Everything Was an AWS Wrapper Too

We used to mock companies as "GPT wrappers." Here is the thing nobody said out loud: everything was an AWS wrapper too. And several of those wrappers became billion-dollar companies.

The big AI labs are running the hyperscaler playbook in plain sight. Anthropic is building a marketplace where committed spend can pay for third-party software running on Claude. OpenAI is doing the same. They are not trying to build every solution themselves — they cannot. They need software vendors building complete products on top of their platforms, exactly the way AWS, GCP, and Azure needed Databricks, Confluent, and thousands of others. Amazon has over 400 native services and still has more partners than services. The platform never eats the whole stack.

Watch what happens when a platform owner does build into your space. AWS built a managed Airflow service. So did Google. At Astronomer, we competed with both — and the clouds still sent us their complex customers. Platform-native offerings are 80 percent solutions built with 20 percent of the effort, because compute is the business and the service exists to sell compute. Anyone with a hard use case falls through, and the dedicated specialist catches them.

The same dynamic is already visible in agents. Claude has background agents. They are a lightweight side project staffed by a couple of engineers, attached to a company whose core business is the model. A team whose entire existence depends on going deep — handling the customer who needs a custom runtime hook on VM startup, the gnarly enterprise request that nobody else will touch — wins that fight on focus alone. We already have customers who tried to go deep on the lab-native tooling and hit the bottom of it. As I argued in the long game in agent companies, the winners in this market are decided by who keeps showing up for the unglamorous depth.

There is a second advantage to choosing the wrapper layer: you do not have to bet on a winner. Models have different personalities and strengths — one is better for research, another more objective and to the point, and new harnesses ship constantly. Nobody is going to lock into one coding agent, because no single coding agent can win. That chaos is the agnostic layer's friend. The wrapper that supports every agent, instead of competing with any of them, gets stronger every time a new contender launches.

The other objection I hear: a strong enterprise team will just build this in-house. We watched this movie during the modern data warehousing era. Companies built in-house pipeline tools because the old stuff did not work — and then Airflow won anyway, not because it was pretty, but because it was complete. There was nothing left to add. In-house tools also share a fatal flaw: even on a team of four, there is always one central figure, and the moment that person jumps to another company, the tool withers fast. Meanwhile, the internal team gets three to six months and a mandate that is really about something else. The biggest hundred companies will build internally. Almost everyone else gets overrun by commercial and open source options playing the long game.

So when someone calls your product a wrapper, remember what the word actually describes — a company that chose its layer correctly. The labs will become the new clouds. The marketplace on top of them is where the next generation of enterprise software companies gets built. The question is not whether to build on the platform. It is whether you are willing to go deeper than the platform ever will.