← 返回

2026.10.06 | TB | Utilities | Utilities Serving Data Centres & AI

原标题：Power & Utilities Sector – Operating Needs to Serve Data Centres & AI Implementation Opportunities

#power #datacenter

Third Bridge · 2026-10-06

Analyst-led • 28 Sep 2026

Companies: NEE US, NRG US, AES US, TLN US, VST US, CEG US

Rajeev Aluru (RA) Managing Director, Innovation & Analytics at Smartbridge LLC

Anthony DeRuijter (AD) — Third Bridge Sector Analyst

Agenda

Key Insights

Introduction

AD [00:00:04]: Let me start today's Third Bridge Interview entitled Power & Utilities Sector – Operating Needs to Serve Data Centres & AI Implementation Opportunities. My name is Anthony DeRuijter. I'm very pleased to be joined today by Mr Rajeev Aluru, Managing Director of Innovation and Analytics at Smartbridge. Rajeev, thanks so much for joining me today.

Before we begin today's Interview, could you please state specifically I agree or I disagree to the following statement, and the statement is: You understand the definition of material non-public information and agree not to disclose any such information, or any other information which is confidential, during this Interview.

RA [00:00:35]: I agree.

AD [00:00:38]: Would you mind briefly introducing yourself?

RA [00:00:42]: Rajeev Aluru, I'm currently working as a Managing Director, Data and Analytics, now it's AI for Smartbridge. I have previously worked for 13-plus years in the utilities, energy sector, including my own entrepreneurial experience as well in power operations, customer analytics, also into battery storage operations, specifically centred around the asset optimisation, using AI, using data. That's my experience. I'm glad to provide my overview and what I've done, what I've gained over the years.

Questions

00:01:34] What are your views on the current state of data centre interconnections and whether we are running into demand issues or bottlenecks on new interconnections that are severe enough to mimic demand issues?

RA [00:01:50]: The current bottleneck, let me step back and give a little bit of numbers here, from what I remember from the last time, what is it, 12,000 projects, approximately almost 1,500GW of generation capacity is currently in the interconnection queues from what I'm hearing. It does appear to be a real bottleneck. On the other side, we also can look at it as a process backlog in my view. Primarily, this is around equipment supply, for example, power transformers, step-up transformers. These lead times have really grown into what is a multiyear, almost three years of lead time for some of these components, and that is what is making this much more defined and much more longer, I would say. The point is with all that said, even if you fix the queue tomorrow, we would still be getting stuck on hardware is what I'm trying to say. The most interesting debottlenecking that I have seen, now going back to what people are doing, is, yes, you can apply some to the interconnection queue, but there is much more if you have to look at it from the entire value stream and a bit more upstream as well is my point there.

00:03:24] When it comes to debottlenecking data centre interconnections, or otherwise streamlining and facilitating improvements in timelines, are there any technological approaches, for example, that you think could make that happen?

RA [00:03:41]: The one thing I've seen was, for example, the hyperscalers or the technology companies, Google, for example, was working with PJM to apply AI to how the queue gets managed and then how can the optimisation happen, including some application of agentic AI principles and then even some of the services, if not products, if you want to call it that. There is some of that going on. Then some others that we can also think about are on the component level, like what I was talking about, which is, for example, the component supply, gas turbines. We recently heard gas turbine components are also being manufactured by SpaceX at this point or at least they're thinking about doing it. In other words, the industry is also looking for where these bottlenecks are happening and then they are going into those. I think that is a very interesting way of solving this problem in multiple fronts.

00:05:03] What are your views on BYOP [Bring Your Own Power] trends and whether new data centres will mostly be expected to bring new generation or underwrite new capacity to serve their load? When we look at what the IPPs [independent power producers] have been doing, we've seen quite a bit of consolidation. Some have clearly been taking bets that they can get their existing acquired plants contracted. How do you think about BYOP and what it means in that context?

RA [00:05:31]: I would say Bring Your Own Power is essentially a CEO-level strategy at this point at the major IPPs. It's not simply a hyperscaler workaround. For example, recently, we heard from the NextEra CEO, the company is even positioning around Bring Your Own Generation and then considers itself uniquely positioned. That's a trend that most of these IPPs are actually betting on in a way. For example, similarly on a consolidation note, Constellation picking up Calpine and the largest dispatchable gas fleet out there. Then as I mentioned before, the xAI situation, putting batteries next to gas turbines and SpaceX building their own turbine blades. That's not something most players will do, but it does show how far some companies are willing to go to control their own supply. Hence, BYOP is playing a very important role in this current situation.

