Before AI Changes the World, Find PMF First | Highlights from Unity Ventures 'Growth' Series Salon

九合创投九合创投··3·4

A Realistic Discussion About PMF in the GenAI Era

"

Generative AI has kicked off a major era of transformation, but bridging the gap from large model capabilities to real product-market fit remains a challenge. Few PMFs have gained broad recognition so far.

The second installment of Unity Ventures' "Growth" series focused on PMF and commercialization in the generative AI era, bringing together founders from Unity's portfolio and industry experts for a practical discussion on finding product-market fit. On the road to AGI, finding PMF first is a critical proposition for startup survival.

"

Xiao Wang, Founder of Unity Ventures

The Rise of Intelligence: Mainstream AI Business Models and Entrepreneurial Opportunities

Unity Ventures has long been deeply involved in the AI space, from autonomous driving company Momenta and "first medical AI IPO" Airdoc, to application-layer companies like LynkSoul and CreativeFitting, and on to robotics and smart hardware — we've been actively deploying capital across the board.

I was involved in building Baidu's first-generation search engine. Later came Toutiao as the second-generation recommendation engine. Both generations centered on information retrieval and distribution. With the explosion of internet data, advances in computing power, and the evolution of Transformer architectures, marked by the birth of ChatGPT, we've entered a new technology cycle.

Large models have become the third-generation intelligence engine, gradually expanding from large language models to multimodal models and vertical domain models, providing intelligent infrastructure for thousands of industries. Vertical industry AI transformation presents significant opportunities — industries have unique data and know-how, and intelligent agents can work alongside humans to provide end-to-end services.

From a business model perspective, in the Mobile Internet era, platforms relied on traffic distribution for commission fees, essentially monetizing through advertising. Core metrics were user scale and retention rates, leading to winner-take-all dynamics and strong headwinds effects.

In the generative AI era, intelligence is the service. The profit model charges for services delivered by intelligent agents, improving human efficiency ratios, with the possibility of multiple wins. Token costs and service revenue determine gross margin levels. The competitive moat lies in using user-generated data for model iteration, forming a self-reinforcing data loop.

The PMF logic differs between the two eras. Rising from niche markets is one of the important trends in AI application development. Products must first meet the needs of a single scenario, establish a data closed loop, improve user experience, refine product functionality, and then iterate to evolve into comprehensive platforms spanning multiple domains. So you can't simply rebuild products using internet-era thinking. Entrepreneurs need to approach from data, and from the intersection of services and data, to think through the underlying business model and value creation.

Breakthroughs in large model technology will bring productivity transformation. I look forward to exchanging with more entrepreneurs, sharing our experience in product, management, and financing, to help everyone more smoothly complete the journey from zero to one.

Jinhui Yuan, Founder and CEO of SiliconFlow

PMF for AI Infrastructure

I started my entrepreneurial journey in 2016. My previous company, OneFlow, built deep learning frameworks. We foresaw early on that models would grow increasingly large and require new underlying architectures, and received angel investment from Unity Ventures at founding. When large models exploded, this direction quickly became consensus. OneFlow was acquired by Lightyears Away, and later I left Meituan to found SiliconFlow.

When the tide comes in, the shovel-sellers make money first. AI infrastructure sits between compute companies and large model companies, connecting several elements into integrated solutions. Infra PMF also requires finding the right market, product form, and business model. From a market perspective, we chose to focus on inference deployment. In the future, inference models may consume hundreds or even thousands of times more compute than training models, with a broader customer base.

PMF first requires finding an entry point. After the first product gains footing, then extend and iterate rapidly — this effectively improves product success rates. The through-line behind our products evolves from meeting developer needs and making developers love the product, to serving enterprise-grade demands with stronger willingness to pay.

Xingyuan Yuan, Founder and CEO of ColorfulClouds

The AI Era: A Methodology for Entrepreneurship from Zero to One

ColorfulClouds Xiaomeng was China's first AI web novel writing app, launching in 2021 — a year before ChatGPT emerged. PMF starts with finding the market. Our insight came from academia; a roughly five-year gap exists between industry and academia. In 2016, if you were following academic papers, you'd notice AI researchers were all working on language. Back then, industry was still focused on computer vision. After Transformer arrived in 2017, it signaled to us that larger models should be applied to language problems.

