AI Needs Power, the Grid Needs AI

Thank you. Thank you. Welcome to yet another Cornell Keynote. I’m your host, Lutz Finger.

I’m a faculty at Cornell where I teach courses on AI, data, and product strategy. I spend quite a lot of time thinking through how AI actually changes. It changes products, markets, and customer behavior. In my own work, I have focused on how digital systems shape what we’ve seen and trusted.

I built up Fisher Analytics, wrote a book about it in the social media area. More recently, I sold a company called R2Decide, a platform where we do fine-tuned models in order to help e-commerce workflows. Now, if workflows interest you, Econet runs a workshop on workflows. This is three hours of.

There are five days in between these workshops. so that you can create templates for companies out of this. So sign up. The link is in the show notes below if you want to join.

Today, we talk about something else. Has something to do with workflows, though. And it has something, obviously, to do with AI. It is about the power part on AI.

AI economic depends on power. Who would have known, right? No, but we often talk about AI in terms of models, chips, applications. But behind all of this sits one harder constraint.

AI needs electricity, loads of it. And as data centers grow and get expanded, the need for AI just accelerates. So the electric grid today becomes a defining bottleneck. And that has, obviously, political implication, growth implication.

Well, we can even say national security implications. And that’s exactly why today’s conversation matters. So I am very, very delighted to be joined by Josh Wong. Hi, Josh.

He’s founder and CEO of Think Labs AI, a specialized AI company focused on exactly this question, critical infrastructure and energy sustainability. And Think Labs AI is building. Yeah. a grid co-pilot.

So no, not the co-pilot that talks to you over your email, no, but a co-pilot that talks about grid planning and operations. And there’s a lot of topics which we have discussed on this keynote before in terms of looping back, feedback loop, true north, and we will get to this. This is a super exciting discussion. Josh brings more than 20 years of experience across grid modernization, solar, storage, microgrids, hydrogen, germs, you name it.

He is your man if it comes to AI and electricity. He previously founded Office One Solutions and later served as a general manager of grid orchestration at a GE part. So Josh, welcome to the show. Thank you.

That’s great to be here today on an amazingly interesting topic and super timely and relevant as well. So really look forward to it. So awesome. Yes.

Like the way we best start this, Josh, is give us a little bit of feedback about yourself. And I think we can start off actually with a huge announcement on the company. So either feedback to you, like talk about yourself or talk about the announcement, whatever is best. Yeah, I think it’s super timely.

So maybe I’ll start off with the announcements and go back to… A bit of my humble origins, but we just announced yesterday closing our Series A. And this is a very critical milestone because this is where we go from AI being a science project and then experiments and looking for product market fit to now we have product market fit. We’re getting more and more demands all the time.

I think people are, especially our customers like utility grid owners and operators are really seeing the need to accelerate AI. So we raised a $28 million A round and led by, I think it’s really the investor group that is significant as well. So led by Energy Impact Partners, one of the world’s largest energy investment funds, backed by some of the, again, the largest infrastructure and grid owners and operators and planners in the industry, along with NVIDIA. So the N Ventures, the venture capital arm of NVIDIA.

So bringing the grid and AI together. The third is significant as well because they are customers. So Edison International, the parent company of Southern California Edison, one of the most progressive, innovative go-to in terms of grid modernization. So they all came together on this round.

Existing investors also participated, including like GE, Vinova, Powerhouse, Blackhorn, Active Impact, Amplify Capital, another large North American utility as well. So it’s a critical milestone. We’re very, very happy and excited. And now let’s just keep running.

Now this is like having raised capital by myself, I know how stressful it is. And I as well know that this is a huge milestone. Putting the right team together, putting the right support together. Well, we know when it’s the right support once you exit, but like that is the first step, Josh, and congrats for this.

So well done. Now, how did you get here? Like who is Josh? Yeah.

I’m not going to go back to my immigrant story, but really, I think when I was going in engineering school, we talked about Cornell, but in engineering school, I was trying to find, so where should I apply my electrical engineering skills? I was looking at like solar and wind and hydrogen at that time, but I fell in love with the grid. So I joined the utility back in- I need a t-shirt for this. I fell in love with the grid.

I don’t know, Josh. So we just talked about the series A story. So when I, and you mentioned us, this is my second startup selling my last one, Opus 1 to GE, now GE, Brnova. But when I was founding ThinkLabs, my investor friends were pinging me and saying, Josh, why are you selling, finding another, like finding another startup to sell to utilities again?

