Understanding the Latest AI Startup Trends in One Article
-
Since ChatGPT took the market by storm, the startup scene has become lively, reminiscent of the mobile internet startup boom in 2014.
Over the past period, we've seen more and more people embarking on AI entrepreneurship, mainly divided into two groups.
One group consists of startup teams led by professors and scientists from universities and research institutes. They were among the earliest in China to research Transformer models and have years of academic expertise in generative models. From improving generation quality to achieving more controllable generation, they have been continuously exploring the boundaries of AI's potential. Recently, it has become evident that executives from major internet companies, professionals from the AI 1.0 era, specialists in niche verticals, and serial entrepreneurs are increasingly joining this wave of AI. Many of them are product managers or have technical backgrounds, possessing flexible thinking, the ability to design products, manage communities, a keen sense of user needs, and expertise in launching products from scratch. On the product front, we are excited to see domestic products like Miaoya making a splash on social media.
However, while being thrilled by the transformative potential of AI technology, it is essential to remain rational. Where does the commercialization path for AI tool products lie? What competitive elements truly serve as barriers for application-focused entrepreneurs? Which scenarios represent genuine necessities? These questions bring us back to the core issue of entrepreneurship: where is the Product-Market Fit (PMF)?
Technology only holds value when it is applied to specific scenarios and needs. Teams with strong technical capabilities also need to incorporate individuals who understand the market and users to ensure long-term success. As Jinqiu Fund, which has been specializing in AI investments for many years, we have always maintained a frontline perspective to observe AI technology trends and entrepreneurial market changes. In the past, we have invested in several leading AI enterprises with global vision that possess both technological innovation capabilities and commercialization potential.
The current AI entrepreneurship wave is in full swing—but is now the right time to enter? Which areas of AI projects are more likely to secure funding? What types of AI products attract more traffic and demonstrate stronger user retention? How has the technological paradigm shifted between AI 2.0 entrepreneurship compared to AI 1.0?
For entrepreneurs preparing to enter or already engaged in the AI field, Zheng Xiaochao, Vice President of Investments at Jinqiu Fund, provides a comprehensive analysis across multiple dimensions including the AI industry, trends, and products. From an investment perspective, he clarifies the two most critical issues for AI entrepreneurship—suitable directions and optimal timing. This article presents the highlights of the speech From an Investment Perspective on AIGC Development Trends by Zheng Xiaochao, Vice President of JinQiu Fund Investment, with some edits:
Below is the table of contents
01 AIGC is Currently in the Technology Trigger and Peak of Inflated Expectations Phase
02 Text and Image-based AIGC Application Opportunities are Emerging 03 3D, audio, and video large models remain in the non-consensus stage
04 Text and image-based AIGC see the highest traffic, while commercial AIGC is rising
05 Chatbots have the highest visit volume, with virtual companion apps showing strongest user retention
06 AI concentration increases, with Copilot-type startups accounting for nearly half of ventures 07 General-purpose AIGC attracts more funding, while vertical scenarios are easier to launch
08 Assessing AIGC entrepreneurship timing through paradigm shifts
09 AI entrepreneurship: from China to Global
AIGC is currently in the technology trigger and inflated expectations phase A new technology typically goes through a relatively slow incubation period during its development. After passing a critical point, it enters a phase of rapid development, followed by a relatively stable growth stage. This is what we commonly refer to as the "S-curve" of innovation diffusion.
In reality, previous technological cycles have all followed this pattern. During the PC internet era, PC shipments peaked in 2011 and then began to decline year by year. By 2012, mobile internet had been incubating for many years, and various mobile internet applications centered around devices like the iPhone had emerged - ByteDance (TikTok's parent company) also appeared around this time. At this stage, mobile internet began to take over from PC internet and entered a period of rapid growth. However, by 2017, smartphone shipments also began to peak, and from 2017 to the present, innovation in the entire mobile internet sector has significantly slowed down.
From the iPhone X in 2017 to the recent iPhone 15 release, even as a devoted Apple fan, I've gradually stopped following these "tech春晚" (technology spring galas), because many innovations have essentially become incremental improvements without particularly exciting breakthroughs. However, the Vision Pro announced by Apple in June this year was truly exciting. The Vision Pro project was initiated around 2015, precisely when smartphone shipments began to peak - marking the incubation period of a new technology. Another interesting point is that the foundational paper for this wave of large models, Attention is All You Need, which introduced the Transformer framework, was also proposed by the Google team in June 2017. After years of accumulation, ChatGPT has now completely ignited the market, and we are entering the early stage of rapid development for large models. Of course, large models may not be the only important development in this cycle; AR and VR are also significant technological variables that cannot be ignored.
