Frontier Forum: The new power map for AI infrastructure

Summary

---

Podcast Summary: Frontier Forum: The New Power Map for AI Infrastructure

Podcast: Catalyst with Shayle Kann (Latitude Media)
Date: October 21, 2025
Featured Guest: KR Sridhar, CEO of Bloom Energy
Host/Moderators: Steven Lacy, Armand

---

Episode Overview

This episode explores the rapidly shifting landscape of power infrastructure for AI data centers. As artificial intelligence propels data center growth to unprecedented levels, the electricity system faces major challenges of scale, speed, and sustainability. KR Sridhar, CEO of Bloom Energy, provides deep insights into why traditional grid models are struggling, the critical role of onsite power (notably fuel cells), and how industry mindsets are shifting toward “AI factories” with bespoke microgrid solutions. The discussion also highlights the economics, policy, and national security implications of this new power era.

---

Key Discussion Points & Insights

1. Explosion in Data Center Power Demand

• U.S. data centers could account for over 10% of total electricity demand within a few years, up from just 2% recently. (00:09)
• The speed and scale of AI infrastructure growth is without precedent, with trillions in investment and hundreds of MWs being planned and deployed rapidly. (00:21, 00:52)

2. Grid vs. Onsite Generation: False Dichotomy

• Traditional central grid power was critical in the 20th century but is now mismatched for data centers’ digital power needs. (01:29)
• Mismatches: Grid supplies AC power; data centers require high-quality DC. Extensive conversion equipment is “just band aids.” (01:54)
• Onsite solid oxide fuel cells (like Bloom’s) offer purpose-built, modular, efficient DC power with lower conversion losses. (02:35)
• The “car vs. horse” analogy: “When you use Bloom, you have a one step process. When you use a conventional combustion technology, you use a six step process.” – KR Sridhar (02:48)

3. Industry Shift Towards Onsite / Off-Grid Solutions

• Sentiment is rapidly changing: a recent Bloom Energy survey showed a jump from 1% (18 months ago) to 29% (now) considering onsite off-grid power for data centers. (03:41)
• The driver: urgent demand and lack of grid availability. “The digital age needs digital electrons coming from a digital source.” – KR Sridhar (03:57)

4. Debate: Does the Grid Still Matter?

• Armand presents the counter-argument: the grid offers scale, cost advantage, and optionality. (04:38)
• KR Sridhar’s nuanced response: If grid capacity is there, use it—“Why would you be stupid enough to put that on your balance sheet?” (05:12)
• But politics and economics: “The notion that the grid, the commons, will somehow build all this capacity so one or two users can use all that capacity at the expense of everybody else is also a political wish, which I don't think will happen in reality.” – KR Sridhar (05:51)

5. The Imperative of Speed

• Building out grid capacity simply cannot keep pace: “You cannot build the grid at that speed no matter what you did.” (06:44)
• Even without permitting delays, constraints like copper, transformers, skillsets, and capital are real bottlenecks. (07:38)
• Onsite solutions aren’t meant to replace but to supplement an “all of the above” strategy for reliability and speed. (08:16)

6. Lessons from Heavy Industry

• All energy-intensive industries (aluminum, steel, chemicals, etc.) run on captive, onsite power for reliability, synergies, and economics. (08:59)
• Data centers have been the exception but will increasingly mirror this model as “AI factories.” (10:38)

7. Future Microgrid Architectures

• Expect diverse, purpose-built microgrids for each site:
  • Baseload: fuel cells for efficiency and zero air pollution
  • Peaking: engines or turbines
  • Storage: batteries, flywheels, ultracaps
  • All tailored to specific loads and with ultra-high reliability, especially in urban environments. (11:27)
• "We are in the dawn of a reimagining electricity and how it is delivered to a customer." – KR Sridhar (13:13)

8. AI Workloads: Training vs. Inference

• Today: 90% of load is for AI model training (centralized, huge data centers).
• In 3–4 years: 90% of load will be inference (smaller, edge data centers, closer to users), but both will grow. (13:36)
• Edge inference centers (5–30 MW each) will pose enormous distribution-level challenges for utilities. (15:17)

