---
title: "Unique Viewpoint | 2024 China AI Startup Investing: Opportunities & Challenges"
author: "Unique Research"
sourcePublication: "Unique Research (FFCAP English library)"
originalPublishedAt: "2024-04-23T16:30:00+08:00"
canonical: "https://ffcap.cn/en/research/20240423a06il400"
source: "https://ffcap.cn/en/research/20240423a06il400"
language: "en"
cnSourceUrl: "https://view.inews.qq.com/a/20240423A06IL400"
---

# Unique Viewpoint | 2024 China AI Startup Investing: Opportunities & Challenges

_Unique Research · Unique Viewpoint · April 23, 2024_

Background: On **January 24, 2024**, Unique Capital and partners hosted the **2024 AIGC Application Development Summit** in Shanghai, with nearly **1,000** AI practitioners. This piece is compiled from the guest roundtable.

## Guests

-   **Junkai Duan** — Managing Partner, Kaitai Capital (凯泰资本)
-   **Yutong Zhou** — Partner, GSR Ventures (金沙江创投)
-   **Guang Yang** — Founding Partner, Yaotu Capital (耀途资本)
-   **Hongyu Lu** — Director & Senior Partner, Detong Capital (德同资本)
-   **Yong Li** — Partner, Chengmei Capital (诚美资本) — moderator

## Q: Please briefly introduce your firm/fund and focus sectors, and recap 2023 investing.

**Hongyu Lu (Detong Capital):** Detong was founded in **2006** — 18 years now — crossing several cycles and many hot themes. The biggest theme of the past two years is clearly AI. We manage roughly **¥18B**; in 2024 we raised a new fund whose main direction is AI — AI infra, foundation models, industry apps, and more.

**Guang Yang (Yaotu Capital):** Yaotu is a young firm founded in **2015**. We focus on next-gen IT across software and hardware. Portfolio compute-related hardware companies saw strong growth in 2023 — not because we foresaw Generative AI, but because we laid out from a compute lens (compute chips, networking chips, etc.). After ChatGPT we also positioned accordingly.

Our hardware companies now better serve large-cluster compute and networking needs; in infra we have done training/inference optimization work; in apps we are trying new projects. On games and education ops we are still learning and hope to talk more with AGI-direction founders.

**Junkai Duan (Kaitai Capital):** Founded **2009**; we manage about **¥10B**, mainly biopharma and AI. Since **2017** we have invested in AI and big data — roughly **40** companies.

We were fortunate three or four years ago to bring **Zhenzhong Lan** onto the team — Google scientist background, first author of the Albert paper, and founder of Westlake Xinchen. In AI we have also backed firms such as DataBaker (标贝科技) and AG Intelligent.

**Yutong Zhou (GSR Ventures):** Founded **2004** among the early VC cohort; now on Fund VIII USD and Fund IV RMB — total scale about **$3.5B** plus equivalent RMB.

Early hits people know best were mobile internet — Ele.me, Qunar, Didi, Xiaohongshu — with strong returns and exits. In recent years we put more energy into AI while continuing to watch consumer go-abroad.

**Yong Li (Chengmei Capital, moderator):** Brief intro — we are a small fund managing roughly **¥3–4B**. Because we are small we seek niches: (1) most core / earlier / lower-layer tech — hardware architecture, AI accelerator cards, cloud security — foundational support around LLMs; (2) profit orientation — AI application scenes, especially MarTech content.

## Q: What is your investment strategy/logic in general-purpose model AI? How did 2023 change that (including evaluation criteria)?

**Hongyu Lu:** The past year changed enormously — from OpenAI’s rise to today’s “thousand-model war.” Internally we share a consensus: foundation models are as important as internet communications tech a decade-plus ago; people cannot live without communications. So we will invest in foundation models.

Whether OpenAI can survive in China is unknown — which itself creates major opportunity for domestic models. On compute everyone is pushing hard — chip clusters, direct chip investment, architectures, etc.

