions, certain areas wherein the issues have cropped up because of the delays in securing the grid power or securing the gas turbines from GE Vernova. I would say these companies are certainly on a very fast learning curve wherein they would want to make sure that they are outsourcing the responsibility of sourcing, construction, designing, engineering to the expert companies, particularly given that the scale is very high and the timeline is very fast. If you think about a traditional data centre 5-7 years ago, a 100MW data centre will take around 24-36 months to build. Today, we are talking about a 1-5GW data centre on a single location getting constructed or targeting to be constructed within the same four-year timeframe.

Now that is not possible if hyperscalers are not entrusting the task to the specialised companies and rather than doing it themselves. When you were asking about the permit and the city providing the approval, a lot of these companies use EPC to do a lot of quality checks, permit checks before the city inspector even gets a chance to visit the site. A lot of these services are being outsourced to the specialised company. Obviously, it comes at a cost that we try to bring some parity in moving forward, but upfront, we are certainly paying a significant amount of dollar to build the data centre site according to the specification and according to the timeline that have in planning. If you remember, when I said 35-40% cost increase for the hyperscaler, some of it is attributed to the accelerated timeline at which are looking to build these data centre sites.

Client 1 [00:46:32]: I understand they're evolving, and people are getting better. Would it be surprising if you read that 10-15GW will be delivered next year? Would it be that surprising if they planned for 20GW but got 10-15GW?

Expert [00:46:59]: Yes, then it will be surprising. It will be definitely very surprising because the 15-20GW, which I mentioned earlier, that has been calculated or has been planned considering all the constraints, supply constraints, manpower constraints, etc, etc, they have been all factored in. If that is coming down to rather 50% of that capacity, then something catastrophically wrong must have happened to cause that such a big difference of the actual implementation.

[00:47:39] What do you make of the regulations turning against some facilities in Texas and Pennsylvania in the US? That seems like another risk.

Expert [00:47:52]: Taxation, have you heard about the Texas Senate Bill 6 or the California Senate Bill 100?

Client 1 [00:48:03]: No. Could you tell me about them?

Expert [00:48:08]: Some of these grid utility companies or some of these regions are putting a lot of local population, I don't want to say, driven, but local population influenced senate bill, which are putting a lot of restriction on the data centre as to how much capacity of the power they can get from the grid, how much power they need to source from their own independent power producer. The point I'm trying to make here is that those local restriction or regional restrictions are also taken into the calculation when a company like Meta or Oracle or Google is building a data centre site. We are certainly not going there blindsided, not knowing what might happen tomorrow by the local government or the state government coming in. Obviously, there could be some new development that can happen. By and large, hyperscalers are doing a lot of diligence before a shovel is put into the ground.

Convergence and consolidation is also happening in the EPC industry, in the construction industry. Earlier, you were asking me about Fluor, and I did say that they are obviously not the leading company. There are certainly some new player or the smaller players who are able to showcase their differentiation and hyperscaler realising that they cannot always and always work with only the top-tier vendor. They certainly have to make sure that they are choosing their EPC partner or their construction partner in a very careful manner. Once they are chosen, it is always a good idea and they will try to prefer to use the same vendors or the same set of vendor for multiple data centre sites that they have in the pipeline, and that could be for next 7-10 years, if not longer.

Client 2 [00:51:55]: Was it like, "Whatever the number is, even USD 200bn-300bn, we expect to do that for one decade"? What was the internal debate like? I understand that the next two years are clear, but it's less clear what the plan is as we go into 2030-2031.

Expert [00:52:38]: I certainly do not believe that election or no election or whichever government comes, I honestly do not believe that will have any real impact on the plan that AI-related infrastructure development is going on or will continue to go on. I specifically did not say Meta or any other company. AI, as a technology, certainly seems to be quite inert to the geopolitical development within USA or even outside of the USA. That is one part I would like to make, in my view, very strongly, I believe in. Second part, I would say, when you said 2028 as a peak, I don't think that's the right timeline. In the bigger scheme of things, 2028 is today, we are talking about 7-8 years, like all the way through 2032-33, wherein approximately another 120GW capacity is going to be added only in USA. That means almost 80-85% of the total capacity that is going to be added in the next seven years, that is going to be in USA and Canada.

