On August 12, Tencent released its latest earnings report. Quarterly capital expenditure had reached RMB 52.8 billion up 176% from a year earlier.
Tencent’s shares fell 4.46% the next trading day.
A week later, Alibaba reported an even larger number. Its quarterly capital expenditure reached RMB 67.7 billion, driven largely by AI infrastructure. Both companies saw free cash flow turn negative as they bought chips, expanded data centres and prepared for the next stage of the AI race.
Investors are now facing a difficult question: how should all this spending change the valuation of China’s two largest technology companies?
I find the comparison useful not because Tencent and Alibaba must defeat each other, but because they are approaching the same market from almost opposite directions. Tencent and Alibaba were the infrastructure builders of China’s internet era. Both companies now have to rebuild parts of their own empires around AI — a process that will inevitably create disruption, uncertainty and painful adjustments.
I call this an “earnings crossfire.” A financial report is not merely a collection of numbers. It also reveals what management wants investors to notice, which questions executives avoid and how they explain difficult strategic decisions.
After reading both reports and earnings calls, the difference is becoming clear.
Tencent is trying to integrate AI into the businesses it already owns. Alibaba is trying to build an entire AI stack, from chips and cloud infrastructure to foundation models and applications.
Tencent’s approach is safer but increasingly fragmented. Alibaba’s is more coherent, but far more dependent on the AI boom continuing.
Tencent AI’s Strategy: Model, Agent and APP #
Tencent’s AI strategy currently revolves around three main products: Hunyuan is the model, WeLM is for agent, and WorkBuddy is the product.
WeLM and Xiaowei are designed to bring AI agents into WeChat. Hunyuan is Tencent’s foundation-model platform. WorkBuddy is its productivity assistant for office workers.
Hunyuan represents Tencent’s infrastructure-level AI ambition.
After former OpenAI researcher Yao Shunyu joined the company, the model’s performance improved considerably. Tencent says it wants Hunyuan to reach frontier-level multimodal capability within 12 to 18 months.
Unlike consumer AI companies chasing user numbers, Tencent expects Hunyuan to improve advertising recommendations, enterprise software, game development, AI workflows, 3D asset generation and WorkBuddy inference.
But Tencent’s greatest AI asset is not its model. It is WeChat. With more than one billion users, WeChat is one of the world’s most valuable consumer platforms. Tencent’s long-term goal is to turn it from a communication tool into something closer to an AI operating layer.
WeLM is being developed partly around this idea. Unlike a conventional cloud chatbot, it places greater emphasis on device-side processing, privacy and low latency. The approach is closer to Apple Intelligence than ChatGPT.That is a very Tencent-style strategy: use a new technology to strengthen an existing empire rather than build a separate one.
WorkBuddy is Tencent’s attempt to build a serious AI productivity product.
Unlike OpenAI’s Codex or Anthropic’s Claude Code, it did not initially target software developers. Tencent went after ordinary office workers, a much larger potential market.
The challenge is monetization. Chinese enterprise software has historically struggled to generate high margins. A large user base does not automatically become a profitable software business.
WorkBuddy’s growth looks promising. Desktop monthly visits exceeded 20 million in June, ranking first among similar products in China.
Tencent strategy chief James Mitchell said gross margins from paying WorkBuddy users and model services were already comparable with Tencent Cloud.
That sounds encouraging until you ask what Tencent Cloud’s margin actually is.
Tencent does not disclose it separately. Analysts often estimate that cloud infrastructure businesses can produce gross margins of around 40 % and EBITA margins of roughly 10 %.
More importantly, Mitchell was referring mainly to paying WorkBuddy users. Most individuals still access the product through free credits.
Good economics among paying customers do not necessarily mean the entire product is profitable.
Tencent still needs to show how many users convert to subscriptions, how much computing capacity each user consumes and whether an AI productivity product can generate sustainable margins in China.
Tencent Is Spending Like an AI Company #
Two numbers disappointed Tencent investors.
Under IFRS, profit attributable to shareholders reached RMB 56 billion, below the Reuters consensus estimate of approximately RMB 61.8 billion.
More unusually, free cash flow was negative RMB 13.8 billion. It was the first negative quarter in Tencent’s reporting history.
Tencent generated RMB 52.7 billion in operating cash flow. That was more than offset by RMB 59.3 billion in capital-expenditure payments, RMB 5 billion in media-content payments and RMB 2.2 billion in lease payments.
A large proportion of the additional spending came from commitments for AI computing resources.
The effect also appeared in operating profit. Tencent reported non-IFRS operating profit of RMB 75.6 billion. Without Hunyuan, Yuanbao, WorkBuddy, CodeBuddy, Xiaowei and its other new AI initiatives, the figure would have been approximately RMB 86.1 billion.
In other words, Tencent’s AI businesses reduced quarterly operating profit by around RMB 10.5 billion.
That is a meaningful change for a company known for producing stable profits from games, advertising, payments and social networking. Tencent is becoming more capital intensive.
The question is whether that transformation produces enough growth.