On thresholds: for foundation models we do not set a special bar; team base capability matters. On applications we mainly invest at Series A, so we want some proof of revenue and scene PMF — not just PPT stage.

Example: most companies at ~**¥1M** monthly revenue are already quite good; vertical/industry-model companies at **¥7–8M** monthly and still growing show strong momentum. The process is angst-filled, but the future is bright.

**Guang Yang:** Strategically we firmly like compute. With AGI, many companies saw explosive growth. Despite U.S. sanctions on China, beyond leaders like Huawei there is still room for many startups.

So we must solve not only compute but networking, bandwidth, and transport. GPT-3.5 needed ~5,000 cards; GPT-4 / 4.5 need tens of thousands interconnected — transmission, bandwidth, and communication problems that China must solve. That is a lower-layer hardware opportunity.

In the middle layer, foundation-model startup opportunity is already small — insufficient compute, insufficient capital on hand; new companies need large raises. More opportunity sits in AI infra and developer tools: optimize train/infer on the infra side; improve GUIs for smoother, more developer-friendly tools.

On applications, two directions:

-   **AI-enable market:** many traditional SaaS companies integrating self-built models with foundation models. They have deep experience in CS and marketing; remaining core functions matter most. Adapt to the new tech wave and they stay competitive.
-   **Native LLM markets:** if investing new projects, watch markets wholly born of foundation models — e.g. auto code generation / autocomplete. We watch this closely.

When founders pitch, I care most whether they have a path to PMF. Software PMF is hard; early investors usually weigh it heavily.

**Junkai Duan:** Economy is the expression of technology, and technology is a source of the economy. First ask: where are the boundaries and effectiveness of this foundation-model wave? Tech democratizes expertise and automates much ordinary work.

Second: how far has commercial application of foundation models gone? What can already commercialize? I think “one and a half”: the “one” is text extract / distill / summarize / generate; the “half” is progress in video generation.

Finally: find applications. A foundation model is tech, not a product. How to put it into products poses more commercial thinking and challenge for next-gen founders. Extending yesterday’s logic rarely yields innovation.

Example: smartphones introduced LBS → Meituan and local-life products; smartphone cameras enabled Douyin-like commercial innovation. Next-gen innovation must combine with hardware — metalenses, next-gen sensors, fused with new chip algorithms — to create newer products.

We watch tech and product innovation, and also business-model innovation. Why did Meituan win group-buy? Why did Douyin win short video? Only when every link in the industry chain can profit — an effective commercial model — does sustained prosperity follow.

**Yutong Zhou:** Over the past year we invested in about **10** AI-related companies across infra and apps. We also firmly believe China will have its own giant models.

OpenAI leads on general-model capability. But startups building general LLMs face both big-tech and open-source competition — commercial value is a real question. So we are relatively cautious on foundation models and more focused on vertical apps and model+app combinations.

Our portfolio splits into two: last year’s early “AI-native” startups; and existing enterprise-service companies that already have mature scenes and customers and can fold AI into products/services for step-change and a second growth curve — interesting, valuable cases.

Example: AI + e-commerce — image/video generation matching real scene demand makes customer value relatively measurable and commercialize-able. AI in medical insurance and social are also watched.

On PMF, startups should ask early: What problem? How to productize? How to land? How to commercialize? Our framework from last year to now has not changed dramatically. AI keeps iterating; we are still learning fast.

## Q: What are the challenges of AI investing in China? Where are the larger opportunities? Which fields do you like most?

**Yutong Zhou:** The biggest challenge — China or U.S. — is that it is extremely competitive. AI is a domain of large consensus: startups, big tech, and category leaders all agree they must embrace and lay out. For startups: what is the overtake path? Vs. incumbents absorbing the tech, what is stronger competitiveness? How to find non-consensus inside consensus? What opportunities will big companies not yet watch? — problems every market faces.

