Original Title: 2025 AI创业真相

Published On: 2025.08.27 

Author: 陶天宇 (Tao Tian Yu)， 郭虹妘 (Guo Hong Yun)

Source: 钛媒体创投家 (tmtpost)

Article Link

A day in AI is a year in the real world. 

Just as the dust on the battlefield of large language models has begun to settle, the Agent wave surges forth. 

As inference costs plummet, the once-imposing technical barriers have crumbled, bringing all those who harbored hopes and aspirations for AI back to the same starting line. 

Entrepreneurs, invigorated by these open waters, are pouring in their creativity and ambitions to seize the first mover advantage. However, beneath the heels of this technological stampede lies critical questions of survival in China's AI ecosystem. 

TMTPost Venture Capital spent a month conducting in-depth face-to-face interviews with twelve independent developers, five AI community moderators, and seven top investors, to compile this article.

These voices paint a picture of the lived reality for China's AI entrepreneurs—not as glamorous as the press conferences, but genuine.

Consumer Habits

Whenever discussing systemic issues in the domestic AI ecosystem, it’s the industry’s consensus that the discussions start with consumer habits.

But just how bad are these habits? From the data we've compiled from reports and news articles, consumers in the North American^[1] market are 3-4 times more willing to pay than consumers in China. The ARR (Annual Recurring Revenue) of leading AI companies is five to one hundred times higher in North America and the average annual procurement budget for North American businesses is nearly 10 times higher.

Besides the data from these reports, the responses from several open-source contributors paints a clear picture. 

One developer stated that an Agentic product that his team in China developed had garnered tens of thousands of users within a month of its launch, but had fewer than ten paying subscribers, leaving him feeling frustrated. Meanwhile, an overseas product with essentially the same functionality achieved millions of dollars in revenue after just three months of its launch. 

Applications with multi-million dollar ARRs are a common sight in the North American market, and their number is on trend to increase rapidly. According to publicly available data, Cursor has reached $500 million ARR. Lovable, another programming tool, achieved $100 million ARR in its first eighth month. Within a few months after its launch, the AI ​​companion app Tolan achieved $12 million ARR. AI ​​video editing tool OpusClip has reached $20 million ARR.

On the mainland, AI commercialization has largely been sobering. 

According to data from QuestMobile, as of March this year, the number of monthly active users (MAU) of AI-native apps in China had reached 270 million, even surpassing ChatGPT's 180 million MAU^[2]. However, few products have truly achieved large-scale monetization. Even Kling AI, which has surpassed $100 million USD in annual revenue, generates over 70% of it overseas. 

"The general consensus is: launch products overseas. Develop it in secret and make money on the down low. Launching products domestically is a loss-making endeavor. If it's truly unfeasible, relocate everyone to Singapore or elsewhere," one developer said.

It is commonplace for consumers to expect products for free in the domestic software market, but the underlying causes are rarely discussed. 

On the consumer side, the differences between Chinese and American web browsing habits result in different expectations for software products. In North America, PC and mobile devices are used at similar rates, while in Asia, mobile usage (over 60%) far exceeds PC usage (32%). Since North Americans typically use PCs to browse the internet, they typically prefer standalone software or apps with diverse functions and complex interfaces, and are willing to learn to use them and pay out-of-pocket for them.

Conversely, Chinese users, who grew up in the era of the mobile web, are more accustomed to highly integrated "super apps" which are typically free and easy-to-use walled gardens serving as a one-stop shop for a variety of services. In this environment, users are naturally unwilling to pay for standalone software.

For businesses in North America, software subscriptions are cheap compared to labor costs. Therefore, even if such subscriptions only improve efficiency slightly, they can still be a worthwhile investment. This contributes to American businesses’ willingness to pay for standalone productivity tools.

In contrast, China's relatively low labor costs make such ROI calculations less convincing, and users are more inclined to choose free products, even if this means accepting advertising, limited functionality, or compromises in efficiency. On the other hand, although China has been catching up in enterprise digitization for a few decades, a large gap remains. This means that the use cases for enterprise service tools are far less mature than in North America. In many industries and for many companies with weak digital infrastructure, the implementation costs of productivity software far exceed the expected benefits, which reduces decision-makers' willingness to purchase said software.

There are others who elevate the issue of Chinese users’ "reluctance to pay" for software to a critique of national values, akin to the lack of respect for intellectual property and contractual commitment. In reality, this is more the combined effects of the social structure and Chinese Internet's "gravy train^[3]" business model.

Perhaps the solution isn’t to mold Chinese users to be like their North American counterparts, and continue to allow them to freeload, exploring an alternate path for China's AI commercialization. The wasted time and R&D spend, however, must then be endured.

