Unique Research · Insight · October 18, 2023 · Chaos Innovation Institute AI series live #6 · Partner Hu Xiaojing (Unique Capital)
“In H1 2023, China’s generative-AI financing totaled RMB 7.399 billion. Silicon Valley’s genAI primary market saw 42 rounds totaling $14 billion — China’s figure is about 7% of that.”
“China’s genAI application stack shows distinctive traits.”
“Take ChatPPT — one-click PPT generation. Paid users already exceed 10,000 with million-RMB product revenue — rare in today’s AI market.” — Hu Xiaojing, Partner, Unique Capital.
Is every industry about to be remade by AI, or is the market early with more foam than substance? Overestimated or underestimated? Subjective opinions abound; what’s needed is data.
H1 2023: China vs U.S. genAI financing
China’s H1 genAI total: RMB 7.399B. June alone showed RMB 5.23B in disclosed deals — a steep chart peak that mostly reflects disclosure timing, not a June explosion. Unique Capital estimates ~30% of deals (including some LLM and strategic rounds) still undisclosed, so 7.399B is conservative. Full-year 2023 China genAI financing could exceed RMB 15B on an optimistic Unique Capital view.
2021 was a prior China AI heat (the “Four Little Dragons” listed). 2023’s genAI wave is accelerating within AI. Among Top-10 AI financings, MiniMax (Mingzhi Dream) stood out as the highest-funded non-giant LLM startup (with Tencent Investment). Active strategic/corporate players: Tencent Investment, Kunlun Tech, Meituan, Ant — a larger corporate-VC share than classic independent VC.
U.S. Silicon Valley H1 genAI: 42 rounds, $14B (~RMB 100.6B). China’s 7.399B ≈ 7%. Read-through:
- China’s foundation layer (chips, infra) is earlier; the U.S. foundation layer is mature.
- China raises across foundation + model + application at once; the U.S. (even aside from OpenAI) spends less on raw foundation build and more on foundation LLMs, vertical apps, and middle-layer tools (vector DBs, AI toolchains, deployment).
- Abroad, relatively mature genAI business models already exist; China is still largely in validation. Near-term job: get infra to a solid “60-point” baseline, then raise precision — then form China-specific application patterns.
Five traits of China’s genAI application ecology
- Colder capital, higher bar. RMB 7.399B looks muted vs U.S. heat. Domestic LPs demand high usability and “10× better”; Silicon Valley often funds R&D/PoC stages more readily.
- Labor-cost structure. Average China labor is cheaper, but some skilled domains (e.g. legal) are expensive — B2B efficiency pressure is intense; software must clear a higher “10×” bar (Silicon Valley often accepts ~5×).
- Weaker open community culture vs Silicon Valley open AI / open source. This wave favors experienced founders (mobile-internet, NLP, AI-native) and industry incumbents with customer scenes and data.
- Richer scene and commercial data from China’s internet vertical penetration — ability to reach “60-point” LLMs fast, then fine-fill for moats.
- Industrial structure. Internet/high-tech are not absolute GDP dominants; smart manufacturing still leads. Align with 14th Five-Year Plan and generative-AI regulation: AI’s China mission is empowering the real economy — different from Silicon Valley’s new-tech-first lens.
Metaphor: raising a child poor vs rich — China is in the “raise lean” phase: identify problems, ship solutions, self-fund quickly, and attach to industry. China’s AI industry scale reached ~RMB 195.8B by 2022 (purple zones on Unique Capital’s map = validated scale apps: internet marketing, risk/security, scheduling, robotics, dialogue). CAST chair Wan Gang’s line — empower the real economy, support social development — matches that judgment.
Case studies: Chinese genAI products that already work
ChatPPT (consumer): one-click decks + deep document layout; >10k paid users; million-RMB revenue; ~20 employees.
Novel Monkey (consumer): web-novel model trained on massive fiction; one-click ~100k-character novels; ~$100k/month revenue; if views hit a threshold, humans edit plot into bestsellers; only 3 employees — tech leveling for tiny teams.
Kuaizi Tech (B2B video): splits commercial video into workflow modules (edit/shoot/cut/post/manage). Example: MINISO store managers need frequent geo-local Douyin/Xiaohongshu posts — AI mashups, voiceover, effects, and recommendation rules cut cost and lift acquisition. Post-ChatGPT awareness, customer and revenue growth accelerated (>50% MoM cited).
Lanma Tech (B2B AI Agents): agents as digital employees — autonomous decisions and resource orchestration vs traditional SaaS. Sales leads can ask in natural language for charts from internal sales/SKU data. Lanma is already working with large enterprises. Takeaway for industry buyers: prepare data storage and knowledge bases so AI products can call them later.
Market skew: more B2B than consumer cases in financing — partly fear that stronger foundation models erode consumer moats; partly limited LP dry powder preferring faster-payback B2B over capital-heavy consumer growth.
Advice for industry + AI companies — and the over/under question
Hu: over- vs under-estimated is less important than case-by-case analysis. Two factors to judge whether a tech wave is over/under-hyped:
- Tech + scene stability. Mobile internet’s stable phone scene spawned App Store empires (Meituan local life in months; Pinduoduo on WeChat to IPO in ~3 years). We’re in AI’s “decade.” Shift from GUI to language/dialogue interface (LA) is real — but dialogue-native apps have not yet mass-adopted; users still mostly search. Application ecosystems typically need 3–8 years once tech is stable and scenes are widespread.
- Clear business model + value network. Foundation/open models + application layer with clear supply–demand links work. Fuzzy zones: simple dialogue tasks many assume foundation models will eventually swallow. Baidu’s Wenxin Fund (RMB 1B) to back Wenxin-based builders is ecosystem strategy — foundation models must define what they won’t do, stay restrained, and light others’ value networks (Chaos’s philosophy).