On “which market”: AI entrepreneurship choosing China vs. overseas is not either/or — it is a global market; choice depends on what you do. If a professional domain, you need sensitive high-quality industry data; if large accounts, China may be more attractive than high go-abroad bars; if a ToC or ToB product, you can go global from Day One. There is no pure “China-market challenge” problem.

Larger opportunities:

1.  **AI across industries** — depends on (a) base general-model capability lifts (GPT-4.5 / 5, hallucination fixes) and (b) combining industry knowledge and high-quality industry data. That is where we watch apps.
2.  **Intelligent hardware** — AI + consumer electronics rising; a domain Chinese founders are strong at — product definition, form, interaction, supply-chain integration, and China’s hardware cost/efficiency advantages. CES showed many AI products (Samsung AI phones, various robots). AI+hardware is a field of opportunity for Chinese founders.

**Hongyu Lu:** Especially at the app layer, two points: (1) Catch population dividends — who do you sell to? Over the past 20 years China caught industrial-worker dividends; over the next 20, aging / silver economy will be important. Aging is bad for society but from a product view a fast-growing elderly cohort is potent. Elders’ learning ability declines — they need AI more. Young people learn well and depend less on AI; elders need it especially for medical, companionship, and emotion.

(2) Products must cut cost and raise efficiency — for B or C you must bring benefit. Domestically most apps we see serve B more than C. OpenAI opened personal memberships, but few individuals buy LLM memberships today. So most apps remain enterprise; in enterprise you must answer cost-down / efficiency-up.

A reminder to founders: if you build here, especially B2B, do not fear. Giants stood ahead in video — then short video beat them, and now short drama rises. Keep innovating, use existing tech, catch markets and opportunities — success will follow.

**Junkai Duan:** We focus on two fields: (1) **marketing** — AI penetrates every industry; content/marketing has always been where new tech lands first; (2) **social** — next-gen social; past social was person-to-person / place-to-place; future social may go deeper into the individual.

We also think about tech evolution and application logic: (1) add AI — combine existing products with AI; Adobe, Microsoft and others already do this well globally; (2) AI creates new products — watch Microsoft Copilot and extend from there; (3) AI creates new ecosystems and models — we need to find and catch those together.

## Q: Does “AI + digital human” have opportunity?

**Junkai Duan:** Digital humans have potential, but they are a single-point tool. In MarTech many companies worked hard to form complete data assets and closed loops. Via digital humans, virtual humans, or video/image tools, they organize user data into private-domain and CDP assets, strengthen marketing, guide ads, and build stronger data assets. Without a full data-asset loop, a single tool’s value is limited.

**Yong Li (moderator):** AI’s end goal is intelligence that increasingly resembles humans — and today’s biggest problem sits right there. Marketing that brings in AIGC: real content is not always a clear “A implies B” logic chain. With emergent ability, will unexpected things appear inside the black box?

Humans’ key trait is not following logic — sometimes flashes of insight. Whether AIGC can make a qualitative breakthrough there is hard. “Cross boundaries; fuse disciplines in R&D.” I strongly suggest R&D/founding teams add more perspectives and capabilities — not only HW/SW/front/back roles, but sociology, psychology, creativity — cross-disciplinary collision that may help iteration.

**Guang Yang:** I also like apps, especially games and education — AGI can play a large role; good for young founders. From an investor view, my technical ability can hardly precisely evaluate early projects there. Early success is hard to judge; tracking the best founders is also hard. So despite huge opportunity, success rates look very low.

In infra, fewer projects are doable. We invested in an inference-optimization company. China lacks enough compute for apps. Even as Huawei Ascend and similar hardware grow, short-term supply still falls short. How to squeeze more performance from existing hardware to support more apps has many opportunities in China’s commercial environment.

## Q: Can blockchain / crypto-finance develop in sync with AI tech change?

**Guang Yang:** From my angle, blockchain or “swearWeb” (sic) more changes production relations — an incentive mechanism. Intrinsically it can complement AGI or be unrelated.

