The "Laodeng" app has topped the AI charts.

Where has the battle for AI applications reached?

If we imagine the entire AI industry as a gold mine: the infrastructure layer (chips, computing power) is equivalent to "selling shovels", providing basic tools for gold mining; the model layer (large model R & D) is like "selling maps", telling everyone where the gold is; while the application layer is directly "diving into the gold rush" to turn the gold into cash.

In the past two years, the number of users of AI applications has been increasing. According to QuestMobile data, as of August 2025, the number of mobile - end AI application users reached 645 million, and the PC - end reached 204 million. Among them, the average month - on - month growth rate of billion - level native APPs was 1.3%.

What is an "AI native application"? It can be understood that from the very beginning of its design, AI is taken as the core driving force. The underlying architecture, operation logic, interaction methods, and business models are all innovated around AI capabilities. Doubao, DeepSeek, ChatGPT, Perplexity, etc. all belong to this category. Because of their purity, they are also considered one of the product types that can best test the market's acceptance of AI.

However, after comparing multiple lists, we found that different from foreign startups continuously incubating leading AI native applications, most of the top - ranked products on domestic lists come from large companies. Some are even upgraded versions of old applications with "AI added", such as Alibaba's Quark and Baidu's Wenku. There are not many native AI applications truly created by startups that have entered the mainstream vision.

The seemingly bustling AI application track is actually still a game dominated by large companies. Why does this situation occur? Is there still a future for startups to develop AI applications?

The explosion of AI applications is still a game for giants

In the past two years, global AI applications have witnessed explosive growth. Although there is no clear statistical data at present, industry insiders estimate that the number of global AI applications has reached hundreds of thousands. They can be roughly divided into two major camps: TOB (for enterprises) and TOC (for the general public).

In order to better observe the long - term development trend of AI applications, we selected the popularity list of domestic AI applications in the first half of 2025 as a reference (data source: Xsignal). The top 20 applications on the list can be divided into three categories: native applications of large companies (12), old applications of large companies + AI (1), and native applications of startups (7). Only about one - third come from startups.

The top three are Doubao, DeepSeek, and Quark. Among the top - ten native applications of startups, apart from DeepSeek, there is only the chat - assisting application Lovekey Keyboard.

An investor even pointed out that strictly speaking, DeepSeek is not a real startup. It relies on the resources and team support of its parent company, Magic Square Quantification. This means that the proportion of startups that have emerged victorious on the list is even lower.

To further verify this trend, we also selected the global Top 100 AI application list released by the famous Silicon Valley VC institution a16z for comparison.

It should be noted that it counts the data of iOS and Google Play. Due to the lack of some applications in the Android market, the rankings of AI applications on the list are different from those on the domestic list. For example, Tencent Yuanbao, which ranks high on the domestic list, and Nano AI Search under 360 did not make it onto the list. Meanwhile, some applications of the "old products of large companies + AI" type appeared in the top ten of the list, such as Meitu Xiuxiu launched by Meitu in 2008 and Xingtu launched by ByteDance in 2020.

But whether it is the domestic or the global list, the results show that the top three remain the same, and applications of large companies are still the protagonists.

The list of the top 100 global AI apps worthy of attention in 2025 released by Pengpai further confirms this point.

The products on the list are mainly from Silicon Valley technology giants and domestic Internet large companies. Among them, the total number of self - owned AI apps of four companies, ByteDance, Alibaba, Tencent, and Baidu, accounts for nearly one - fourth. ByteDance ranks first with 12 self - owned AI apps.

Further analysis of these lists shows that the leading advantage of large companies in AI applications is, on the one hand, due to the type advantage.

Currently, the popularity of AI applications is clearly differentiated. Chatbots lead by an absolute margin. Almost all domestic large companies have placed heavy bets on this track. For example, ByteDance's Doubao, Tencent's Yuanbao, Alibaba's Tongyi, and Baidu's Wenxiaoyan.

Only some of the six leading large - model companies have launched chatbot products.

An industry insider explained that the general nature of chatbots allows them to attract a large number of users, but it also requires a large investment in computing power, data annotation, and algorithm optimization costs. This is one of the important reasons why some startups have not developed chatbots.