9. Customer Pain Points

• Power volatility: AI loads can swing from 10% to 100% of rated power in milliseconds, stressing traditional systems. (16:35)
• Solution: Direct DC delivery at proper voltages (800V DC) minimizes conversion losses and copper requirements. (17:50)
• Cooling: “Liquid cooling in racks” becomes mandatory; Bloom’s solutions help with integration. (18:29)

10. Sustainability vs. Speed

• Sustainability now ranks lower—but customers want future-proofed solutions that allow for net-zero add-ons later. (19:11)
• “Without natural gas we are not going to be able to power the AI revolution. That's a fact.” – KR Sridhar (19:51)
• Bloom systems are compatible with carbon capture and can transition to green molecules (e.g., hydrogen, ammonia) when available. (20:18)
• Carbon capture at scale for electricity will precede widespread green hydrogen/adoption. (21:47)

11. Supply Chain Resilience

• Bloom’s early decision to geographically diversify supply chains is now paying off amid global trade tensions. (24:16)
• Native DC power eliminates need for transformers and reduces critical material bottlenecks. (26:12)

12. Gas Availability

• U.S. has sufficient gas, but local pipeline buildout remains a potential hurdle, varies by region. (27:04)

13. Customer Deployment Approaches

• Most customers still want grid backup if available (“insurance for almost nothing”), but off-grid solutions are increasingly viable and reliable. (28:35)
• Bloom’s tech has run “islanded” mission-critical sites (like eBay’s center) without grid support for over a decade. (28:57)

14. Using AI for Power System Optimization

• Bloom leverages AI and digital twins for real-time health monitoring, performance optimization, and continuous improvement of each fuel cell stack. (30:12)
• AI enables differentiated, per-unit performance management and predictive maintenance. (31:17)

15. AI as National Priority

• Winning the AI race is an economic, technological, and national security imperative; failure would risk U.S. prosperity and geopolitical leadership. (32:28)
  • “If we lose our technology superiority, we will lose our economic superiority that hinges on AI.” – KR Sridhar (33:55)
  • “There'll be one thing that every government is going to agree is we cannot afford to not win the AI race.” (34:55)

---

Notable Quotes & Memorable Moments

• “This AI movement is forcing data centers to finally say the digital age needs digital electrons coming from a digital source.” – KR Sridhar (03:57)
• “If you have excess capacity in the grid … why would you be stupid enough to put that on your balance sheet and take that responsibility yourself? Absolutely yes. … But ... that grid had excess capacity, so the data centers were able to use it. Now ... that politics is never going to fly.” – KR Sridhar (05:12–06:15)
• “You cannot build the grid at that speed no matter what you did.” – KR Sridhar (06:46)
• “There is not a single industry that builds factories that have very high power demands that doesn't have captive power...” – KR Sridhar (08:59)
• “We are in the dawn of a reimagining of electricity and how it is delivered to a customer.” – KR Sridhar (13:13)
• “Without natural gas we are not going to be able to power the AI revolution.” – KR Sridhar (19:51)
• “If we lose our technology superiority, we will lose our economic superiority that hinges on AI.” – KR Sridhar (33:55)

---

Timestamps for Key Segments

• 00:09 – Data center electricity demand explodes
• 01:29 – Centralized grid model vs onsite generation debate
• 03:41 – Industry’s shift toward off-grid solutions
• 05:12 – Economics and politics of grid expansion
• 06:44 – Speed as the driver for off-grid power
• 08:59 – Captive power in heavy industry as precedent
• 11:27 – All-of-the-above microgrid architectures for AI data centers
• 13:36 – The rise of inference at the edge
• 16:35 – Customer pain points: load swings, voltage, and cooling
• 19:11 – Tension between speed and sustainability goals
• 21:47 – Carbon capture vs. green hydrogen timelines
• 24:16 – Supply chain resilience & material choices
• 27:04 – Gas pipeline and local infrastructure constraints
• 28:35 – Real-world deployments: grid-anchored and islanded
• 30:12 – Digital twins and using AI for system optimization
• 32:28 – The U.S. AI race: economic, security, and policy stakes

---

Conclusion

This episode provides a comprehensive look at the new “power map” being drawn by the demands of AI infrastructure. The convergence of decarbonization, urgent capacity needs, and geopolitical factors is driving rapid innovation and disruption—particularly in how and where power is made and delivered for data centers. KR Sridhar’s insights make clear that the era of “AI factories” will fundamentally reshape both technology and policy in the years to come.