00:06:52] If you expect the trend towards BYOP, do you think that will be the rule more broadly, or are there regions or markets where there will be more flexibility to contract with existing generators? It seems that ERCOT [Electric Reliability Council of Texas] might have more flexibility for behind-the-meter arrangements for existing generators compared to PJM, for example.

RA [00:07:13]: I would say it really depends on the ISO, as you were pointing towards, because of the way how the ISO, the region, the interconnection, the load centres. There's a lot that goes into the ISO. I'm glad you pointed out, really. PJM, for example, is going through quite a bit of lay-off (ph 07.36), if you want to call it that, because of the density, because of, again, the current constraints, and the interconnection queues are very different. Yes, BYOP will vary depending on the region. Even though we had all these, let's say, structural bottlenecks from the process components and all that, even if that gets cleared, region ISO by ISO, I think we will see some differences, to your point.

00:08:13] What project models do you think are going to dominate in the long term when it comes to capturing data centre demand? Do we need integrated end-to-end solutions where the project doesn't just cover the generation or power needs, but the actual data centre infrastructure? How should we think about certain considerations around energy mix, backup capacity, redundancy considerations or anything else that you think is going to define the long-term successful data centre power projects?

RA [00:08:51]: What we have seen here is a shift at least in the plans and strategies from treating these data centres as standalone real estate. What they are doing right now is they're vertically integrating the deals that will secure the power and digital infrastructure together and then for that matter, actually, colocation of a data centre with battery storage. Obviously, the data centre itself comes with it. Maybe we can actually talk about how they are looking at infrastructure. The point is it is becoming its own, call it, mini power plant in a way. That's the way how people are looking at it as opposed to just serving a certain market or an opportunity. In a way, it's a shift in investment thesis for that matter. There are some estimates out there where approximately USD 4tn-5tn, I think, is projected to be built out. Hence, when we are putting this much of investment, I feel like it is absolutely critical to think of it that way. What will dominate long term, these will become these integrated power islands, if you want to call it, that deliver to various needs, including minimising the problems, if not totally preventing them with load curtailment, interconnection demand and then even the shifts, for example, grid reliability.

There are multiple things that I think will come out of this infrastructure investment. That's my view. Hence, the new project models, that will really look at how to use grid in the future in many ways. That is the way how I look at it. Some of these might not even exist today. Once the infrastructure is there, depending on what the needs are, the models will emerge, but there is already a very clear shift in the investment thesis that I see.

AD [00:11:28]: What does this imply for capital needs or the average CAPEX or CAPEX intensity per project, as well as operating complexity?

RA [00:11:42]: Actually, the operating complexity is a tricky thing for now. Most of these data centres, they're just coming on board in a way. There are a lot of projections, a lot of builds that are happening. What I also see is some of the asset management, meaning utilities, might even open up their own operations perhaps. Maybe they will be asked to manage these assets. That can be one. The complexity is, in a way, twofold. One is the CAPEX side of it, but then these are massive multibillion-dollar investments. The utilities are actually pretty much structured, and they have the experience managing these things. The complexity implications for a hyperscaler might look like a very much uphill and a very complex thing, but if we put the process right, we can have either some of the existing utilities that are looking for newer challenges, newer opportunities or even newer players might come up and then they can actually help with helping us with the power and the digital infrastructure all together now.

It's infrastructure at the end of the day, and I am expecting some changes, as I said, with the utilities as well as some newer players who might help with long-term service agreements, even power purchase agreements. For example, the contracting dynamics might change because now you have a different player. Your PPAs might get longer. Reliability might become more and more important perhaps. At that point, hopefully, it becomes more like a commodity as you have new players and the existing players over time. That's the way how I look at it.

00:13:53] Are hybrid architectures – grid-connected generation, paired with on-site or behind-the-meter assets – becoming a distinct project model separate from pure grid-connected or pure BYOP?