The remaining 15% we are talking about in western Europe, some of the Asia-Pacific country and potentially some of the Nordic country also. When these companies like Google, Microsoft, Meta, etc, they are talking about in a public domain that they certainly want to put such a high, I would say, higher than USD 2tn of the CAPEX in the next 5-7 years, they are backing their statements because of the intelligence and the data analytics that they have at their disposal. I do not really believe that these guys are getting driven by the political compulsion or the political influence or geopolitical reasons. I certainly believe that they have a very strong roadmap driven by their opportunity to expand and use AI as a primary mechanism to run their business. I certainly believe that they are not going blindfolded or have blinded into this level of CAPEX spend.

[00:55:39] What’s the most misunderstood point that people bring up or debate with you that is most likely perception vs reality? People in our world have one view, but the reality on the ground might be very different.

Expert [00:56:06]: There are two things which I can certainly mention are quite not well aligned to how the hyperscalers think about it. The number one is how much dependency these hyperscalers have on OEM supplier, manufacturer of the companies outside in China, Malaysia, Taiwan, Japan, etc. There's a lot of talk about manufacturing a lot of semiconductor parts, GPUs, etc, etc, in USA. Even though TSMC is building a data factory in Texas, that should not lead us to think that a company like Meta or Google or Nvidia or AMD or Intel, they will start doing all the chip manufacturing in USA alone and China will be completely cut off, for example. That is not going to be the case. I think this is one of the biggest misconception. Even in the beginning of this century, even then there used to be a lot of talk about whether Apple should buy the memory or the processor from the China, Taiwan-based company or should they buy it only from Micron, which is a US company and not Samsung or Hynix, which are Korean companies. Even after 20 years, Samsung and Hynix remain the number one, number two company in the world for the memory.

The reason I specifically mentioned these two companies or the memory as a reason is memory is the number one commodity at this moment, whose supply is the tightest. Hyperscalers have absolutely no option but to purchase these components from Samsung and Hynix as a leading provider, even companies such as Western Digital, Seagate, who build the hard disk drive, solid-state drive, they have a strong manufacturing capability outside of USA and Canada. This is one of the biggest misconception that hyperscaler will start purchasing all their supply from mainland North America. That is a big misconception in my view.

The second category, which I believe, given that almost 80-85% of the CAPEX that is going into North America, the how are we going to secure the power? There is some level of thinking that a company like ERCOT or PJM or MISO, these utility company, they will be in a dominant position. The cost of the power will go up and there will be a lot of resistance from the population because now the people will have to pay a higher price of the electricity. I believe that's a very strong misconception because hyperscaler realised that not only they cannot get the power capacity from the grid, given their timeline, given the pace at which they're looking to build their data centre, they have to secure the power coming from behind-the-meter. In the behind-the-meter, I'm talking about natural gas-based gas turbine, solid oxide fuel cells, solar power, wind power, even the nuclear power, which is projected to be available in a large volume only as early as 2032.

That means the process of securing the nuclear power in a noticeable capacity has not even begun yet. Company like Meta, Google, they have been signing contracts with a company like Constellation Energy or Oklo or other big player to secure the nuclear power, which is expected to become one of the big way of securing the capacity. We are talking about next 15-16 years for which the demand for energy and the demand for power is going to remain on a positive slope before it slows down or before it becomes more flatter. I think those are the two major areas in which there is some level of misalignment between how the hyperscalers think about vs how there's a public perception or a general perception in the industry.

[01:00:50] Real projects include building a data centre, a copper mine, a refinery and an offshore oil platform. Doing it fast creates risks in many different ways. How did that get managed internally? Meta was relatively new to building data centres pre-AI, and demand was limited, but suddenly, everybody is running as fast as they can. People would have planned for at least 2-3 years, but in some cases it was just in time, almost as if they were ordering some components. What was the internal tension regarding moving fast?