Looking at the specific development cycle of large models, the Gartner Hype Cycle illustrates the typical stages new technologies go through. Currently, AI 2.0 based on AIGC is in the stages of technology trigger and peak of inflated expectations. We have observed that ChatGPT's traffic began to peak in May this year and has been declining in recent months, so the overall market expectations are gradually becoming more rational. For this new wave of technology led by large models, we remain optimistic in the long term but maintain a relatively cautious stance in the short term.
Opportunities for text and image-based AIGC applications are beginning to emerge At different stages of each technological innovation cycle, the position of the value chain varies, which is referred to as the transfer of the value chain. As the value chain transfers, the market scale expands exponentially. Although the application layer has the largest market size, its market concentration decreases. In contrast, the underlying technologies, hardware frameworks, and operating systems (OS) tend to be more concentrated (otherwise, it would be a disaster for application developers). Generally, the upstream market is highly concentrated, with two or three companies occupying more than half of the market share, while the downstream market becomes more diversified, with numerous applications emerging.
For example, during the PC internet era, the companies that initially captured the value of the entire industry chain were hardware manufacturers like IBM and Apple. Later, the value shifted to OS providers like Microsoft and chipmakers like Intel, and eventually reached the application layer, such as Google and Facebook. Similarly, the mobile internet has gone through a similar cycle, from smartphones to Android and iOS, then to ARM or Qualcomm, and finally to applications like ByteDance and Meituan. For large models, we believe that companies like OpenAI in the large model space and Nvidia in the chip industry play a foundational role in the technological wave of large models. With the maturation of large language models and text-to-image models, we've seen the emergence of excellent applications like Midjourney and Character.ai. One key takeaway here is the importance of doing the right thing at the right time.
When underlying technologies are not yet mature, you can focus on technological innovation because technical advantages can provide strong product competitiveness. For example, ChatGPT is powerful because it has its own underlying large model, and Midjourney's impressive results are also due to its proprietary underlying large model, making their products stand out. However, when underlying technologies have already matured, entrepreneurs should focus more on seeking innovation opportunities at the application layer. If you try to innovate in large language models at this stage, it will be highly competitive and quite challenging.
3D, Audio, and Video Large Models
Still in the Early Stages of Technological Development Breaking it down, the current technological development levels of different modal large models vary. People may currently focus more on text and image models, both of which have entered a relatively mature stage with applications beginning to flourish.
Large Language Models
For large language models, the development can be divided into several stages. The first stage was from the publication of the Transformer paper in June 2017 to the emergence of the Bert model in 2018. This period was a non-consensus phase—people knew about it but were unclear about its practical applications. In 2018, the Bert model was introduced and began to be applied in scenarios such as translation. At this point, it was discovered that Transformer-based pre-trained models could improve the efficiency of translating various texts, but they were still primarily seen as tools for enhancing task performance. By the third stage, with the emergence of GPT-3, a new consensus began to form: the idea that a general-purpose pre-trained large model could accomplish almost any task. As long as the prediction of the next token becomes increasingly accurate, more intelligence could emerge—a phenomenon known as "emergence." This is the first principle of large model AI, and previous small models and architectures may gradually be abandoned. Currently, large language models have foundational models like the closed-source GPT-4 and open-source models such as Llama 2. From a technical perspective, large language models have become relatively mature, with killer apps like ChatGPT educating the market and users, boasting over 100 million monthly active users. Large language models have entered a phase of rapid application explosion, where various enterprises are likely using AI tools and Copilots to integrate with business scenarios, creating corporate knowledge bases, chatbots, game NPCs, etc. These developments are already happening.
Image generation has also reached a relatively mature stage. The foundational theory of diffusion models was proposed in 2020, followed by OpenAI's DALL-E in 2021. By the first half of 2022, the closed-source model Midjourney V1 was released, though it was still very rough and difficult to commercialize. In August 2022, Stable Diffusion released an open-source model, and text-to-image generation quickly entered a rapid development phase. By this year, Midjourney's V5 version has achieved image generation quality that alarms designers. Currently, Midjourney's Discord community has over 17 million users, making it the killer app in text-to-image generation. In the field of image generation, both closed-source and open-source foundational models have emerged, ushering in an explosive phase of applications, with numerous apps like Miaoya and Lensa springing up like mushrooms. Compared to large language models and images, other modalities are currently at a relatively earlier stage of development and receive less attention, such as in the fields of 3D generation, video generation, and audio generation.
The foundational theory of 3D generation, NeRF, was proposed in 2020 and entered an accelerated development phase starting in 2022. Numerous models have emerged, such as the early DreamFusion and Magic3D, but so far, no universally recognized foundational model has appeared. However, through continuous research, we have found that there are many explorations of 3D generation applications in areas such as game assets, metaverse assets, and digital human assets.
