By | Deep Flow Research Institute
In the past few years, it seems like everyone has been holding the same "map" and searching for the "New Continent" within the AI industry.
This "map" was born at the end of 2022. At that time, just two months after its launch, ChatGPT reached 100 million monthly active users, becoming the fastest-growing consumer-grade product in history. It seemed like everyone felt they had found a "treasure map": the AI era, like the mobile internet era, would ultimately see value converge in a new super-entrance—the Chatbot.
Therefore, the industry widely believed that whoever built the strongest Chatbot first would be seizing the next era. Several years have passed, and the players who bet on Chatbots have found that this "map" did not lead them to the "New Continent."
OpenAI built a Chatbot with over 900 million weekly active users, but it's still losing money. According to The Information, as of Q1 2026, the company loses $1.22 for every dollar of revenue it takes in. Looking back domestically, C-end monetization for Chatbots is still being explored. On May 4th, Doubao, the top Chatbot in China by monthly active users, updated its pricing plans to three tiers, while its basic features remain free. That day, "Doubao charging" trended into the top three on social media, generating significant user reaction.
Anthropic, walking a different path, instead sees the dawn of the "New Continent." In April 2026, Anthropic's annualized revenue exceeded $30 billion, surpassing OpenAI's approximate $25 billion in the same period. The two companies' revenue structures are completely different. According to data from the US business payments platform Ramp, approximately 85% of Anthropic's revenue comes from enterprise clients, while about 85% of OpenAI's revenue comes from individual ChatGPT subscriptions.
As early as April of last year, Anthropic studied about 4.5 million Claude conversation records and found that dialogues involving emotional communication accounted for only 2.9%. The vast majority of uses were work-related. Those who chat with AI all day long remain a tiny minority; most people use AI as a work assistant. A month later, Claude Code, focused on AI coding, officially launched. By early 2026, its annualized revenue had reached $2.5 billion. The "Agent fever" ignited by OpenClaw, which has continued from the beginning of the year until now, also indicates that users don't want a dialog box that chats better, but an executor that can actually help them get the job done.
People are beginning to realize that Chatbot is merely a corridor leading to AGI, not the destination.
1. The larger the DAU, why does it lose more money?
The Chatbot product form became the focus in the past few years largely due to the shock brought by ChatGPT. It allowed ordinary people to see the shape of AI's general capabilities for the first time through a familiar dialog box.
And this dialog box is too similar to a search box: an input field, typing, hitting enter, and getting results. The capital market's initial imagination about Chatbots was built on this similarity. In the internet era, many big businesses were based on entrances, like Google for search and Facebook for social networking.
When ChatGPT looked like the next search box, the market instinctively used the previous script to construct the future: the super-entrance of the AI era has appeared, and whoever occupies it will be the final winner.
But years later, the market began to realize things didn't follow the script. According to QuestMobile data, as of September 2025, native app user scale was 287 million with a Q3 compound growth rate of 3.4%; In-App AI user scale was 706 million with a Q3 compound growth rate of 9.3%, both the scale and growth rate of the latter are larger than the former. In other words, AI may not need a new independent container.
The "super-entrance" was a product of the PC and mobile internet eras, established on the premise that information or services must pass through a unified container to reach users. However, whether the AI era requires a new independent entrance remains questionable. This is because AI is not a revolution at the distribution layer, but at the capability layer; it can seep into all existing products like electricity.
Another iron law of the internet era is also failing for Chatbots. In the past, the market generally recognized that traffic equaled value, meaning the larger the DAU, the bigger the business. This iron law relied on the superposition of several mechanisms: marginal costs approaching zero, network effects, and data flywheels.
The marginal cost of traditional internet products is almost zero. The broadband and server costs consumed by a single search or page load are so small they can be ignored, and serving one more user basically has no incremental cost. Chatbots are the opposite. Each model inference burns real money in computing power; the more people use it, the higher the cost.
Taking OpenAI as an example, user growth is rapid, but so is cash burn. HSBC analysts estimated at the end of 2025 that to support its massive computing needs, OpenAI would need to raise at least $207 billion by 2030, believing OpenAI would continue to incur losses within the next decade, requiring constant financing to subsidize users and pay the high fees to data center owners.
Looking at network effects: in traditional internet products, the addition of the Nth user makes the experience better for the previous N-1 users. For instance, one more person playing a mobile game allows faster team matching; one more merchant on an e-commerce app gives all buyers more options. However, User A writing a thousand prompts has no impact on User B's conversation in a Chatbot.
For Chatbots, the data flywheel also turns weakly. Douyin, Taobao, and Meituan become better with use by feeding user behavior data back into recommendation algorithms. But Chatbots are driven by large model pre-training. User conversation data needs to go back into model training, which involves a long chain, high collection costs, significant noise, and issues of privacy and latency. Moreover, a single Chatbot's user conversation data has limited impact on model capability improvement.