And they are great, but they are very, very challenging customer group to sell and work with. And I said, I love the grid. That’s it. I have unfinished business and I see the need to keep modernizing the grid.

Yeah. So, so yeah, in mid 2000, I joined a utility. It was interesting because I was actually the first engineer they hired in 13 years. So there was a major hiring freeze, major skills gap.

And so when I joined about 50% of the 1700 utility workers can retire, in the next five to 10 years. So new kid in the block and really learned a lot from going out in the field to, well, that was the very early days of smart grid 1.0. So I hopped onto that, that trend and really became the leader of the smart grid units. Again, another incubation type of unit, but looked at a lot of cool, like modernization activities.

So self-fueling switching, battery storage, electric vehicle charging, data analytics, and metering. So that’s the onset of smart metering infrastructure. So I ran that for about five years, supported regulation and legislation as well. Wrote a 25 year roadmap, which I think 20 years now into the roadmap, but still fairly consistent.

So the vision stays the same. The mission is persistence. Yeah. So I, after running that, I founded my first startup Opus One Solutions.

We became known as awards number one, Durham’s company or DER management system company. Focused on grid modeling, but physics-based modeling and large scale optimization of grid power flow. And yeah, in 2021 that was exited to GE. And that’s how I went on the next phase, which is looking at the next 10 years of the grid within GE, the grid of the future.

And we realized a strong thesis that we need to help the grid or the utilities accelerate their pace and depth of analysis and performance. And that’s what we’re doing. So we’re looking at the next 10 years of this in decision-making. So we coined that autonomous grid or autonomous grid intelligence.

For example, a few weeks ago was actually NVIDIA’s GTC conference, lots of talk around autonomous driving, but again, I love the grid. So we really believe that autonomous grid is a more pressing and perhaps even bigger opportunity than driving. And so we then lean into AI to drive that autonomous grid vision. Yes.

So let’s talk about in, let’s talk about your love. Let’s talk about grid a little bit. So because, you know, I don’t think that most people understand the complexity autonomous driving is easy, right? Like you look on the street, you say I have drunken folks on the street.

Like, wouldn’t it be way better if an AI drives, but what is the grid? Let’s start with the basics. Not everybody in the audience probably knows exactly what a grid is or could describe it before you do that. There are一定一定一定一定一定一定一定一定一定一定一定一定一定一定一定一定 that like quick announcement like always if you’re new to the show um um the other folks know it but like always you can um ask audience questions i will monitor them so at any point in time see below there is this window where you just add ask anything you want to know if you don’t understand grit we are a university at the end so ask us and we will try to make sure that you get the right answers so um but let’s channel the audience question here nobody asked this but by the way if you put in your name i will mention you if you don’t put in your name i won’t mention your name but i still ask your question if it fits into the flow so i channel my inner audience what the heck is the grit like let alone that you fall in love with it but like what is it i’m thinking about the movie tron but uh the grid is basically a network so we we actually call it the mankind’s largest and most complex machine some of the computing folks might uh refute that but i think actually that was uh named by the national academy of engineering uh that it is a very large it connects the entire north american everybody basically 100 customer penetration unless you are in world areas and it is extremely complex because it is a continuous supply and demand balancing mechanism so just like any network or i always call it the grid it’s a very large network and it’s a very large network compared to the roads that’s an electron highway or highway for electrons it connects between supply generation whether it’s nuclear wind solar gas etc with demand our homes ev charging data centers etc and it goes in various levels of voltages higher voltage or like transmission lines uh big long towers um and then uh the last few miles might be the distribution system so a web network is a network that’s connected to the road and a web that connects from substations or key hubs to your homes that is the grid but how electronic may i one one more point to it because i think that’s uh what people might not realize it is not a one-way street meaning electrons which gets pushed in from um the generation source needs to be used they cannot just linger around maybe explain this a little bit because that’s yeah challenge yeah i think uh when smart group came people thought we can maybe perhaps packetize electricity it doesn’t work like yeah we can’t assign an ip address and route it easily it follows the laws of physics which is from high potential to low potential it’s like the river system so uh we don’t have stop signs we have we can open and close switches uh but we don’t have something hey electrons wait here until you turn left or right so it is also very dynamic so each all the way to each cycle or or you know you can you can you can select It has been for the past decades, typically a one-way streak.

But what we’re doing now is we have distributed resources as well and more and more solar and wind. So it’s becoming a two-way streak, which becomes extremely complicated to manage. It is also very, very opaque, meaning there’s not a lot of sensors and monitors and camera equivalents to the grid to see exactly what’s going on. And given how old it is, most of the grid was built during the post-World War II period.

