By | Autumn Water Pen Talk

In Q1 2026, Counterpoint Research's global large language model revenue ranking unveiled a new power landscape. Anthropic, with 134 million users, captured 31.4% of the global AI revenue share, with an Average Revenue Per User (ARPU) as high as $16.2. OpenAI had 900 million users, but its ARPU was only $2.2. ByteDance's Doubao, with 345 million monthly active users (MAUs), ranked first domestically but was conspicuously absent from this revenue ranking. Another internet giant, previously criticized for conservative investment and lagging R&D, surprisingly appeared as the top revenue-earning Chinese AI company.

This data reveals a glaring fact: the largest user pools contribute the least revenue, while the smallest user bases capture the largest shares. Moreover, every inference call consumes real computing power, and each new user means a higher bill.

The iron law of zero marginal cost from the internet era has crashed into the "non-zero marginal cost" wall in the AI age. The old logic of burning money for scale is being replaced by a new rule: "logistics lines and monetization efficiency determine survival."

In June 1942, the Battle of Midway began. The Japanese Combined Fleet held advantages 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 Americans were precisely the opposite: the base network in Hawaii and the industrial capacity of the mainland meant their supply lines grew thicker the more they fought.

Today, China's AI industry has reached its own "Midway." China's share of global large model inference calls frequently exceeds 50%. DeepSeek V4 leverages global developers at one-fiftieth the cost, but China's overall AI revenue share is squeezed into single digits globally, its total less than that of a single American company. Behind this data lies not only the strategic games of companies but also the contest between two national industrial approaches.

Two fronts have thus opened: the Eastern Front, where the logistics lines of three corporate giants are under extreme pressure—whose ammunition runs out first, whose line collapses first; the Western Front, a more covert global AI strategy contest—how to win an infinite AI war with finite resources. These two fronts, from the outset, have echoed each other in a pincer movement.

Eastern Front: The Game of Logistics and Monetization Efficiency in a War of Attrition

Tencent, Alibaba, and ByteDance have chosen three distinct paths but face the same test: in a war of attrition with non-zero marginal costs, whose logistics lines are more reliable, whose monetization efficiency is higher?

Tencent: The Efficiency Race in Scenario Monetization

Among Chinese AI companies, Tencent, whose strategic foresight and R&D capabilities have been much debated, boasts the highest monetization efficiency. In the global ranking, Tencent ranks first in China with 114 million users and a $2.9 ARPU, more than double that of Baidu and over four times that of Alibaba, though still nearly half of Microsoft's $5.0. Its secret lies precisely in not making money by selling AI itself.

In Q1 2026, Tencent's revenue was 196.46 billion yuan, up 9% year-on-year. But a more critical comparison reveals: excluding the impact of new AI product investments, operating profit grew 17% year-on-year to 84.4 billion yuan. The new AI product line alone consumed about 8.8 billion yuan in profits that quarter.

Where did this money go? Marketing service revenue grew 20% year-on-year, with AI-driven ad recommendation models as the core engine; enterprise service revenue grew 20% year-on-year, with demand for AI-related cloud services as the main driver.AI is not a revenue item on Tencent's books; it is the catalyst that accelerates existing revenue streams—it makes ads more precise, cloud services easier to sell, and has driven over 20% growth in user time spent on Video Accounts.

The true secret of Tencent's AI competitiveness lies not at the model layer, but in the closed-loop efficiency of "scenario-data-monetization." Upgrading recommendation models doesn't require the strongest general-purpose large model; it requires the most intimate understanding of Tencent user behavior data—this data moat is in Tencent's own hands.

Tencent's strategy is clear and pragmatic: core business profits are the main logistics line. Advertising and games provide ammunition for AI; AI in turn makes ads more precise and gaming experiences better—this is a proven virtuous cycle. Tencent President Martin Lau systematically elaborated Tencent's "AI Economics":"In the AI context, every time we deliver intelligent services to users, it incurs considerable costs." The core strategy is to "find high-value scenarios," not to "blindly acquire massive daily active users."