---

Transcript

[Steven Lacy]

This is a frontier forum brought to.

[KR Sridhar]

You by Latitude Studios.

[Steven Lacy]

There's a lot of debate about where exactly data center electricity demand is headed, but we know it's steeply up and could account for more than 10% of U.S. demand in the next few years. That's up from 2% a couple years ago.

[KR Sridhar]

That's an unbelievable leap. And that leap really, if the US has to stay in the AI race, has to happen in the next four to five years.

[Steven Lacy]

KR Sridhar is the CEO of fuel cell maker Bloom Energy. About a third of the company's solid oxide fuel cells are sighted at data centers. We're talking hundreds of megawatts with that number growing quickly. So KR knows the industry well and he's never seen anything like the current AI infrastructure boom.

[KR Sridhar]

It's the speed and the scale. I truly believe that this is a circular trend that's gonna continue for well over a decade.

[Steven Lacy]

Data centers are now being planned and built faster than the electricity system can handle. That has sparked another debate about the role of on site generation and whether data centers will be built off grid. Now, some folks argue that co located behind the meter resources will be rare and that the smartest economic move for data centers will be to connect to the grid, not walk away from it. KR believes deeply in the importance of the grid, but he also thinks that on site solutions and including fuel cells will be critical for building infrastructure at the pace of AI expansion.

[KR Sridhar]

What the centralized power plants and the electric grid did to humanity in the 20th century is nothing but unbelievably amazing. That model still has a place because electricity is the ultimate perishable and having that large flywheel called the grid is very important. So it's an and and, not an or. However, when you go to data centers, here's what happens.

[Steven Lacy]

There are many misalignments between data centers and the grid. A data center consumes DC power. The grid sends AC power. And the quality of the power on the electricity system, which shifts depending on loads, turning on and off and changes to the generation mix, is not anywhere near what a data center can accept.

[KR Sridhar]

What do you do? You put all these band aids, make it high quality using UPS quality power, step it down from high voltage to medium voltage to low voltage. All the transformers, all the power conversion devices, they're all band aids because they are taking a mechanical age electricity and trying to adapt it to a digital age computer.

[Steven Lacy]

A fuel cell is a solid state power device that KR says is purpose built for a digital facility like a data center. He compares today's power strategies to trying to build a faster horse and buggy when we really need a car for the AI era.

[KR Sridhar]

When you use Blum, you have a one step process. When you use a conventional combustion technology, you use a six step process. The horse is consuming a lot of water, it's consuming a lot of hay and a lot of other things to walk the mile. We consume very little to walk that mile.

[Steven Lacy]

I'm Steven Lacy and this week we have a conversation with Bloom Energy CEO K.R. sridhar about rapidly changing power strategies for data centers. This conversation was recorded live as part of Latitude Media's Frontier Forum series. The data center industry is investing trillions in infrastructure while up against permitting timelines and grid constraints. I talked with KR about customer pain points, the unique power demands of AI data centers, and why on site generation will be critical for winning the AI race.

[KR Sridhar]

So we did a survey two years ago and now 18 months ago the survey respondents said 1% of the survey respondents said they will use on site power that is not connected to the grid. Two months ago when we did the same survey, 29% of the respondents said they will use on site power without being connected to the grid. The sentiment is changing enormously simply because of lack of availability of the power and understanding that we have to use this. So this AI movement is forcing data centers to finally say the digital age needs digital electrons coming from a digital source. I think once they get used to this and understand all the advantages of this digitally purpose built on site power, they are not going to switch back.

[Armand]

Let me take the other side of that. I think that people who believe that most of these data centers will be grid connected would say the grid offers scale and optionality that no off grid system could potentially match. It's much more cost effective. You have to overbuild the system for an off grid data center and that.

