ChinAI #349: Tokens Made in China?

2022-03-07 · Source: ChinAI Newsletter · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Public Policy & Governance · Depth: Intermediate, quick

Summary

On February 24, 2026, OpenRouter data revealed that Chinese models MiniMax M2.5, Moonshot AI's Kimi K2.5, and Zhipu GLM-5 were the top three most popular models on its API aggregation platform, favored by individual developers and AI start-ups. This popularity stems from their superior price-to-performance ratio; for example, MiniMax M2.5 achieved 80.2% on a software engineering task compared to Claude Opus 4.6's 80.8%, but cost only $0.30 per million tokens versus Claude's $5. This cost advantage is attributed to China's 40% lower electricity prices and algorithmic efficiencies like DeepSeek V3's MoE architecture, which reduces inference costs by approximately 36 times compared to GPT-4o. The article highlights how this trend, particularly in "agentic flows run by U.S. firms," represents an "invisible" export of Chinese computing power and electricity, challenging the narrative of U.S.-China tech decoupling.

Key takeaway

For AI developers and start-ups prioritizing cost-efficiency, you should evaluate Chinese large language models like MiniMax M2.5 and Kimi K2.5, especially for agentic workflows. Their significantly lower token costs, driven by cheaper electricity and algorithmic advantages, offer a compelling alternative to more expensive Western models. However, be mindful of potential data residency and compliance issues for sensitive corporate data, as API requests are processed through Chinese data centers.

Key insights

Chinese AI models are gaining traction among U.S. developers due to their compelling price-to-performance ratio.

Principles

• Cost sensitivity drives AI model adoption.
• Algorithmic efficiency reduces inference costs.
• Tokens bypass traditional trade barriers.

Method

Chinese models achieve cost efficiency through lower electricity prices and advanced architectures like Mixture-of-Experts (MoE) that activate only a subset of parameters during inference.

In practice

• Consider Chinese models for cost-sensitive AI agent development.
• Evaluate models based on price-to-performance benchmarks.
• Explore MoE architectures for inference cost reduction.

Topics

• Chinese AI Models
• OpenRouter Platform
• Price-Performance Ratio
• AI Agent Workflows
• US-China Tech Relations

Best for: Machine Learning Engineer, NLP Engineer, CTO, AI Engineer, AI Researcher, Policy Maker

Related on AIssential

• What’s next for Chinese open-source AI — Chinese open-source AI models offer near-frontier capabilities at a fraction of the cost, rapidly reshaping global AI innovation.
• Intelligence Per Dollar — AI model evaluation is shifting from raw performance to cost-efficiency, prioritizing "intelligence per dollar."
• TAI #185: China's Open-Weight Holiday Blitz; GLM 4.7, Minimax M2.1 & MAI-UI — Open-weight models are closing the intelligence gap with proprietary models but show a widening safety gap in reliability.
• When Americans choose Chinese AI — Cost-effective Chinese AI models are gaining traction among U.S. developers, challenging established American providers despite geopolitical concerns.
• How are Chinese models so strong with so little investment? — Chinese AI models achieve high performance through distillation from Western models and a robust national AI ecosystem.
• ChinAI #350: Around the Horn (24th episode) — Chinese open-source AI models are often processed via US-based data centers, challenging assumptions about data sovereignty and developer lock-in.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ChinAI Newsletter.