The longer-term bet is on WeChat Agents, but the timeline has been repeatedly postponed from "full launch in Q3" to "no launch in the short term." This gap measures the depth of the cracks in Tencent's AI supply line: the full potential of WeChat Agents depends on a "significantly better" next-generation Hunyuan model; and Hunyuan's progress depends on the allocation game of computing power among seven or eight projects—Hunyuan training, WeChat AI, Yuanbao, etc. Tencent rarely made a high-profile rebuttal to rumors that its AI "Number One" Yao Shunyu left due to "WeChat taking part of the computing power"—the rumors precisely touched the most fragile link: when limited computing power must be allocated among multiple business lines, who fights for ammunition for the long-term future?

Beneath this crack lies a deeper strategic vulnerability for Tencent. The "hybrid use of models" strategy, where core model capabilities partly rely on external sources, reflects Tencent's consistent pragmatism: leveraging external model capabilities to buy time and develop its own AI application portals. But this also means that once the generational gap in foundational model competitiveness widens, the portals, services, and ecosystems built on these models may shift accordingly. With the lifeline in others' hands, even the strongest muscles can become useless overnight.

Tencent is clearly aware of this issue. Chief Strategy Officer James Mitchell stated on the earnings call that to prioritize internal scenarios, Tencent has "actively delayed the external commercialization of cloud computing power," and "all computing power is given to ourselves." Concentrating computing resources on foundational model R&D and high-value scenario monetization is precisely Tencent's current urgent task.

Alibaba: The Cost Gambit of Full-Stack Self-Reliance

In Q4 FY2026, Alibaba Cloud Intelligence Group revenue was 41.626 billion yuan, up 38% year-on-year; AI product revenue was 8.971 billion yuan, exceeding 30% of the total for the first time, marking the eleventh consecutive quarter of triple-digit growth. CEO Eddie Wu clearly stated: "Full-stack AI technology investment has entered a positive cycle of scaled commercial returns." Yet in the same quarter, Alibaba's adjusted EBITA plummeted 84% year-on-year, with operating profit turning from profit to loss. The food delivery war and the AI arms race are happening simultaneously. Between the sprint of ARR and the cliff-fall of profits lies the real test of "how long is the darkness before dawn."

Alibaba's supply line is "infrastructure depth." The official shareholder letter released on May 20 provides the clearest strategic blueprint for Alibaba's full-stack gamble. The model strategy shifts from single-point breakthroughs to a combined arms approach of agents, world models, and multimodal models. It bets on a core logic: only by achieving end-to-end control from chips to applications can inference costs be reduced to the threshold sufficient for scaled services. Alibaba's full-stack gamble essentially aims to replicate the "Android moment" of the AI era. Controlling the base indirectly controls all upper-layer portals that grow upon it. Google took a decade to turn Android from a cost center to a profit engine. Whether Alibaba can endure the darkness before dawn depends on whether it possesses strategic patience of similar magnitude.

The letter further explicitly lists instant retail as a "core strategic pillar for upgrading the Taobao Tmall platform," with Taobao Flash Purchase becoming a key scenario for AI-driven new user growth and enhanced stickiness. The QWen App for C-end users is deeply integrated with various applications within the ecosystem, including Taobao Tmall, Taobao Flash Purchase, Fliggy, Damai, Amap, Alipay, etc., holding significant resource advantages in mobilizing daily life, service, productivity, and entertainment ecosystem services. Combined with the Wukong enterprise-grade AI work platform forming a B-end C-end dual thrust layout, this likely poses a real threat to Doubao's ambition to build a super app.

A deeper challenge also lies in the organizational battle over which business gets computing power allocation. The widely circulated internal review meeting minutes following the departure of Tongyi Qianwen technical lead Lin Junyang revealed a rift, exposing the computing power shortage for this strategic-level product. As a heavy-asset cloud service provider, Alibaba Cloud must balance guaranteeing its own large model R&D, supporting the group's internal e-commerce AI transformation, and selling computing power to external customers. Resource allocation and the complex multi-business line coordination game objectively exist.

This conflict reveals a structural contradiction in Alibaba's AI: even if the supply chain is built long, if segments block each other, supplies still cannot reach the front lines.