However, in other types, the performance of large - company AI applications is generally better than that of startups. For example, in the fields of virtual characters and video generation, ByteDance's Maoxiang and Kuaishou's Keling outperform similar applications of startups. Even though MiniMax's virtual character application Xingye performs well, it still ranks behind large - company applications.

On the other hand, some new AI stars are actually "veterans making a comeback". Large companies are using AI to make their old - fashioned applications take off again. For example, Alibaba's Quark, Baidu's Baidu Netdisk, and Baidu Wenku. They originally had a large user base. Now, with the "new guise" of AI, they have attracted another wave of popularity.

After Quark was upgraded to the "AI Super Box", it emphasizes AI dialogue, in - depth search, and AI tool integration, covering scenarios such as writing, PPT generation, and problem - solving, which is popular among students and office workers; as a cloud storage tool, Baidu Netdisk has functions such as one - click image classification and subtitle generation after adding AI, improving the efficiency of users; Baidu Wenku is one of the most commonly used work tools for office workers. Its revenue in 2024 was even higher than that of WPS. In May this year, the access volume of its "Intelligent PPT" once ranked first in the world.

Why do startups lack stamina in developing AI applications?

Actually, the pattern of AI applications was not "monopolized" by large companies from the very beginning.

In the early stage, the performance of native applications of startups was not inferior to that of large companies. The "Six Little Tigers of AI" were once very imposing with their rapid iteration and flexible strategies. Kimi once outshone similar products of large companies such as Wenxinyiyan, Doubao, and Tencent Yuanbao in terms of popularity.

But now, the track has reached a watershed. Many industry insiders believe that large companies have gradually taken the lead due to three main reasons: technological iteration, business models, and ecological entry points.

Zhao Jiangjie, a senior person engaged in research on the algorithm direction of Agent applications, explained that developing an AI application generally involves three aspects: the algorithm side, the front - end (interaction interface), and the back - end (integration environment). Among them, the most important is the construction of the algorithm - side ability, that is, the large - model ability.

Now, the ability of large models is undergoing the third iteration.

The birth of ChatGPT represents the 1.0 stage (general dialogue) of large models; then, the post - trained strong - reasoning models combined with reinforcement learning are the 2.0 stage, starting with OpenAI's o1, and the popularity of DeepSeek is a landmark event; with the explosion of Agents this year, large models have entered the 3.0 stage, focusing on Agent ability as the key breakthrough direction of the model. On the basis of the LLM + reinforcement learning route, the model's reasoning ability continues to be expanded to make it more generalizable and versatile, so as to promote the implementation of Agents in more tasks in real - world scenarios.

Zhao Jiangjie said that the abilities of large models of various companies were almost the same in the 1.0 stage. In the 2.0 stage, OpenAI and DeepSeek took the lead in the closed - source and open - source fields respectively, but later other large - model manufacturers also caught up. Currently, the 3.0 stage is still in the early stage of development, facing problems such as the engineering challenge of large - scale RL training, long - term planning, and the construction of a verifiable training environment. However, models like OpenAI's gpt - 5 and Google (DeepMind)'s Gemini deepthink have begun to show extremely strong reasoning abilities and have successively won gold medals in high - difficulty international mathematics competitions such as IMO and IOI, approaching the top human level.

This means that when facing users' regular tasks, all major models can well meet users' needs, and differences only appear in complex tasks. On the premise that the abilities of large models are similar, scenario mining and high - quality data directly determine the upper limit of AI applications. The core data of domestic TOC scenarios are concentrated in the hands of Internet large companies.

As users' demands for complex tasks increase, startups are gradually losing speed.

Meanwhile, the unproven business models have also magnified the pressure on startups.

Zhao Jiangjie said that the costs of AI applications can be divided into early - stage personnel development, mid - stage marketing and promotion, subsequent operation and maintenance, and the computing power consumed when users use the applications. Among them, API calls are the largest expenditure. Although large - model manufacturers are competing on price, it is still a high cost to support a large number of user requests.

However, currently, the willingness of users in the TOC market to pay is generally low.