RA [00:14:11]: I would say, yes. It's now defining as its own category, not just grid only, not fully off grid, so grid connection plus on-site generation plus a control layer managing both. Hybrid is certainly a very viable option from what I'm seeing. There's also an example of Santa Clara, where there is a facility blending grid power with even hydrogen. Some of these modern technologies, they're also taking hydrogen and gas and then they're running workloads while the neighbouring sites are still stuck waiting on interconnection. That's one example. Then everybody knows about Microsoft Three Mile Island. They're actually bringing that plant back up is another thing which is also connected to the grid. The hybrid architecture is a good solution for some of these problems that we already talked about and hence, I would say yes to that.

AD [00:15:29]: These hybrid architectures sound quite capital-intensive in the sense that you're building for grid connection, but you also have on-site capacity. I don't know if that's going to be necessarily grid-scale generation behind the meter on-site, or if it's going to be more modular equipment, but in either case, you're basically building for both being connected to the grid or doing it yourself. It seems that a lot is being spent. How do you think about the capital intensity of going hybrid vs committing to grid connection or behind the meter on-site?

RA [00:16:11]: There is that additional spending for a good reason. One is the reliability that you have from the grid, especially with the one equipment vs a shared equipment producing electrons on the grid. That is one thing that's obvious. The other thing is if there is additional capacity, and over time, we talked about how these power islands can actually serve the grid, so those are also options, meaning you can generate revenue especially with super dynamic pricing in states like California. California ISO and even PJM, those can serve additional demand. Yes, you're spending CAPEX, but then is there a creative way of serving other opportunities on the grid is the way how I think people will look at this. It's not just a one-way flow of electrons to the data centres. When you step back and look at it, it is more about a two-way street, and then the financial model makes a lot more sense. There is need for the grid as well. Hence, the location matters. You would put this hybrid not necessarily for just the data centre centric, but if you are bringing that closer to a load centre, I think there is a lot more value.

00:17:58] Regarding data centres adopting hybrid projects or hybrid power solutions, do you think they will hold on to on-site or behind-the-meter power in the long term, or is that primarily a bridge solution to get them to the point where they're grid connected and the grid has the slack to be reliable enough for both primary and backup functions?

RA [00:18:31]: It did stem from the current backlog, for sure. Over time, will it slow down? Perhaps. As the queue gets smaller or minimal, is there a need? Probably not, unless there are other business opportunities and evolution in the marketplace for the hybrids to use that CAPEX and then if the recovery time is much smaller. It's so hard to say yes or no, but these are the parameters that will really dictate whether the hybrid is going to slow down, to your point. It is currently solving a problem. Over time, time will tell, based on how these new business models can help the hybrid generation for that additional CAPEX. Even CAPEX probably is going to go down, the amount of money that you're investing today vs three years from now, because there is no queue, there is no demand, not as such, not as we talked about. There are multiple things like that that play into it. Hence, we can't really give a yes or no.

00:19:54] How are data centres impacting contracting dynamics in the industry compared to more traditional PPA [power purchase agreement] structures? When we talk about key aspects such as contract duration, pricing and so on, those would be important clauses, but is there anything more bespoke or custom that you've noticed making its way into data centre power contracts?

RA [00:20:26]: To an extent, yes, in short. Traditional PPAs typically are fixed-price, 10-year contracts, maybe 20 depending on how the person is, but most commonly, they are used in the past with refineries, with cities and then mostly for power generation, but you can combine heat and power. That's where the traditional PPAs were known to us. Then obviously, there's renewable energy and relatively standard terms. Now, what's changing here is regulators and utilities are now pushing the full cost burden and even the stranded asset risk onto data centre. I think even that happened in the last week. There was actually a pushback. It was from one of the regulated utilities where there was no clear explanation of stranded asset risk, so the public service commission did not even approve. That is active right now. As a result, the contracting dynamics are definitely changing, to that point. What does that mean? There is a minimum billing demand that people are looking for. There is an exit fee for early departures. All new infrastructure costs have to be borne by the data centre customers.

In a way, the traditional PPA where I bring all the CAPEX and you pay this for a time-based recovery, I don't think that is happening. It's more about how do I present this case to PSC so that the risk is very minimal. That is what is driving it. Then it also depends on public service commission to public service commission, not even ISO. Some states are very, very conservative in some ways as it should and then that is what is driving these contracting dynamics and the traditional PPAs.