## Q: Advice for Chinese AI founders?

**Guang Yang:** Though we are based in China, look global. This is a super-large opportunity — do not limit the lens to domestic. Polish product at home; also plan to go abroad — especially important for Chinese ToB. China often has users without paying customers; overseas users and customers are both more prosperous.

**Junkai Duan:** Open-system thinking matters. Many commercial logics and models differ from the past — including R&D, design, marketing, and go-to-market. Take OpenAI: from founding through foundation models, open source, APIs — internally they likely have many ideas on how to commercialize, productize, and service-ize. That logic differs from many past software/hardware products; the commercial path differs. Hope founders give this more attention.

**Yutong Zhou:** If you have firm belief or entrepreneurial passion for this field, find ways in early. The industry iterates fast; this may be a transitional product stage — a second or third try may find the right direction. Early founders should focus when they find a direction — not try everything.

With limited people, resources, and capital, a precise entry point that fits your resources, gifts, and advantages matters. Some domains have AI opportunity — but is it the founder’s or the incumbent’s? Think carefully.

Also watch PMF and commercialization early. Though AI is hot, primary-market capital is still tight — do not rely wholly on fundraising; start making money soon. VCs also weigh long-term sustainability and competitive advantage more — think what the critical factors of durable growth are.

## Q: Views on compute carbon-reduction services (e.g. data-center energy savings)?

**Hongyu Lu:** We have invested in multiple data centers — once called IDC; today they position as intelligent compute centers. Compute as infrastructure necessarily involves efficiency — not everyone can build a data center; SMEs buying many compute cards may not fully utilize them. Governments nationwide push local compute centers — Shanghai, Beijing, Guangzhou, Shenzhen and other coastal economic hubs plan facilities like national supercomputing centers. Leasing those centers will be an economic, feasible path.

Given national conditions we also face constraints: if purchase volume is too small, NVIDIA will not even consider you — let alone other suppliers. But I believe we will solve this, even if it takes more resources and higher energy burn — chip process still has limits. Eventually we will find solutions.

## Q: Views on domestic AI application development issues?

**Guang Yang:** This is mainly about OpenAI. Most teams polishing products may first wrap OpenAI with other features for pilots and POCs — but when formally going live you should switch to domestic foundation models. Using OpenAI early to polish product is fine.

**Hongyu Lu:** Models alone are not enough — data is critical. For founders this is crucial and delicate. Today’s model is “garbage in, garbage out.” How to get good data? We once built a financial data bureau; in the future the big companies’ advantage will be the data they hold.

For startups, matching big-company data is very hard — JD will not share with Alibaba and vice versa. If an SME wants a shopping-recommendation model, where does data come from? How to make the model better? A series of problems follows. Watch: how to get better data? With whom to partner for better data? That is vital to the company’s future.

**Yong Li (moderator):** On data, one more point: when we talk “data,” do not only see surface “numbers.” If possible, separate “numbers” from “data.” Numbers are what we clearly see — buyers, return rates, likes in marketing. Data is the hidden whole logic system; the same numbers mean different things under different logics.

Example: outdoor camping was hot for a stretch — lots of broad “data,” but most people only got “numbers” — how many bought gear, traveled, spent how much. Deeper data differs: your relationship with nature — conqueror posture vs. fusion with nature? Those two “data” produce wholly different consumption, product sales, and design.

“Conqueror” buyers may prefer stronger functional products; “fusion” buyers may care more about eco and sustainability — different design and sales strategies, affecting production, supply chain, marketing, and channel choice. Understand these differences. Founders need multidisciplinary lenses on the team — otherwise you think everything can be generalized.

How far do general big data and models apply across different people’s functions, needs, and scenes? Where they cannot generalize, how do we bridge? Those are questions AI founders must keep thinking about at the application layer.

— END —

---

Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20240423a06il400
On-site reading page: https://ffcap.cn/en/research/20240423a06il400