The Model Isn't Your Moat

My technical partner, Daniel Englebretson of Elynox, recently crossed 100 billion tracked tokens in his own AI usage telemetry — 232 days of building, 16 billion of those tokens in the last seven days alone. He is someone who relentlessly measures what moves the needle because he builds with AI every day.

One line from his recap stuck with me as a marketer and founder working to build an AI-native business: hard things are still hard, no matter how much you work with AI.

He's getting more done, but he's still working just as hard — his standards moved right along with the tools.

As marketers everywhere evaluate their AI stacks and technology approaches, many are still fixated on the wrong question: which model is best?

Should open weight really matter to marketers?

Let’s start by exploring this question a little more deeply. Discussions of open weight AI models are everywhere, but do most marketers really need open weight or are out of the box options good enough?

There are three main benefits open weight models can bring a marketing team: data control, cost control and fine tuning.

• Data control — self-hosted, your prompts and outputs never have to leave your own infrastructure. That matters if you're working with unreleased product details, customer data or anything else you can't hand to a third-party API.
• Cost control — no per-token vendor pricing once it's running on your own hardware, though your compute and ops costs largely replace it (more on whether that trade actually pays off below).
• Fine-tuning — you can adapt the model's actual weights to your brand voice, your taxonomy, your historical content — not just prompt around the edges of a general-purpose model.

Open-weight models — Llama, Mistral, DeepSeek, Qwen and others — release the trained parameters themselves. You can download them, run them on your own infrastructure, and fine-tune them on your own data.

What you don't get is transparency, auditability and community governance. You're still trusting a vendor's training choices and safety work sight-unseen, same as you would with a fully proprietary model.

Self-hosting isn't automatically cheaper — tuning is what actually pays off

So, if the model isn't differentiating, is the hardware where marketers need to focus?

There's a common assumption that self-hosting simply trades a subscription bill for a compute-and-electricity bill. That trade is real, but the biggest cost factor isn't which model you run — it's how disciplined you are about tuning that model to your actual workflow versus using its default settings.

Let that sink in. Your chosen model and the architecture that supports it aren't what matter most. It's the marketer's proprietary methods and how well she can articulate them that changes the game - whether we are talking output quality, energy cost or customization.

Daniel shared a concrete example of this with me with regard to efficiency: The graphics card he runs defaults to up to 256 concurrent processes. It is built for the broadest possible customer, since that's what a hardware vendor has to design for. His actual workflow never requires more than four processes running at once. By re-tuning to his real use case, he was able to achieve a 20% electricity savings. The operator's judgment about his actual workflow is what dictated his cost — not the hardware, and not the model.

Who actually does the "tuning"?

It's worth clarifying that none of this requires a marketer to become a hardware engineer. When I asked Daniel who's actually capable of the kind of tuning he described, his answer was reassuring – marketers already have the underlying instinct required to tune their workflows to perform in any model. They already think in A/B tests and are well versed in explaining the same finding five different ways to five different audiences. They already have to ask what the numbers actually say, whether or not they're naturally good at it. That instinct — knowing what to test, and why — is the exact thinking a model needs directed at it.

It's no coincidence that marketing has been hit by AI disruption harder and faster than most fields: the discipline was already built for this kind of thinking.

Proprietary context is the real constant

However helpful, none of this actually answers the competitive question marketers care most about: how do we get and stay ahead?

Whichever model you choose, it's ultimately a wasting asset. Frontier models get replaced on a predictable cycle. Pricing keeps falling. Whatever edge "we use the best model" bought you evaporates within a release cycle or two, because your competitor is one API call away from leveraging the same model.

What doesn't commoditize is what you feed the model: your customer evidence, your proprietary research, the positioning you've refined over years of actually talking to buyers. That context is yours. The model is rented.

A real example of the marketer’s moat

I asked Daniel for a concrete case where the craft, not the system, turned out to be the moat. His answer went back to Kaizen — the manufacturing discipline built on going to where a problem actually happens, watching it occur, and helping the people closest to it see what they couldn't see themselves.

Sit with a team long enough, and patterns jump out that have nothing to do with any tool: a messy desk is a sign of a messy process; a part that travels 100 unnecessary yards across a floor is a sign the process itself needs fixing. Nobody can articulate all of that in a manual — it's earned pattern-recognition built from experience.

The system you build around a team is replaceable, but the hard-won judgment that spotted the problem in the first place isn't.

Every competitor in your category has access to roughly the same handful of frontier models that you do. They don't have is your specific customer evidence, your positioning, your proprietary datasets, or your team's understanding of what's true in your market. That context is the asset that compounds.

This also lines up with the most critical point in Daniel's post: LLMs cannot do anything truly novel.

Anything genuinely new that comes out of working with one is entirely owed to the human directing it. The model supplies leverage, but the judgment, the context and the nuance are still yours to bring – and that’s where the savings, efficiency and competitive edge come from.

Build the moat: auditing your AI stack

Here's a useful gut-check for any marketing team building on AI right now:

• Are we leveraging our proprietary assets in what we're building — customer evidence, positioning research, brand voice, structured workflows?
• Are we investing more in this context infrastructure — the knowledge bases, prompt libraries and internal vocabulary — than the model relationship?
• If the model we use today were swapped out tomorrow for a better or cheaper one, would our output quality survive the swap?

If yes, your moat is context. If no, you don't have one yet — you have a subscription.

As Daniel loves to remind me: The model is the easy part. Your context is the moat.

That's every marketer's opportunity and value in technical conversations.

Let's not forget it.