Unique Research · Ranking / industry deep dive · May 17, 2023 · Full rewrite — not a stub
Hottest 2023 VC theme: AI. OpenAI/Google killer apps → Baidu et al. chasing Chinese foundation models; Wang Huiwen, Wang Xiaochuan, NLP veterans enter. By May 16, 255 AI firms/products in “Generative AI Circle” mini-program — more LLM startups pouring in. Where is the real opportunity?
Unique Research’s months of founder/expert/investor interviews: first “race ticket” goes to firms with foresight + self-trained model engineering + vertical scene commercialization. Deep interviews with four such companies:
- BeYou Tech (必优) — AICGC (AI-Control Generate Content); 30+ vertical content SaaS + one-stop vertical model build (train/API/low-code).
- Langboat (澜舟) — cognitive intelligence; Mencius pretrained model → search/gen/translate/dialog engines + vertical apps.
- Metaso (秘塔) — automate repeat knowledge work; ~10M users; Beijing+Chengdu R&D; text AIGC, doc assist, legal.
- Ruiqi Tech (睿企) — vertical multimodal LMs + cognitive decision brain; gov/legal/finance/edu/marketing; compliance review & case-analysis bots.
02 Why self-train?
Definition: Transformer-based (Vaswani 2017) pretrained with self-supervision — BERT/GPT/T5 class; industry self-trains tend to be privatizable / openable.
vs API apps: Closer to base capability → more exposed as foundation models race ahead. Jasper AI cautionary: early GPT-3 API fine-tune marketing winner (>100k users, $1.5B val in 18 months); ChatGPT ate mid/small customer base functions for free → pricing pressure; rumored secondary sales talk. Self-train is harder/costlier (months–years) — BeYou’s SLLM (CTG niche) took 14 months — but commercial closed-loop beats alchemy for survival.
Langboat founder Zhou Ming maps routes like AV levels: L1 self-train general LM (GPT-4 class); L2 industry LM on sector data; L3 task fine-tune/prompt on industry/LLM. Most startups at L1; few at L2. Langboat (Sinovation / Sequoia China etc.) self-trains Mencius L1→L3 with commercial loop — finance clients need privatizable, customer-data-trained models; big-tech LLMs won’t custom fine-tune. BeYou CEO Zhou Ze’an: controllability — general LM hits 60 fast; self-train takes core scenes 60→90. Ruiqi chair Yu Wei: foundation models ≈ good high-school/undergrad; production needs professional private data — Ruiqi trains industry NLU/NLG since 2018.
03 Industry models = clearest startup opportunity
Cost: GPT-3 175B ≈ 10k+ A100s, ~$4.6M machine-hours / ~$12M spend. Zhou Ming: 10–20 people + 500–1,000 GPUs ≈ ≥¥50M/yr for ~10B-param model; 100B-class ≈ ¥200–300M. Yu Wei: only MSFT/Google-class keep funding general LMs. ≥30 China LLM teams (Baidu, MiniMax, Zhipu…) — not most founders’ game. Industry models fine-tune open/API bases; smaller params → $100ks not $Ms per train → multiple domain iterations + affordable customer pricing. Metaso COO Wang Yiwei “impossible triangle”: cost × diversity × trustworthiness — even OpenAI can’t max all three. Customers often want cheap-and-fit, not 100B vanity.
Moat = scene + data: “General LM is omnipotent — an illusion” (Zhou Ming). C-end mass/personalize favors general; high-stakes B (finance/legal/healthcare) needs private high-value data + workflow. RLHF quality feedback from vertical users lifts depth. Ruiqi: public security / prosecution / courts — heavy document load, high compliance, big budgets + quality data → multimodal vertical brains (compliance review, case intake bots). Metaso: legal can’t use raw LLM (hallucination); choose high-pay, high-trust niches; single-task text-to-text formal style. BeYou: WCCG + RFKL paradigm; ~90 scene models tried → focus HR hiring + Office; data flywheel from high-value feedback.
Coexist with LLM: China will have domestic foundation models (big tech); C-end squeezed by traffic/scale; B-end large buyers prefer custom private deploy. Metaso: most C-end text-AIGC users (~10M); thousands of law firms/legal depts. Ruiqi: challenge is business+data+model fusion + becoming category leader.
Last mile of China industrial digitalization: Zhang Yong — remake every industry with LMs. US OpenAI path = capital + compute + talent “force miracle”; China weights application depth (Lu Qi). 13th FYP spent trillions on public-security AI → AI listings; 14th FYP LM spend narrows US gap. China still early digital transformation vs US since 1990s — vast traditional industries needing AI efficiency. Info→digital→intelligent arc; industry models in key scenes can force data sedimentation elsewhere and accelerate the whole wave.
Charts/images in CN; attach on ingest. Full rewrite of interviews + thesis.
— END —