The Eastern Front of Chinese AI: From Yan'an to Midway
In Q1 2026, the global large language model revenue list released by Counterpoint Research unveiled a new power landscape. Anthropic, with 134 million users, captured 31.4% of the global AI revenue share, and its Average Revenue Per User (ARPU) reached as high as $16.2. OpenAI had 900 million users, but its ARPU was only $2.2. Doubao of ByteDance, with a monthly active user base of 345 million, ranked first in China, yet it was absent from this revenue list. Another internet giant, which has been criticized for its conservative investment and lagging R & D, unexpectedly topped the list of Chinese AI companies in terms of revenue.
This set of data exposes a glaring fact: the largest user pool contributes the least revenue, while the smallest user group grabs the largest share. Moreover, every inference call consumes real computing power, and every new user means a higher bill.
The iron law of the internet that marginal cost is zero has hit the iron wall of "non - zero marginal cost" in the AI era. The old logic of burning money for scale is being replaced by the new rule that "supply lines and monetization efficiency determine survival."
In June 1942, the Battle of Midway broke out. The Japanese Combined Fleet had an advantage in tonnage and experience, but its supply lines stretched thousands of nautical miles from the homeland. Every sortie consumed fuel and ammunition that were difficult to replenish. The US military, on the contrary, had a base group in Hawaii and the industrial production capacity of the homeland, which made its supply lines stronger as the battle went on.
Today, the Chinese AI industry has reached its own "Midway." China accounts for more than 50% of the total global calls of large AI models. DeepSeek V4 leverages global developers with one - fiftieth of the cost. However, the overall revenue of Chinese AI in the global share is squeezed to a single - digit percentage, and the total is less than that of a single US company. Behind this set of data lies not only the game of corporate business strategies but also the game of two national industrial routes.
Two battlefronts have thus emerged: on the eastern front, the supply lines of three corporate giants are under extreme test - whoever runs out of ammunition first will see their defense line collapse first; on the western front, a more covert global AI route game is unfolding - how to win an infinite AI war with limited resources. From the very beginning, the two lines have echoed each other in a pincer - attack formation.
Eastern Front: The Game of Supply and Monetization Efficiency in the War of Attrition
Tencent, Alibaba, and ByteDance have chosen three completely different paths, but they are facing the same question: in the war of attrition where marginal cost is non - zero, whose supply line is more reliable and whose monetization efficiency is higher?
Tencent: The Efficiency Competition of Scenario Monetization
Among Chinese AI companies, Tencent, which has been controversial in terms of strategic foresight and R & D ability, actually has the highest monetization efficiency. On the global list, Tencent, with 114 million users and an ARPU of $2.9, ranks first in China, more than twice that of Baidu and more than four times that of Alibaba, but still nearly half the distance from Microsoft's $5.0. However, its secret lies precisely in not making money by selling AI itself.
In Q1 2026, Tencent's revenue was 196.46 billion yuan, a year - on - year increase of 9%. But more importantly, here is a comparison: if the impact of investment in new AI products is excluded, the operating profit increased by 17% year - on - year, reaching 84.4 billion yuan. The new AI product line consumed about 8.8 billion yuan in profit in a single quarter.
Where did this money go? The revenue from marketing services increased by 20% year - on - year, and the AI - driven advertising recommendation model was the core engine; the revenue from enterprise services increased by 20% year - on - year, and the demand for AI - related cloud services was the main increment. AI is not an income item on Tencent's ledger but a catalyst to accelerate the operation of existing income items - it makes advertising more accurate, cloud services easier to sell, and the user time on Video Accounts increase by more than 20%.
The real secret of Tencent's AI competitiveness lies not in the model layer but in the closed - loop efficiency of "scenario - data - monetization." Upgrading the recommendation model does not require the strongest general large model but a data closed - loop that understands Tencent users' behavior best - this data barrier is in Tencent's own hands.
Tencent's strategy is clear and practical: the profit from core business is the main supply line. Advertising and games provide ammunition for AI, and AI in turn makes advertising more accurate and the game experience better - this is a proven positive cycle. Tencent President Liu Chiping systematically elaborated on Tencent's "AI economics": "In the AI scenario, every time an intelligent service is delivered to users, it will incur considerable costs." The core strategy is "to find high - value scenarios" rather than "blindly acquire a large number of daily active users."