AD [00:22:47]: If the shift is towards hedging against offtaker or counterparty risk, what does that mean for the pricing of it all? I would imagine that if you are contracting with a data centre and you write in language that protects you from early exit or any stranded asset risk, there's some penalty on the pricing vs whatever benchmark you want to use. Is that something you've seen? Is that the case?

RA [00:23:23]: I haven't done a comparison, but I would think it is logical to assume that there will be a price difference as well because the capital expenditure and the upfront risk is being minimised, so I would think there is going to be that. I have not done a clear comparison though. These are public documents, so obviously, the PSCs keep these documentation very transparent. One can go in and check that, I would think. Logically, I would think it does come down to the cost reduction because of the risk that they are actually sharing to your point.

00:24:08] When we talk about contracting structures for data centres and power specifically, are there material differences between new builds and the existing generation? NRG Energy is a big IPP [independent power producer] in the space and has been talking a lot about data centres. The company highlighted that for its first BYOP project in Texas with an undisclosed hyperscaler offtaker, it has priced that new-build BYOP project in terms of both CAPEX and OPEX recovery, where I believe 95% of the project's estimated cash flow is meant to come from the CAPEX payments or pricing for capacity, essentially things that are not variable. Do you broadly see that being the case, where these IPPs or anyone trying to power data centres are going to price their contracts to essentially recover their capital in structured payments?

RA [00:25:17]: That is certainly something that I have come across as well. That is one way of a structured payment recovery over years, is one way, but then that might not fly everywhere. For example, what I've also seen was an upfront capital investment, whatever percent, 35%, 50%, whatever, that the hyperscaler, in this case, can put together depending on the public service commission's regulation. Then they are talking about structured payment in the early years, for example. It's a time-based mechanism so that the risk is much, much, much more minimal for the utility customers and the rate payers as they call it. That's what is driving these things. Hence, I think there will be a PSC-by-PSC, public service commission by public service commission, or even a state-by-state model catering to their own risk levels. That would be my view of the market emerging right now.

00:26:45] In terms of contracting dynamics, what are we seeing with regard to demands from the data centres to the power providers? I've heard about the classic five-nines reliability, but what else have you seen with regard to what the data centre customer wants in terms of redundancy or backup capacity? Do they have any energy mix demands, concerns or assertions over dispatch strategy or anything such as that?

RA [00:27:16]: Now, five-nines, if you actually look at that, what is it, five minutes approximately per year. It is a long-standing data centre industry benchmark. When you actually take it down to the critical component level, it does tie to the critical components, for example, N plus one, meaning the number of turbines plus one. There is definitely the methodology and the component-level infrastructure backing that. That is there. That is one part of it. The next thing is, in addition to that, the top-rated reliability, yes, there is the redundancy where a grid-connected redundancy would be highly beneficial and then also the overall quality, the power quality, which is very much known to the industry because part of it is about managing the entire grid in that frequency, for example, that is known to the utility industry, and they have done that phenomenally well over multiple decades. The power quality is another major ask from the data centres as well. I think those are the three primary ones, power quality, redundancy and reliability among long-term maintenance and then cost-effectiveness, like what we talked about.

00:29:15] My understanding is that data centre consumption can ramp or change significantly in exceedingly short periods of time. From the operator's perspective, how do you think generators are going to match these load swings? Or if it was a utility, how would they adjust for this issue, thinking about grid planning, for example?

RA [00:29:40]: As a generation mix, for example, there is the newer demand being imposed from the grid side rather than requested by data centre, like what we talked about going into the previous one. Reliability and redundancy remain the baseline ask, but the newer demand is being imposed from the grid side as well rather than requested by data centres because of the consumer demand, rooftop solar. There are other factors. Flexibility as the price of fast interconnection. The other side of it is transmission access for flexible large loads and then that is where some utilities are already piloting that matter. I don't know if that's what you are referring to, but that is the other side of the coin that we should also be looking at for the data centres and then their interaction with the grid today. These are currently playing or coming from the energy mix, the dispatch as well, if not the location and how close they are, geographical proximity to the load centre perhaps.

