The Vendor Lock-in Is in the Vectors

The decision seemed obvious. In a conference room with a whiteboard full of cost-per-token calculations, the engineering lead makes the case: switch the company’s AI chatbot from OpenAI’s latest model to a new, cheaper alternative from a competitor. The API calls look similar. The benchmarks are promising. The CFO will be thrilled. Then the principal architect, quiet until now, clears her throat. “What’s our plan for the embeddings?” The room stops. The projected six-figure savings evaporate.

This is the conversation happening inside product teams right now. It’s a technical trap that was set months ago, and it’s just starting to spring.

To make an AI understand your company’s private data—your support tickets, your legal documents, your product specs—you first have to convert that data into a format the machine can process. You run it through an embedding model, which spits out long strings of numbers called vectors. These vectors are mathematical representations of the original content. Your entire corporate knowledge base, once a collection of human-readable documents, becomes a vast numerical library. When a user asks a question, your application converts that question into a vector and searches the library for the closest numerical matches. This is how retrieval-augmented generation, or RAG, works. It’s the engine behind almost every “Chat with your data” feature on the market.

Here is the problem nobody talked about in the rush to ship: those vectors are not portable.

Embeddings generated by OpenAI’s models are tuned to the specific mathematical universe of OpenAI’s models. The vectors created by Cohere live in a different universe. The ones from an open-source model like Sentence-BERT inhabit another still. They are not interchangeable. Switching your primary large language model provider isn’t just a matter of changing an API endpoint. It means you have to throw away your entire vector library and re-compute it from scratch.

For a small startup, that might be an afternoon’s work. For an enterprise with terabytes of indexed documents, it’s a disaster. It means millions of new API calls to the new provider’s embedding model. It means weeks of processing time. It means a costly, complex, and risky data migration, all while your product’s core feature is frozen. You are locked in, not by a contract, but by the math.

We have seen this script before. It’s the ghost of Oracle’s proprietary database formats from the 1990s, or the subtle incompatibility of AWS and Google Cloud APIs in the 2010s. The deepest moats are not built with features; they are dug with data formats. By encouraging the world to embed its proprietary knowledge using their proprietary models, the major AI labs have quietly turned every company’s most valuable asset into an anchor.

The freedom to choose the best model for the job—the cheapest, the fastest, the most accurate—is the central promise of this competitive AI market. But that promise is an illusion if your data is chained to the first provider you chose. The most important architectural decision was never which chat model to use for the final answer. It was which model you used to turn your library into numbers. That choice was made, often casually, by a single developer months ago. The bill is only now coming due.