The LLM Competition Is Won at the Data Layer
The competition between large language models is usually told as a rivalry between two camps. On one side, the US makers, whose original models have become the frontier: the most capable, and the most expensive. On the other, the Chinese models, which the US makers accuse of distillation, training cheaper systems on the output of the frontier ones.
Both accounts are about the model. That is where the attention goes. It is also where the attention is least useful.
My own path ran through both. I started building with the frontier models, the way most engineers did. Recently I moved much of my work to the Chinese models. Not only for the cost. For the design decisions they pushed me toward.
The price gap is real
Search the topic and the headlines agree on one point: the Chinese models are taking share in software development, mostly on price. That part is accurate.
Here is what it looks like in practice. Ask a frontier model to build a CRM in a single prompt and the run can cost over US$20. Ask a Chinese model to do the same and it can cost under US$1. The frontier model is more likely to get it right on the first attempt. The cheaper model usually needs a few rounds of iteration to reach the same place. Even so, the price difference is large enough to pull a lot of engineers across, and it has.
So the incentive is real. But price is not what decides the quality of what you build.
Design decides more than price
Set vibe coding aside for a moment. That is exploration: one prompt in, one output out, and there the model does most of the work.
Building an actual AI application is a different story. If you construct a good data layer to support the prompt, then instruct the model to analyze and respond against that data, a less advanced model can produce results comparable to a frontier one. The model is one component. The data layer, the retrieval strategy, and the synthesis framework are the parts that shape the output.
The price comparison hides this. Two engineers using the same cheap model will not get the same result. The one who built a design specification, a logical workflow, and a data foundation beneath the prompt more reliably, and can explain why. The one who only picked the model cannot.
Less from the model, more from your evidence
There is a deeper reason to care about the layer beneath the model, and it is what separates building an application from using a chatbot.
A consumer chatbot relies on parametric knowledge: the information built into the model during training. You ask, and it answers from what it absorbed. That is fine for general questions. It is a liability for work that has to be correct, because that knowledge can be outdated, generic, or wrong, and you cannot see which.
When you build an AI application, you want the opposite. Less reliance on what the model memorized. More output grounded in evidence you supply from your own data. The model stops being the source of the answer and becomes the reasoning engine that works over your material. Handled this way, the raw capability of the model matters less, because you are no longer asking it to know things. You are asking it to reason over things you already trust.
That is why a smaller model on a strong data layer can match a frontier model running on its own memory. The intelligence moved. It went from inside the model to the structure around it.
The harder audience is the end user
Explaining this to a developer is one task. Explaining it to the person who used to search on Google is harder.
That user opens a chatbot and asks a question the way they used to type it into a search box. They have no reason to know what separates an evidence-based AI application from a chatbot, or what parametric data is versus a dataset supplied at the point of use. What they see is a clean answer, delivered fast. From there the conclusion is easy: the chatbot is more advanced than Google, and Google is old-fashioned.
Look at what each system actually does, and that conclusion inverts.
Google builds a model of relevance before it answers you. It crawls and indexes the web, ranks pages against each other with a system that began with PageRank and now includes a knowledge graph of how entities relate, and then shows you the sources and why one page sits above another. You can inspect the reasoning. The work is visible.
A chatbot answers first from its pre-trained data. When it does reach the live web, the method is often opaque, and in the weaker case it does not build a model of relevance at all. It relies on the one Google already built. This is the part worth emphasizing. Reporting has shown that several leading AI systems lean on search engine results to ground their answers, at times reaching them by scraping the results page through third-party services. The chatbot is not discovering relevance. It is borrowing it, then presenting a synthesized answer with the sourcing out of view. Is this not the same objection the frontier makers raise against distillation, one system taking the value of another's foundational work without reproducing it?
So the user rewards the fluent answer and penalizes the system that showed its evidence. That is a perception bias, not a judgment of quality. And it sits above the model competition without being touched by it. Frontier or Chinese, US maker or Chinese maker, none of it changes what the user adapts to. They adapt to the interface and the instant answer, not to what was built underneath.
Openness decides reach
There is a second reason to look at the Chinese models, and it has nothing to do with price.
The two sides have taken opposite positions on access. The US frontier labs keep their strongest models closed, and access itself has become a policy lever. Most leading Chinese models run the other way. They are open-weight: free to download, adapt, and run on your own servers, in any region.
For a product built to reach the world, that difference outweighs a benchmark. An open-weight model can be deployed on any cloud, in any country, or on your own hardware, without depending on one vendor who can be cut off by a single government decision. A closed frontier model gives you the vendor’s reach, and the vendor’s limits. An open one gives you your own. It runs anywhere. For a global build, that is the part you can plan on.
The competition worth watching
The model makers will keep overtaking each other. A cheaper model closes the gap, a frontier model opens a new one, and the benchmark tables keep changing every month. That contest tells you little about what your own system will produce, and even less about whether an answer in front of you can be trusted.
The competition that decides quality is not between the models. For the developer, it is between a system with a real data layer and a system that only wraps a model. For the user, it is between an answer that shows its evidence and one that hides it. The same distinction, seen from two places.
Pick the cheaper model if the economics favor it. Most of the time now, they do. Just do not confuse that decision with the one that matters. Everyone competes on the model. Fewer build the foundation underneath it, and that is the part that separates the results, whether you are building a system or only reading what it returns.
