Mozilla CTO Raffi Krikorian quietly plugged a Chinese LLM into Firefox last month. The model, MiniMax’s GLM-4-0520, costs $0.50 per million tokens versus OpenAI’s $10 for GPT-4o. It joins a wave of Chinese open weights models like Qwen2.5 and DeepSeek-V2 that match or beat US peers on benchmarks while skirting US restrictions. Beijing backed this play: the $1.4B Big Model Fund bankrolled MiniMax and others, and the China AI Industry Alliance now lists 239 models, 78 of them open source.

Price is the new feature. At one twentieth the cost, Chinese models turn AI from a luxury good to a commodity, eroding the pricing power of US incumbents. Open weights let downstream developers fine-tune without permission, accelerating a race to the bottom on margins. Meanwhile, Washington’s export controls on advanced chips only push Beijing to optimize for efficiency, not performance. The result is a classic innovator’s dilemma: US firms chase frontier performance while Chinese rivals undercut them on cost and openness.

MiniMax, DeepSeek, and Alibaba win as their models get adopted by cost-sensitive US startups. Mozilla, Hugging Face, and Perplexity win by arbitraging the price gap. Nvidia wins on hardware either way. OpenAI, Anthropic, and Google lose on pricing power. The US loses on influence as global developers default to Chinese stacks.

Watch for a US response: tariffs on inference APIs or a CFIUS review of cloud providers hosting Chinese models. Expect a bifurcated market where US enterprises pay for compliance and everyone else uses Chinese open models. Next benchmark milestone: a Chinese model beats GPT-4o on Arena Elo while costing 5x less.