[Steven Lacy]

There'S a lot of grid capacity left.

[Armand]

That you just need to get creative to unlock that grid capacity. What do you say to those arguments about why off grid data centers may not blossom in the way that you're painting?

[KR Sridhar]

I'm going to surprise you by saying that I would agree with those arguments for a very simple reason. If you have excess capacity in the grid and you can get that without putting that on your balance sheet, why would you be stupid enough to put that on your balance sheet and take that responsibility yourself? Absolutely yes. That's exactly what all the data centers have been doing so far. But here is why. The commons, which is the US Taxpayer and the US government build this excess capacity in the grid that was paid for by the ratepayers one way or another, either through taxes or through their bills. That grid had excess capacity, so the data centers were able to use it. Now, if the grid were to build excess capacity purely to meet this huge demand and expects the taxpayer to pay for it, so the hyperscaler can benefit, that politics is never going to fly. So the idea that somehow data centers will go and build their own on site power if the grid were available is a fallacy. The notion that the grid, the commons, will somehow build all this capacity so one or two users can use all that capacity at the expense of everybody else is also a political wish, which I don't think will happen in reality.

[Armand]

And then what about just speed in general? How much is speed to power? The phrase of the moment, how much is that going to drive? The off grid approach, that is going.

[KR Sridhar]

To be the key. The truth is this is a true race. It's a race between corporations trying to get supremacy. It's a race between the western world and the rest of the world competing for AI supremacy. One thing that everybody will agree is being number two or number three is not an option. Speed is going to be everything. And there it's just simple knowledge for people who understand the grid that you cannot build the grid at that speed no matter what you did. You know, the common refrain is if only permitting and permissions go away. Okay, yes, that is a big large issue. But imagine somebody waves a magic wave like wand, the king and I can build high voltage transmission lines running straight over your house. And you have no right to object to it. So what? What's the amount of copper needed? What's the amount of transformers needed? What are the leads times for it? Where are you going to get that from? Where are the skill sets to build these things? How much capital is required? How fast will that happen? Realistically, over a five, six, seven year period, it cannot catch up with the demand base that exists out there. That doesn't mean that we shouldn't upgrade the grid. That's why I started off saying the grid has a very important place. It's an all of the above strategy. We should be investing heavily in optimizing the grids. The things that you talked about. Can we use AI to make the grid better and extract more out of less? Right. Can we make it more efficient? Can we figure out other ways of optimizing it? Can we add more transmission? Can we add better technologies to remove the bottlenecks? Where they exist and get more out of the grid that the answer is yes. But is that the solution? No, it is part of a solution.

[Armand]

Tell me about what other industries you look to where this is a model. Are there large industrial loads that you see that have accomplished this?

[KR Sridhar]

You know, when I look around, there is not a single industry that builds factories that have very high power demands that doesn't have captive power. Let me walk you through it. You take metals, whether it's a steel mill or it's aluminum. If you go to an Alcoa plant in Indiana, they'll have a captive power plant that's a few hundred megawatts. Because that kind of load they cannot depend on the grid. If they were to do a shutdown for maintenance and you suddenly drop 100 megawatts to 200 megawatts off the grid, what happens to the grid? You cannot do that. It has to be captive power. You go to cement, large cement plants, wholesome. If you go and look at it, they will have their capital power plants. But then you go to refineries, same thing. But here's what happens. Here's another reason why they do it. It's not just because the load is so large. Number one, you get the reliability and resiliency you need when you have a capital power plant and not when you depend on the grid. These are continuous processes. If they shut down for whatever reasons that happen in the grid, that cost of in process inventory, everything is too high. Right? Number one, reliability and resiliency. Number two, there are synergies that come from keeping that captive power plant right where you use the power. Here's why. A refinery for example, will use the off gases that are coming out and as a fuel into their power plant. And they can get better economics. A steel mill, a cement plant, same thing. Then you look at paper mills. Paper mills do the same thing. You go and look at chemicals, pharmaceuticals, it's the same anywhere. You have seen an extremely large load in a factory. It's a capital power plant. We are finally building high electricity intensity AI factories. These data centers are AI factories. They have been the exception and not the rule. They are now going to have to follow the rule.