00:31:09] In terms of what data centre consumption profiles mean for IPPs, generators and utilities, what key trends do you see emerging? Would there be any uptake in energy storage to smooth generation profiles or have built-in redundancy? Would there be any shift to a reliability-first fleet as an operator rather than going with what is the lowest cost or the typical merit order of generation?

RA [00:31:55]: The current data centre load is also opening up a few more factors, like what we talked about, load curtailment, for example, curtailing the data centre load for as little as even 0.25% could probably roughly open up, let's say, 50-70GW of new grid capacity. The system savings can certainly be in hundreds of billions of dollars over the next decade or so. That's what some of these reports are indicating. Then as a result, the load is becoming a flexible, curtailable resource in addition to or as opposed to pure demand. When we look at the summer reliability assessment, for example, the demand forecast for some of these ISOs is growing by 3-5%. Now, explicitly, that is because these data centres can now be curtailed by the grid operators. What am I trying to indicate here is this load curtailment and the load balancing will be another thing that as the players are in the grid, the data centres have to be ready for and the grid operators have to be also looking forward to, for example.

Now, next question is how do generators match the fast data centre load swings? I think that was another part of your question there. That is essentially happening within the infrastructure itself for the data centre. For example, there are supercapacitors in addition to the battery storage systems and then there is also MEMS that are orchestrating these set points continuously to gas, whether it be gas, whether it be supercapacitors, whether it be battery storage. As the workloads are coming in, they are ideally doing this optimisation and then load balancing right then and there. That is what is happening on a high level, just to name a few things that are constantly changing and some of those are here for us to stay.

00:34:34] You mentioned data centre load potentially being more flexible rather than being a fixed demand-growth issue. The idea that data centres can be curtailed or are more capable of flexible demand response runs counter to what we've seen in the media. We've seen considerable data centre pushback against being categorised as flexible loads or any attempt by ISOs [independent system operators] to have them designated as curtailable or otherwise more flexible. In reality, can data centres be flexible, given the amount of pushback against the idea?

RA [00:35:26]: That is the current situation. What we are also indicating is how they can potentially be part of this ecosystem. Whatever you're saying is correct, but then this provides a case for us to think of strengthening the infrastructure is one way to look at it. Will it happen this year, next year? Probably not, because there's already enough of load that needs to be met, but over time, this will become handy is the way how I look at it, the load curtailment and the data centres, yes.

00:36:09] How does simultaneous ramping across multiple large data centres on the same transmission network affect regional grid stability and planning beyond single-site load matching?

RA [00:36:27]: I think this is becoming very realistic. I don't think this is theoretical anymore. There was also NERC. This was this May, I think, there was highest severity alert that was issued on the computational load risk to the bulk power system. That is happening. Actually, even a couple of years ago, this is also in the PJM area, there was a transmission fault that actually caused a huge loss of about 1,500MW of data centre load across multiple substations, 20-plus substations. As a result of all this, as we're talking briefly, the system frequency is supposed to be at 60Hz and then it briefly pushed the system frequency to slightly above the levels that they wanted to keep it under. If you look at the NERC reports, there are definitely reports like those. As a result, the transmission network is affecting the grid stability and these are proven. These are already there.

Now, what FERC is doing in addition to that to mitigate that is they have added mandatory reliability standards for these data centre loads and the compute workloads, assessing this directly. In addition, they're also talking about load behaviour across colocated data centres, which can erode the natural load diversity and then degrading the frequency, stability margin as well. That is what we are seeing. Definitely not theoretical. It is actually impacting us right now.

00:38:33] In the case of new build projects, how are data centre speed-to-power demands impacting the urgency to get capacity up and running?

RA [00:38:43]: It's one of the biggest drivers, I would say, the project economics, then cost alone, the speed to power. The equipment lead times are now the binding constraint, if you want to call it. For example, we talked about the step-up transformers alone are taking, what, three years or so, three-plus years. As a result, the new build projects are taking significantly longer unless there are some debottlenecking measures that are being implemented. Even those, nothing will happen today. I've also come across some other interesting VC-funded programmes where there are some other things. I know now we are going into other debottlenecking issues, but the point is the speed to power urgency is really, really opening up various options for the industry, and they are literally willing to try anything and everything, including the AI-based load forecasting, AI to actually unlock the data centre capability or anything to do with debottlenecking the component within that specific module, if I may. Very much affecting. Over time, this is supposed to be minimised. Everybody is on the toes, including these equipment manufacturers and also the used market, for that matter. They're also looking at capturing the pain point or capturing the opportunity behind the pain point to minimise that impact and making it easy for the new build projects to be finished.