We don’t have enough sensing, monitoring, control technologies on an increasingly congested and increasingly two-way, system. So that’s where we are facing this, not only an asset wall, meaning aging infrastructure, things are failing naturally, but also congestions, violations, factors that might destabilize the system and leading to more wide-scale blackouts. Totally. So this is super fascinating.

I mean, I like the river example. If you suddenly get many new rivers into a system of water, that might clog up and might have very unexpected chaotic consequences. Now, why should the general audience care about the grid? How does it impact economic or the AI and the grid?

How does it impact economic or society or political levers? Yeah, we commonly say people take the grid for granted until now. I think for the past decades, we assume we flip the switch, we turn the lights on and we keep the lights on. And that has been the case and load has been fairly steady until you lose it.

When we have a blackout, everything gets into chaos. And so now it’s even more important because of the overall macro trend around electrification. So buildings are getting electrified, cars are getting electrified, and is a core part of economic growth. Yeah.

These are major investments to fuel industries and livelihoods and ways of living, standards of living. Now, AI brings this to a whole new level. And it’s, I think, if I recall correctly, around 40% of low growth is driven by AI and related data centers right now. And I think ICF just wrote out a report around before by 2030, by year 2030, we would be expecting 25% low growth in the US.

That’s a lot of growth. And I think that is phenomenal, knowing load has been steady, if not declining over the past few decades. And so that’s really what’s fueling it. If you look at, for example, Nvidia’s five layer AI cake, I don’t know if you have been citing that.

I have certainly been citing that more and more. So the whole AI stack is powered by the energy foundation before you hit infrastructure and then models and applications, etc. But that foundation is extremely shared. Yeah.

And that foundation is shaky. And if that foundation is shaky, we won’t be able to power AI and all the associated industry. So that’s why it’s super critical for us right now. Meaning it’s critical on one part, especially in a time where the, because of the current war, we see energy prices rising.

Suddenly energy, electrical energy has as well a totally different cost. And now if everybody thinks about it, but I have an EV and I can plug it into the grid, doesn’t work this easily. Right? So there is this consumer level challenge.

Then there’s obviously the economic challenge overall. Energy cost is rising. And then there is a political challenge. We, as the US, would like to be the leader in AI, but in order to be the leader in AI, we need power to actually empower AI.

Exactly. So AI is now a matter of energy security. As you may have seen. Yeah.

In recent articles, whoever wins the energy race, wins the AI race. And I think that is really, really real today. Is the, is the, the grid in the US is a little bit outdated to say the least. How are other countries in terms of grids?

I think there’s opportunities to leapfrog because brand new infrastructures, they don’t have to take it over, afford to overbuild a bit. There’s also new unemployment unemployment unemployment unemployment unemployment unemployment unemployment unemployment unemployment unemployment unemployment unemployment unemployment at. I think leading to where AI now comes in is we already have a lot of infrastructure, we have a lot of capacity actually, and we utilize that capacity more effectively with intelligence. So sometimes we say intelligence or software AI driven intelligence is the cheapest form of electricity because it actually optimizes the use of what you have today by understanding your grid more, but also understanding supply and demand more and how we can line up these by time and location.

Got it. Meaning to reformulate this, certain countries are way more advanced in grid setup because they had the ability to build greenfield and the US has an existing grid and now AI could help them to leverage it to actually build greenfields. So I think that’s a really good point. leapfrog the greenfield part.

I believe so. Yes. And so what that means is whoever, again, I might be repeating myself, but whoever wins the grid intelligence race, hence I would, that’s the AI race for the grid as well. If we just take a small bit back to the five layer cake, a small bit of that AI capabilities at the top layer of the cake to fuel and power for AI to power the grid, the energy foundation itself, this is the flywheel.

effect that we can leverage. So winning the AI race for the grid also wins the race for the energy security. Awesome. Now talking about flywheel, because, and we on the show before we talked a lot about data flywheel and feedback loops and all of this, how do you measure grid success?

And how do you as a company optimizing for the success actually have access to this flywheel information? Let me say this differently. A flywheel always works when you have, you have a true north, you have something you want to achieve, then the AI optimizes this for it by doing so it learns or for Google and Meta, it’s not the AI, but it’s a human optimizing for it. And then an AI learns about those humans and then you get the new true north again.

How does this work in the grid? What is your true north and what do you optimize for? Yeah, we can tackle this from various levels too. So let’s poke at it from different directions.