Video Generation Video generation is a crucial direction, built upon image-based Diffusion models, which stitch together different frames to create coherent and connected videos. Since last year, there has been a surge in research papers and products in this field, with Runway leading the pack, now valued at $1.5 billion. We believe that the commercialization of video generation is still in its early stages.
Audio Generation
In the realm of audio generation, new Transformer-based pre-training paradigms represent a significant leap forward compared to traditional TTS (Text-to-Speech) technologies. However, this field remains in the early exploration phase, primarily due to the scarcity of audio data compared to images and text. Notably, Microsoft and Meta have introduced some innovative audio generation models this year. Similar to ChatGPT's prompt-based approach, these models can clone a user's voice from just a 3-second sample, replicating even the emotional tone. Overall, both 3D video and audio generation technologies are still in their early stages of innovation. Text and image AIGC receive the highest traffic
Commercial AIGC is on the rise
To gain a clearer insight into the AI industry's entrepreneurial landscape, we analyzed the traffic of overseas AI products in August this year. Excluding major applications like ChatGPT, the overall traffic distribution of AI products in August aligns with the patterns we described above.
Currently, text applications receive the highest traffic, followed by image applications, and then commercial applications, code generation, audio, and video. We've found that large language model products are most frequently accessed for virtual character interactions, similar to character.ai, followed by general writing purposes. In the image domain, usage is primarily concentrated in image generation and editing. The commercial sector mainly involves various vertical industry applications.
This traffic pattern indicates that text and image-based AIGC products currently find product-market fit more easily, but it also suggests increasing competition among entrepreneurs in these fields.
Chatbot services receive the highest traffic
Virtual companion applications demonstrate the strongest user retention From a dynamic perspective, let's examine the traffic changes of the top 100 AIGC products over the past three months.
It's evident that the overall market peaked in May and began to decline slightly from June to August. The most visited products remain ChatGPT-like chatbots, although chatbot traffic has been declining for three consecutive months. In contrast, search-related products like NewBing have maintained relatively stable traffic, while image generation products saw a slight increase from June to August.
Analyzing from the perspective of average user session duration, virtual companion applications, despite not ranking high in traffic volume (only fifth among all applications), boast the longest average user session duration at 25 minutes—and this figure is still rising. For instance, character.ai currently has an average user session duration of 34 minutes. Productivity assistants and document creation tools rank second and third in terms of user session duration, respectively.
AI Concentration Increases Copilot-like Startups Account for Nearly Half
YC Startup School serves as a bellwether for North American entrepreneurship. By examining data from Y Combinator's winter and summer batches this year, we can observe trends in overseas AI startups.
First, the AI density is rapidly increasing, indicating fiercer competition in AI entrepreneurship. In the first half of this year, 37% of projects were AI-related, but by September this year, the proportion of AI projects had risen to 60%. Secondly, there has been a noticeable shift in the distribution and categorization of different types of AI projects. Startups focusing on foundational large models have decreased from 14% to 6%, with more entrepreneurs shifting towards the application and middleware layers, particularly concentrating on Copilot technologies.
In the first half of this year, Copilot implementations were mostly superficial ChatUIs, but now deeper integrations with workflows have become a hot trend for startups. Taking the legal industry as an example, Copilot in this sector acts as a partner to professional lawyers, assisting with tasks such as contract review, meeting minutes, and internal corporate knowledge base searches, rather than just serving as a legal Q&A dialog box.
General Scenario AIGC Attracts More Funding
Vertical Scenarios Are Easier to Launch From the perspective of AI startup funding amounts, there are notable differences across various sectors.
In B2B applications, startups focusing on general-purpose scenarios have secured significantly more funding, while vertical-specific B2B applications receive approximately only one-fifth of that amount. B2C applications are primarily divided into general personal-use categories and personal productivity tools. As of September 17th, the number of funding rounds for general personal applications is about twice that of personal productivity tools.
The same trend is observed in average single-round investment amounts. General-purpose B2B tools receive about 4 times more funding on average compared to vertical-specific solutions. Similarly, general B2C applications secure 2.5 times more average funding per round than specialized single-point B2C tools.
From this, we can conclude that general-purpose scenarios tend to be more capital-intensive as they require knowledge and capabilities across different domains while targeting broader audiences. In contrast, vertical-specific or single-point application tools can typically launch with relatively less funding. ## The Timing of AIGC Entrepreneurship from the Perspective of New Paradigm Shifts
After summarizing the new trends and changes in past and overseas AI products, we return to the core question: For entrepreneurs, is now the right time to start an AIGC venture? Here are some perspectives to share.
