Transcript
Greg Fallon: The early versions of ChatGPT, I think the numbers that I’m available to get were something like 45 terabytes of data that they are trained on. To give you a context of the amount of data that’s being produced in the industrial world, a single oil field can produce up to 5 terabytes a week.
Daniel Darling: Welcome to the Five Year Frontier podcast, a preview of the future through the eyes of the innovators shaping our world. Through short insight-packed discussions, I seek to bring you a glimpse of what a key industry could look like 5 years out. I’m your host, Daniel Darling, a venture capitalist at Focal, where I spend my days with founders at the very start of their journey to transform an industry. The best have a distinct vision of what’s to come, a guiding North Star they’re building towards, and that’s what I’m here to share with you. Today’s episode is about the future of the industrial economy. In it, we cover generative engineering, digital twins, AI for the physical world, data that dwarfs the internet, Saudi Aramco, the tech giant, and the digital future of heavy industry. Guiding us will be Greg Fallon, CEO of Geminus, a company at the forefront of simulating and automating industrial operations through cutting-edge AI technology. Geminus is revolutionizing how industries operate by bringing unprecedented speed and scalability to AI model deployment with successful implementations in energy, oil and gas, manufacturing, and semiconductors.
With heavyweight investors and partners like SLB and Lam Research backing them, Geminus is a key player in the digital industrial landscape. As a proud early investor in Geminus almost 6 years ago, alongside our friends at The Hive, I’ve had the privilege of watching this company innovate the industrial complex over the years. Greg, who holds a master’s in science from the University of Virginia, has a wealth of experience, having held senior executive roles in product and commercialization at both Autodesk and Ansys, two of the most influential companies in industrial software. Hey, Greg. So nice to see you.
Greg Fallon: Hey, Daniel. Good to be here.
Daniel Darling: So, Geminus has spent the last 5 years helping industrial companies embrace digital transformation, and in particular, the adoption of cutting-edge software. So what’s the innovation you’re bringing to your customers?
Greg Fallon: So we are a very special kind of AI company. We’re an AI company focused on the physical world, and we bring a set of highly tailored and highly efficient machine learning techniques into the AI space to help companies make AI models or digital twins of their machines or large systems. And so that’s something like a power plant. You might make an AI model of your power plant and have the AI model make suggestions about how to improve the performance of that plant.
Daniel Darling: Amazing. And, you know, this concept of digital twins is a really fascinating one—of simulating the physical environment in the digital. Can you tell me what sort of value that brings to these kinds of complex companies?
Greg Fallon: It’s a really interesting problem that not a lot of people get exposed to unless they’re an engineer. I think a lot of people in the world think that engineers get everything exactly right, and they’re full of precision. And that may be true in the way that we interact with folks. But the way that engineering is typically done is in a relatively imprecise, knowledge-based way.
And so, for example, when a power plant is designed, there’s a lot of engineers crunching a lot of numbers, trying to figure out different aspects of the performance using mathematical models. They may be using CAD models to see how different pieces of equipment fit together. But there’s a lot of judgment that comes into play simply because there isn’t enough computing power or the ability for really complex mathematical models to compete quickly, right? So people have to make mathematical approximations, and engineers are really trained in school. We spend most of our time learning how to make mathematical approximations.
And so what we do at Geminus is use artificial intelligence combined with some sophisticated computational techniques to allow engineers to make more sophisticated models that can answer questions instantly. So it’s this idea of superintelligence for an engineer. And it’s also this kind of real-time simulation of the physical, isn’t it? It’s an ability instead of having to test something in the physical world, a change, an adaption to a machine, an introduction of a new material, you can do that in a simulated environment with all the same parameters, constraints, and intelligence from that.
Daniel Darling: Is that one of the big promises of the digital twin? It’s very expensive to make a change to, let’s say, a power plant unless you’re absolutely sure it’s going to work, but you really, as an engineer, want to experiment and try things.
Greg Fallon: So a digital twin allows you to very quickly, in real time, explore different things, different changes, different operating conditions. You might want to, in a power plant, change out a pump and see how that impacts the plant. You might want to turn a valve.
Digital twins have different levels of maturity. And at their kind of peak maturity, you would have a digital twin ideally that would be highly intelligent. And the digital twin may look at the physical object and say, “Hey, given what’s going on right now, these are the things that you should be doing to improve its performance.” Or, you know, “This is the thing that may happen next if there’s a problem.”
