Distribution vs. Frontier: The Central Tension in Chinese AI

The Chinese AI economy is defined by a fundamental question: Does platform distribution or frontier capability determine market outcomes?

The Distribution Thesis

Big Tech platforms leverage super-app ecosystems with 1B+ daily touchpoints, zero-margin inference economics, and data flywheel effects. Distribution wins under three conditions:

• AI is a feature, not a product—embedded in existing workflows
• Output quality is hard to evaluate—subjective chat experience makes differentiation difficult
• Switching costs are high—data, habits, and integrations create lock-in

Primary Evidence for Distribution

Doubao recaptured market leadership. ByteDance’s AI app reached 157M MAU, with ~40% of DeepSeek churners returning. The “app factory” distribution muscle absorbed the shock.

Ernie hit 200M MAU through ecosystem embedding. Baidu achieved this by integrating into search, then connecting with JD.com, Meituan, and Trip.com.

Super-app integration economics favor incumbents. Douyin: +11% session time with AI features. Taobao: 22% increase in merchant conversion. WeChat Work: +9% productivity gains.

Zero-margin inference is sustainable for platforms. Chinese providers offer inference at $0.20-0.40 per million tokens vs. $5-15 for US providers.

The Frontier Thesis

Efficiency breakthroughs create periodic disruption windows. Distribution weakens under three conditions:

• AI becomes the primary interface—agentic commerce replaces app navigation
• Output quality is measurable—binary success/failure on task completion
• Interoperability increases—model-agnostic platforms reduce switching costs

Counter-Evidence

DeepSeek’s January shock proved disruption is possible. Within weeks, it achieved global phenomenon status. R1 trained for ~$6M versus GPT — as explored in the intelligence factory race between AI labs — -4’s $100M.

Model-agnostic hedging reveals platform vulnerability. Baidu’s Ernie now lets users choose DeepSeek models. Tencent’s Yuanbao integrates DeepSeek R1.

Industry-wide price cuts followed immediately. Commoditization pressure on the model layer is real.

Temporal Framework: Three Phases

Phase 1 — Chat Era (2023-2024): Distribution wins. Evaluation is subjective. Habit formation dominates.

Phase 2 — Transition (2025): Contested. Agentic features emerging. Platforms hedging with model-agnostic strategies.

Phase 3 — Agentic Era (2026+): Frontier may resurge. Task completion reliability becomes the primary metric. Switching costs drop with interoperability.

This is part of a comprehensive analysis. Read the full analysis on The Business Engineer.