00:40:48] What are some AI implementation use cases in the power and utility sector, especially as it pertains to facilitating data centres and related load growth?

RA [00:41:05]: That's where I spent a good amount of time, especially with respect to AI applied to dispatch strategy, generation profile optimisation. These two are very much value-generating use cases or programmes that most of the utilities out there are already currently doing. This also works indirectly with the reliability as well, meaning the dispatch strategy and the generation profile optimisation indirectly can also bring the outages, the outage times. The SAIDI and SAIFI, some of the industry metrics are also getting better. All in all, there is value being generated by these things. For example, the plant performance optimisation itself can drive anywhere from… Depending on the capacity, a 1% capacity improvement or a generation improvement can result in USD 50m-100m depending on the size of your fleet. Hence, the value is directly there. Historically, the heat rate optimisation is also very, very common.

Using AI, you can actually do AI-driven heat rate optimisation and then that is reducing the heat rates almost 1-2% with no capital investment required. There is a lot of application that is happening not only in generation, but also in the dispatch, outage duration minimisation, plant value optimisation, parts, condition-based outages, various ways. In addition to using AI to identify an outage, minimise the outage, correct the outage, there are even self-correcting devices that are placed by the utilities that are heavily using AI to do whatever they can do for the self-healing mechanisms to work in real time.

00:43:45] How should we think about the implementation of AI in operations for day-to-day things, such as dispatch strategy or generation profile optimisation? How should we think about debottlenecking, improving uptime and capacity, or even higher outputs related to that? How would an IPP think about unit cost reductions at the plant? How does AI play a role here?

RA [00:44:21]: These are all interconnected. If we break it down, it comes down to a plant with the capacity of X number of megawatts and that capacity will be declared. Then when it's time to dispatch, you better have that, let's say, X equal 100, you better have that 100MW plus or minus 1%. It's really that simple. When you actually are bringing that level of accuracy across thousands of assets… In this case, I gave you an example of 100, but then I know we have plants that go up to 1,000MW, 1,500MW even. Now, you have temperatures playing into it, your weather, humidity, all of those. When you look at it, it's a mathematical problem of, one, declaring and then estimating how much and then on the real dispatch, you are delivering that many megawatts precisely. Otherwise, there are not only the grid-related stability issues, but you will be fined for compliance if that happens repeatedly. There are multiple things in every place, the entire value chain, the example that I gave you all the way from the asset with the capacity factor determination, all the way to generation and then how much AI is used all over the place.

It's very much time-related data and then there are multiple years' worth of data, not only at the IPPs, but also at the ISOs. Everybody is working collectively to understand how to make it better. There is a huge element of using… Even before we used AI, there are traditional machine learning techniques that are used very, very heavily and now with the advent of generative AI and the conversational ability, you are having much more closer, quicker and easier interaction with the data that makes the job easier for these things. That's the way how I'd summarise the use of AI without going way too deep.

00:47:14] Do you see any commercial upsides with AI implementation in the power and utility space – for example, managing counterparty risk or unlocking additional revenue levers?

RA [00:47:32]: Absolutely. There is the heat rate improvement, for example, and the higher output at a similar marginal cost that we talked about it. Then AI-driven forecasting is actually more and more used, and this is not new by any means, optimising the bidding across the full revenue stack. When I say full revenue stack, you're talking about the energy, the capacity, ancillary services and then various pricing mechanisms that you have, including the hedging part of it, if you're allowed to do that. In the entire value stream and the revenue stack, AI can be used for commercial and margin enhancements, and people are doing this. This is not new by any means as I was saying. Traditional machine learning has been used in this. It's only getting faster now, again, for this. The opportunity is there. People are using it. People are doubling up on those opportunities. Then there are acquisitions happening as well to unite this real-time ancillary market forecasting with automated trading execution. For example, there was a recent acquisition that actually was specifically centred around that, where the ancillary market can be forecasted and then also be added with automated trading. The operations are getting more and more streamlined, combined, even agentified as well with the AI additions. That is certainly happening.

