We recently raised a round. If you have done that, you know the word that comes up all the time: moat. Investors look for a moat. So a big part of my job is creating conviction around mine.

Moat is a concept I find hard to put into words. Our industry is complex, and the moat itself keeps moving. The way I would have described ours when we started is not how I would describe it today. There are good frameworks (Helmer's seven powers, the network-effects literature) and they work as a checklist. But a checklist is not conviction.

The software industry is being disrupted. Building a competent application used to take a ten-person team a year. Now it takes just a few smart people and less than a quarter. When the marginal cost of a feature trends to zero, any moat that was really just "we shipped first and it is hard to rebuild" is a head start, not a moat. Fast followers have AI tailwinds, and buyers know it. So the question is not "what are the moats." It is which ones survive cheaper software building, and what replaces the ones that do not.

My lens is obviously energy-applied. I use energy because it is the industry I love and work in, and because physical, regulated, high-stakes markets make the surviving moats easier to spot. The principles might generalize. Maybe you have some feedback on that.

‍

The moats that might die

Implementation and integration complexity: that held when integrations were hand-built and migrations were suicide missions. Models now read APIs, map schemas, and write connectors with little friction. Software talking to software is exactly the work AI just made cheap. (Hold that thought. There is a physical-world cousin that gets stronger.)

The static data moat: a dataset that was expensive to collect once is a depreciating asset. The test is simple: does the data keep getting better because you operate, or did you just get or buy a big pile of it?

UI and workflow lock-in: a familiar UI is a habit, not a barrier, and AI is dissolving it fast. Once an agent drives the workflow, muscle memory for your specific screens stops mattering. The interface you polished for years becomes a quite replaceable part of the stack.

My colleague Seba, who runs product at Companion, cut me off right here when asked for feedback:

That is a narrow definition of UI. Pixels on a screen, sure, those are getting cheap to make "good enough." But UI is the interface between your model and the rest of the world, and agents are users too. A crappy interface used to cause human errors. A crappy interface today hands those errors to an agent that happily multiplies them at scale.

He has a point. Kill the pixels-on-a-screen moat, not the discipline of designing a clean interface between your model and everything, human or agent, that has to act on it.

Regulatory and market access: this is the one people are most sure of, and I want to push back. In energy, more and more prevalent market-access-as-a-service models now let you rent the license and plug in through someone else's balancing perimeter. The moment access is a product you can buy, it is a line item, available to your fast follower on the same terms. If your edge is a license someone will rent to your competitor, you do not have a moat. You have a supplier. Again, this one is likely quite energy-specific. A banking charter, an FDA approval, or a defense clearance is a different animal, with no rent-a-license shortcut, and there regulatory access can be one of the most durable moats there is. You can figure it out for your own industry.

None of these are worthless, but neither are they a wall. Speed bumps, maybe.

‍

The moats that get stronger

Compounding context: the defensible version of the data moat is not data, it is context that deepens through a feedback loop. You operate, you learn how specific systems behave and which edge cases matter, better decisions win more operations, more operations generate more context. Raw data is cheap, but the context for how a specific slice of the physical world behaves is not. LLMs make it trivial to collect Fluvius data. Turning that into a correct savings number takes a real data framework, serious compute, and domain expertise. That gap is the moat.

As Erik, part of our engineering team, puts it:

The companies that stay safe post-AGI are data companies, not software companies. And not because they own some proprietary dataset. The moat is being able to juggle that data, operate on it, and turn it into value.

Bridging software and physical reality: models write connectors now, and I am not going to pretend a pile of clever API work is what makes us truly defensible. The hard part is the other kind: reliably acting on the physical world. Hardware behaves differently in the field than on the datasheet, sensors drift, readings go missing, and assets fail in ways no schema anticipates. A wrong number can have real, physical consequences. Getting software to make good decisions against that mess, and to know when not to act and raise a flag instead, is slow, unglamorous, and genuinely hard. That difficulty does not compress the way code does. Reality does not have an API just yet.

Counter-positioning: Helmer's sharpest idea in my opinion, and AI makes it more relevant, not less. Build a model the incumbent cannot copy without hurting their existing business. An incumbent whose margin does not fully hold in a world of ultimate transparency cannot become the transparent, optimization-first player without repricing their own book.

Portfolio and network economies: some moats only exist at aggregation. A portfolio of energy assets gets diversification that lowers imbalance risk and access to volume-gated markets a small operator cannot touch.

Process power (the real last mile): in the enterprise segment, the last mile is getting the thing actually used inside an organization that usually resists change. Encode the unwritten rules of how a corporate function runs, and over years your opinions about how they should operate become inseparable from how they operate.

Brand: brand-as-reputation is real and hard. When you ask a utility to let you dispatch their assets and trade their position, the deciding factor is whether they trust you not to blow up their P&L. A fast follower cannot buy that. In our market, brand is just the trust layer wearing a nice suit.

‍

Trust and lineage

Where software takes financial positions or controls safety-critical assets, fully autonomous, black-box AI is not ready. Pretending otherwise is the fastest way to lose an enterprise customer. When you are trading a customer's position or dispatching their battery, "the model said so" is not an answer a CFO will accept. They need auditability, traceable data lineage, deterministic guardrails, and a human in the loop where it counts.

The trust layer is a moat, and genuinely hard. Yet it is exactly what enterprises demand before they hand over control.

So the moat is not just the intelligence. It is the context you have earned, and the machinery to act on that context autonomously, in a way the customer trusts. Cheap intelligence makes that possible. Earned context and a trust layer are what make it defensible. This is the position I would stake out for any company like ours, one that should interest a generalist fund as much as a specialist. Someone has to translate model output into an action a real-world system will accept, safely and within the rules, wired into hardware and markets that took years to reach. The scarce thing is decreasingly "thinking," and increasingly trusted action in the physical world.

‍

What I would tell a founder

Nothing. I hate unsolicited advice, so I will spare you mine.

But if I were talking to myself, I would say that in energy it is unlikely to win by digging one deep moat. You win by owning the less-sexy, hard-to-copy work between a good model and a customer with real spend and assets on the line. The intelligence is becoming a utility. The layer that safely connects it to the physical world, and derives value from it, is not.