Alibaba's new AI segment shows the model lab losing $2 billion a quarter

On 20 August Alibaba reported results for the quarter ended 30 June 2026, and in the same filing changed how it reports its business. The reorganisation is the more interesting half. For the first time, the company's AI model labs, its consumer Qwen app and its QwenWork enterprise agent have been pulled out of a catch-all bucket and given their own segment, AI Labs and Applications. What that produces is one of the clearest public disclosures so far of what it costs to run a frontier model business — and the answer is that it costs a great deal.

AI Labs and Applications took in RMB3,338 million (US$492 million) in revenue for the quarter, up 16% year over year. Against that, it posted an adjusted EBITA loss of RMB13,861 million (US$2,043 million). The same collection of businesses lost RMB3,224 million a year earlier, so the loss grew 330%. Alibaba attributes it to "our increased investment in AI capabilities, and higher inference cost related to Qwen app."

For scale: that one segment's loss is worth roughly half of Alibaba's entire consolidated adjusted EBITA for the quarter, which came to RMB27,329 million. The segment spends about RMB4.15 for every RMB1 of revenue it brings in.

The other half of the split makes money

The same reorganisation combined Cloud Intelligence Group with T-Head, Alibaba's chip design arm, into AI Cloud and Compute Services. That segment reported RMB48,437 million (US$7,139 million) of revenue, up 45%, with adjusted EBITA of RMB5,628 million — up 133% year over year, and an EBITA margin of roughly 12%. AI-related product revenue inside it reached RMB12,376 million (US$1,824 million), which Alibaba says is the twelfth consecutive quarter of triple-digit growth for that line.

So the shape of the business, once you separate the two, is fairly stark. Selling compute and AI infrastructure to other people is profitable and accelerating. Building the models and running a free consumer assistant is not. None of that is unique to Alibaba. What is unusual is that Alibaba now reports the two separately, in a filing, with a number attached to each.

Inference is named as the cost

The wording in the filing is worth reading closely. The loss is attributed in part to "higher inference cost related to Qwen app" — serving, not training. That is a different kind of expense. Training runs are periodic; serving is continuous, and it scales with every additional user and every additional turn.

Alibaba also says 250 million users have had their first AI-driven shopping experience through Qwen app's agentic features since the app launched, with the assistant wired into Taobao, Tmall and Taobao Instant Commerce. Agentic flows are precisely the expensive kind of usage: a single shopping task that browses, compares and checks out consumes far more tokens than one chat turn. The company is describing a product that gets more costly the better it works.

Capital spending, and the chip bill

Capital expenditure for the quarter was RMB67,678 million (US$9,975 million), up 75% from RMB38,676 million a year earlier. Alibaba gives three reasons: procurement cycle timing, more CPU-compute capacity in anticipation of customers adopting AI agents, and "higher pricing of a broad range of chip components." That third one is the line worth noting for anyone modelling future inference prices.

The cash effects follow. Free cash flow was an outflow of RMB44,670 million (US$6,584 million), against an outflow of RMB18,815 million in the same quarter last year. Group net income fell 75% to RMB10,444 million (US$1,539 million). Income from operations fell 57% to RMB15,161 million, a figure that includes RMB8,541 million of goodwill impairment and other charges. Alibaba still held RMB474,505 million (US$69,933 million) in cash and other liquid investments at the end of June, so this is a company choosing to spend, not one running out of room.

What it changes if you build on Qwen

Three practical things follow from the split.

1. Alibaba is at once the subsidiser and the open-weights supplier. The filing says it launched flagship model Qwen3.8-Max in August and "opened its model weights with 2.4 trillion parameters." A matching 2.4-trillion-parameter repository is public on Hugging Face as Qwen/Qwen3.8-2.4T-A95B, created on 8 August and ungated; its config describes a mixture-of-experts model with 512 experts and 10 active per token. If your stack sits on Qwen weights, the economics above are your supplier's economics.
2. The profitable layer is the one that sells compute and APIs, not the one that gives away an assistant. That is where the commercial pressure points — toward metered API usage and paid enterprise agents such as QwenWork, which Alibaba says is integrated with Alibaba Cloud and DingTalk.
3. Component prices are going the wrong way, by Alibaba's own account. It is reasonable to stop assuming that per-token prices only ever fall.

What to treat carefully

Everything above comes from Alibaba's own results announcement and the Form 6-K it was filed under. These are the company's reported figures, and adjusted EBITA is a non-GAAP measure that excludes share-based compensation and amortisation of intangibles; the loss looks different depending on which line you pick.

Several supporting claims are the company's own and are not independently established. The statement that Alibaba Cloud leads China's AI cloud market with a 38.1% share is drawn from an Omdia report cited in the announcement. The assertions that Alibaba's frontier models deliver "top-tier performance" and that its Zhenwu M890 processor has been adopted by more than 650 external customers across over 20 industries arrive without accompanying independent evaluation. Read them as what a vendor says about itself.

The disclosure that is hard to spin, though, is the one Alibaba had no obvious reason to volunteer: it now costs roughly two billion dollars a quarter to build the models and serve the assistant, and the bill is growing faster than the revenue attached to it.