[Armand]

So I definitely take your point for the all of the above approach. And you've said that this is not a winner take all market. If we apply that to the actual groupings of on site technologies for these data centers, what do you expect to see? When we look at fuel cells, gas turbines, renewables, eventually nuclear. How do you imagine them all fitting of System designs are you currently seeing?

[KR Sridhar]

These are, you know, it'll be all of the above, It'll be all of the above and more. Okay. New things that come along. And so what I would say is microgrids supporting such infrastructure will definitely have a multiplicity of generation sources. Each one comes with very different attributes. And you may need particular attributes depending on what you're in and need is. So as, as soon as you start purpose building for your needs, you're going to exactly tailor it for the needs that you have. And so you can imagine for some extremely peak loads, there being some engines or turbines for a steady baseload, there being a fuel cell, because we'll be the most efficient from a fuel perspective and not create pollution, zero pollution. Right. Air pollution. That's super important for an AI factory because unlike a refinery, unlike a cement plant, which will by definition only be located far away, and inherently they create so much pollution that having just a captive power add to that pollution is de minimis. Unlike that AI factory being built in a population center, especially for inference, cannot afford to have air pollution. So that becomes an extremely important aspect. Then you got the load swinging up and down for that. You're going to have a variety of storage technologies, from batteries to flywheels to ultracaps to super caps. They will all become part of the microgrid. How you adapt it, what particular technologies you use will depend exactly on what kind of a data center it is, what does its load profile look like, and it has to be tailored and met perfectly. So I think we are in the dawn of a reimagining electricity and how it is delivered to a customer.

[Armand]

Yeah. That brings me to another question. It's kind of a two part question. One was on the differences between training and inference workloads. And is it possible that there is even more upside to growth than we.

[KR Sridhar]

Previously thought based on historical data, including our conventional CPU data centers? And if you go back to any form of human use of a new technology, a new tool, what you see in history is while things start in a very centralized fashion far away, ultimately the value is only created and the growth comes when that technology comes to the edge, closer to where people are, closer to where machines are, closer to where people are. And that is the predominant use. That applies to computing, that applies to telephony, that applies to everything else. So what do I mean by that? Today, most of the AI load we are all focused on and are talking about is the hundreds of megawatts to a gigawatt kind of a data center for training. While the total capacity of that will keep increasing over the next eight to 10 years. In terms of percentages, while today the training is 90% of the AI load and 10% is inference or somewhere in that neighborhood three to four years from now, I predict that 90% of the load will be inference and 10% will be training, even though the 10% in training then will be significantly larger than the 90% training today. That's what I mean by proportionality. They'll both grow, but proportionally this is where it'll be. The inference Data centers for AI, unlike the CPU data centers of today that are 1 to 10 megawatts, typically on the edge, are going to be somewhere between 5 to 30 megawatts on the edge. So if you live in Austin, if you live in Dallas, if you live in Chicago, if you live in Manhattan, if there are hundred inference data centers that need to come that need 10 to 30 megawatts of load, God bless you. If you went to your local utility and asked them for that load and they're able to provide that to you. If you thought transmission was a problem, wait till you figure out what distribution is. That's the surface streets, right? Transmission is your highways. If you think the highways congested, wait till you get to the surface streets to see what the surface streets are doing, right? So you need on site power for that. So I completely agree with Armand that that's going to be a huge growth area. And the beauty of our technology, which is modular, scalable, with no compromise in efficiency or performance or reliability, whether you do it in 5 megawatt chunks or 50 megawatt chunks, or 500 megawatt chunks, that advantage that we bring, again, is the car, unlike the horse.

[Armand]

What are the other pain points you're seeing from customers when it comes to power supply and delivery? Are there any others that we haven't touched on that you think are particularly acute?

