China’s open-weights AI strategy is winning

TL;DR Highlight

Models have no moat, so open Chinese AI is argued to be beating closed US AI.

Who Should Read

A backend developer deciding whether to stay locked to one LLM API vendor or switch to Chinese open-weights models

Core Mechanics

Models alone have little moat
- The author says models have only brand loyalty and shallow switching costs.
- The moat sits in contracts, enterprise integrations and enterprise-friendly features.
API models are easy to swap
- In engineering you can swap the API and keep the same prompt.
- So picking the best model and vendor and moving when another gets better makes sense.
China turned a weakness into distribution
- US GPU export controls and data rules keep Chinese firms from running global centralized services.
- Open release turns that into a distribution advantage and commoditizes the layer where US firms earn money.
Open weights are portable though not open source
- You can host them where you want and modify them.
- Being permissionless makes them a natural center for innovation.
The performance gap is closing

Community Reactions

Doubt about the 80% startup figure
- A commenter interviewed at several startups and all used US models.
- Their core usage was Claude and Codex subscriptions.
Self-hosting a Chinese model cut costs, per one commenter
- Their general cluster runs the most accurate and robust Chinese model.
- Cost per token was about 1/20 of Anthropic and 1/15 of OpenAI in their testing.
- US/EU models scored a few points lower on accuracy and are offered to customers who require them, for an extra fee.
Counterpoint that enterprises care about data retention, not openness
- Commenters say enterprises care about zero data retention and staying with their vendor.
- Another says there is not yet enough evidence that the US is losing or that closed weights are the reason.
Doubts that Chinese labs will keep releasing weights
- Training costs tens of millions, so free release is hard to justify, commenters say.
- Others ask what the VC-funded labs' business model and exit are.

How to Apply

Run your prompts on another model
- Pick 5 core prompts and run the same inputs on a Chinese open-weights model and your current one.
- Compare accuracy, latency and cost per token in a table.
Put a model adapter layer in your LLM calls
- Hide vendor SDK calls behind one interface.
- Move model names and endpoints into config to cut swap time.
Route models by customer requirements
- Classify customers with data region or vendor restrictions first.
- Send unrestricted traffic to cheaper models and restricted traffic to US/EU models.

Try It

copy & run

# Paste into Claude or ChatGPT as is (a simple prompt to check the article's claim)

Our service uses an LLM API for {{use case}}. The current model is {{current model}}.

Build a checklist for how swappable our model is, using these criteria.
1. Parts where we can swap the API without changing prompts
2. Parts with vendor lock-in (proprietary features, contracts, integrations)
3. Traffic that can't change models due to data region or retention rules
4. Five evaluation criteria for comparing accuracy and cost across models

Next step: ask it to apply the criteria to {{20 sample questions}} and summarize in a table.

Terminology

open-weightsReleasing the trained model weight files. It's like handing over the finished dish, unlike open source, which also gives the recipe and ingredients (training data and code).

moatA barrier that keeps competitors from easily taking your customers. Like the water around a castle.

export controlsGovernment rules that stop certain chips or technologies from being sold abroad. Here, US limits on GPU exports.

zero data retentionA vendor policy of not storing your input data. Like a consultant who takes no notes.