"Unless the Lord builds the house, those who build it labor in vain." — Psalm 127:1
There is a phrase on a great many pitch decks right now: AI-native. It signals that a company was built from the ground up around artificial intelligence — not a legacy business bolting on a feature, but something new and intelligent at its core. It pairs with words like proprietary and defensible and our technology. It is meant to make you believe the company owns something rare.
Open the hood on a large share of these companies and you find the same thing: the intelligence isn't theirs. It's an API call to one of a small handful of foundation-model providers. The "proprietary AI" is a prompt, some plumbing, and a monthly bill to a company that could raise the price, change the terms, cut off access, or build the exact same feature and ship it to everyone tomorrow. The startup did not build the house. It is renting a room and calling the building its own.
What "AI-Native" Often Actually Means
Be precise about the architecture, because the precision is the whole point. A meaningful number of "AI-native" products are, structurally, a user interface wrapped around someone else's model. The company designed the experience, wrote the prompts, built the integrations — real work, often good work. But the part that does the actual thinking, the capability the entire pitch rests on, is licensed from elsewhere and reachable only as long as the bill is paid and the provider remains willing.
The tell is a question you can ask any "AI-native" company in one sentence: whose model is this, and what happens to your product if they change the deal? If the honest answer is "it's OpenAI's, or Anthropic's, or Google's, and we'd be in serious trouble," then the company's core intelligence is rented, and everything described as proprietary sits on top of a foundation owned by someone else. That's not necessarily a bad business. But it is a fundamentally different business from the one the words "proprietary" and "defensible" describe, and the gap between them is exactly the thing the marketing is built to hide.
The Concentration Underneath Everything
Now widen the lens, because the dependency isn't just a single-vendor problem — it's an industry-wide one. The infrastructure required to train and run frontier AI models — the compute, the chips, the energy, the capital — is controlled by fewer than a dozen entities on the planet. The thousands of "AI-native" startups, for all their variety, mostly trace back through a few model providers to the same tiny set of compute owners at the bottom. It's a wide, colorful canopy growing out of a very narrow trunk.
This means the apparent diversity of the AI economy is, in an important sense, an illusion. A thousand companies that all depend on three models that all depend on a handful of compute providers is not a thousand independent bets. It's one bet, wearing a thousand logos. When the trunk moves — a pricing change, a policy shift, a capacity constraint, a strategic decision at the bottom of the stack — it moves the entire canopy at once. The startups experience as "innovation" and "competition" what is, structurally, shared dependence on infrastructure none of them control and most of them could never build.
The Hypocrisy of "Independent" Innovation
Here's where the language does its quiet damage. The whole story the industry tells about itself is one of scrappy independence — disruptive startups, bold founders, a thousand flowers blooming. The reality for many of those companies is near-total dependence on the very incumbents the disruption narrative pretends to challenge.
The "independent" startup is often a reseller of an incumbent's capability with a markup and a nicer interface. Its existence depends on the incumbent's continued goodwill and stable pricing. And the incumbent, who can see usage patterns across the whole ecosystem, is perfectly positioned to identify which "independent" products are working and absorb their best ideas into the platform itself. The startups are, in effect, doing unpaid product research for the companies that own the layer beneath them — and calling it a war of independence while they do it. That's the hypocrisy: an innovation economy that markets autonomy and runs on dependence, where "we built something new" frequently means "we found a clever way to resell something three companies own."
Why This Is Your Problem, Not Just Theirs
If you're an operator choosing tools — rather than a founder building them — this is not an abstract concern. It determines whether the products you depend on will still exist, at a price you can afford, in two years.
When you build your operation on an "AI-native" tool, you inherit its dependencies. If that tool is a thin wrapper on a model it doesn't control, then its pricing, its reliability, and its survival are all hostage to decisions made several layers below it, by companies you have no relationship with and no leverage over. The tool can get more expensive overnight because its provider did. It can degrade because its provider changed a model. It can vanish because its provider decided to compete with it. You took on all of that risk the moment you built on it, and none of it appeared on the pricing page.
How to Evaluate What's Actually Underneath
You don't need to audit anyone's source code. You need to ask where the durability lives, and a few questions surface it fast.
What does this company actually own? Strip away the rented model and ask what's left. A proprietary dataset, a hard-won workflow, a real relationship with you, accumulated domain knowledge, the ability to run on more than one underlying model — those are real assets. A clever prompt on top of a public API is not. The owned part is the part that will still be there when the rented part changes.
What happens if the layer below changes the deal? Ask the vendor directly how exposed they are to their model provider's pricing and policies, and whether they can switch providers if they need to. The quality and honesty of that answer tells you how much risk you're quietly inheriting.
Is the differentiation in the model, or in everything around it? The most durable tools rarely win on having a smarter rented model than the next wrapper. They win on the things that are genuinely theirs — their data, their judgment, their fit to a specific job, their willingness to let you own and move your own work. Differentiation you can't rent is the kind that survives.
The Point
"AI-native" and "proprietary" are doing a lot of unearned work in this market. Sometimes they're accurate. Often they're a costume over a reseller's economics and a borrowed brain. The distinction matters because it predicts which tools will still be standing, and at what price, after the companies at the bottom of the stack make their next move — and they will make it for their own reasons, not yours.
Unless you know who actually built the house, the old line warns, you may be laboring in vain on a foundation that isn't yours. Before you bet your operation on something that calls itself intelligent and independent, find out whose intelligence it is, and what's left of the independence when the rent comes due.
Sources: industry analysis of foundation-model dependency and "wrapper" startup architecture; reporting and commentary on the concentration of AI compute and frontier-model capability among fewer than a dozen providers; standard platform-risk and vertical-integration dynamics in technology markets.