[KR Sridhar]

Yes. The data center customers, the AI practitioners, know this extremely well. The AI load, unlike the CPU load, is very volatile. It has milliseconds, seconds and minutes variation that go from 10% of a load to 100% of the load. So you're talking about, if it's 100 megawatt data center, as an example, you're talking about the load swinging from 10 megawatts to 100 megawatts up and down. And some of those variations happen in millisecond frequency. Some of them happen in seconds frequency sum in minutes frequency, the amount of backend band aids that they are putting to deal with this, as well as the harmonics it creates because of converting AC to DC have all become very large pain points. Number one, with our architecture of being able to provide dc, it becomes easier. That's one pain point. The second pain point that they are working on as we speak is shifting the mindset to a 800 volt DC bus going into Iraq. Let me explain this to you. When you are a large power plant that produces 50 megawatts, 500 megawatts worth of power, because so much power comes out and think of it as a pipe, the wire is a pipe. Think of water flowing in it and think of the total power coming in as the amount of water, the mass of water coming out. You can either increase the pressure or increase the diameter to increase the flow. Because copper gets very large if you keep it at low voltage, they only bring it out at high voltage. You have to convert them all. And that process in our system is completely obviated. Because we can natively provide 800 volt DC going to the grid, we think that's where the world will go. So that's a huge thing. That is a pain point. But we have an elegant solution to solve that problem. The third pain point that they see is cooling. The amount of cooling load necessitates that it's a liquid cooling that has to happen in the data center racks. And there again we have an elegant solution for them.

[Armand]

So when we look at the needs of customers, what they're looking for, do you see any tension between the speed to power and sustainability goals and do they have any concerns about a gas strategy?

[KR Sridhar]

I would say for pretty much all the data center operators today, the sustainability goal has shifted from being in the top three to being in the bottom three. However, if you offered them a pathway to I can give you the quick power today, but I can future proof you and I can give you add ons that get you to net zero or extremely low carbon footprint. They're super interested, they still want to do it, but they will not do it at the expense of waiting for power. That's where the sentiment is. That's where Bloom plays really well. Because we consume less fuel with natural gas, we have a lower carbon footprint compared to any other on site generation technology. Without natural gas we are not going to be able to power the AI revolution. That's a fact. That's a fact. We are lower comfort, however, because of the uniqueness of our technology. And your viewers can go watch that on our website. There's a very good explanation of carbon capture we put out a very clean stream of carbon dioxide coming out. So our users can capture the carbon dioxide coming out of our system and sequester it. And I think the world will get to that place very soon. So that's one way of zero carbon should a green molecule become available in the future. Hydrogen, ammonia, rng, biogas, you name it. And it'd be available in large quantities. Our systems are already pre wired for that. You don't even have to tell us you can feed that gas in and we will happily accept it and produce the electricity so we are future proofed. So in other words, think of us as we are giving you a 5G phone now, even though you don't have 5G in your neighborhood, but you don't have to wait for 5G to come to your neighborhood to use it. You can be using your LTE or 3G right now and when 5G comes, it automatically switches over. Our systems are built for a zero carbon molecule, but that molecule, when it comes, the customer can switch, but they don't have to wait for it to come.

[Armand]

And I remember when Bloom first emerged, there was a lot of talk about the potential of using green hydrogen in these fuel cells. And green hydrogen has been very difficult and expensive. There's limited supply. And you talked about carbon capture, this partnership with Chart Industries I which comes first, carbon capture at scale or green hydrogen?

[KR Sridhar]

Carbon capture at scale, hands down for electricity generation. So it depends on what you're asking for, right? The green hydrogen going into green ammonia coming from certain places, going to produce fertilizers and other hard to abate industries. That would make sense. It would not make sense to make a green ammonia and use that for electricity generation because the world is short on electricity. Green hydrogen is produced by bottling electricity. Why would you bottle something when you have current shortage? You don't save money if you can't feed yourself Today.

[Armand]

Who is asking for carbon capture? Are they corporate customers with sustainability goals? Are there other reasons? Is it primarily the folks who have clean energy targets and are now burning more gas?

