If you're using Copilot, Gemini Code Assist, or Amazon Q to write code, the m...

If you're using Copilot, Gemini Code Assist, or Amazon Q to write code, the model is steering your architecture toward its owner's cloud. That's not a hunch anymore. Researchers measured it.

They tested 10 provider-affiliated models against 3 independent controls on real integration decisions: picking a database, choosing a cloud provider, selecting an auth library. The stuff that shapes your architecture for years. They call it "Vertical Integration Bias" (Nguyen et al., published this week on arXiv). The numbers aren't subtle.

When the model writes a single file, the bias toward its owner's ecosystem runs about 19 points. When it plans and executes a multi-step project (the "agentic" mode everyone's selling you), that jumps to +39 points. Once the agent made a vendor-affiliated choice early in a workflow, that decision persisted into downstream files 90% of the time. One biased pick at the top of the chain infected the whole project.

I made a whole video about harness lock-in back in March how the tools you build around compound faster than you realize. But this is a layer deeper.

The tool is quietly making architecture decisions that make switching more expensive. Your agent doesn't just prefer its owner's cloud. It builds your entire stack around it, file by file, before you've consciously chosen anything.

The practical question for anyone using AI to write code (or evaluating tools for a team): who's actually making your infrastructure decisions? Because if you're not asking, the answer is probably a model with a conflict of interest.

None of this means stop using AI coding tools. It means stop trusting them like they're neutral. Review the dependency choices, not just the code quality. And if you're standardizing your engineering org on one provider's agent, know what you're buying... because the agent already decided for you.