After following the trend, wait for inspiration from user insight. In 2019, while building ColorfulClouds Xiaoyi, we discovered that 60-80% of people using Xiaoyi for translation were actually reading novels. This made us realize that novels represent the mainstream consumption of Chinese text. So we pivoted from Chinese-to-English translation to continuation writing from context to context, with excellent results.

After launch, from meeting user needs to achieving PMF, there's a significant hurdle: user retention. Currently, a major factor affecting AI app retention is insufficient product intelligence — large model capabilities still need further improvement. Beyond accepting market and user feedback, you yourself must become a user, breathing the same air as your users, to better make product feature trade-offs and meet users' real needs for AI applications.

Yong Qin, COO of Airdoc

The Journey of Exploring Medical + AI Products

Airdoc leverages AI technology to provide early auxiliary diagnosis and health risk assessment through retinal imaging, delivering integrated hardware-software solutions. The keyword for our target market selection is chronic disease. As of 2024, China has approximately 500 million chronic disease patients and about 600 million people with myopia. According to projections, by 2030, the retinal AI market will reach 34 billion yuan, and the myopia prevention and control market will hit 210 billion yuan. Meanwhile, the imbalance in domestic medical resources creates massive demand for AI products.

We chose the retina as our entry point for chronic disease detection because it's the only part of the human body where blood vessels and nerves can be observed non-invasively. Retinal images serve as direct medical evidence for assessing the progression of multiple diseases. For product form, we adopted an integrated hardware-software solution. The devices are more portable, and most importantly, patients can complete self-testing, dramatically lowering the barrier to use.

For commercialization scenarios, we cover clinical departments in hospitals, health examination centers, and other medical institutions, as well as insurance companies, pharmacies, and optical centers in the broader health ecosystem. Depending on the scenario's needs, we generate different test reports. For insurance companies, our health risk predictions help improve conversion rates and determine premium levels based on risk tiers. Beyond B2B scenarios, our myopia prevention products also enter the consumer market, serving more families.

Panel Discussion

Panelists:

Jinhui Yuan, Founder of SiliconFlow

Bihao Wang, Co-founder & COO of LynkSoul

Zhi Zan, Founder of Longmao Data

Liaoyuan Ning, Co-founder & CTO of TTC

Xingyu Li, Chief Ecosystem Officer of Enflame

What metrics do you generally use to validate product PMF, and what situations or problems do these metrics reveal?

Jinhui Yuan: For AI infrastructure products, PMF validation metrics include token generation volume, GPUs used for model inference, and number of representative users. Token and GPU volumes represent market demand, while we also look at how many high-traffic companies in each vertical are using our products.

Bihao Wang: We mainly look at three dimensions of data. First, virality — without it, consumer product PMF doesn't hold. We also look at whether users are willing to drive裂变; we generally do this through UGC. The biggest problem with UGC now is that content consumption value isn't very high, and still needs exploration.

The second metric is MAU and DAU, including new user retention and active user retention. New user retention reflects channel health. The third metric is commercial rate, including repeat purchase rate from existing customers and conversion rate for new users.

Zhi Zan: We focus more on hard PMF — whether the company is actually making money. On one hand, whether the product and service represent sustainable demand that can generate orders and revenue. Second, you need to calculate gross margin clearly, factoring in sales, customization development, and delivery personnel costs. Some large orders have slow payment cycles, and profits can be slim or even negative.

Liaoyuan Ning: At the core of PMF is market. Our choice was to enter a relatively mature market, like recruiting, so we only needed to adjust the product — combining AI with professional headhunting services to meet client needs. Unless your product is exceptionally outstanding, I suggest avoiding markets with high customer education costs. Providing better solutions is far faster than cultivating a market.

Xingyu Li: From a compute metrics perspective, when companies shift from training compute to inference-dominated, it proves products are beginning to reach the market. Because training is internal R&D while inference faces users. If inference compute grows quickly, it proves user scale is growing and PMF is within reach.