But on the grid side, we always say the true north is safe, reliable, resilient, affordable energy. So if we have an abundance of resilience, affordable electricity, or energy in general, because there are fuel switching options as well, while maintaining the safety and security of the grid, that’s what’s power, it powers the economy that we are looking for and the standards of living that we are. To get to that, we need to understand really how the grid works, meaning how we can optimize for capacity on the grid to make sure we’re not overloading the lines. We have to understand all the voltage valuations, etc.

Not to get too deep into power system engineering. But in terms of what the AI can do, I think let’s start with the AI. Let’s start with the AI. into what the outcomes we want the AI to do.

one is we want the AI to be trusted. That’s the primary outcome. This is not just putting a whole bunch of SCADA measurements and smart metering information into ChatGPT or Microsoft Copilot or Gemini and say, hey, give me how the grid works. Give me Powerflow, basically.

It doesn’t work like that. So we leaned into something called Physics Informed AI. So we basically train our AI models using physics. It’s physics-driven, physics-based, and physics-accurate.

And so basically we’re saying, hey, Cornell, I’m sure you run a lot of engineering courses. Say, can we teach AI how the grid works? And can I reproduce that engineering analysis faithfully? So that’s number one.

Second is the outcomes of how much can the AI accelerate and deepen our understanding of the grid to reach the safe, reliable, resilient, affordable outcome. And so for us is, can we turn traditional studies, which is a core part of the utility’s internal processes, that tends to take the order of weeks to months to do a study. And that is actually a primary bottleneck into people connecting to the grid today, data centers, car buildings, et cetera. So can I turn studies from months into minutes?

And can I go from onesies and twosies? One feeder at a time into full system analysis. So that’s another key outcome. Let’s put a bow on this because this is so important for us to understand.

There is an excitement about AI and the market. And that excitement is driven by every consumer now chatting to Gemini or ChatGPT and kind of getting answers that sound smart and sometimes are trustworthy. Mm-hmm. about how does AI kind of create a drug it’s not chat GPD creates it is a workflow that we are having workflow that workflow follows certain heuristics is old is trusted and gives them the feedback so and Josh I paraphrase you because it’s such an important point for us to understand how the grid works is a heuristic approach it’s electrons behave based on physics but it’s complicated and therefore it’s not so easily to map out in rules therefore we use an AI to test it out and follow the heuristics is that correct I think so I think I’m definitely aligned and maybe two two ways of taking this further as well so one is the grid today as you mentioned is very complicated so the rules are just basically worst case scenario build it and it will never break type of deal but that’s when we over invest and bring up people’s electricity bills but the grid hasn’t been very data driven and data availability data quality has been a huge huge complex point point of complexity for the grid and so now with AI we can finally be a!

lot more data driven let me give you one case in point so when I started my career used to be a power system planner and a joke that we always have is planners are always right because they can blame it on the assumptions but planners are always wrong because the assumptions are never right and that’s exactly what’s happening on the grid and that’s why I’m so excited to be able to talk to you about this because I’m so excited to be able to talk to you about unemployment grid because we have all these amazing incredible engineering software but we feed it with a ton of assumptions on data that are not correct not not confident we’re not confident in its in its accuracy to date right exactly so the moment you spend your time running first principles driven scientific analysis the results will be outdated already because it takes so long to run an analysis and it’s not what I would call snapped with reality We see back to the electron highway example with AI now, when we open up a driving app like Google Maps or others, the primary sources of data is actually what’s coming from the cars. It’s not the width of the street. It’s not the duration, the probabilistic duration of those traffic lights. It’s the cars that’s driving the feedback loop.

In this case, it’s the smart metering information or the telemetry through SCADA signals that’s reinforcing how the grid can improve its accuracy of how it interprets the grid. The second back-to-depth feedback loop is around reinforcement learning. So in the past might be mathematical optimization, which like nonlinear mathematics, and that can be hard to find a global optimal and you can optimize for one objective at a time. Here.

We have such a complex grid. I would compare this to like a millionth dimensional Rubik’s cube. My son’s a cuber. And so, but we only have capacity to look at one corner at a time.

And so we hack it. We corner twist, we take it apart. But here is we are actually having AI to model the entire system in its complexity. And through things like RL, we can finally optimize for the entirety of the grid.

Explain RL. For the audience for a second. Yeah. So it’s reinforcement learning.

Is that feedback looped as you said? So let’s play a big game and put in the right reward and penalty functions so that we can arrive at the optimal outcome. Yes. You have human reinforcement learning, meaning there is a human in the loop.