Daniel Darling: Awesome. So then there starts to become this monitoring and predictive element into that. And just a question, where is this data coming from? Is this from a whole wide network of IoT sensors, etcetera, like that to give you that picture of what is happening in the physical world?
Greg Fallon: It’s a combination of a bunch of things. So there is traditional static data that exists in archives, things like blueprints of plants. They’re usually digital. As you just alluded to, there are live data streams coming off of the plants. It kind of depends on the machine and how old it is. Older machines may have fewer sensors. Newer machines have lots of sensors.
With things like power plants, you can add sensors, and so they’ll be streaming into some sort of data historian, which is collecting the data. And then you have other types of synthetic data that might be coming from simulations.
Daniel Darling: Underpinning a lot of why they would choose Geminus over, let’s say, a generalized model is that you’re able to operate in this kind of physics-constrained world or real understanding of what is happening in the world of atoms, really, versus just the world of bits. Can you please just unpack a little bit around that innovation and why it’s important?
Greg Fallon: The problem in the engineering world is multifold. First of all, the machines live in a very dynamic environment. So in the physical environment, let’s just say, go back to the example of the power plant. Across the power plant, there’s lots of things happening. There’s fluid flowing. There might be steam going through a pipe.
Well, that steam is changing constantly. There’s lots of turbulence going on in the steam. The temperature is constantly changing. And so part of the problem that you run into when you’re creating an AI digital twin of a power plant is you have to capture all those changes, and you can’t really have an infinite number of sensors, right?
So, you can have a lot of sensors, but it’s often what happens between those sensors to collect data that is meaningful. And so that’s one big problem. To create a really good predictive machine learning model, you need data capturing those dynamics, and you can’t do that with physical data. That’s one thing. The other thing that is different about machine learning models of physical systems is context.
The patterns that you really want are the patterns that are defined by the laws of physics, right? Well, if you have the laws of physics, why not start there? And when you imbue a machine learning model with the laws of physics, you get this predictive magic within the machine learning model. And that really gets you towards a much more accurate and predictive machine learning model.
Daniel Darling: And let’s talk about that for a minute because, in the machine learning model for the physical world, when we talk about big machines like power plants, the requirements for accuracy and precision are very different from the requirements that we might have for using ChatGPT, right?
Greg Fallon: Exactly. If ChatGPT gives me an answer that is a little bit off, it’s not a big deal. If you have an AI model of a power plant giving the wrong answer and an action is taken on that answer, the plant could have catastrophic implications, right? And so precision is a really big deal. And so those requirements have led to a field emerging, an academic field emerging over the last 15 years where engineers like myself and computational scientists have gotten together and applied data science techniques in the physical world. We have a geeky name for it. It’s called AI-augmented computational physics, but it’s a very specialized field.
And that’s where we draw our experience in IP from. My cofounder, Karthik Duraisamy, was one of the early players in that field.
Daniel Darling: Amazing. That makes a ton of sense as to why you need such a specialized approach in such an environment. And who are the other players involved in this area? What does that ecosystem look like from a software perspective?
Greg Fallon: Daniel, it’s wide open right now. So even though digital twins have been around for almost 50 years in different aspects, now with the invention of cloud and machine learning, as you can imagine, a new ecosystem is developing. So there are lots and lots of players. There are IoT companies that are creating sensors. There are the companies that collect and integrate data from the IoT sensors.
So there are companies like Cognite out there and Seeq that are really interesting players in that particular space. There are the CAD companies that create digital versions or drawings of the plants themselves. So those are companies like Autodesk. There are the simulation companies that make tools for the engineers to simulate what’s going on in those plants. And so those are companies like Ansys, AspenTech.
Schneider Electric has a company called AVEVA. There’s about 150 different simulation companies out there or more. And then there’s the new generation AI companies like Geminus, and we are already seeing two or three generations of AI companies. Right? So we had C3.ai come out in the 2012-2014 time frame.
That was kind of industrial AI version 1. There’s companies like SparkCognition. And then, you know, only in the last two or three years have you started to see companies like Geminus appear where we’re really focused on this particular physics-informed AI space.
Daniel Darling: How exciting. It sounds like a really interesting moment in time that’s coming together for you to layer on top of a lot of the work that’s been done in the digital transformation space to date.
And, you know, we’ve all been amazed by OpenAI’s work, which was based on generative AI. And now you’re bringing this new concept that I’m excited to talk about, which is called generative engineering for the industrial economy. Can you share with us what that is and why it’s such a big deal?