But change is happening. QWen and Taobao Tmall have completed full two-way integration, systematically funneling 166 million MAU users into Taobao's pool of 4 billion products. The B-end AI customer service product "Dian Xiaomi" has already pioneered a paid closed loop. Instant retail, as a new battlefield for AI+e-commerce integration, is extending the "stitching surgery" from customer service tools to core transaction scenarios. Whether this stitching surgery can prove its value during the 618 shopping festival will be the most direct pressure test for Alibaba's ATH strategy.

Also, the bug mentioned in a previous article about outputting useless outlines hasn't been fixed by QWen to this day, which is bizarre. Neglecting C-end user experience is also a problem.

ByteDance: The AI Stress Test of Traffic Logic

ByteDance's approach is the inertial continuation of the "app factory" model from the mobile internet era: simultaneously launching over 20 AI applications across C-end and B-end, covering chatbots, virtual characters, social, images, tools, and more. The logic is straightforward: use traffic to nurture hits, use hits to seize portals, and plan monetization after portals are solidified.

This methodology was repeatedly validated in the mobile internet era, relying on the industry rule of near-zero software replication cost.The AI era has broken this rule: every model call incurs real computing power consumption; the larger the scale, the higher the cost. Doubao's 345 million MAUs represent 345 million active costs burning money every day. This is the deepest dilemma of ByteDance's AI strategy. Meta's situation on the global ranking is an even starker reference: 1 billion users, ARPU only $0.1. It's easy to retain users with free AI; it's hard to make money from free AI.

The scale of C-end losses is far more severe than the absence from the revenue ranking. A reference point: OpenAI's Q1 revenue was $5.7 billion, but its operating loss was as high as $7 billion, losing $1.22 for every $1 of revenue. Its C-end user payment ratio is about 5.5%. In contrast, the payment conversion rate for domestic C-end AI applications is generally below 1%. Some institutions estimate that even if Doubao could achieve ChatGPT's 5.6% payment rate, its annualized revenue might barely cover operational costs; at the actual domestic conversion rate of less than 1%, annual revenue might be less than 10 billion yuan, a drop in the bucket compared to the C-end battlefield burning hundreds of billions in costs per quarter.

A more pertinent question is whether ByteDance is "actively revolutionizing" or "passively defending" in this AI race. Doubao Phone Assistant attempts to take over user operations from the system level—this is precisely ByteDance's most anxious proposition: when users no longer open Douyin to scroll through videos but directly tell AI "find me something fun," will the advertising revenue foundation of the old empire collapse before the new empire is built? Moreover, the "run first, fix guardrails later" model is depleting an asset more precious than traffic—trust. A mistake by an AI Agent could leak your bank password.

A more serious crack is its internally distorting organizational culture. Former ByteDance Seed team researcher Zhang Chi publicly criticized thebenchmaxxing culturewithin Seed after leaving: team leaders evaluate performance based on the benchmarks they are responsible for, everyone is chasing scores, "but this doesn't translate into a good experience in actual use." Furthermore, ByteDance reportedly needs about six months to complete a round of large model training (pre-training plus post-training), while Google is rumored to need only three months. This means the gap might still be widening, not narrowing.

ByteDance's supply line is traffic and main business cash flow, but this line is narrowing—2025 net profit fell over 70% year-on-year, with AI investment voraciously consuming profits. Between open source and closed source, ByteDance is the most unique: Doubao is not open source, but achieves widespread global developer adoption through extreme low pricing. This is the logic of low-price closed source: not open source, yet using price wars to achieve the effect of open source. But price wars have an end. When cash flow is continuously consumed by AI investment, lacking both the ecosystem moat of an open-source community and the premium ability of high-end closed-source clients, how far can this middle-ground strategy go?

The international performance of ByteDance's AI product matrix is also strong. Dola's Q1 2026 downloads exceeded 72 million, cumulative downloads broke 200 million, already ranking among the top global AI assistant applications. AnyGen,对标 Manus, is testing paid subscriptions; Trae is positioned as an AI programming tool, but also faces the supply line test: the more Dola users, the higher the cost of calling external models; the deeper the paid products go, the fiercer the competition with OpenAI and Meta. Overseas paid products currently contribute almost negligibly to the supply line—Gauth's annual revenue is only $14 million, AnyGen is still burning money to acquire users, Dola is completely free. Overseas markets have higher payment willingness, perhaps a future variable, but at least for now, overseas is far from a granary; it's another bottomless pit burning money.