And here we have a technical RL, which is essentially you test something out and you get feedback. I want to point out that this is a huge area in the whole industry, not only for energy or healthcare, where we will see models thrive. So if you’re in the audience thinking about which company to join or where to invest, then listen up. Anything which has the ability to create an RL is helpful.

Very early on, one of the things everybody talks about is AlphaGo Zero. The RL there is computer plays Go, gets a feedback loop within the framework of allowed Go moves, gets feedback whether this works or not. If we do this for the grid, Josh has a model of the grid. Computer starts adjusting parameters in the model and then gets feedback.

Is this grid now better or not? Yeah. Like for Josh, Josh’s son, who is a cuber, the model would try to say, does some changes in the cube and figure out is this a smart strategy. So that kind of RL, this reinforcement feedback loop is exactly where we will see a lot of new business models coming to market.

We had here on the show actually a discussion about recursive learning for coding, for workflows, and now for the grid and for healthcare, right? So you have always this kind of allow the computer to test out things and to learn from it. Yeah. And I’ll give an example on what that is today and along the grid.

So today let’s look at the status quo without reinforcement learning, all this AI-based optimization. We are game planners. We make assumptions. We run analysis.

But the simulation engines give us only problems and not solutions. Hey, we run, we connect this data center. Every, we get alarms everywhere. We have congestions, traffic jams, voltage violations, but the solutioning comes from, honestly, trial and error.

It’s the human engineer, which they can be very experienced saying, I think I can add a line here and alleviate. Let’s try this. Let’s then each try runs like hours. If not weeks of analysis, I get some feedback.

Then I can tweak something else. So it’s very much a human based in the loop trial and error, which is great. There’s a lot of creativity and intelligence with that. But that process is very, very time consuming.

And I think it’s up to how many scenarios we can run as well, which is typically one to five scenarios. That’s it. But now with reinforcement learning, we are running millions of simulations and scenarios, pre-training and getting ready to push forward. And now we’re getting ready to propose the right solutions, not just a trial and error.

So the type of answers or solutions we are generating include, well, we still need some lines. Yes. But where do we put them and how large? And so we basically generatively designing the grid.

Second is, well, a big battery storage is a big thing these days, especially in the past five years or so. Can we put in a battery storage unit? Now we can finally pause electrons right on the road. Store them for a while before you release them back into.

Into the grid again. And third is actually a huge relevant topic today around demand flexibility. So if we have a large loads like data center connects, can we flex them? Meaning can they consume more power sometimes and consume less power these these other times?

And can those signals be generated by the grid’s optimization? This is where orchestration comes in. So all of these are actually part of the feedback loop for the AI to learn and generate these solutions. Which is fascinating because if you think about the AlphaGo example, Go has a defined space and defined rules.

In the grid, you could actually now think about arbitrage models. And I know that Cornell has a lot of financial savvy folks. So here comes a business opportunity. Maybe it is worthwhile selling more electrical cars because it optimizes.

Yes. The ability of the grid. Now that needs modeling to understand. Is this the right policy to do in order to support where event your model comes in?

I like we have loads of questions here from the audience and quick reminder. If you have questions, put them down. I see them all. And I would like to pick up Daniel from Kansas city.

Um, he talks about, um, Daria, uh, uh, Amodei and Sam. Outman committing to massive CapEx into data centers. And, um, then he talks about that. There is discussion of that.

There’s an energy risk, right? And we alluded to this in the beginning, meaning the grid might be a constraint to our growth in the AR market. If you look at this, is anybody modeling this out? Are you guys modeling this out?

Um, There is also a heightened uncertainty in our machine regional bulk power system planning these days. So we’ll ask Daniel from Kansas City. The shortcut that’s happening right now is for all these data centers to bring their own power. I think that is a shortcut.

So meaning they can bring their own batteries to buffer their electricity from the grid. They can bring their own gas turbines, for example. And the big topic around, can they bring their own flexibility to the grid as well? Now, that’s good.

But it’s also islanding the data center to themselves. And as we learn from the grid, everybody in it for themselves is not the most economically optimized. The whole point why we are connecting everybody is that we’re able to share resources and capacities. So the ability to reinforce the grid and look at how to best where.

And how to best bring on these large loads is an absolute area of critical studies. We are doing these now, for example, looking at where there is capacity. Can we optimize? Can we shorten the quote unquote time to power?