Over the next six to 12 months, Hunyuan needs to become more competitive with Qwen and Kimi. WorkBuddy needs more paying customers. Xiaowei needs to prove that agents inside WeChat can create measurable economic value.
Until then, Tencent is spending like an AI infrastructure company while continuing to operate like a traditional internet company.
What Exactly Is Tencent Building? #
Every AI company needs computing capacity. The question is whether owning that capacity becomes a business in itself.
Tencent President Martin Lau described three possible uses for the company’s computing resources: training internal models, supporting AI applications and renting capacity to outside cloud customers.
James Mitchell added that previously purchased hardware could itself generate returns because computing resources ordered earlier may now be worth considerably more.
The logic is reasonable. The disclosure is not.
Investors still cannot answer a basic question: what exactly is Tencent’s AI infrastructure for?
Is Tencent trying to become a larger cloud provider? Is it building Hunyuan into a frontier model? Or is it mainly buying capacity to support WorkBuddy, WeChat agents and other internal applications?
Tencent Cloud does not provide enough financial detail to answer these questions. After spending tens of billions of yuan, Tencent continues to present AI as one combined story.
Investors cannot clearly separate cloud revenue, model-development costs, application revenue or the economics of renting computing capacity.
Without that distinction, it is difficult to know whether Tencent is building a new business or merely increasing the cost of its existing one.
Alibaba Is Building the Entire Stack #
Alibaba is taking almost the opposite approach.
While Tencent is adding AI to its existing ecosystem, Alibaba is reorganizing itself around AI infrastructure. In its latest results, Alibaba placed Qwen and Qwen Office under “AI Labs and Applications,” moved semiconductor subsidiary T-Head into “AI Cloud and Computing Services,” and integrated instant retail more closely with e-commerce.
Compared with the company’s previous “Cloud Intelligence” structure, the new organization sends a much clearer message.
Alibaba does not see AI as merely a model business. It sees AI as an infrastructure business.
The strategy increasingly resembles Amazon’s approach to AWS. Amazon does not need to dominate every application. It wants to provide the computing infrastructure on which everyone else builds.
Alibaba wants to occupy that position in China. It has one additional advantage: it controls both the infrastructure and one of the country’s leading foundation models, Qwen.
Spend First, Profit Later #
Alibaba Cloud generated quarterly revenue of RMB 48.4 billion, up 45 % from a year earlier.
AI-related product revenue reached RMB 12.4 billion, up 38 %. Annual recurring revenue reached RMB 49.6 billion.
For the first time, Alibaba also revealed part of the financial burden created by Qwen and Qwen Office. Its AI businesses produced approximately RMB 13.9 billion in quarterly losses, mainly because of inference costs and infrastructure investment. CEO Eddie Wu remained confident.
Alibaba has committed RMB 380 billion to AI infrastructure over three years. Wu described the current spending as a temporary investment cycle rather than a permanent increase in costs.
His argument is counterintuitive: the more Alibaba spends today, the lower its future AI costs may become.
The assumption is that demand for computing capacity will continue to exceed supply before 2030.
If an accelerator has a depreciation period of five years but earns back its cost within three, Alibaba can continue generating returns from the hardware for another two years. Older Nvidia V100 and A100 chips remain heavily utilized even after newer generations arrive. Alibaba therefore believes that today’s infrastructure can continue producing value for years.
It is an attractive argument, but it depends on one critical assumption: AI demand must remain strong.
If computing demand keeps growing, early investment creates an advantage. If demand slows, the same infrastructure becomes a very large depreciation burden.
T-Head May Be Alibaba’s Hidden Advantage #
Alibaba’s most important AI asset may not be Qwen. It may be T-Head.
The semiconductor subsidiary allows Alibaba to design its own accelerators, reducing its dependence on outside suppliers and potentially lowering computing costs.
Alibaba says its second-generation domestic AI chips can support large-model training without a significant loss of performance. Known manufacturing partners include Hua Hong Semiconductor, while Alibaba-designed chips have already been deployed in China Unicom’s Qingyang project.
Tencent and ByteDance cannot currently do the same at comparable scale.
ByteDance is reportedly exploring customized CPUs with Qualcomm, but the project remains at an early stage. Tencent has not publicly disclosed its AI-accelerator suppliers.
Huawei has built a more independent technology stack, but its architecture requires developers to make larger adjustments. Alibaba’s GPGPU-based approach may fit more easily into existing AI-development workflows.
The weakness is scale. T-Head remains tiny compared with Nvidia’s ecosystem, and Alibaba does not yet produce enough chips to satisfy its own demand.
But if production expands, T-Head could become one of the most valuable parts of Alibaba’s AI strategy.
Alibaba Cloud Needs Customers #
Alibaba Cloud is already China’s largest public-cloud provider.
After integrating AI services and computing infrastructure, cloud revenue grew 45 % from a year earlier, slightly faster than Microsoft Azure’s reported 43 % growth.
Alibaba’s AI Cloud and Computing Services segment produced an EBITA margin of 11.6 %.