Greg Fallon: You have generalized generative AI models, like from OpenAI and Anthropic that are training off of basically all the potential data that they can get available, all the publicly available data, all the data they can acquire. And those are very generalized models using LLMs, and they can essentially complete sentences and are very, very interesting.
Then there’s these specialized models that are starting to get into kind of domains. And so you interviewed, Poolside recently. That was a super interesting specialized model. We’re taking specialized models a step further and really focusing on the general engineering space. And the goal of Geminus is to essentially automate engineering.
Right? To be able to take all of the concepts involved in engineering, particularly engineering that involves the physical world, and create intelligent agents that will assist engineers in making really robust decisions very, very quickly and eventually help them start to create things.
Daniel Darling: Incredible. And, you know, huge amount of value for the world if that’s the case. And you mentioned the amount of data that OpenAI is trained on, which is essentially training on the entire Internet corpus of data.
Can you give me an idea of the amount of data available to your company or potential companies that are looking at working in the industrial complex?
Greg Fallon: Yeah. It’s super interesting, Daniel. So the early versions of Chat GPT, I think the numbers that I’m available to get were something like 45 terabytes of data that they are trained on. To give you a context of the amount of data that’s being produced in the industrial world, a single oil field can produce up to 5 terabytes in a week.
Right? So you’re talking about a single oilfield, and big corporations have multiple oilfields, big oil companies. You know, so you’ve got, like, nine days to create the amount of data that is equivalent to all of the data on the Internet. It’s a massive amount of data. And as you can imagine, collecting, organizing, and using that data is a huge opportunity.
And I think that’s something that really excites me about the next five years, the sheer volume of proprietary information that AI can train on within private companies. And you’re applying it to industrial companies, which as you point out, have incredible amounts of data that isn’t being leveraged at the moment, but they are producing it. So you can start to train, tune, and specialize a model onto that to start to have real value within that context versus a generalized model that is just trying to mimic a generalized understanding. You know, one of the things on the horizon that a lot of people want to save from AI is a sense of unlocking multi-step reasoning. So how you can send an AI off to address a complex problem and you’re dealing with some of the most complex and technical problems in the engineering world that involve multiple steps, multiple versions of how you get to that answer, a real process that maybe would take a human, you know, a couple of weeks, if not a month, to do.
Daniel Darling: Are we starting to get to a point where we can leverage AI in the next couple of years to do such complex reasoning over a longer period of time?
Greg Fallon: Yeah. We’re getting there today, Daniel. And we have some really exciting examples of projects we’re working on. We’re using intelligent agents to orchestrate other intelligent agents and bring that information together.
In a unified format so that we can make detailed conclusions across incredibly complex systems. So that’s kind of a specialty of ours, and it’s an area that we’re going after aggressively. We have a lot of specialized IP in that particular area. And the way that we’re thinking of the world and going after this world is almost the antithesis of the way the hyperscalers are approaching the world, right?
We don’t have the advantage of large amounts of data set across a huge number of companies for applications, right? And we certainly don’t have access to the hardware that the hyperscalers have. So the way that we are viewing the world is a structure of intelligent agents connected to each other and sitting on top of each other. And so you’re building up towards what I would call enterprise intelligence.
And in order to do that, you need to be able to have agents kind of work in tandem so they can work together to reason and solve problems. That requires sequencing because you may solve kind of the first step of the problem that gets passed to the second agent, then the third agent. And we’ve been able to create some really impressive networks of agents that have been able to advise engineers running systems that have formally just been operated by experience and knowledge that’s been gained over decades. And we’re able to give them incredibly precise answers that are having huge impacts on their businesses. So a company that is putting a lot of effort into this is Saudi Aramco. So Saudi Aramco has announced a multiple tens of billions of dollars investment fund around digital technologies.
And they’re working to make an entire digital twin, if you will, of their entire operation, from the ground up. And in tandem with that, they’re taking those technologies and the infrastructure that they’re using to do that, and they’re creating an entity that’s providing those services to others.
Daniel Darling: Can I just drill down on that? Because I think that’s huge from that perspective. So we’re talking about the largest industrial company in the world, Saudi Aramco, investing tens of billions of dollars into creating a digital version of itself. And then you said also unlocking that for other industrial companies. Can you explain, like, what is the motivation? What do they see as the potential upside of doing something like that?
Greg Fallon: Saudi Aramco has created an entity called Aramco Digital. So Aramco Digital is providing digital services to digitize Aramco itself, and then they’re providing those digital services to other players first in the energy sector, but I imagine they will go out into other industrial sectors.
