How an American AI Company Uses Chinese Open-Weight Models
- Minerva's Daniel Saeedi: build on frontier closed models, run production on cheaper open-weight Chinese models like Qwen
- Fine-tuning costs have collapsed, with some training runs landing near $90 versus Harvey's reported $6 million legal model
- Warning on 'abliterated' DeepSeek models at roughly $0.30 per million tokens, about 1/100th frontier pricing, as a security risk
- Compute math: one GPU per AI-native knowledge worker against 5-10 million deployed GPUs and 100 million US knowledge workers
If production inference keeps migrating to open-weight Chinese models, the premium pricing case for US frontier labs erodes while chip demand stays intact. Saeedi's knowledge-worker math argues the world is still short compute, which cuts against the bubble narrative pressuring Nvidia, and Washington's proposed disclosure rule on Chinese model use is the variable that reprices both sides.