So instead of connecting or building data centers and connecting them every five to seven years ahead, can we connect them now? But there might be a bit of bridging solutions, which is some flexibility required while we reinforce the overall grid. So there’s a lot of variables and options. And again, going back to reinforcement learning, that’s why it’s good because we can go beyond studying one data center at a time or one solution at a time.

But we can look at studying the entire grid, what we call the queue, the interconnection queue of both loads and generators who wants to come onto the grid and optimize them and restudy, adjust that study continuously. So they are not. One off studies, leave them for five years and restudy again. But these are like continuous daily studies to the point of if an executive or Meta or Facebook or Microsoft, Google, et cetera, want to connect here and we will turn around these studies on a daily basis.

So that’s the outcome we are achieving. Because one data center might be planned in isolation, but if you don’t know what another company is doing, suddenly you have two data centers. And you can’t do that. So you have to have a data center in a closed setup.

And now the question is, can the grid handle it? Exactly. Now, Alexander and James, both asked a very similar question. And I thought I’d bring it directly after Daniel because it is the same only at a different end of the grid.

So Daniel asked about consumers of energy data centers. How do we plan for this? Alexander and James asked the opposite. It’s kind of like, hold on.

We have solar. We have renewable energies. Optimizing the grid, does this mean that those are more effectively or less effectively? Does this have an impact?

Will we be stuck with gas and oil forever? And I think I know the answer to it. But tell us. Yeah.

It goes back to, let’s say, energy is very fluid. It’s very dynamic. We need to balance them. We need to balance them continuously.

The shortcut in that case is to put batteries everywhere or energy storage everywhere. But that is also not very cost effective and has to be carefully planned. And batteries also degrade over time. They are an asset.

So the best way is actually going back to the cheapest form of electricity, which is intelligence. So can we understand their dynamic forecasts? For example, we are working with NVIDIA’s Earth 2 on probabilistic weather forecasting as well. Which translates directly into energy.

So if we say, oh, I want to put solar and wind into the probabilistic output forecast of solar and wind. Now if we just forecast solar and wind outputs and push all these analyses into a black box model or a traditional mathematical engineering model, it’s not going to become actionable. So where we play is, can we drive from the solar and wind probabilistic forecast into so that when we make a decisional action, it’s risk adjusted. And I think that’s where the industry is going right now.

It is a huge topic to go from fixed worst case scenario to now probabilistic risk adjusted actions. So just like people like to keep things firm with like gas or nuclear and on the generation side and data centers want firm power, but that’s actually not optimizing the infrastructure. So we need to start injecting some flexibilities, which drives us directly into risk-based analysis. Yes.

Meaning because you have AI for the grid, sustainable energy is actually a feasible option. I’m living in California and obviously like it’s very, like a lot of people have here solar cells, solar power, but they have a lot of problems to actually inject it into the grid because the energy companies in California say, we cannot take that energy. As we discussed earlier, it’s like water. Suddenly it’s there.

It has to flow. So the energy companies are not able to bring it into the grid. So a lot of solar power from consumers get wasted because the companies cannot effectively plan. They are, they are not able to do that.

So we have a plan for, look, I mean, if the sun is not shining, your solar station doesn’t work and therefore still people still need electrons. So we focus on the fixed setup and not on the flexible setup because we cannot model it. So the very idea of using AI for the grid is probably the biggest driver for the empowerment of sustainable energy. Yes.

And I would add to that as well. That’s when we say the, the utility company wouldn’t allow that solar to back feed, for example, it’s not only, or many of the time, most of the times it’s not because they can’t take that power, but more so because they don’t know, they don’t know the impacts of that real time dynamic power on the grid. And what AI allows is I would compare it to the enlightenment era. So we are enlightening the utility plan is an opportunity to see deeper, to see the power of the grid.

And I think that’s the biggest thing that we can do. There’s also these but also more dynamic capacity onto the grid. Now, let me ask you something in terms of time horizon, how fast and how flexible you are, so that we get a feeling for it. We are unfortunately in a war in the Middle East.

We unfortunately have seen that energy prices have gone through the roof. That has an impact on our industry, and that has an impact on our productivity. So has the White House been knocking on your door and kind of saying, model this out for us? What would we need to do in order to supply additional methods of energy?

How flexible are we here? I think, well, the administration definitely has been involved, especially through the Department of Energy, but we work directly with the utilities. So even if it’s with the administration, it’s through the utilities as well. But modeling these scenarios is an absolute priority and goal.