And if you look at the way that Saudi Arabia is working to diversify its economy, one of their big investments is in training around digital services like AI. So there’s a huge effort within Saudi Arabia to build the universities, to create digital competency, especially around AI, and then leverage those to move their economy away from a minerals-based economy.
Daniel Darling: That’s such an exciting trend because, essentially, you’re taking a company that has built wealth over extracting petrochemicals from the ground and now saying we’ve got a huge amount of data that we’re sitting on from doing that and a huge amount of dollars to invest into taking that data and building a digital infrastructure around that. And now they’re saying we can have that as the next generation of how we can bring this company forward and propel the economy forward.
That’s a really huge idea. And I know you and I have riffed in the past around the concept of how we’ve seen this in other areas. For example, with Amazon and AWS, where Amazon essentially had an e-commerce environment that is powered by a cloud computing environment. And instead of just making its own cloud computing proprietary, it decided to open that up and create Amazon Web Services, where other services could start to use its cloud, invest those dollars into its own infrastructure, and build out a massive part of its business that powers Netflix and all these other companies around the world, not just Amazon. Is that kind of where you’re seeing others mimic potentially this process of what Saudi Aramco is doing in the digital world as well as the industrial world?
Greg Fallon: Yeah. Interestingly enough, I’m seeing it a lot in the energy sector. So, one of the companies who is collecting and aggregating data within the oil and gas services sector is Cognite. And Cognite was spun out of a company within the energy sector.
They were specifically mandated with creating a digital infrastructure for this company. They then spun out and are now providing services to every energy company around the world.
Daniel Darling: Incredible. And what other kind of moonshots or innovations over the next couple of years are exciting you?
Greg Fallon: From an AI perspective, which is where we are, you know, I talked about automating engineering. That’s probably one of the most exciting things I’m thinking about. I think about combining that with space technologies. And so, you know, we’re starting to look at applications in space, applying artificial intelligence to help satellites, for example, navigate through space. And so I get very excited about space technology. I think that the advances in space technology that are coming are really interesting.
I mean, it’s all physics-based, and there is a huge amount of simulation that’s being done. Space is obviously an environment where it’s really hard to test, so the use of digital tools and things like simulations are pervasive.
Daniel Darling: Amazing. That’s exciting. And what are the kind of real-world impacts that you’re hoping to see as a result of all of this innovation?
Greg Fallon: The one that I am most excited about is really the impact of these technologies on climate resilience and hopefully slowing climate change. And so, we’re working in the energy sector. A lot of the work that we’re doing is helping energy companies produce energy much more efficiently. So even if we’re working, for example, in an oil field, we are helping those energy companies produce the oil using much less energy and with fewer carbon emissions.
I think the last time I looked, the carbon emissions coming from the energy sector just extracting oil from the ground was a huge amount of emissions. It’s something like 5% of GHGs a year, coming just from extracting and refining oil. Then there is the transition to new energy. So as we think about transitioning to new energy sources, like wind, like nuclear, like hydrogen, you have a lot of engineering that has to get done in a very short period of time. And so, the only way that we’re going to be able to compress the engineering development timescales is through the use of tools like advanced AI and machine learning for the physical world.
To give you an example of what that might mean: let’s just take a wind farm development. So we’re developing wind as quickly as we can. You know, I’m sitting here looking out at the Atlantic Ocean and from Boston, and there’s huge amounts of wind installations going in off of Boston. When you site a wind turbine, aerodynamics are really important. And there’s an infinite number of potential changes to wind direction. And so just understanding those aerodynamics and how they can affect the wind turbines upstream and downstream of a given turbine can have a major impact on turbine efficiency and field efficiency.
I think the numbers I’ve seen are to the tune of 5 to 15%, right? So that’s an example of using these digital technologies to help get more out of renewable energy. We can look at things like electrolyzers to create hydrogen, right? So the whole engineering space around electrolyzers is evolving. So the need to evaluate different design options needs digital twins in order to do that.
And if you can accelerate those digital twins through machine learning, you’re able to get to the best answer a lot faster. And then when you get to operating those electrolyzers, there are a million control decisions that need to be made to make sure those electrolyzers are running at their performance criteria. And so having AI give them the right control set points is a huge opportunity. So that’s just in decarbonizing the world. If we look at climate resilience, there are implications for the types of things that we do on weather prediction.
