For the last few years, the AI conversation in the US was mostly about one question - who has the smartest model. The assumption was that American labs would stay far enough ahead that nothing else mattered.
That assumption is quietly breaking.
Chinese-built AI models are gaining real traction among US companies. They have narrowed the performance gap with leading American rivals to the point where, for a large share of everyday tasks, the difference no longer shows up in the output. What does show up is the bill. These models remain significantly cheaper to run.
Most of them are open source or open weight, and that matters more than it sounds. An open weight model is not something you rent through an API with a metered price attached to every token. You can host it yourself, route to it selectively, and control exactly where and how it runs. As companies deploy AI to build new products and squeeze out internal efficiencies, engineers are increasingly reaching for these cheaper options - and the most capable of them are made by Chinese companies.
This is happening at an awkward moment politically. The US administration is increasingly looking to regulate its most powerful AI models, and at the same time trying to figure out how to slow the rapid adoption of alternatives coming from overseas. The two goals pull against each other. The tighter the leash on domestic models, the more attractive the cheaper foreign ones become.
The clearest signal comes from Coinbase. The company reportedly cut its AI spend roughly in half by routing tasks to lower-cost Chinese models based on task complexity. Instead of sending every request to the most expensive frontier model, it built automated routing that picks a model based on how hard the task actually is, what each option costs, and how well it caches. The result - AI spending cut in half, even as total token usage went up. They are doing more work for less money.
That last detail is the whole story. The interesting move is not "switch to a cheaper model." It is stop paying frontier prices for tasks that don't need frontier intelligence. Every request quietly asks the same question - is this worth the premium path, or will the cheap and fast one do just as well?
If you are building AI search, this is the tradeoff that will define your cost structure. Search is not one task. A query might need a quick classification, a rerank, a summary, or a full reasoning pass over a pile of documents - and those steps have wildly different difficulty. Sending all of them to the most expensive model is the easiest way to build a product that works in a demo and bankrupts you at scale.
Why It Matters:
The Coinbase pattern is the answer, and it is becoming standard practice. Match the model to the task. Use cheap open weight models for the bulk of the pipeline, save the expensive ones for the moments that genuinely need them, and lean on caching wherever the same context gets processed again. The performance gap has closed enough that this no longer means shipping a worse product. It just means shipping a cheaper one.
The AI race everyone talks about is still about intelligence. The race that actually decides whether your product survives is about cost - and right now, cheaper Chinese models are reshaping what that number looks like.