And I think that’s a criticality. Now you mentioned about electricity as well. So electricity compared to like gas and oil pricing, it is far more resilient localized. So these are areas where we have a lot more sovereignty over the power that we can generate, consume and the reliability that we have as well.

The speed is an interesting one though, because we have been always looking at grid planning, especially resource planning, in a 10 to honestly 40 year horizon for a gas turbine. But now to give you a sense of how fast AI is impacting or learning, since our spin off from GE into last year, we have been focused mostly on distribution systems. Within six to nine months time, AI is already learning to a deeper level transmission networks to the point where 80% of our work today is actually on transmission. And so because AI is rapidly training on not just one or two scenarios, but millions of scenarios, the pace of learning and speed for AI to impact and the outcomes and the applications that we have is hyper accelerating the growth and modernization of this industry beyond devices and telecommunications, like smart meters and wires, et cetera.

Very interesting. As you know, I’m German, right? And as the audience probably has figured out by now. Here we go, Josh.

And how active are you in Europe slash China? China is probably more homogeneous because they have more greenfield grid. But if you look at Europe, they kind of were against like Germany was against nuclear energy. Now they’re buying nuclear energy from, from France.

So there is a huge mix of different grids all coming together and interacting. Is this technically all same, same for you or is this a way more complex problem because different grids are involved? Yeah, we thought I’ll have my three layer cake, not five layers. Hold on five layers before three layers.

Now let’s actually put the layers down. What’s the five layer cake of Nvidia and the three layer cake of Josh. Here we go. Yeah.

So the five layer cake when the Nvidia is basically going from the energy foundation to the infrastructure and CPUs and cloud, et cetera, to the AI models and all the way up to the applications layer. So that’s the Nvidia five layer AI model. I think here is, I see AI learning the grid, but not just the grid, but the utilities, the, regulatory environments and the people and process there, right? Cause it’s always people process systems that transform, the industry for AI to learn, let’s say the 60 Hertz North American grid versus the 50 Hertz European grid.

And it’s, it’s topologies and subtleties and asset base, et cetera, is the same. So for AI to learn and transfer that learning from here to Europe to APAC is the same thing. We’re not concerned about that at all. Nonetheless, the more complex side is to learn the people process and regulations.

So meaning taking it from, what we call the AI digital twin layer. So the AI models representing the grid to the agentic workflows around the AI models. So how do we take these high performance grid simulations into the desk and use of everyday users? So how do we plan the assumptions, standards procedures?

What are we optimizing for? And how do we bring the human? I mean, you mentioned that co-pilot, concept. So how do we bring the human elements into part of the AI learning as well?

Something we always say is there’s two sides of the screen as a, as a power system planner or operator, there’s the grid side, but there’s also the person facing side of the screen. So we need to bring the two together. And I think for the European market, that other side of the human factor policies, regulations will be more complicated and will take some time. Super fascinating because, and we discussed this quite often, a workflow or an AI autonomous workflow has obviously the AI component, which is heuristic or with a feedback loop.

And then it has a human interaction. You have a one more layer, which is a regulatory setup for Europe. If certain things are, allowed in France, it doesn’t mean that it’s allowed in Germany. And therefore there’s a regulatory structure in the U S probably as well.

Meaning looking at an AI workflow and autonomous AI workflow, we need obviously the AI, we need the feedback loop in terms of what is the truth more true North, which is a heuristic model or a feedback model whatsoever. And we need the hand over to humans in order to take decisions and we need to continue to work with humans. So we need to take decisions and we need to continue to work with humans. There’s also looking at machine learning in general, part i stress this because um interestingly enough this is exactly the same workflow no matter whether you have an ai working on emails or whether you have an ai working on cars or whether you have like an autonomous car or whether you have an ai working on an autonomous weapon we will have in like urge you to sign up we will talk to schwimmer um and the drone company there recently went public in another keynote where we go through the autonomous workflow and if you want to build your own workflows again remember cornell has a workshop on this and i’m happy to walk you through so all in all it is the brain um the heuristic feedback loop or the rl the handover to the human and regulatory or permission document around it it’s always the same structure more cakes more layers more cakes more layers i have another question and that question is from tim so tim first of all congratulates you to the funding and i think that’s well deserved absolutely then um but then he talks about uh how do you do your engineering how do you hire how do you set it up so there is a very straight technical question um which is essentially do you hire um and um how can you apply but for me from a more structural point of view there’s another question of how has your engineering changed and um how do you hire and everybody like you probably use cursor and codecs and other tools how has this changed tell us like let’s look at another hood a little bit yeah that this is what i wish my engineering teams here so they can give you the the uh the the core tech stack that we have which we can open it up but um number one is tim thank you uh and uh we are hiring um we are small company but we are actively hiring i think in the in the world of ai it’s not the number of people we’re hiring as the just communicating my my hiring philosophy it’s now the quality because i think there is going to be so much more productivity with each individual working with ai as well so i think as we are very conscious in trying to blend basically two or three worlds together one is the power systems engineering world and the other is the power systems engineering world so about half our team are power system phds that actually has done great work in the ai space so applying blending those two together the other are ai architects hyperscalers machine learning folks and then we have the data scientists and data engineers so bringing and blending all three domains together in a harmonious culture is really what makes think lab special and i think it’s a big defensible moat as well with many of our competitors many times it’s a larger incumbents as well because for them to in a smaller agile setting bring these skill sets together it’s critical now do we use ai in our workflows absolutely yes but now our core differentiation is i think ai brings the speed of development et cetera but the core differentiation is still around these understanding deep technical modes around power systems with ai and how bring them together.

