On August 20, Alibaba reported results for the quarter ending June 30. It is the first report after the April reorganization – and the first one in which the AI business is accounted for on two separate ledgers. The result? One half of the puzzle makes money, the other burns through it faster than ever.
What the split revealed
Alibaba merged its cloud business (Cloud Intelligence Group) with the chip unit T-Head into the AI Cloud & Compute Services segment, and bundled the model lab, the Qwen app and Qwen Office into AI Lab & Applications. The difference between the two books is dramatic:
- Cloud: revenue of 48.4 billion yuan (+45% year over year), adjusted EBITA of 5.6 billion yuan (+133%), margin of 12% versus 7% a year earlier. Driven by higher-margin MaaS, in-house chips and better operational efficiency.
- Lab and apps: revenue of just 3.3 billion yuan, but an adjusted EBITA loss of 13.9 billion yuan (3.2 billion a year ago) – the hole widened by more than 10 billion yuan in a single quarter, mainly on model investments and the Qwen app’s inference costs.
- Both segments combined: minus 8.2 billion yuan.
This is the first time the Chinese giant has shown the world the bill for keeping a flagship model alive. Previously Qwen’s costs sat in the “all other” bucket and outsiders could only guess at the scale.
Qwen free, money collected in the cloud
Behind these numbers is a clear plan: open models, free app, revenue collected in the cloud. August brought Qwen3.8-Max (2.4 trillion parameters) – the first time Alibaba has opened the weights of its largest model. Qwen has now passed 3 billion cumulative downloads and 300,000 derivative models, and the Qwen app (launched in May) claims 250 million users who completed their first AI-driven purchase through an agent.
August also saw an open platform inside the Qwen app: partners such as SF Express, Ziroom, Hello and Midea build their own agents in the chat window, and the model shifts from free conversation to taking a cut of transactions. Add the in-house Zhenwu M890 AI processor from T-Head (20+ industries, more than 650 external customers) and AI products that have grown at triple-digit rates for 12 consecutive quarters – already 25.6% of the cloud segment’s revenue.
There is one more detail that never lands in this ledger: China’s version of Apple Intelligence has completed registration with Qwen, using an architecture that combines on-device and cloud processing. Qwen is entering hundreds of millions of devices, but in Alibaba’s accounts this shows up only as inference cost, not revenue. It is a continuation of the story we covered when Apple started training its own model on Alibaba’s infrastructure – see the earlier post.
Where the catch is
The 12% margin belongs only to the compute-selling line. AI as a whole remains deeply in the red, and the margin itself was earned at a point when depreciation has not yet caught up with capital spending. Group-wide, revenue grew to 269 billion yuan (+9%), but adjusted EBITA fell 30%, net profit fell 38%, and capex jumped 75% to 67.7 billion yuan in one quarter. Free cash flow turned negative (-44.7 billion yuan), though the balance sheet – 474.5 billion yuan in cash and liquid investments – still allows for a good night’s sleep.
Those machines will come back in future quarters as depreciation. CEO Wu Yongming pledged 380 billion yuan over three years for cloud and AI infrastructure back in 2025, and in May said it will be “far more” – at the current pace that is roughly 270 billion yuan a year. The same pattern shows up at the competition: Tencent spent 52.8 billion yuan in Q2 (+176%) and posted negative free cash flow for the first time since 2005. China’s big three are all racing on compute, and the bill arrives in the depreciation tables.
Why developers should care
Qwen stays open and cheap because Alibaba can subsidize its models with profits from the cloud and e-commerce (39.7 billion yuan in the quarter). That is not a permanent state of nature, it is a deliberate strategy: open model, free app, revenue in the cloud. Building on Qwen? Keep a plan B in case pricing policy shifts. And for the industry at large, this is one of the first readable accounts of what it really costs to keep a top-tier model alive at a big player – the market is starting to see the true price of the AI race.
This content was generated using artificial intelligence and reviewed by a human editor.
This article is also available in Polish (PL).
Sources
- NetEase Tech (tech.163.com): 阿里AI的两面：云赚56亿，千问烧掉138亿, 2026-08-21
- ReadAI Time: 卖算力赚 56 亿，做模型亏 138 亿：阿里第一次把 AI 拆成两本账, 2026-08-21
- Sina Finance: 阿里AI的两种生意：算力云收入484亿赚56亿，实验室为何亏138亿？, 2026-08-21
- China AI Weekly Brief (2026-08-16 to 2026-08-23)
Leave a Reply
You must be logged in to post a comment.