PIVOT! What the Moving Guy Taught Me About AI Moats in OT Security

My furniture was delivered on Saturday as part of a personal move to get back into the warmer weather I grew up with in the South (from New Hampshire to Georgia!).

One of the guys making multiple trips through my door (let’s call him Marcus, because that was actually his name) had the kind of forearms that come from two years of hauling furniture up staircases. At some point between the couch and the dining table, he asked what I did for work.

I told him I work in OT cybersecurity. I help protect the kinds of industrial systems that keep the lights on, water flowing, and refineries running. I mentioned that I work to try to prevent AI from disrupting everything.

Marcus set down the table, wiped his hands on his jeans, and said something I haven’t stopped thinking about: “Well, AI’s not taking my job anytime soon.”

He’s right. But here’s what’s interesting: in my world, that’s no longer true.

In a world where AI-”powered” OT Cybersecurity is beholden to third-party technology, if that 3 party pulls the plug or makes a change, you may need to… change direction quickly. What’s the word I am looking for?

“PIVOT!”

The moat question nobody’s asking out loud

For the last two years, every OT security vendor has been racing to slap “AI-powered” on their product. Detect anomalies faster. Triage alerts automatically. Predict vulnerabilities before attackers find them. It’s compelling marketing, and some of it is genuinely useful.

But here’s the question CISOs in critical infrastructure should be asking that almost no one is: what happens when the AI is the commodity?

We got a preview of this recently with events like the Mythos export controls, a stark reminder that a company’s AI-dependent capabilities can be regulated, restricted, or simply cut off at the infrastructure level. If your OT security vendor’s core differentiator is running on top of a model that can be export-controlled, sanctioned, or eventually open-sourced into irrelevance, that’s not a moat. That’s a dependency.

Marcus knows his job is safe because no AI can carry a couch up three flights of stairs and solve a problem in a narrow hallway in real time. What’s your vendor’s equivalent of the narrow hallway?

What a real moat looks like in OT security

Let me be specific, because “we use AI” versus “we are AI” is a distinction with huge strategic consequences.

Architecture is a moat. Models are not.

A vendor whose value is “we run threat intelligence through a large language model” has a defensible product until GPT-5 is open-sourced and every competitor runs the same model. The vendors who will still matter in ten years are the ones who built something that can’t be replicated by plugging in a better model.

What does that look like in OT? It looks like software-defined perimeters that make your assets invisible to attackers in the first place. Not AI that detects the attacker after they’re already in your network. It looks like network architectures that are fundamentally harder to exploit, regardless of which AI the attacker uses to probe you.

The dirty secret of detection-first security is that it’s an arms race, and AI makes the attacker’s side of that race dramatically cheaper. If you’re playing defense by watching for bad behavior, and the attacker is using AI to generate novel behavior at scale, you’re on the wrong side of a cost curve.

Data is a moat. But only specific data.

There’s a version of AI in OT security that is actually defensible: proprietary knowledge of OT protocols, device fingerprints, and attack patterns that no one outside deep operational experience could have. The vendors who have spent a decade watching how Modbus traffic behaves on a real power grid, or how a specific PLC responds under stress, have data that can’t be scraped from the internet.

The ones who don’t (who are essentially running general cybersecurity threat intelligence through a general-purpose model and calling it OT security) are going to find their advantage much thinner than it looks when a well-funded competitor (or an open-source community) catches up.

Regulatory integration is an underrated moat.

Here’s something Marcus doesn’t have to worry about: NERC CIP compliance. But your OT security vendor’s ability to be genuinely embedded in the compliance frameworks that govern critical infrastructure (not just “compatible” but mappable to CIP-005, IEC 62443, and TSA Security Directives in ways auditors actually accept!) is stickiness that no AI upgrade cycle can disrupt.

The vendors who’ve done that work, who are referenced in the compliance documentation for major utilities and pipeline operators, have switching costs that have nothing to do with whether their model is better than a competitor’s.

Profitability, predictability, and the “AI premium” problem

There’s a financial angle here that boards are starting to scrutinize.

A lot of OT security vendors are burning significant GPU compute resources to run AI inference on high-volume sensor data from industrial environments. That’s expensive at scale in ways that traditional software isn’t. When a vendor’s gross margin depends on the cost of compute they don’t control, and compute costs are set by hyperscalers that are also building competing security products: that’s a structural risk, not just a technical one.

Compare that to a vendor whose product is fundamentally a software-architecture play: the value lies in the design of how access is granted and how assets are hidden, not in the ongoing cost of running models on every packet. The margin profile is different. The predictability is different. The exit multiple, if it ever comes to that, is different.

CISOs don’t always think about their vendors’ business models, but they should. The vendor whose profitability depends on an arms race they can’t control is a different kind of third-party risk than the one whose value is embedded in how your network is structured.

What I told Marcus

I didn’t give him the full version of this on Saturday. I told him AI was going to change a lot of knowledge work, but that physical problem-solving in unpredictable environments would be the last thing to go.

What I didn’t say (cause the couch was getting heavy!) is that many OT security vendors share the same vulnerability he doesn’t: they’ve built their value on something someone else can replicate, regulate, or take away.

The ones who’ve built the moat into the architecture? Those are the companies I’d want protecting the things that can’t afford to go down.