Treat AI like infrastructure, not the end product. Build the indispensable layer on top. The real winners will look more like Databricks than another “AI for X” wrapper.
Most companies buy AWS to build their real product.
Most buy AI as the product they then polish and sell.
This single difference explains why AI moats look so weak.
Durable AI businesses need what the AWS winners had: deep integration, proprietary data, switching costs, or distribution control. Pure model upselling has almost no moat.
Most AI revenue today is just rent on someone else’s model. Sustainable wins need proprietary data loops, workflow ownership, or distribution not pure polish. Pure AI wrappers are racing to the bottom.
AI upselling is fragile because models improve for everyone at once. Your “moat” of prompt engineering disappears with the next release. AWS-polish plays win through data gravity, compliance, and switching costs.
Exceptions exist. Perplexity created a daily browsing habit. Cursor changes how engineers work. These prove AI can build real usage. But they’re rare compared to the ocean of me-too copilots and summarizers.
Snowflake, Databricks, and Datadog didn’t just use AWS they productized it. They added deep engineering, abstractions, and charged premium. Most AI startups stop at the thin wrapper layer and call it a product.
Everyone buys AWS as raw infrastructure. The game is what you build on top. AI is different: companies mostly buy models to repackage and resell as their own product. Success depends on how well you hide the commodity.
People think the AI battle is about smarter models. I think it’s about deployment.
Factories, logistics, and infrastructure are still fragmented and messy. Bezos keeps positioning there quietly. Am I overestimating that shift?
The loudest AI companies optimize for visibility. Bezos keeps leaning toward logistics, automation, and operational systems instead.
The visible AI race may not matter most. Deployment into the real world might. Who’s closest to solving that?