Wu argues that Alibaba’s scale and internally developed chips will reduce costs, allowing it to offer more competitive prices while improving margins.
The problem is that server rental alone cannot support Alibaba’s ambitions.
Basic infrastructure rental may generate margins of around 20 to 40 %. Model-as-a-Service can earn more. Enterprise AI agents could earn more still.
Alibaba therefore needs customers to buy higher-value services such as models, agents, databases and deployment tools. It cannot simply rent accelerators.
If token prices fall sharply, ByteDance’s Volcano Engine starts another price war or China eventually builds too much GPU capacity, Alibaba’s economics will become much weaker. Alibaba is effectively betting that AI infrastructure will remain scarce.
Qwen Is Powerful but Still Needs a Business Model #
Qwen is one of China’s strongest foundation models.
Alibaba says Qwen-related products have reached around 250 million users. The model has been integrated into Taobao’s search and recommendation systems, as well as enterprise applications.
What Alibaba has not disclosed is how much direct revenue Qwen produces.
That matters.
Alibaba e-commerce chief Jiang Fan said AI had improved Taobao search, recommendations and merchant tools. Those improvements may strengthen the existing business, but they do not necessarily create a new source of revenue.
Accio Work looks more promising. The AI assistant for global business-to-business commerce has attracted around 50,000 paying users. It helps companies research markets, find overseas customers and automate workflows.
But one successful enterprise product cannot carry the cost of a national-scale foundation model.
Alibaba faces the same problem as every AI company. Users love AI products. Turning that enthusiasm into recurring revenue is much harder.
Alibaba Does Not Need to Win the Chatbot War #
Eddie Wu does not appear to believe that selling model APIs will become the industry’s main source of profit. Alibaba instead wants to combine chips, cloud infrastructure, computing services and foundation models. Under this structure, Qwen is not necessarily the final product. It is one capability inside a much larger platform.
This again resembles Amazon. AWS does not need every customer to use Amazon’s own models. It only needs those customers to rent computing infrastructure.
Alibaba wants the same position in China.
Customers can use Qwen or third-party models through Alibaba Cloud. From Alibaba’s perspective, the infrastructure generates revenue regardless of which model the customer chooses.
Wu said most revenue-generating models on Alibaba’s MaaS platform still belong to the Qwen family. He expects MaaS to reach RMB 30 billion in annual recurring revenue.
But this creates another tension.
If Qwen mainly exists to attract cloud workloads, Alibaba does not need it to become China’s dominant consumer assistant.
If Alibaba wants Qwen to compete directly with ChatGPT, Claude or ByteDance’s Doubao, it needs a much stronger consumer strategy.
Right now, Qwen sits between those two identities.
Alibaba’s Consumer AI Problem #
Alibaba’s consumer AI products have struggled to create the same momentum as their competitors.
Qwen’s office products have updated more slowly than Tencent’s WorkBuddy. The Qwen app reportedly has around 251 million monthly active users, below ByteDance’s Doubao, which has exceeded 300 million.
But user numbers alone can be misleading.
Some Qwen users were attracted through promotions, including discounted products and drinks. That creates activity. It does not necessarily create a willingness to pay.
ByteDance has faced the same problem with Doubao. A large free-user base can help improve models, but it is also expensive to maintain.
If Alibaba wants to control Qwen’s losses, it may eventually have to reduce its spending on consumer applications.
Alibaba’s Biggest Risk Is Financial #
The greatest risk to Alibaba’s AI strategy is not technical. It is financial.
AI infrastructure requires enormous upfront investment. Alibaba needs its traditional businesses to keep generating enough cash to support it.
But its e-commerce engine is slowing.
China commerce-management revenue declined 7 % from a year earlier, while overall e-commerce growth remained around 4 %.
The situation is not disastrous. Alibaba and JD.com both improved after China’s aggressive food-delivery subsidy competition began to cool.
Jiang Fan said customer-management revenue and EBITA growth improved from the previous quarter. Instant-retail losses are also expected to stabilize or narrow.
That provides stability. But stability is not the same as growth.
E-commerce may prevent Alibaba from facing additional financial pressure. It may not produce enough new profit to fund an unlimited AI infrastructure race.
Alibaba therefore needs its AI investments to generate returns before its traditional businesses weaken further.
Two Companies, Two Kinds of Risk #
Tencent and Alibaba represent two different ways of building an AI business.
Tencent believes AI should become another layer inside an existing ecosystem. Alibaba believes AI infrastructure can become the ecosystem itself.
Tencent has the safer foundation. Games, advertising, payments and social networking continue to generate enormous cash flow.
But Tencent still needs to answer a basic question: where exactly is its AI business going?
Alibaba’s strategy is easier to understand. Build chips. Buy computing capacity. Expand Alibaba Cloud. Put Qwen on top. Move customers toward higher-margin AI services.
It is coherent. It is also much more dependent on the AI boom continuing.
Tencent is asking whether its existing empire can absorb AI.
Alibaba is asking whether AI can build it a new one.
Investors are not merely choosing which company has the better model. They are choosing which assumption they are willing to finance.