The bigger bet is on the WeChat intelligent agent, but the schedule has been repeatedly postponed from "full - scale launch in Q3" to "not to be launched in the short term." This gap measures the depth of the crack in Tencent's AI supply line: the full potential of the WeChat intelligent agent depends on a "significantly better" next - generation Hunyuan model; and the progress of Hunyuan depends on the allocation game of computing power among seven or eight projects such as Hunyuan training, WeChat AI, and Yuanbao. Tencent rarely and publicly refuted the rumor that Yao Shunyu, the "No. 1 person" in AI, left because "WeChat took away some computing power" - the rumor touched on the most vulnerable part: when limited computing power needs to be allocated among multiple business lines, who will fight for ammunition for the long - term future?
Under this crack lies Tencent's deeper strategic hidden danger. The "mixed model - using" strategy, where the core model ability partially depends on the outside, reflects Tencent's consistent pragmatism: borrowing external model capabilities to gain time and develop its own AI application entrances. However, this also means that once the generational gap in the competitiveness of basic models widens, the entrances, services, and ecosystems built on these models may all shift. If the lifeblood is in others' hands, even the strongest muscles may become useless overnight.
Tencent clearly realizes this problem. Chief Strategy Officer James Mitchell admitted in the earnings conference call that to prioritize internal scenarios, Tencent "actively postponed the external commercialization of cloud computing power," "giving all the computing power to itself." Concentrating computing power resources on the R & D of basic models and monetization in high - value scenarios is currently Tencent's top priority.
Alibaba: The Cost Gamble of Full - Stack Self - Development
In Q4 of fiscal year 2026, the revenue of Alibaba Cloud Intelligence Group was 41.626 billion yuan, a year - on - year increase of 38%; the revenue from AI products was 8.971 billion yuan, accounting for more than 30% for the first time, and it has achieved three - digit growth for the eleventh consecutive quarter. CEO Wu Yongming clearly stated: "The full - stack AI technology investment has entered a positive cycle of large - scale commercial returns." However, in the same quarter, Alibaba's adjusted EBITA dropped by 84% year - on - year, and the operating profit turned from profit to loss. With the food delivery war and the AI arms race going on simultaneously, between the soaring ARR and the cliff - like drop in profit, there lies a real question of "how long is the darkness before dawn?"
Alibaba's supply line is "infrastructure depth." The official shareholder letter released on May 20 provided the clearest strategic map for Alibaba's full - stack gamble. The model strategy has shifted from single - point breakthrough to group - army operations of intelligent agents, world models, and multimodal models. It bets on a core logic: only by achieving full - link control from chips to applications can the inference cost be reduced to the critical point for large - scale services. Alibaba's full - stack gamble is essentially a replication of the "Android moment" in the AI era. Controlling the base indirectly controls all the upper - layer entrances growing on it. It took Google ten years to turn Android from a cost center into a profit engine. Whether Alibaba can survive the darkness before dawn depends on whether it has the same level of strategic patience.
The letter also clearly listed instant retail as the "core strategic pillar for the upgrade of the Taobao and Tmall platforms." Taobao Flash Sale has become a key scenario for AI - driven new user growth and enhanced stickiness. The Qianwen App for C - end users is deeply integrated with various applications in the ecosystem, including Taobao, Tmall, Taobao Flash Sale, Fliggy, Damai, Gaode, and Alipay. It has significant resource advantages in mobilizing ecological services for users in daily life, services, productivity, and entertainment fields. Together with the Wukong enterprise - level AI work platform, it forms a layout that focuses on both B - end and C - end users. It is likely to pose a real threat to Doubao's plan to build a super entrance.
The deeper challenge also lies in the business organization game of computing power allocation. The widely circulated minutes of an internal review meeting after Lin Junyang, the technical leader of Tongyi Qianwen, left the company revealed a crack and exposed the shortage of computing power for Qianwen, a strategic product. As a heavy - asset cloud service provider, Alibaba Cloud needs to make multiple trade - offs between ensuring the R & D of its own large model, supporting the AI transformation of the group's internal e - commerce, and selling computing power to external customers. The resource allocation and the collaborative game among a large number of multi - business lines objectively exist.
This conflict reveals the structural contradiction of Alibaba's AI: no matter how long the supply chain is built, if there are obstacles between different sections, the supplies still cannot be delivered.
However, changes are taking place. Qianwen has completed full - scale two - way interconnection with Taobao and Tmall, and 166 million monthly active users have begun to be systematically introduced into Taobao's pool of 4 billion products. The B - end AI customer service product "Dian Xiaomi" has taken the lead in running through the paid closed - loop. As a new battlefield for the integration of AI and e - commerce, instant retail is extending the "suture operation" from customer service tools to core transaction scenarios. Whether this suture operation can prove its value in the 618 promotion will be the most direct stress test for Alibaba's ATH strategy.
By the way, the bug of generating useless outlines, which was mentioned in a previous article, still hasn't been fixed by Qianwen. It's really strange. The neglect of C - end experience is also a problem.