00:49:33] When we think about margin enhancement, such as the example you gave in terms of using AI to drop heat rates, what have you seen there? Have you seen a specific heat rate improvement from AI implementation? Could you provide a rough range? How should we compare that to the complexity and cost to implement that AI solution?

RA [00:50:02]: Heat rate optimisation is a well-known subject even before the pre-ChatGPT era. For example, simple things such as weather-driven optimisation is very commonly done across the industry. There are two times that they do, one before summer, one before winter, because the density of the air directly plays into it and hence, the heat rate. Those are very manually driven methods to tune the machine. Now, with AI, what you are doing is you are actually operating within that envelope of the machine and even within that operating envelope for that season, are you maximising it is what it is. There are recommended control set points that are driven for these machines, and can you run as close as possible to that control set point is really the ultimate AI-driven optimisation in a nutshell, in a simplistic way. The industry records indicated somewhere around 1-3%, depending on the age of the machine, where you are, how tuned the machine is and then the type of fuel that you're using, meaning how clean the fuel is and then how well the machine was maintained. There are many things that go into it to give you the exact number that you would get from the optimisation, but it can go as high as that 2-3%, in that range, based on the industry norms and the studies out there.

00:51:54] You mentioned that you've seen a 2-3% improvement in heat rate, which, for a more modern CCGT [combined cycle gas turbine], is roughly 7,000 Btu [British thermal units] per kilowatt-hour. How should we think about the cost to do the heat rate optimisation you talked about? Is there a big CAPEX or running OPEX associated with that? How should we think about the lasting benefits to that heat rate? Is that 2-3% improvement essentially permanent?

RA [00:52:34]: Good question. That does not require any capital investment as I was saying. Essentially, it continuously tunes hundreds of set points for each and every parameter, each combination and then they take that over time. Hence, the recommended control set point changes are provided and implemented. That's entirely AI-driven software and data-driven, all using the existing infrastructure that you have without any capital investment as such. That specific thing we talked about, the heat rate optimisation, that is consistently done beyond the seasonal tuning as well. There are even proven reports out there. For example, the recent one that I saw was an IPP, Vistra, for example, rolled it out to multiple units. There is an average of 1% gain is what I read. This resulted in a 1% gain, USD 20m-plus approximately, something like that. Again, this is a cited study that was done by an IPP, all using their existing control system and the data points that they have accumulated over time, and you are not putting any more infrastructure.

Now, the second part of the question, how long will it stay? Since there is no capital as such, there is the tendency to actually go down as the season changes, but now the seasonal tuning and the continuous optimisation with the age of the machine and with how it was operated under condition can possibly keep it as long as it can be. That is what dictates how long will it stay beyond the season. Maybe towards the end of the season, it probably slowly comes down maybe to the half or whatever that percentage is depending on the age of the machine, but then with proper tuning, again, you might be able to take it back and then do a continuous optimisation using the current data for that time or for that age of the machine at that time.

00:55:29] Have you seen an increased collaboration between the power sector and its prospective technology customers and how the partnerships there can create opportunities for AI implementation and related debottlenecking? Are you seeing any other opportunities?

RA [00:55:51]: I saw some partnerships that formed already. For example, there was Google partnering with the PJM. I already mentioned that. Google also, for example, has partnered with Florida Power & Light, or even AWS and Duke Energy, for example, there are actually recent studies about them partnering. Meta and Entergy was another one that is also in the news for partnerships. There are definitely partnerships that are forming so that they can bring their expertise, but also solve their own problem in a way, because these utilities are expected to provide these additional loads, the load demand. There is definitely a lot of interesting and very valuable partnerships that are forming across the industry here in the US, at least. Not to mention, I also read some very interesting partnerships that are being looked at between the Southeast Asia, for example, even Japan, looking at some of the hyperscalers here in the US as well. We work very closely even as an industry. Before, my experience indicates that US, Europe and Japan, at least, they work very, very closely. They share the lessons. They cooperatively bring the bar higher and higher. I am expecting a similar thing to happen as well.

AD [00:57:54]: We will now conclude the Interview. Thanks so much for joining me and sharing your insight on the power and utility sector. It's been very much appreciated.

Transcription ends at 00:58:18 of the recorded material