According to LatePost, in early 2025, ByteDance CEO Liang Rubo stated at a company-wide meeting that Doubao had not shown the internet product characteristic of "the more people use it, the better it gets". This company, renowned for its growth engine, also acknowledged its engine was hitting a wall in the Chatbot business.
Ultimately, a Chatbot is something that looks like an internet product but has completely different underlying economics.
2. A Low-Barrier Business
Currently, ChatGPT's commercialization path resembles the traditional internet company logic of "entrance + traffic": first establish the largest-scale general user entrance, then implement tiered monetization on this entrance, such as personal subscriptions, advertising, e-commerce commissions, etc.
The subscription model ChatGPT first tried hasn't been validated yet. Among ChatGPT's 900 million weekly active users in 2025, personal subscribers numbered about 50 million, accounting for only about 5%. A Deutsche Bank research report pointed out that since May 2025, consumer spending on ChatGPT in Europe had already stagnated, suggesting ChatGPT paid user growth might have peaked.
In the Chinese market, this difficulty is multiplied by 3 to 4 times. According to media synthesizing data from a16z, Bessemer, and other institutions, the C-end payment rate for AI products in the North American market is about 15%–40%, while in China it's only 3%–13%, a gap of 3 to 4 times.
Under the long-term influence of the "free + ads" internet model, domestic users haven't developed the habit of paying for standalone software. This May, when Doubao tested subscription plans, "Doubao dumb and still charging" trended. The negative user feedback shows that most domestic users believe Chatbots should be free. According to the latest news from 36Kr, Doubao will officially start charging at the end of June. Proceeding despite the criticism indicates that after massive investment, it's time for chatbots to prove their commercial viability.
The difficulty of the subscription model essentially lies in the low user migration cost of Chatbots—it's a low-barrier business.
One of the moats for internet products is user migration cost. For example, the social graph on WeChat, transaction preferences on Taobao, the service networks built by local merchants on Meituan, etc.
However, the switching cost for Chatbots is very low. The default state of a Chatbot is that users can leave and return anytime, and using two or three Chatbots simultaneously is also possible. Chatbots don't require configuration, learning, or data import. The questioning methods mastered by ordinary users are universal across all Chatbots.
Looking back, the shock ChatGPT brought to the world actually came from the model itself; the real moat of a Chatbot is model capability. A Citi Innovation Lab survey of 1,800 users in March this year also showed that among users willing to pay, 63% listed "access to more advanced models" as the primary driver.
Three years ago, GPT-4 was the most powerful model users could access, with a visible generational gap in capability. But now, various companies' model capabilities are iterating and strengthening. As model capabilities become infrastructural, temporary advantages are less obvious. The shelf life of the most powerful model is getting shorter. When the gap in model capability narrows to the point where ordinary users can't perceive it, Chatbots may degenerate into a cost-performance contest of "whichever is free, use that one."
In a business that requires continuous cash burn, where users can leave anytime, and whose moat is being eroded, it's hard to dig for "gold."
3. The Attention Economy Fails
OpenAI's CEO Sam Altman once called advertising ChatGPT's "last resort."
With the paid subscription path blocked, ChatGPT is no longer holding back. Starting in February this year, ChatGPT began showing ads to users on its free version and lowest-priced paid tier. On May 5th, OpenAI officially launched its self-serve advertising platform, Ads Manager, allowing advertisers to place ads on ChatGPT directly or through agencies.
ChatGPT is referencing the search advertising path. Google made a fortune from search ads. The year before ChatGPT launched, Google's 2021 ad revenue was $208 billion, accounting for 81% of its parent company Alphabet's total revenue.
In February 2023, Microsoft integrated ChatGPT to launch New Bing. Bing's homepage, originally featuring a thin search bar, was replaced by a large dialog box reading "ask me anything," essentially handing the search engine entrance over to a Chatbot. Microsoft CEO Satya Nadella once said, "we're going to make Google dance." Microsoft's public challenge to Google was precisely eyeing the advertising monetization potential of Chatbots.
However, the search advertising potential of Chatbots hasn't been as high as expected. Data from Statcounter shows that from 2024 to April 2026, Bing's global search share increased only from about 3.4% to about 5.1%.
The premise for search advertising is that when users search, they have clear purchase intent; search results are a list where multiple ad slots can be inserted; users don't necessarily expect the answer to be correct, just relevant.
Chatbots lack all three of these premises. User interaction with Chatbots is more about answering, explaining, emotional responses, etc., naturally lacking purchase intent. Secondly, Chatbots provide a single answer, leaving no room to insert additional ads.
This is also why OpenAI's advertising strategy initially used CPM (cost per thousand impressions) and later introduced CPC (cost per click). According to The Information, ChatGPT's initial target CPM was as high as $60, comparable to premium ad slots like streaming TV, but some advertisers actually paid CPMs of only $15 to $25, possibly reflecting too few buyers bidding for ad space. Advertisers are accustomed to performance-based payments and precise targeting, and the conversational nature of Chatbots is difficult to fit into the traditional digital advertising framework.