ByteDance: The Big Test of AI under the Traffic Logic
ByteDance's approach is an inertial continuation of the "application factory" model in the mobile internet era: it has launched more than 20 AI applications simultaneously on the C - end and B - end, covering multiple categories such as chatbots, virtual characters, social media, pictures, and tools. The logic is straightforward: use traffic to cultivate popular products, use popular products to seize entrances, and then seek monetization after the entrances are stabilized.
This methodology was repeatedly verified in the mobile internet era, relying on the industrial law that the software replication cost tends to zero. The AI era has broken this law: every model call consumes real computing power, and the larger the scale, the higher the cost. The 345 million monthly active users of Doubao mean 345 million active costs that burn money every day. This is the deepest dilemma faced by ByteDance's AI strategy. On the global list, Meta's situation is a more severe reference: with 1 billion users, its ARPU is only $0.1. It is easy to attract users with free AI, but it is difficult to make money from free AI.
The scale of losses on the C - end is far more serious than its absence from the revenue list. A comparable benchmark is that OpenAI's revenue in the first quarter was $5.7 billion, but its operating loss was as high as $7 billion, with a loss of $1.22 for every $1 of revenue. The proportion of paid users on its C - end is about 5.5%. The paid conversion rate of domestic C - end AI applications is generally less than 1%. Some institutions estimate that even if Doubao can reach the 5.6% paid rate of ChatGPT, its annualized revenue can only barely cover the operating costs; if calculated based on the actual conversion rate of less than 1% in China, the annual revenue may be less than 10 billion yuan, and this income is a drop in the bucket compared with the cost of burning tens of billions of yuan in the C - end battlefield every quarter.
What is more worthy of questioning is whether ByteDance is "actively revolutionizing" or "passively defending" in this AI race. The Doubao mobile assistant tries to take over user operations from the system level - this is exactly the most worrying proposition for ByteDance: when users no longer open Douyin to watch videos but directly tell AI "find me something interesting," will the advertising revenue foundation of the old empire collapse before the new empire is built? Moreover, the model of "running first and then fixing the fence" is over - consuming an asset more precious than traffic - trust. A mistake of an AI agent may leak your bank password.
An even more serious crack is the gradually deformed organizational culture within the company. Zhang Chi, a former researcher of ByteDance's Seed team, publicly accused the benchmaxxing (score - brushing) culture within Seed after leaving the company: team leaders evaluate performance based on the benchmarks they are responsible for, and everyone is chasing scores, "but this cannot be translated into a good experience in actual use." Moreover, it takes about half a year for ByteDance to complete a round of large - model training (pre - training plus post - training), while Google is rumored to only take three months, which means the gap may be widening rather than narrowing.
ByteDance's supply line is traffic and the cash flow from its main business, but the supply line is narrowing - in 2025, the net profit decreased by more than 70% year - on - year, and AI investment is devouring profits crazily. Among companies in the open - source and closed - source camps, ByteDance is the most unique one: Doubao is not open - source, but it has achieved wide access for global developers through extremely low prices. This is the logic of low - price closed - source: not open - source, but achieving the effect of open - source through price war. However, price war always has an end. When the cash flow is continuously devoured by AI investment, without the ecological moat of an open - source community and lacking the premium ability of high - end closed - source customers, how far can the strategy in the middle go?
The international performance of ByteDance's AI product matrix is also strong. Dola's download volume exceeded 72 million times in Q1 2026, and the cumulative downloads have exceeded 200 million times, ranking among the top global AI assistant applications. AnyGen is testing paid subscriptions, competing with Manus. Trae is positioned as an AI programming tool, but it also faces the test of the supply line: the more users Dola has, the higher the cost of calling external models; the deeper the paid products go, the more intense the competition with OpenAI and Meta will be. Currently, the contribution of overseas paid products to the supply line is almost negligible - Gauth's annual revenue is only $14 million, AnyGen is still burning money to attract users, and Dola is completely free. The overseas market has a higher willingness to pay, and it may become a variable in the future, but at least for now, the overseas market is not a granary but another bottomless pit for burning money.
No matter how long the supply line is, it cannot replace the monetization efficiency of "gaining resources from the enemy" - the former determines how long you can hold on, and the latter determines whether you can win.
Tencent collects rent, Alibaba builds roads, and ByteDance measures land. The essential differences among the three models are not only the distance between AI and money but also strategic choices based on their own resource endowments.
Tencent's AI hides behind advertising and cloud services and is closest to money; Alibaba's AI sells infrastructure, and the consumption of Tokens is booming, but to turn computing power into profit, there is