Since mid-summer, the AES Corporation and the AI Fund have been quietly working on building artificial intelligence startups for the energy sector. And in under six months, the two have taken one all the way to the funded startup stage. It’s currently in stealth mode, but is on track to be deployed internally at AES in 2025.
That partnership, announced in June, is a new twist on an approach to incubating energy startups that AES has been leveraging for several years. Most notably, the energy company created storage company Fluence as a joint venture with Siemens, and spun it out in 2018.
And the two have been moving rapidly. AES sifted through a list of more than 100 ideas sourced from internal stakeholders, before ultimately paring that list down to just 12 that they pitched to AI Fund, a venture studio dedicated to AI applications across many industries.
AES and the AI Fund are still working their way through prioritizing that truncated list. But already, in addition to the company that is already funded (by the AI Fund), they have several others in incubation.
I sat down with Julia Lundin, who leads digital solutions at AES, and Abid Siddiqui, a principal at the AI Fund, at Latitude Media’s Transition-AI conference in Washington, DC this week, to get an inside look at their unique approach to creating built-for-purpose AI tools for energy. Our conversation has been edited for clarity and brevity.
Why did AES opt for this different structure to incubating AI companies specifically?
Julia Lundin: The reason why we decided to do it this way, as opposed to most of our other ventures which were incubated in-house for years before we sought external partners and spun them out, is that the pace of AI really demanded a different approach. It was just a matter of the pace of change — and the need to work with an external partner that was closer to that pace of change and could really help us manage that.
And then the second reason is that AI is a general purpose technology, and we wanted to both have access to the way that AI was being applied across the energy industry, and be able to learn from what other industries were doing. And so this partnership with AI Fund was really a way for us to leverage that.
The AI Fund invests across a range of sectors, which means that energy is competing for financing against non-energy companies. In that context, what makes an AI solution for energy venture-backable?
Abid Siddiqui: It’s an interesting time in the market. We’re seeing use cases that were historically not venture-backable become venture-backable.
The proliferation of agentic technologies has taken its small market share, and added in the fact that we can now have AI operate at the junior analyst level within a lot of different organizations. That fact creates larger market opportunities that we see a lot of our LPs, such as AES, take advantage of — and build and launch these companies that historically may not have been as attractive to venture capital.
And the second thing is, we’re seeing a lot of focus on productivity tools. Cutting the administrative burden on a particular task is a very common use case. But what we’re looking to see now is really not just productivity, but also the top line output efficiency. Is it increasing yield or throughput for a particular plant?
How do you decide whether to incubate a concept within energy directly, or to build it out in another sector and then translate it over to energy?
Abid: I have a very simple answer, which is who has the “hair on fire” problem? That becomes one of the core aspects of what we do: understanding where an ideal customer has this pain point, and how that resonates across other sectors, in manufacturing, construction, etc. What we look for is a very specific ideal customer profile that is experiencing this pain point. That typically guides us on a good market interest strategy.
What are some examples of where you’ve run into challenges with access to data, and how do you work around those limitations when training models?
Julia Lundin: When it comes to automating workflows, one challenge we run into is that our systems of record don’t capture intermediate states. So when we think about something — like status, for example — moving through a workflow, a lot of the other fields [with potentially useful information] get overwritten as it goes through the stages.
How do you recreate that data to give the AI model the same information that the human [operating in the system] had on that date? We’ve put in place now more “snapshot” functionalities in our systems of record to be able to start capturing that data, to enable us to do more.
Abid Siddiqui: When we do run into data challenges, there’s a number of different techniques you can use. Nothing is better than real and clean data, so that’s obviously the Tier One option. But there’s also other alternative strategies for creating synthetic data sets, such as using physics-based modeling to try to come up with a representative data set to be able to train your model.
The other approach we commonly see in the AI Fund — because we do invest across sectors — is that we have models that are pre-trained in other sectors that are very relevant to the use cases that we’re talking to with AES. So we can take pre-trained models and have at least an initial starting point that can then be finetuned to a particular use case with AES.
When it comes to AI startups for energy, is it more important to be fast? Or to be right the first time?
Abid Siddiqui: As a studio, we want to be fast. We want to hire a founder that can iterate, pivot when we need to, and capitalize the company appropriately. So I think for us, moving fast is important. The market is moving quickly. So we are very, very biased towards fast.
Julia Lundin: For AES, we want to come up with the right answer. We have more of a deliberate process [than AI Fund] where we don’t want to end up with a bunch of orphaned ideas that started and disappeared, and it’s not documented why they disappeared. And so if we kill an idea, we want to kill it thoroughly and make sure that we really have given it a chance.
So that’s been a tension, but a good tension, in how we balance those two processes. And one change we’ve made is in the idea generation phase, AES has taken on more of the work forming ideas on our side, so that when we bring something to the AI Fund, we’re at the point where we’re ready to move at their pace.