Capital markets generally recognize PMF through continued financing and expansion, but currently there are relatively few widely recognized PMFs. What's behind this phenomenon? What are the differences between generative AI entrepreneurship and the Mobile Internet era?

Jinhui Yuan: The explosion of AI applications is tied to the maturity of model capabilities, dependent on improvements in model reasoning abilities, and technology penetration takes time. Meanwhile, with the development of smaller models, infrastructure layer optimization, and declining chip costs, compute costs will also decrease. I have no doubt about the rise of AI applications.

Bihao Wang: The AI era offers far more opportunities than the Mobile Internet era, because this is a technological capability that changes all humanity. We'll see the vigorous development of three super applications: autonomous driving, robotics, and AI assistants. For AI to continue evolving, the consumer side is definitely crucial. In the future, terminals that everyone can participate in will emerge, generating massive amounts of data that will drive large model development.

The current problem is insufficient user scale on the consumer AI side, but we can first mine value for this current cohort of AI users, then expand. Second, apply AI transformation within your strongest industry. We need to think clearly about whether we're solving new problems and new needs, or meeting existing needs more efficiently.

Zhi Zan: Generative AI differs from the Mobile Internet era — network effects are less pronounced, so you can't expect to start making money only after scaling to massive user bases. You need to look at service costs and revenue earlier. In the AI era, developer costs can be very low, and serving some niche needs can be profitable.

In the Mobile Internet period, many companies burned money to expand user scale, but now the economic environment and era are different. You should charge users early, obtain revenue, then raise financing to expand validated business — rather than raising financing first, then burning money to scale. That's no different from drinking poison to quench thirst. If users truly aren't willing to pay, you may have built the wrong product.

Liaoyuan Ning: Financing is now mainly concentrated in top-tier large model companies, but actually some under-the-radar AI companies in China are already profitable. They're not publicizing their AI apps hitting rankings — they're making money quietly. As long as you keep your head down and work, you can feel AI's rapid development. I don't recommend directly analogizing generative AI with the Mobile Internet. You need to more directly observe what the essence of this technological shift is.

Xingyu Li: AI entrepreneurship in China today differs profoundly from the Mobile Internet era — B2B entrepreneurship far outnumbers B2C. First, B2B represents the new quality productive forces encouraged by national policy. Second, large model regulations naturally constrain B2C development. Third, from a technical difficulty standpoint, B2C product development is harder than B2B — B2B operates in relatively closed vertical domains. Right now China has many vertical AI products on the B2B side; while not as eye-catching as before, they already have certain commercial scale. Eventually, AI applications will develop toward B2C, just with a longer cycle.

What capabilities do new-generation AI companies value more in talent?

Liaoyuan Ning: We've observed that many startups, beyond capability, also heavily value candidates' alignment with company mission and vision. If candidates focus more on comparing salary differences, it negatively impacts the company's impression of them. Companies prefer to see candidates'认同 and passion for the company vision and the事业 they're building — finding partners to fight alongside, not just employees joining to work. If candidates only value salary, they're likely to leave for higher pay, which causes greater loss to the company than not hiring that person at all.

Xingyu Li: First, AI talent can't pursue certainty so strongly — they need to tolerate development in uncertain environments. Second, they need growth orientation. AI companies' requirements for talent may constantly adjust. Compared to past skill sets, dynamism and growth matter more.

Bihao Wang: We also value mindset and growth. Professional capability is relatively easy to assess in interviews; growth is harder to judge. Only those who can learn from failure have growth potential.

Zhi Zan: I focus more on candidates' innovation and entrepreneurial capabilities. The AI era is constantly changing, and past experience may not help future performance. You need to autonomously learn new technologies and solve problems with new methods. For role division and collaboration, I want salespeople to understand technology, and technical people to understand customer needs. If you only stay in your own domain, it's often hard to make correct decisions.

Jinhui Yuan: In the AI wave, growth and innovation capabilities are indeed crucial. From an AI model perspective, each of us is also a large model — with past experience, while continuously reinforcing learning. Especially in startups, you need to rapidly iterate yourself.