First, pre-trained large models have brought about a shift in paradigms, significantly lowering the barrier to using AI technology for application development. If you're in the B2B sector, this also leads to a substantial reduction in delivery costs. The principle is simple. In the AI 1.0 phase, to build an AI application, you needed computing power and extensive data preparation, such as data cleaning, labeling, and structuring. Then, you had to decompose the task into smaller parts, each potentially requiring training separate small models, followed by model maintenance and deployment to support your application. For B2B solutions, you often had to rebuild models for each client, making it nearly impossible without a team of dozens.
However, in the AI 2.0 phase, the development threshold has changed significantly. All these complex tasks are now handled by a general-purpose large model. You can either use APIs of such models or fine-tune an open-source model yourself. If your team is capable, you can even train your own large model, which can solve problems that previously required 10 smaller models combined.
Of course, this is a simplified picture. You can further enhance it with vector databases, chain-of-thought, tree-of-thought techniques, or create AI agents to tackle more complex problems. But regardless, compared to the AI 1.0 era, the application development threshold has been reduced by an order of magnitude. Second, large models bring a new paradigm of human-machine collaboration.
In the past, AI was a simple tool that didn't offer much assistance. Now, we are developing AI Copilots, making AI a partner in your work. You set the goals, and AI quickly executes tasks, which you then modify, adjust, and confirm before AI completes the task.
We are now transitioning from Copilot to agent models. In the agent model, AI acts as a company employee, capable of forming its own team. You assign a task to an AI agent, and it can automatically delegate tasks to product managers, architects, operations staff, and programmers. They observe, think, act, and then complete the entire task. This model will give rise to many super individuals or super teams, with humans playing the role of setting goals for AI, providing resources, and conducting checks, evaluations, and feedback. In reality, we've found that many AIGC teams no longer need as many human members. Behind the scenes, numerous AI robots are handling the work. For instance, a 10-person AI startup team might have you as the CEO, GPT as the CTO, and Midjourney as the CMO.
The next phase of development may lead us into the AI Society model, where humans and intelligent agents operate as relatively equal partners. Both can autonomously propose requirements and mutually provide resources, though this remains a long-term vision.
The development threshold for AIGC products has significantly lowered, meaning AI products can now be launched with minimal funding. According to the latest statistics from A16Z, among the top 50 most visited AIGC companies, over twenty have never raised funding—a scenario unimaginable in the previous internet era. Here, we observe a shift in the internet's burn-rate logic: it's now feasible for entrepreneurs to launch AI products with limited funds, thanks to the reduced development barriers and the productivity boost from new human-machine collaboration models. On the other hand, the good news for AI entrepreneurs is that consumer willingness to use and pay has also significantly increased. The chart on the right shows that when comparing the organic customer acquisition of AIGC products with non-AIGC products, the lowest quartile of AIGC products only has 2% of traffic coming from paid sources, with 98% being organic growth. Meanwhile, 90% of the top 50 AIGC companies are already profitable. In contrast, non-AIGC products have a lowest quartile where 70% of the traffic is paid.
Therefore, compared to the past internet and mobile internet eras, AIGC products have seen a much higher willingness among users to adopt and pay for them. I believe the underlying logic is twofold: first, AIGC products enhance productivity, and people are generally more willing to pay for "revenue-generating" solutions rather than "cost-saving" ones. Second, products like Miaoya have qualitatively improved in meeting consumers' aesthetic and social needs compared to previous technologies.
Returning to the initial question, is now a good time to start an AIGC venture? My conclusion is Yes, as AIGC is still in the early stages of a new innovation cycle. AI Entrepreneurship: From China to Global
The middle image is a screenshot I took in February this year, which I believe represents the pinnacle of China's mobile internet era: four out of the top five apps in the North American App Store rankings were from China. The first was Temu by Pinduoduo, the second was CapCut (also known as Jianying), the third was TikTok, the fourth was Google, and the fifth was Shein. This demonstrates that Chinese teams are fully capable of creating world-class applications.
The image on the right is from a Stanford study, which shows that China's AI research output surpassed that of the US in 2020, indicating that China has also accumulated AI talent. However, it's important to recognize the gap: while China produces a large number of AI research papers, they often lack originality and impactful studies. Although the current AIGC wave is once again being led by the US, with OpenAI and others having paved the way, catching up is not an insurmountable challenge for us. I believe Chinese entrepreneurs should have global ambitions. In technologically mature niche sectors, they should compete globally in applications because China has the best programmers, product managers, and operations teams. In less mature technological fields such as 3D, video, audio, and AI agents, Chinese entrepreneurs also have the capability to pioneer and become the state-of-the-art (SOTA) in more industries.