And that’s where we bring a lot of the machine learning data engineering expertise. Totally. And this is super important. And I would like to bring this to somebody like into a structure, which is also understandable for somebody who is not a power engineer.

If you go today to your favorite coding agent, let it be from OpenAI, which is like, which would be called Codex, or we do cloud code, or like you plug in a tool into a cursor, you could easily ask something like, okay, and I do this in the workshop, which I announced a few times now. But like, you could ask, for example, something like, find the right tags for my email. Now, there are two problems. Problem number one, what are the right tags for your email?

That is dependent on your personal workflow. For the grid, what is the right throughput for the grid? That depends on the knowledge in the grid. So this is not easily creatable.

Even if you can easily write it to Codex, OpenAI, or whatsoever, you need that knowledge. The second problem is, even if I go and say, no, I know my labels, and you tell it, then Codex probably will first do an AI. Using keywords. And that’s wrong technical approach.

It’s not scalable. It’s fast, but very error prone. And Codex or any large language model will tell you, I succeeded, I did exactly as you wanted me to. That’s not usable.

It is a side note, which is more of a concern for an engineer and a product manager. So the workflow has changed. Yeah. But you still need the knowledge about, in this case, the grid, as well as the knowledge about the AI implementation.

So I like your mesh structure, Josh. Yeah. And maybe one more point on that as well is, I think given how many AI tools, especially productivity driving tools, are there out in the market today, the general trend, especially for those who are looking for jobs, is that general AI is becoming more and more democratized. So it’s also being a lot more competitive in those positions.

Now, specialized AI, especially domain specific specialized AI, and in many of our discussions with NVIDIA, it is physical AI. So bring AI with the physical world. That is the opportunity today. And if you look at even talk to a lot of investors, that’s where they’re keenly developing their investment thesis, maybe a year ago to, I would say, the foreseeable next year.

And that’s where they’re going to be in the next three to five years. That’s where the growth is. Awesome. I think this is super fascinating.

We could go on for hours. I would like to end with one very open question. Where do you see AI going in the future? What is for you the next big horizon that you see?

And obviously it has something to do with NVIDIA because NVIDIA invested in you. Joking apart, what does the future hold? Yeah, I think future is relative timeframe. So before AI, the future has always been the next, like if you talk about power plants in the next 40 years, if you talk about grid software, we typically look at a 10 year horizon for next generation of software to be developed.

For AI, the future is in two years. That’s my time horizon. Yeah. So, even we are at the forefront, but we are advancing AI for the grid, called the AI powered grid, right?

Grid powers AI, AI should power the grid. Basically every single two week experience that we have, things are moving very, very quickly. Where we see these two years bring us is really the onset of the next leapfrog in autonomy. That’s where I think the grid is going.

It doesn’t mean hands off, go drink your coffee, fall asleep and the car will drive for you type of deal and play video games. It’s not that. It’s that you’re going to be able to do that. You’re going to be able to drive around and play video games while you’re driving.

I don’t think that’s the future that we desire, but the grid is fundamentally getting so complex and the people planning and operating the grid needs support. So I think the autonomous grid intelligence to basically have this grid aware physics informed agent that analyzes the grid for you, continuously learns it with reinforcement learning and proposes options and solutions for you. I think that is really where we hang. Very, very cool.

Thank you so much, Josh, for coming on the show. Thank you for all the questions from the audience and stay tuned. There are more keynotes coming and talk to you soon. Thank you, Lutz.

Thank you, Cornel. Bye.