Hidden Barriers 

Some have suggested that the AI startup boom is reminiscent of the vibrancy of the early internet. Product announcements and funding rounds fill news feeds daily, with high-profile conferences, eager investors, and a proliferation of related influencers and startup communities. Even the long-forgotten Zhongguancun Startup Street^[4] is gradually filling with people and regaining its former glory.

Data shows that nearly 5,000 new AI companies were created around the world in the first half of 2025, or an average of 27 per day. Of these, 1,380 were founded in China, or an average of seven per day. In the same time frame, global AI startups secured approximately $140 billion USD in venture capital, double the approximately $70 billion raised in the same period of 2024. AI's share of global VC investment rose to an all time high of 53-58%, compared to just 25-30% in the same period of 2024. In the first half of 2025, China saw 938 AI funding rounds, totaling 597 billion^[5] RMB^[6]. Compared to 2015, the golden age of internet entrepreneurship, where China's tech sector saw 3,932 funding rounds, totaling approximately 178 billion RMB^[7]. That year, Zhongguancun Startup Street alone incubated 1,791 startups, or an average of 4.9 new startups per day^[8].

Based on the raw numbers, the VC boom appears to be approaching its past highs, even surpassing it in terms of total investment. However, the details belie subtle differences.

A veteran independent developer stated: the AI industry does not have the "low activation energy" of standard tech startups, because exorbitant hidden costs are much higher than one might anticipate. Low-cost interfaces, network channels, computing power, high-quality training data, and a support ecosystem are all extremely scarce resources. These resources are even more crucial to developers than technology, funding, and product architecture. They are essential for these entrepreneurs to transition from early-stage startups to scalable AI companies.

Only a handful of entrepreneurs, typically from top AI companies and university labs, have the deep technical background and personal networks to access these resources. This results in the rumor that the top 10% of teams capture 90% of the capital.

Several top AI investors stated that their investment thesis maps are similar. The consensus is that there are fewer than 200 founders on their wish-list all hailing from the world's top AI companies. Those who can reach top talent, build trust, and secure shares have the clear advantage. While this may sound simple, it's closely tied to a variety of factors like an investor's education, work history, network, social skills, and even age. Thus, the ability to secure these core founders is the crucial differentiator between AI investors.

Being an AI VC investor is a game for the select few.

Of course, this doesn't mean that "second-tier developers" are out of luck. A growing number of investment institutions and incubators are actively reaching out to them with an endless stream of competitions, developer communities, AI boot camps, and free compute credits in an attempt to discover talent from the source. Some investors have also realized that excessive capital flowing toward the top can lead to a bubble, while niche teams that may have untapped potential. However, this path is necessarily an arduous one.

In the Shadow of Colossi

China's tech giants appear out of their depth on the AI ​​stage.

In 2024, while the four largest US tech companies announced capital allocations of 1.7 trillion RMB^[9] for AI expenditures, China's seven top tech companies only allocated 630 billion RMB^[10]. "It's important to note that this already dismal ratio is after the launch of DeepSeek in the fourth quarter of 2024 when the companies try to catch up."

This isn't just a numerical disparity; it's also a difference in strategy. Chinese giants appear more willing to invest in stock buybacks, dividends, and debt repayments rather than build AI infrastructure. While US chip export restrictions have had some impact, ultimately there is a lack of commitment from within Chinese tech companies. It is a gamble, on short-term profits rather than long-term results.

The result is a generational gap between Chinese LLMs and those made by frontier US companies. "Compared to domestic LLMs, foreign models like Claude Opus 4 can maintain logical consistency and have lower error rates in complex reasoning chains, across domains, and branching conditional structures. They also boast context lengths in the millions. When outputting structured documents, code, and JSON, they exhibit exceptional levels of stability that are currently difficult for domestic models to achieve."

Beyond the disparity in direct investment, the major domestic companies' AI ecosystem build out differ significantly from those of their American counterparts. One media outlet bluntly stated, "the US is using trillions of dollars of 'asset-heavy' investments to turn AI into the next generation’s water and electricity, while China is pursuing AI with an 'asset-light' mindset, still focused on monetizing views."

Some independent developers believe that while major domestic companies may appear to be investing heavily in building AI infrastructure, they are often mired in the inertia of monetizing internet traffic. Their main concern is whether AI features can mitigate traffic loss on their platforms, whether it will impact the existing business models, or whether it can stabilize market capitalization, rather than driving technological breakthroughs and industry disruption. As a result, when building AI ecosystems, major domestic companies tend to invest the bulk of their resources and technologies in internal projects, rather than opening them up to external developers and corporate partners. This model of "large companies lead, while small companies tag along" not only inhibits the growth of startups but ossifies the entire industry.

The most direct impact of a weak AI ecosystem is the loss of AI talent. According to a report by the Paulson Institute, 47% of the world's top AI researchers graduated from Chinese universities, but only 51.35% choose to pursue further studies in China, and even fewer remain afterwards. This loss of talent deprives the domestic AI industry of a valuable source of innovation.