[KR Sridhar]

Look, I think most corporations want to be responsible. I've not met a CEO yet who's waking up in the morning saying I want to add to climate change. That's not the goal of any corporation. When we tell them that we have this solution, they say as long as it meets the economics, they're willing to pay a premium. As long as it meets the economics. And it can be within scale. If you don't slow me down, I'm willing to buy that solution from you that we have heard across the board, from everybody across the board, from everybody. What we are seeing is every data center customer is super interested in our carbon capture solution. If they are deploying that solution in a region where the geology is suitable for sequestration, they're all interested. Now what they also want is speed. And the solution we offer is carbon capture and sequestration is a bolt on to our technology. We will provide you the power with the lowest carbon footprint today. Two years from now, when everything else is ready and available, like the Class 6 wells, all the pumping, all the technology with chart ready to be implemented, we will bolt that on for you and they're very happy to sign on to a solution like that.

[Armand]

We are of course in a wildly different trade environment now. And I know that you had worked over the years to move your supply chain away from China. Can you talk about your advantages and any risks you have in your supply chain currently? And are any of the major policy changes changing the way you're procuring materials and equipment?

[KR Sridhar]

So when I started Bloom 24 years ago, the idea was very simple, right? Centralized systems with centralized directive managed by very few is a command and control communistic kind of approach. The distributed is democracy. We were building a democratic electrical system with a distributed system. That wasn't our objective for that reason. That thought process permeated through everything we did. We said we want to build a supply chain where no one single region can solely choke a supply chain for growth. So we only picked materials, we only picked processes, we only picked supply chain partners. As long as we could geographically diversify across the board and not create any tensions in that system. Because the goal of the company was one day to be able to provide this around the world and create energy abundance, it was built at the very lofty goal. Therefore, we set it up that way and never constrained ourselves to any one country from the day we started. Today when we look back, that was a very good strategy. From where the world is going and the interdependence. And that interdependence becomes unhealthy. If that interdependence is on only one person and it's a one way, that's not interdependence anymore. We don't have that, number one. Number two, the question you asked, unlike other energy generation technologies, if you take turbines as a class, engines as a class, you go to the second third tier suppliers, you will find common bottlenecks. Unless that second third tier is able to grow, the rest of them can't grow, it's only a matter of allocation of between the people who consume it, who's going to get more and who's going to get less even as they grow. We are decoupled from that supply chain because of the completely different way that our technology works. And if you think of copper and transformers and things like that, because we natively produce DC power at the right voltage, we are just eliminating the need for a lot of those materials that are in short supply. So if you're trying to take the limited amount of materials available and say where can I get the most bang for my buck? You will see that our technology and our supply chain and how we think about that entire value stream will provide a lot more power with the same amount of copper compared to any other technology.

[Armand]

And are you seeing sufficient gas pipeline capacity?

[KR Sridhar]

The gas pipelines and the availability of gas is there. In fact, as you very well know, we are exporting a lot of LNG out of the country. So clearly we have excess of that. Now in terms of the pipeline coming to exactly where your location is, from the medium pressure pipeline to the last mile distribution, there could be issues there in terms of adding that infrastructure. Depending on what state you live in, it can be a matter of weeks or it can be a matter of years. If you happen to live in New York City or the state of New York two years ago where it was completely anti pipeline, that would be very hard to get. But in other states it's very easy to get. But I think that's all changing. That attitude is changing. Getting to reality of this is the cleanest way we can do it and better cannot be the enemy of good.

[Armand]

Let's talk about customers a little bit. You've deployed a lot of capacity with customers like Equinix and aep. You've inked a deal with Oracle to deliver on site fuel cells for cloud data centers in I think only 90 days if that's right. And that's unthinkable when we think about traditional grid connections. What are customers asking for? Are they currently supplementing with grid power or are they truly looking for these on site off grid solutions?

[KR Sridhar]

All of the above. Anytime a customer can connect to a grid, they would want to and we would recommend that they connect to the grid because that's insurance for almost nothing. Why would you not take it? It's an option. Why would you not take it? Right. So wherever the the grid is available, even if we are primary power and we have all the reliability, it costs almost nothing compared to the entire data center spend to be able to get that connection. But when that connection is not available, our solution offers a highly reliable option where they can go without it compared to any other solution out there. So there our customers. So today I would say we have a mix of both. Traditionally, over the last few years, if you looked at it, even though we were the primary source, even though we were mission critical, in many cases they were connected to the grid. But we have an example, 2013, when we did ebay, of having run that data center without losing power even once for over a 10 year period of islanded solution, where we were the only solution. The grid was connected, but it didn't have the capacity that they needed. Bloom offered that capacity. So for us, islanded mode, that's what the system was designed for. It's very easy. It's not an afterthought.