More crucially, users expect Chatbots to provide correct answers. Once an answer contains an advertisement, users' trust in every response is discounted. This trust is the core of the product itself, making advertisers feel conversions are impossible.
Perplexity has already proven this path is hard. In 2024, this Chatbot-powered search engine company launched ad formats like Sponsored Follow-up Questions. However, Perplexity's ad revenue that year was about $20,000, less than 0.1% of its total revenue of $34 million. In February this year, Perplexity formally abandoned its ad model.
Essentially, Chatbots break the dependency path of the attention economy's monetization in the mobile internet era. In the past, attention was scarce, and content supply was cheap. But Chatbots reverse this structure: each answer costs computing power, making supply expensive. Meanwhile, a single session only takes a few minutes; users ask and leave, making attention less valuable. The more expensive the supply and the shorter the attention span of a business, the harder it is to survive on advertising.
However, AI advertising is not without opportunity. As of Q3 2025, Google AI Overviews had covered over 2 billion users, and AI Mode had over 75 million daily active users. Both features embedded ads. In the same quarter, Alphabet delivered its first-ever quarter with revenue exceeding $100 billion, with Google Search & other revenue growing 15% year-over-year to $56.6 billion. This is one method currently proven viable for AI ads: embedding AI into an already established commercial system, rather than starting a separate dialog box.
Currently, domestic Chatbots haven't attempted to integrate ads. Investor Zhuang Minghao discussed the reasons in a recent podcast with guests. They pointed out that existing ad systems are based on keyword matching from search. To form associations with user inputs involves data desensitization issues, facing significant regulatory pressure.
Additionally, Chatbots are exploring e-commerce shopping monetization. Following Alibaba's Qianwen integrating with Taobao for AI shopping features, according to 36Kr, Doubao will also connect with Douyin's e-commerce next, attempting to close the AI shopping loop. As early as last September, ChatGPT launched an "Instant Checkout" function but canceled it five months later. Similar to search ads, shopping within Chatbots faces issues like consumer demand and user trust. However, while ChatGPT integrated with scattered third-party e-commerce, Qianwen and Doubao integrate with their own complete e-commerce ecosystems. Whether domestic Chatbots can succeed on this path remains an open question.
4. Chatbot is an Intermediate Form of AI Development
In Q1 2026, ChatGPT's month-over-month active user growth rate was 6.78%. A year earlier in the same period, this number was 18%.
The domestic situation is similar. QuestMobile data shows that by March 2026, monthly active users of AI-native APPs reached 440 million. Industry monthly average usage frequency and duration per user were 87.1 times and 173.3 minutes, respectively. Based on this calculation, the average daily usage duration per user across the entire industry is less than 6 minutes. In the same report, Douyin's average daily usage per user is 1.5 hours, over ten times the former.
The development potential of Chatbots may have been overestimated. The value of a Chatbot lies in providing "general conversation." This means many AI capabilities cannot be expressed within such a product form.
Chatbots structurally confine AI's capabilities within a turn-based cage. An NBER study based on 1.5 million ChatGPT conversations showed that up to 49% of user interactions with Chatbots fall under the "Asking" category. User asks, AI answers, session ends, state resets. It's a passive response mode, unable to execute multi-step tasks, call external tools, or work continuously in the background. Yao Shunyu, who has worked at both Anthropic and Google, recently lamented in a podcast that AI's capabilities are so powerful, yet people only use it to ask questions.
The aforementioned NBER research also indicates that 40% of user interactions with Chatbots are starting to move toward "Doing." When users discover AI can do more and more things, they tend to explore more of its uses. Therefore, one evolutionary direction for Chatbots is "Doing." This means Chatbots need to develop Agent capabilities, such as multi-step execution, tool calling, background operation, memory, goal orientation, etc.
But the paradox is, once it develops these capabilities, it is no longer a pure Chatbot. And a harsher reality is that not all Chatbots can complete this transformation, as it requires simultaneous upgrades in underlying models, Agent architecture, ecosystem integration, and other capabilities.
A more distant imagination is that the future of AI might not even need a standalone native App.
For example, AI will embed into existing Apps. The integration path of OpenClaw already hints at this. Its interface is WeChat, WhatsApp, etc., which people use daily. Users send messages to the Agent within these apps just like they would to colleagues.
Or, AI will embed into operating systems. For instance, the personal intelligence system Apple Intelligence launched by Apple in April this year for iPhone, iPad, and Mac. AI might even embed into hardware. Just last September, Meta released the Ray-Ban Display AI glasses with a screen, where users don't need to open an App or use a phone.
The industry once thought only native AI applications were the future. But when AI starts embedding into social Apps, OS, and various hardware, more possibilities emerge for how AI truly lands.
In the AI era, if you still hold the "old map," you won't find the "New Continent." Only by updating the map can you possibly find a truly valuable continent.