"Looking at domestic platform - type enterprises, from iQiyi, Youku, and Tencent Video (advertising + membership), Douyin (advertising), Taobao (bidding ranking) to Meituan (merchant promotion fees), advertising and traffic monetization are still the mainstream business models. Getting users to directly pay for AI applications is still a big problem." A senior investor in the AI industry said bluntly that currently, the path for domestic consumers to pay for AI TOC applications is almost unworkable.

Image source / pexels

Startups such as OpenAI, Anthropic, and Perplexity can create multiple leading native AI applications, which is closely related to the relatively high acceptance of AI services by their C - end users. OpenAI once said that ChatGPT is expected to achieve nearly $10 billion in revenue this year.

In contrast, large companies can regard AI applications as part of their overall AI strategy and do not need to care too much about profitability in the short term.

The ecological entry point is another trump card of large companies.

You Yang, the founder of Luchen Technology, said that large companies have brand advantages and a large traffic pool. For example, ByteDance's short - videos, Tencent's social media and games, and Alibaba's e - commerce naturally have a large number of distribution channels. In the short term, they can also ignore the input - output ratio and obtain free users through a large amount of traffic promotion.

More importantly, the non - AI native applications of large companies do not "start from scratch". They have already accumulated a group of users. They only need to carry out "AI upgrades" on the original products to gain users on the existing products and even quickly reach the top with the leading traffic entry points.

For example, after promoting the AI concept, in December 2024, the MAU of Baidu Wenku AI reached 94 million, a year - on - year increase of 216% and a month - on - month increase of 83%.

For startups' native applications to succeed, they rely entirely on "cold start". For example, Kimi once became popular briefly through advertising, but lacked long - term user retention methods and finally could not maintain its popularity.

Last year, Kimi was known as one of the "Three Strong in AI Native Applications" along with Doubao and Wenxiaoyan. According to AppGrowing data, Kimi's single - month advertising expenditure in October last year was 220 million yuan, and the total expenditure in Q4 reached 530 million yuan. However, in Q1 this year, Kimi's download volume decreased by 3.9% month - on - month, with an average monthly download volume of 8.338 million. On the one hand, it was related to the emergence of DeepSeek, and on the other hand, it was also related to the decrease in advertising intensity. Its advertising volume in Q1 was only 150 million yuan.

You Yang said bluntly: "Maybe the single - month advertising and marketing expenses of one product of a large company are higher than the total financing of many startups."

In such a reality, it is imaginable how difficult it is for startups' AI native applications to break through.

Is there still a chance for startups in the AI application market?

Even though AI applications have witnessed explosive growth and the pattern has become relatively stable, overall, its user scale and monetization potential are still in the early stage.

The "2025 AI Application Market Insight Report" released by Sensor Tower shows that in the first half of this year, the global downloads of generative AI applications (AI assistants + AI content generators) reached 1.7 billion, a month - on - month increase of 67%; the in - app purchase (IAP) revenue was close to $1.9 billion, a year - on - year increase of more than 100%; the cumulative user usage time was 15.6 billion hours, equivalent to 86 million hours per day, and the total number of conversations was 426 billion, with an average of about 50 times per person.

These data indicate that both the user stickiness and the willingness to pay for AI applications are significantly increasing, and the entire market is far from reaching its ceiling.

Meanwhile, the lowering of technical thresholds has also given startups more opportunities.

Amin, an industry insider, introduced that with the support of large models, a small development team or even an individual developer can develop a fully - functional AI application in just a few days.

Zhao Jiangjie also mentioned that this year, AI Coding products represented by Cursor, Claude Code, and Codex have developed very rapidly. Benefiting from the improvement of the model's Coding ability and Agent ability, the current AI Coding products can basically cover full - stack development. Through prompt input, one - sentence programming (vibe coding) can be realized to complete the overall project construction of medium - difficulty complexity or below, which can greatly shorten the development cycle of AI applications. Maybe a demo can be made in one or two days.

In this context, it is not impossible for startups to get a share of the pie.

Zhao Jiangjie believes that startups have short decision - making chains and act quickly. They can focus more on polishing a single product and respond to market changes and user feedback much faster than large companies. This characteristic enables startups to easily avoid direct confrontation with large companies and instead