The model is not your moat

Most executives asking “which AI should we buy” have already lost the plot. The model is the one part of the stack that will never be yours, and they’re treating it like the whole decision.

I watch this happen in boardrooms and on sales calls. A leadership team spends a quarter evaluating foundation models like they’re choosing a database vendor in 2004. Bake-offs. Benchmark decks. A procurement matrix with forty rows. Then they sign a contract, roll out a chat window, and call it a transformation.

Six months later a competitor with the same subscription is pulling away, and nobody can explain why.

Here is the uncomfortable math. The frontier labs leapfrog each other roughly every quarter. The model that tops the benchmarks in March is second place by June and third by September. Whatever careful analysis your team did to pick a winner has a shelf life shorter than the procurement cycle that produced it.

That’s not a criticism of the labs. It’s the point of them. They’re locked in the most expensive capability race in the history of software, and the output of that race is a commodity that gets better and cheaper on a schedule you don’t control and don’t pay for.

Which means “which model” is a question about a depreciating input. Nobody built an empire on picking the right electricity provider.

What actually happened when I swapped the engine

I’ll tell you what convinced me, because it wasn’t a thought experiment.

I run Vidloft as the only engineer on the team. The systems I’ve built around AI produce output I’d previously have needed about twenty people to ship, across engineering, sales operations, and research. And in the time I’ve been building this way, I’ve swapped the underlying models repeatedly. Different labs, different generations, sometimes mid-project.

You know what changed about my competitive position each time? Nothing. Not once.

Because the value was never in the model. It was in everything built around it. The workflows that break a fuzzy business problem into steps a machine can actually execute. The judgment encoded into checkpoints, so bad output gets caught before it ships instead of after. The taste, which is the unglamorous word for knowing what good looks like in your specific domain, at your specific quality bar, for your specific customer. The feedback loops that make the whole system a little sharper every week.

When a better model arrives, I drop it into that machinery and the machinery gets faster. My advantage compounds. My subscription didn’t get smarter. My system did.

The instrument is not the performance

The analogy people reach for is plumbing. Models are the water, you build the pipes. It’s fine, but it undersells what’s happening, because water doesn’t get twice as good every year.

Try this instead. A frontier model is the finest instrument ever manufactured, and every one of your competitors can buy the identical instrument tomorrow morning. The instrument is astonishing. It is also, competitively speaking, worth nothing, because ownership is universal.

What isn’t universal is the ability to play. The arrangement, the ear, the years of knowing when a note is technically correct and still wrong. Orchestration is the actual craft, and the word means exactly what it sounds like. Deciding which instrument plays what, in which order, listening to what comes back, and having the standards to send it back when it isn’t right.

That craft doesn’t depreciate when the next model drops. It appreciates. Every improvement in the instrument makes the person who can play it more dangerous, and does nothing at all for the person who bought it and left it in the case.

The thirty-second diagnostic

Here’s a test you can run on your own organization right now, no consultants required.

Imagine a frontier lab releases a model tomorrow that’s twice as good as anything available today. Genuinely twice as good, at everything. Now ask: does your competitive position improve, or does your competitors’?

If your honest answer is “everyone benefits equally,” you do not have an AI advantage. You have a subscription. You are paying the same rate as everyone in your industry for the same commodity, and calling the invoice a strategy.

The companies that should smile at that hypothetical are the ones with something proprietary wrapped around the model. Workflows the new model plugs into on day one. Evaluation criteria that immediately show where the new capability clears their quality bar and where it still doesn’t. Institutional judgment that’s been written down, encoded, and made executable, so the improvement propagates through the whole operation instead of dying in one enthusiast’s browser tab.

For those companies, a better model is leverage. For everyone else, it’s a price change.

Where the work actually is

None of this is an argument against the models. I use them constantly, promiscuously, and with real gratitude. It’s an argument about where your scarce attention should go.

The hard question was never “Claude or GPT or Gemini.” The hard questions are the ones your team is avoiding because they’re about your business and not about technology. Which decisions in your operation actually require human judgment, and which just have it by historical accident? What does good output look like, specifically, and who in your company can articulate it precisely enough that a machine could be graded against it? Where does work sit in a queue for three days waiting on a person whose contribution takes four minutes?

I spent a decade producing video before I built software, thousands of projects, some of them for the biggest brands in the world. The clients never asked what camera we used. They were right not to. The camera was the commodity. The judgment about what to point it at was the product. Nothing about AI changes that logic. It just raises the stakes on it.

Answering those questions is slow and unglamorous and it doesn’t produce a press release. It produces something better. It produces an operation where model upgrades land like free horsepower instead of like another tool nobody adopted.

The labs will keep leapfrogging each other. Let them. Their race is your tailwind, but only if you’ve built something for the wind to push.

The model is the commodity. The orchestration is the company. Every quarter, that distinction gets more expensive to ignore. The executives who understand it aren’t asking which AI to buy. They’re asking what to build around whichever one is best this quarter, knowing the answer to “which one” will keep changing, and the answer to “what we built” won’t.