So the same physics-informed machine learning techniques that we’re using, let’s say, to make a power plant more efficient, are just as applicable to weather prediction. How do you take weather models that have been created over decades of experience and data and have those models adapt to local scenarios in areas where the climate is changing very, very quickly?
Daniel Darling: And timeframe-wise in your mind, for a lot of these things, you’re dealing with the industrial industry. It’s massive, it’s complex, it’s slow to roll out and adopt. When do you start to think that some of these digital innovations will be adopted at scale, from a time frame perspective?
Greg Fallon: Two years from now, you’ll probably see the top industrial companies all having deployed these types of digital technologies specific to Geminus in the field in some way, shape, or form. And then once they’re in the field, I think you’ll start to see them accelerate very, very quickly. So five years out, I would imagine that you have at least an early majority of industrial companies running these types of codes.
That brings me to another point, too, which is the scale-up of these technologies. Right? So one of the things that we have spent a lot of time thinking about is how quickly you can scale AI technology. It’s one thing to be able to create a highly effective predictive AI model. It’s another thing to create a thousand of them.
So if we’re working with a large energy company like Saudi Aramco, they have tens of thousands of industrial assets that need to be digitized and then connected to one another. So we’ve been working really, really hard on getting to scale. Today, up until really last year, a lot of the large AI deployments were, I would say, one-offs—pilots that took a lot of effort upfront to get them to deploy. We have shown that we’re able to get to a deployment about an order of magnitude faster than the state of the art this year, and we’re looking to cut that down by another order of magnitude over the next 12 months.
And what that means is the technology is now ready to be deployed at scale without a tremendous amount of effort.
Daniel Darling: What kind of speeds are you starting to see at this moment in time?
Greg Fallon: We are getting that down to a month, under the right conditions. And we can get an effective pilot going in a matter of weeks. We’re looking to cut that down even more through automation.
Daniel Darling: What an exciting time. And are there any other big hurdles that your industry has to sort of break through to really start to see this inflection point accelerate, or are we waiting for the first kind of in-production case studies from the industrial complex to then prove out to the rest of the industry that this has value and this works?
Greg Fallon: Yeah, there are some significant other bottlenecks as you can imagine. I would say that data privacy is a major one. So it’s something that we think about all the time, and we face it constantly with our customers. In the industrial space, the data is IP, intellectual property. And the industrial companies that we work with are very concerned about their IP somehow escaping in any way, shape, or form.
And so from an AI perspective, there is a fear of your data training an AI algorithm that could somehow be used by one of your competitors. And so there’s this idea of AI privacy within these entities. It’s one that we’ve tackled head-on. The way that we work with our enterprise customers is we don’t come in with any pre-trained AI model. So we always start from scratch with every customer. And even doing that, we’re able to get these disruptively short deployment times going. So there’s the data privacy issue.
There are other issues as well that relate to data sovereignty that we’re dealing with. What a data sovereignty law would dictate is, let’s say, that you’re operating in the country of Kuwait, data can’t leave that country even for processing. So you can’t have any sort of data center outside the country. And so, the hyperscalers are addressing that. Google has put data centers in the Kingdom of Saudi Arabia. They have a partnership with Saudi Arabia. Microsoft also has partnerships with the UAE and Saudi Arabia. So that problem is starting to get solved, but it will be around for a while.
Daniel Darling: Thanks so much for sharing your vision for what you guys are building and how you’re really transforming such a big sector of the economy. It really feels like we’re coming to an inflection point in that adoption and that change. I appreciate you coming on, Greg, and chatting with us.
Greg Fallon: Yeah, thanks so much, Daniel.
Daniel Darling: It was truly enlightening to glimpse the future of our industrial world through Greg’s perspective. What stands out is the immense potential AI holds for the industrial sector, particularly with its unparalleled access to a huge amount of proprietary data and the capability to train highly specialized models. Industry giants like Saudi Aramco are already capitalizing on this opportunity, not just to enhance their own operations but to offer digital infrastructure and AI-driven intelligence to their peers and the industry at large.
This trend of collaboration and innovation is incredibly powerful. I’m particularly excited to see Geminus’ specialized AI models enter production at scale, poised to usher in a new era of generative engineering. The future of heavy industry is being rewritten, and we’re just at the beginning of this transformative journey. To follow Geminus’ and Greg’s progress, head over to their website - geminus.ai. I hope you enjoyed today’s episode, and please subscribe to the podcast to listen to more coming down the pipe.
Until next time, thanks for listening. Have a great rest of your day.