However long the supply line, it cannot replace the monetization efficiency of "living off the land"—the former determines how long you can hold on, the latter determines whether you can win.

Tencent collects rent, Alibaba builds roads, ByteDance surveys land. The essential difference between these three models is not just the distance of AI from money, but the strategic choices based on their own resource endowments.

Tencent's AI hides behind ads and cloud, closest to money; Alibaba's AI sells infrastructure, token consumption is exploding, but turning computing power into profit depends on crossing the threshold of scale effect; ByteDance's AI faces users directly, farthest from money. Being far from money means the highest risk, but also the greatest imagination space—but turning imagination into income precisely tests time and a viable payment closed loop.

Beyond monetization efficiency, the AI super app ByteDance values most, apart from users, may find ecosystem value a more critical variable—and in this aspect, Tencent and Alibaba hold more significant advantages—Who can turn AI conversations into real transaction closed loops can truly convert the "traffic value" of the portal into "rent-collecting ability." It tests not users' willingness to pay, but AI's ability to orchestrate real-world service chains.

WeChat Agents hold the best cards. The WeChat Mini Program ecosystem has accumulated service capabilities from millions of merchants—ordering food, hailing rides, booking appointments, paying bills—theoretically, an AI conversational portal could call all these capabilities, completing the full closed loop from demand to delivery within the chat interface. Calling mini programs isn't hard; the difficulty lies in the AI accurately understanding the user's real intent in specific scenarios and making the correct choice among hundreds of similar mini programs. One wrong recommendation costs the user time and patience; one wrong payment costs trust. Tencent's caution is less a technical issue and more a sober recognition of the error tolerance for 1.4 billion users.

ByteDance's Doubao Phone Assistant attempts another path—skipping mini programs, directly taking over user operations from the system layer. This sounds more radical but also more fragile. The system-level floating ball can recognize screen content, simulate clicks, fill forms, but it faces two fundamental obstacles: first, permission barriers from phone manufacturers—no OS will easily let a third-party app take over system-level interactions; second, the inherent短板 of security trust—an AI that can read your screen and simulate your clicks is also the AI that can scare you the most.

After Alibaba's QWen and Taobao completed two-way integration, users can complete product selection, price comparison, and ordering within conversations, with payment and logistics seamlessly handled by Taobao. This is currently the closest among the three to an "AI+transaction" closed loop, but its ceiling is equally clear: the range of services QWen can mobilize is almost limited to the Alibaba ecosystem.

Whoever occupies the future super portal for human-computer interaction—the new layer replacing browsers, apps, and search boxes—holds the greatest option value.

But imagination also needs to be tested. Its falsification conditions are equally clear: if Doubao Phone Assistant fails to significantly increase user reliance on AI commands within 12 months; if WeChat Agents still have no clear launch timetable by the end of 2026; if after QWen's integration with the Alibaba ecosystem, the transaction order volume and repurchase rate completed by users within conversations fail to grow steadily—then the option value of these "super portals" will be heavily reassessed.

Imagination is not an eternal talisman; it is merely an option with an expiration date. If not exercised by expiry, the premium goes to zero. Option pricing is always an art, not a science, before exercise.

Western Front: The Strategic Divide of "Preserving Land" vs. "Preserving People"

As Tencent advances model open source, as Alibaba's QWen captures over 50% of global open-source downloads, as DeepSeek and Kimi sweep global developers with extreme low prices, the strategic choices of Eastern Front companies are converging into the Western Front's route contest. The logistics line game on the first front is mutually supporting the standards contest on the second front.

In March 1947, Hu Zongnan's 250,000-strong force launched a blitz on Yan'an. At that time, the Northwest Field Army had only 26,000 troops. Mao Zedong decided to withdraw proactively. On the eve of withdrawal, Mao Zedong left a statement later repeatedly quoted:"When our army fights, we do not focus on gaining or losing a city or a piece of land. Preserve the people, lose the land, and both people and land can be preserved; preserve the land, lose the people, and both people and land will be lost." He also gave a vivid analogy: "The enemy advances into Yan'an with a clenched fist. Once in Yan'an, he must spread his fingers, making it easier for us to cut them off one by one." A year later, Yan'an was recaptured. In less than two more years, New China was born.