[Armand]

I want to turn now to the actual use of AI to model and optimize these systems. Can you talk about how you're using AI for digital twins and what that does to the actual performance of the systems and how you're modeling them at these large data center sites?

[KR Sridhar]

That's a great question. So if you just think about how we planned it and when we look back, we are actually very happy that we did it that way. Each one of our Lego block modules. I'm sure your viewers have seen what a Bloom power module looks like. They look like big refrigerators. They produce about 65 kilowatts and there are 64 little fuel cell stacks sitting inside. Think of them as your chips. Every chip is wired and it's providing its health data regularly to us wherever in the world it is. That's how we did it from the day we shipped our first system in 2008. So our 1500 megawatts, or 1.5 gigawatts worth of installed base is shipping us billions of data every day on how it's performing. Very lately we started calling them digital twins because the world, the world created that terminology. So we have a digital twin for every fuel cell stack that goes in every system feeding us data. We used heuristics and machine learning and other knowledge to be able to extract and learn from it. And that was part of our innovation cycle. Every system we installed was also a lab experiment for us, giving us data, allowing us to learn and become better. We are a learning organization now. Look at what AI has done with that data today. Not only can we learn what is working and make something better, we can probably real time manage the performance of each One of those little refrigerators slightly different from the next one, just customized for what it can do and extract more out of it. I think it's a phenomenal opportunity and we are all over it.

[Armand]

I want to go back to the urgency that you expressed earlier about why we need to do everything we can to dominate the AI race. Talk about the consequences of not scaling fast enough. How do you see the US stacking up globally and what are the positive outcomes if we can get this right?

[KR Sridhar]

So when you look at the AI race, it's not just about intelligence and information and number of rote manual processes we do today that can be automated and therefore improve productivity. If you just take that productivity gain along, we can improve our GDP growth by a couple percentage points. That would mean thousands of dollars in the pockets of every average American, every average person that lives out there. So economically, just imagine what it creates from a GDP growth and the trillions of dollars it puts into our economy. So that's one simple way to think about it. The next thing to think about is not just business processes, but we are living in a sensor ubiquitous world. Every physical device has sensors today and it's affordable to put these sensors IoT the ability to use IoT properly using physical AI and how to make everything perform better, become more predictable, last longer changes every aspect of every industry. Whoever wins this race is going to dominate in manufacturing, is going to dominate in materials processing, is going to dominate then using all that information in the speed of R and D for any future technology. I think your viewers will agree that if we lose our technology superiority, we will lose our economic superiority that hinges on AI. Now take another layer. National security wars are going to be fought and won by AI superiority, not by boots on the ground superiority. So it's a national security issue, it's an economic prosperity issue. It's about who controls the numbers issue. Now the world pretty much is between a free world and a not free world. Who wins this race is going to dictate which way the rest of the world is going to be dominated by. And that's a scary thought if you think we can be number two. It's absolutely scary to me that we can be number two. That's where I would say policy notwithstanding, doesn't matter who is in Congress, who's in White House. There'll be one thing that every government is going to agree is we cannot afford to not win the AI race.

[Armand]

KR Sridhar is the CEO of Bloom Energy. What a moment we are in. Thank you so much for sharing your thoughts.

[KR Sridhar]

Thank you. It was a pleasure.

[Steven Lacy]

KR Sridhar is the CEO of Bloom Energy. To learn more about how fuel cells and onsite power can help your business stay ahead, read Bloom's 2025 white paper by clicking the link in the show notes. And to find out more about how Blooom is bringing fast, reliable power to data in the AI era, go to bloomenergy.com this is an edited version of a Frontier forum recorded in front of a virtual audience. We took lots of live questions from the audience, and there's a ton of technical detail on Bloom's fuel cells. So if you want to go deeper, you can click the link in the show notes to watch the full [e-mail] events.

[KR Sridhar]

SAM.