01.ai, DeepSeek, and other Chinese companies are reducing costs to create AI models by focusing on smaller training data sets, as they deal with export controls
Eleanor Olcott / Financial Times : LinkedIn: Kai-Fu Lee LinkedIn: Kai-Fu Lee : Driving down inference cost is the most important thing to enabling great AI-first apps. @YI01_Yi's new model Yi-Lightning is ranked #6 (https …
Context & Ripple Effects
Earlier coverage showed Chinese AI startups pairing monetization efforts with more efficient code and smaller models after losing access to the most powerful chips; this report identifies smaller training data sets as a concrete extension of that response to constraints (the earlier push toward efficient code and smaller models).
The approach also foreshadows the later DeepSeek narrative: the lab said its DeepSeek-V3 used fewer chips for training, while subsequent coverage described a broader wave of low-cost models and services from Chinese companies (DeepSeek's lower-chip training claim; the later low-cost AI-services push).
First-order effects
- 01.ai, DeepSeek, and peers shift model-development priorities toward reducing training inputs and costs rather than relying on larger data sets and greater compute availability.
- Export controls become an engineering constraint as well as a procurement constraint, increasing the immediate value of efficient training and lower-cost inference for AI-first applications.
Second-order effects
- Lower development and serving costs can let Chinese providers price models and AI services more aggressively, pressuring domestic rivals to match efficiency rather than compete solely on model scale.
- The strategy reinforces a market in which deployment economics matter alongside benchmark performance, especially for customers deciding whether AI features can be offered at sustainable cost.
Third-order effects
- If smaller-data, compute-efficient models remain competitive, restricted access to leading hardware could shift durable advantage toward data efficiency, software optimization, and inference economics rather than raw training scale alone.
- The pattern points to a more fragmented AI supply chain: export controls may not halt model development, but can steer it toward alternative technical paths and cost structures.
The trend: Export restrictions are accelerating an efficiency-first AI model race in which the cost of training and serving useful models becomes a core competitive variable.