Odd Lots: Why One AI Startup Runs Production on Chinese Open-Weight Models

As seen on the 24/7 Wall St. homepage on November 20, 2026.

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.

Continue ReadingShow less

Daniel Saeedi, co-founder of Minerva, lays out a bifurcated model strategy that is quietly becoming a template across the AI industry: build and experiment on frontier closed models from US labs, then push production inference to cheaper open-weight Chinese alternatives like Qwen. The economic logic is hard to argue with when fine-tuning costs have collapsed to roughly $90 for some training runs.

The price gap at inference time runs in the same direction. Abliterated DeepSeek models, which Saeedi flags as a security risk, can run at about one-hundredth the cost of frontier pricing. That spread is what makes the migration attractive to cost-conscious engineering teams, even as it raises real concerns about data sovereignty and model provenance that Saeedi discusses in the episode.

Sponsored

Compute demand is the part that matters for Nvidia investors. Saeedi's math posits roughly one GPU per AI-native knowledge worker, set against the 100 million knowledge workers in the US alone, and the argument is that the world remains acutely short on compute regardless of which model vendor wins the inference race. That framing cuts directly against the bubble narrative that has pressured the stock.

The wildcard hanging over the entire dynamic is a proposed Trump administration rule that would require US companies to publicly disclose their use of Chinese AI models. If that disclosure requirement becomes law, it would force a reckoning for any business quietly routing production traffic through Qwen or DeepSeek, and it would reprice the competitive equation for US frontier labs that have struggled to hold the premium pricing line against cheaper open-weight alternatives.

Mentioned: NVDA, UBER, COIN