🧠 Food for thought

Implications, context, and why it matters.

Subsidies and token pricing drive China’s AI agent growth

• Model pricing sets the pace. Anthropic’s Claude Opus 4.6 charges US$25 per million output tokens, while MiniMax’s M2.5 costs US$1.10, up to about a 22 times gap 1.
• That gap matters for AI agents such as OpenClaw, an autonomous program that can perform tasks on a user’s computer and access deep system permissions. Running it on pricier models can rack up hundreds of dollars in daily API fees 2.
• Local governments add direct support. Wuxi offers subsidies up to 5 million yuan (~$720,000) for major breakthroughs tied to robotics or embodied AI (AI systems designed to act in the physical world through machines like robots) 3.
• Other perks reduce overhead. Shenzhen’s Longgang district offers up to two months of free accommodation and up to 18 months of preferential office space for newly registered or relocated one-person startups 3.

China exports low-cost tokens with tradeoffs

• Cheaper Chinese models are shaping global usage. OpenRouter data found Chinese models made up 61% of token consumption among the platform’s top 10 models, and they surpassed U.S. models during a referenced February week 1.
• Adoption reaches beyond China. A European studio runs 80% of routine reasoning on Kimi K2.5, then uses expensive models like Claude for the hardest 20% of work 1.
• Cost edges tie back to operations. China benefits from an integrated supply chain plus industrial electricity prices that run about 40% lower than in the US 1.
• Some analysts treat tokens as ‘digital electricity’, which sends the value of power and compute across borders through API calls 1.