Several leading developers remarked that if they were to start a new project, they most likely would not choose the AI ​​ecosystem of a major domestic company, nor would capital from market-driven funds be their top choice. Instead, they prefer Open AI’s incubator because it offers both financial support and a robust AI ecosystem.

Major domestic companies are undergoing rapid change. Substantial increase in AI investment, the world's most active open-source ecosystem, and the gradual opening up of core resources are sending positive signals. As long as domestic companies are on the right path, we should give them more leeway.

AI Hardware Rising

While there are noticeable gaps between the AI ​​software startups ecosystem in China and the United States due to the reasons mentioned above, China possesses unparalleled advantages in AI hardware.

Globally dominant tech hardware companies such as DJI, Xiaomi, Insta360, Dreame, and Ninebot have, over the years, developed a robust supply chain and a vast talent pool in China. The AI hardware companies’ ability to compete abroad align perfectly with the strategic need for Chinese tech companies to "go global" in the current political climate.

More importantly, the hardware narrative has been fully adopted in the secondary market. Both the STAR market in Shanghai^[11] and the Hong Kong stock markets have responded extremely enthusiastically to smart hardware companies, fostering positive expectations in the primary market. Against this backdrop, a wave of AI hardware startups emerges in China.

As of June 2025, there are approximately 1,180 operational/profitable AI smart hardware companies in China. According to IT Orange data, 312 new companies were registered in the first half of 2025, a 73% year-on-year increase from the first half of 2024. These companies cover cutting-edge fields such as robotics, AR/VR, smart cars, and wearables. During the same period, the industry disclosed 178 financing rounds, averaging one per day, with a record 42 in April alone. Total financing reached 17.6 billion RMB^[12], a 2.1-fold increase as compared to the first half of 2024. This figure does not include the embodied intelligence sector which has been receiving a lot of buzz.

Several leaders of RMB commercial investment firms remarked that in the past six months AI hardware projects have become a hot commodity in the RMB investment community. The Greater Bay Area, centered around Shenzhen, is the most active, with the most projects, the strongest supply chain, and the fastest pace of iteration. Founders from leading domestic hardware manufacturers are increasingly favored by investment institutions, especially those with work experience in key positions at companies like DJI and Xiaomi. 

Some software developers express skepticism over the "AI content" of AI hardware, believing it doesn’t make much use of models and — in the worst case — are simply basic algorithms overlaying traditional hardware. However, this view overlooks the incremental nature of AI hardware development. With advancements at the bleeding edge computing capabilities and the development of lightweight LLMs, more and more AI hardware will become truly "intelligent." In this regard, China's AI hardware technology is on par with its international counterparts, and, in some niche domains, may even lead.

Among the different aspects of the AI ​​startup ecosystem, Chinese companies are well suited for AI hardware. Unlike the US AI model, which focuses on software and algorithms, China, leveraging its manufacturing base and unrivaled supply chain, has a unique advantage. This advantage may provide an opportunity for China's AI innovators to overtake their competitors.

Summary

In this ecosystem full of paradoxes, some see adversity while others glimpse opportunity.

While developers in Silicon Valley struggled with prompt engineering, developers in Shenzhen are already deploying Transformer on smart glasses.

Perhaps, a veteran investor puts it best: "China's AI ecosystem never lacked talent; what it lacks is the nourishing substrate allowing them to flourish."

The cultural reluctance to pay for services may be difficult to undo in the short term, but as thousands of AI hardware companies emerge in the Greater Bay Area, history may offer a new opportunity to China's AI innovation ecosystem.

1. ^{^}

   This is the literal translation, but typically the authors just means the US or the US market. 

2. ^{^}

   I'm not sure where the reports got this number. Most reports I've seen online suggest that ChatGPT has 700 million MAU in August 2025. 

3. ^{^}

   The literal translation of this idiom is "sheep's wool from a pig" and is a play on the idiom "sheep's wool from a sheep". The later is something akin to "no free lunch" where the users get a product for free/cheap but are being taken advantage of in some way. The former means taking advantage of VC-funded start-ups to gain benefits, e.g. taking subsidized Uber rides when it was offering a lot of discounts to capture market share.  

4. ^{^}

   This is a part of Beijing containing Peking University.

5. ^{^}

   Source: CB Insights' "2025 H1 Global AI Report"

6. ^{^}

   ~$82 billion USD (2025 September rates)

7. ^{^}

   ~$27.8 billion USD (2025 September rates)

8. ^{^}

   Source: Zero2IPO Research Center's "2015 China Equity Investment Annual Report"

9. ^{^}

   ~$238 billion USD (2025 September rates)

10. ^{^}

    ~$88 billion USD (2025 September rates)

11. ^{^}

    Shanghai Stock Exchange Science and Technology Innovation Board

12. ^{^}

    ~$2.5 billion USD (2025 September rates)