China's AI industry stands at a similar historical crossroads. Accounting for over half of global inference calls, yet revenue squeezed into single digits—the root of this contradiction lies in China's AI choosing a business model fundamentally different from America's. QWen is open source, DeepSeek is open source, Kimi is open source, Hunyuan is open source. Kimi K2.6 costs only $4 per million tokens, six to eight times cheaper than Claude; Tongyi Qianwen captures over 50% of global open-source model downloads. Low-price closed source and even lower-cost open source are essentially doing the same thing: encircle the cities from the countryside.

The US follows a closed-source premium route. Anthropic, with a $16.2 ARPU, locks in the enterprise high-end market. Foreign Policy's assessment hits the mark:"The real AI race is not a hardware arms race that can be won with the most advanced chips, but about whose models become the default choice in emerging markets."

This is the strategic divide between "preserving land" and "preserving people." The American-style "preserve land" builds high walls, defending the profit high ground of B-end top clients with monopolistic performance. The Chinese-style "preserve people" treats global developers as the core asset, using open source and low prices to encircle people in the vast emerging markets, among SMEs and independent developers, accumulating ecosystem momentum in the countless crevices of long-tail scenarios.

The clenched fist of closed-source models in the high-end market will spread its fingers one by one in the low-end market, while open-source models let developers fine-tune autonomously in countless long-tail scenarios like healthcare, agriculture, cross-border e-commerce, cutting off the spread fingers of closed source one by one. Once these developers' startups grow into giants, path dependency in the tech stack will lock them into China's open-source ecosystem. A deeper possibility is that open-source large models themselves could become the "Android" of the AI era—not as a user portal for direct monetization, but as foundational infrastructure, making all upper-layer applications and Agents naturally run on the Chinese model's tech stack. Whoever controls the base of the developer ecosystem indirectly controls all upper-layer portals created by developers. Android's success wasn't in licensing fees, but in making Google Search, Gmail, and the Play Store the default options for billions of devices globally. The commercial closed loop for open-source large models may mostly come not from the models themselves, but from the cloud services, application markets, and distribution channels that grow on top of them. QWen has been used by the Singaporean government to build national sovereign AI, proving open source can become a standard export.

The profit erosion of American closed-source giants is not the "future tense," but an accelerating "present continuous." After models like DeepSeek and Tongyi Qianwen caught up to GPT-4 level performance, their extremely low token prices directly destroyed the high premium pricing power of closed-source giants. The day when the combined traffic of open-source models comprehensively surpasses that of closed-source models in global mainstream cloud vendor call rankings will be the moment closed-source profit margins permanently decline.

However, whether "preserving people" leads to "preserving both people and land" depends on whether the open-source ecosystem's commercialization closed loop can be realized. The logic of "models free, tax computing power" holds only if developers eventually migrate model calls to the same cloud platform for training and inference. Reality is, open-source models can be deployed on any cloud—QWen's open-source downloads exceeding 50% does not mean Alibaba Cloud's revenue increased by over 50% in sync. Meta's Llama is also a main force in the open-source camp, but Meta itself hasn't gained scaled cloud revenue from it. Ecosystem capture is the first step; ecosystem monetization is the second. The gap between these two steps is precisely the most easily overestimated aspect of the open-source route.

The root of the gap is not technology, but structural differences in payment willingness. Anthropic cut into high-value enterprise productivity scenarios—code generation, long-text analysis, deep reasoning—customers pay for output, not per token. Microsoft Copilot similarly, its $5.0 ARPU is backed by the workflow closed loop within the Office ecosystem. Chinese AI monetization still primarily relies on consumer-grade free scenarios. B-end token consumption is exploding, but the unit value is far from reaching the level of "premium for results."The transition from "pay-per-use" to "value-based pricing" is the essence of the monetization efficiency chasm between China and abroad, and the hardest stretch of the last mile for Chinese AI moving from a token factory to a value highland.

