China's AI Startup Drought | Generated by AI
Question: Why do many people in China now discourage building startups, especially AI education, AI courses, or AI agent products?
Answer:
Several converging forces make this a genuinely hard environment right now, not just pessimism.
1. The Price War Has Made Margins Near Zero
DeepSeek-V3 dropped API costs by ~90% in early 2024. Every major Chinese cloud — Alibaba, Tencent, Baidu, ByteDance — entered a subsidy war. API costs approach free. This is structurally brutal for anyone building AI products on top: your core input cost is weaponized against you by companies with infinite balance sheets. An AI agent product that was viable at $0.01/1K tokens becomes commoditized when the platform charges $0.0005.
2. AI Education / Courses Is Saturated and Commoditized
The moment ChatGPT hit, thousands of “AI course” businesses launched in China. The KOL-driven training market (知识付费) peaked ~2020-2022. Now:
- Free content on Bilibili, Zhihu, and official model docs is better than most paid courses
- Companies like Kimi, DeepSeek, Moonshot publish their own educational content
- Buyers are burned out — they paid for 5 AI courses and got generic slides
- The actual learners who need depth (engineers like you) don’t buy courses, they read papers and build
3. The Giant Threat Is Real and Fast
ByteDance, Alibaba, Tencent, Baidu, Huawei are all doing:
- AI agents (Doubao, Tongyi Qianwen agents)
- AI education (Xueersi/TAL has unlimited capital + distribution)
- AI tools bundled free into existing products
A startup building an AI agent product faces a timeline problem: by the time you have PMF, a giant has shipped a free version bundled into WeChat/DingTalk/Feishu.
4. Funding Market Has Contracted Sharply
Post-2022 China VC is structurally different:
- US-China decoupling scared institutional LPs out of cross-border funds
- RMB funds face stricter exit requirements (A-share IPO window tightened)
- “AI” as a category got overhyped and many 2023 cohort investments are underwater
- Angels and seed funds are being much more selective — they want defensibility, not demos
Many VCs now explicitly say: “We won’t fund consumer AI apps or AI education.”
5. The Regulatory Overhang
China’s Generative AI Regulation (生成式人工智能服务管理暂行办法) went into effect Aug 2023. Any public-facing AI product needs:
- Filing with MIIT/CAC
- Content moderation compliance
- Occasional surprise audits
For a small team, this compliance overhead is disproportionately painful. Large companies have legal departments for this; a 3-person startup does not.
6. The “AI Group” Business Is Especially Fragile
Building a paid AI community (知识星球, WeChat group subscription) around AI tools/education specifically has these problems:
- Half-life is ~6 months — members churn when they feel they’ve “learned enough” or the tool changes
- You’re selling access to information that’s free elsewhere
- No compounding moat — every new ChatGPT/Claude release disrupts your content
The Honest Assessment for Your Situation
The people warning you are not wrong about the market conditions. But the framing matters:
| Bad bet | Better bet |
|---|---|
| Generic “AI course” | Niche technical training with real depth (your notes + GPU experience is a real asset) |
| “AI agent product” for consumers | AI agent tooling for enterprise/bank workflows (your bank AI engineering background is defensible) |
| AI education community | Consulting/contracting → productize later once you have paying clients |
| B2C in China | B2B or international (your English + technical depth opens non-China markets) |
Your specific background — bank AI engineering, model training, CLI agent building, 1B token consumption — is actually well-suited for B2B AI tooling or international developer tools, not domestic consumer AI education. The people warning you are seeing the commodity layer; you could potentially build above it.
The real question isn’t “should I build a startup” but “which layer of the stack has defensibility that a giant can’t trivially replicate or subsidize away.”