Client 2 [01:00:50]: We were under the impression that the hyperscalers are big, sophisticated organisations with phenomenal forecasting skills and procurement departments. They seem to have collectively missed every shortage that they run into. Memory is a good example. If you took Nvidia’s chip specifications and looked at the memory attach rates and the HBM [high-bandwidth memory], it was obvious, even in 2024, that we were going to run into a ‘memory wall’ in early to mid-2026.

If you add up GPU chip times gigabytes of RAM x3, it's HBM, like there wasn't enough memory. There's not enough capacity in the market. Power was the same. You could add up Siemens Energy, GE Vernova and Mitsubishi Heavy Industries. They only have so many turbines, and then all of a sudden, people are scrambling because they have chips and shells, not power.

When Exxon Mobil Holdings or a copper company designs its mines, you FID [final investment decision] and order all the equipment simultaneously. It's unheard of for a refinery to wait for a generator for long if it was ordered seven years ago. How does this happen? It's like somebody building a house and then realising, "Oh, I don't have any windows", and there's a window shortage, "Oh, we're just going to, like, tape it over for two years until all the windows show up." It’s surprising because these are large organisations.

These are enormous capital budgets, but that doesn't always happen in the real world. I contrast that with if Uber goes to London and says, “It's illegal, who cares, we're going to do it.” The customers will write letters. That's not what this is. This isn't social media, Uber or delivery. These are huge capital projects with real government oversight and overlapping agencies.

Expert [01:05:38]: Everything you said, the way you said, I can totally relate to that because every single word. The only thing which I would like to say is that these companies like Meta or Google or Microsoft or AWS, these are, really speaking, technical companies. They are not Chevron, they are not Exxon. They are not industrial companies. Even today, though Meta is developing its own MTIA, Google is developing its own TPU, that is more out of compulsion because these are not hardware companies. They have no choice. When I say that that means securing the power, securing the cooling also is something that they have to do or they have to manage, something they don't have the necessary expertise in-house. They have to really run helter-skelter to develop the intelligence or hire the people or hire the third-party company.

What really has happened is the example you take of constructing a house and not knowing how to get all the material, this is exactly how I will say it is that suddenly, the AI has created like a step function from how the data centre were operating using a traditional server-based data centre, suddenly, the AI has created a demand-step function. That demand-step function has created a lot of turmoil across power, cooling, equipment level, even the construction level, civil construction, permitting, etc. There is a lot of disruption that has been created, and it has happened in a very short amount of time. Think of it like a growing pain that none of us were used to, particularly in the technology industry. Maybe the people in the heavy-duty industry like a Chevron, those guys, they are much more smarter knowing how to manage the procurement of the turbine, etc. These technical companies don't have that level of intelligence or level of experience compared to them.

Client 2 [01:08:13]: Oracle is the best example of this, where it's almost as if the CEO had a meeting and said, "We're going to build 10 GW of data centres, go." Everyone is like, “Hold on, that doesn't work that way. You can't just snap your fingers and deliver power to OpenAI.” Everyone is scrambling and wondering, “What sites? Where are we going to go?” It's done in a very haphazard way. I think there’s a spectrum. Meta, Amazon and Microsoft have done so for a long time and are much more organised.

Expert [01:09:16]: I totally agree with you. There are a lot of the surprise element. This surprise element is not going to become normalised anytime soon. A lot of learnings are being made, like Meta is making a lot of learning. Google is making a lot of learnings. Obviously, it is very disruptive in the overall power, etc, and the construction industry.

Client 2 [01:09:37]: When people ordered turbines from GE or Siemens, the company hadn't secured the E&C [engineering and construction] labour for installation. Kiewit and others are full, so they may even have the generator on site, but no crew is qualified to work on the turbine.

Expert [01:10:07]: I totally agree with you.

Client 2 [01:10:10]: This environment has been difficult, and it's happening so fast that people are scrambling.

Expert [01:10:21]: Very true. This is why using somebody vertically integrated. Sometimes, we actually ask.

Client 2 [01:10:29]: Yes. It's vertically integrated so it can manage those conflicts internally.