It must also be acknowledged that this chasm cannot be filled by the AI industry alone. Microsoft Copilot can sell for a $5.0 ARPU by leveraging Office 365's hundreds of millions of existing subscribers and B-end channel network—enterprise customers aren't buying AI, but the "AI+Office" workflow closed loop. China lacks an enterprise software ecosystem of comparable magnitude, making the leap from "pay-per-use" to "value-based pricing" miss the most crucial springboard.

The overseas enterprise market has decades of accumulated B-end payment inertia—the US SaaS industry's annual revenue exceeds $300 billion; enterprises are already accustomed to paying for software subscriptions; AI merely adds a new charge on top of this inertia. China's SaaS market still hovers around the $10 billion level; enterprises' willingness to pay for software is far from widespread. AI monetization is tilling thinner soil.

But monetization efficiency is not the whole AI narrative. OpenAI's ARPU is only $2.2, yet its valuation once far exceeded Anthropic's—the capital market was paying for the option value of ChatGPT as the "new portal for human-computer interaction," not for current subscription revenue. Similarly, ByteDance's Doubao, though absent from the ranking, with its 345 million MAUs and system-level positioning of the Phone Assistant, is currently the most aggressive asset closest to the "super portal" imagination among all Chinese AI companies.Monetization efficiency measures "who is making money today"; portal imagination answers "who will have the right to define the rules tomorrow." The two are not a trade-off. For Chinese AI to move from a token factory to a value highland, it needs both Anthropic-style monetization discipline and OpenAI-style portal ambition—without the former, it won't reach the endgame; without the latter, even at the endgame, it can only collect rent, not define rules.

Convergence of Two Fronts: From Token Factory to Value Highland

Chinese companies lack neither technology, nor users, nor scenarios; what they lack is the ability to integrate these three—making the model strong enough, the scenario deep enough, and users willing to pay for the results. On this point, Anthropic and Microsoft have provided a phased answer.

Several signals will determine the direction of the battle in the next 12 months: Can Tencent's API price hike launched in March push its ARPU towards Microsoft's $5.0 line in Q2? Can Alibaba's QWen-Taobao integration convert C-end traffic into effective transactions and retention during its first 618? Can ByteDance's payment tests allow it to escape the "Others" category in the next global ranking?

But the ultimate fate of China's AI industry depends on the combined capabilities of these companies and the DeepSeeks, depends on whether the Western Front's "encircle the cities from the countryside" route contest can navigate the "last mile": making global developers not only use Chinese tokens but also willing to pay for Chinese AI. "Preserving people" is the first step; "preserving both people and land" is the endgame.If this mile cannot be traversed, Chinese AI will forever be a factory, not a brand.

Traversing this mile requires at least three conditions. First, at least one Chinese company runs a viable enterprise payment closed loop—customers willing to pay continuously for AI-driven productivity improvement. Alibaba is limited by the ceiling of its e-commerce scenarios; ByteDance hasn't crossed the chasm from "free users to paying customers." Second, the ecosystem advantage of Chinese open-source models needs to transform into a standards advantage—when global developers become accustomed to QWen's or DeepSeek's toolchain, switching to Chinese cloud services is no longer a "trial" but a "default." Third, whether a Chinese presence can emerge in the global enterprise software market, providing a Microsoft Copilot-style springboard for AI monetization. For Chinese AI's "last mile," the first 500 meters can be run by AI companies themselves; the last 500 meters depend on whether generational upgrades across the entire enterprise services market occur simultaneously.

Midway was only the beginning of the shift from defense to offense, not the endgame. After that, the US military still endured the brutal stalemate of Guadalcanal, the heavy casualties of Iwo Jima. The victor of 1942 took two and a half years to reach Tokyo Bay. Similarly, China's AI industry has a long voyage from "entering a stage of stalemate" to "winning the war."

In the subsequent Battle of Guadalcanal, the later victors had to face not only the Japanese army but also malaria, tropical downpours, and severed supplies. The victor of Waterloo,the Duke of Wellington, said, "Victory is the ability to fight five minutes longer than any other army in the world."And on a longer time axis, this war tests not only whether supply lines are smooth and robust, whether monetization by 'living off the land' is effective, but also strategic resolve and patience.