China’s AI Demand Is Real. The Returns Are Already Splitting.

Alibaba’s AI demand is not the weak point in its latest numbers.

Revenue from AI Cloud and Compute Services rose 45% year on year to RMB48.437 billion in the June quarter. Revenue from external customers grew at the same rate. Adjusted EBITA for the segment rose 133% to RMB5.628 billion, taking its adjusted margin to roughly 12%.

That is meaningful contrary evidence to a simple overcapacity thesis. Customers are buying compute, the cloud segment is growing, and its reported operating economics are improving.

But the rest of the full-stack picture moved in a different direction. AI Labs and Applications revenue rose 16% to RMB3.338 billion while its adjusted EBITA loss widened from RMB3.224 billion to RMB13.861 billion. Alibaba said the increase reflected heavier investment in AI capabilities and higher inference costs for the Qwen app.

The distinction matters. AI infrastructure can become profitable before the applications consuming that infrastructure prove they can support the whole capital structure.

The profitable layer and the costly layer

Alibaba now reports AI Cloud and Compute Services separately from AI Labs and Applications. The cloud segment includes T-Head, Alibaba’s chip-design operation, so the figures are not a pure public-cloud measure. Adjusted EBITA is also a non-GAAP measure. Those limits prevent a clean valuation of either layer.

They do not erase the directional split. Cloud revenue grew nearly three times as fast as application revenue, while cloud adjusted earnings more than doubled and application losses expanded more than fourfold.

A full-stack strategy can rationally tolerate that pattern. Applications may be subsidised to drive adoption, create distribution and generate workloads for the cloud. Early inference losses can fall as models become more efficient and usage scales. Alibaba’s applications may also strengthen commerce and advertising economics that do not appear in the AI application segment.

That is the strongest countercase, and it is credible. The losses do not prove failed economics.

They do, however, identify the return test. The full stack works financially only if application adoption eventually produces enough direct revenue, ecosystem value or durable cloud demand to exceed the cost of serving it and the capital required to build ahead of it.

The capital burden is still rising

Alibaba spent RMB67.678 billion on capital expenditure during the quarter, up 75% year on year. It attributed the increase to AI infrastructure, procurement timing, anticipated demand for AI agents and higher component prices.

Operating cash flow rose 11% to RMB22.945 billion, but free cash flow was a RMB44.670 billion outflow, compared with an RMB18.815 billion outflow a year earlier. Alibaba said the deterioration was mainly due to cloud-infrastructure expenditure.

Six days after the results, Alibaba completed an HK$80 billion placement of 710 million new ordinary shares. It allocated approximately 60% of the net proceeds to global computing infrastructure and 40% to hyperscale AI data centres and traditional-cloud upgrades.

The placement is not evidence of distress. Alibaba reported RMB474.505 billion of cash and liquid investments at quarter-end, and issuing equity can preserve flexibility while demand is accelerating. The more useful conclusion is that the company is asking shareholders to fund a much larger infrastructure base before the economics of the application layer are mature.

What this changes about the China investment case

China does not need to win every model benchmark to change AI economics. Alibaba’s own results show a more immediate route: strong external cloud demand, improving infrastructure earnings, inexpensive and widely distributed models, and heavy investment in applications that may deepen usage before they earn acceptable standalone returns.

This shifts the analytical question from capability to allocation. Which layer captures the profit? Which layer absorbs inference costs? How much application demand is incremental, and how much is created by subsidised distribution inside an existing ecosystem? Can cloud margins keep expanding after depreciation and capital intensity are fully reflected?

The evidence currently says Alibaba’s infrastructure economics are improving. It does not yet say the full-stack return is settled.

What BoomRisk will watch

• External cloud growth. Whether the 45% pace persists and remains driven by unaffiliated customer demand.
• Application losses. Whether Qwen-related inference and development costs begin to scale down relative to revenue and usage.
• Cash conversion. Whether operating cash flow catches up with the infrastructure build after the current procurement cycle.
• Capital intensity. Whether cloud margins remain durable once depreciation from the new capacity becomes fully visible.
• Equity-funded capacity. Whether the HK$80 billion buildout produces external revenue and returns fast enough to avoid repeated capital calls.

This paper does not change the authoritative BoomRisk score. At publication, Capex & Cash Returns, AI Monetisation and Open-model Pressure each remain at 4/5, while strong semiconductor demand continues to hold the composite below the high-risk threshold.

The conclusion

Alibaba has supplied real evidence that Chinese AI demand can support fast cloud growth and improving infrastructure earnings.

It has also supplied evidence that the applications consuming the capacity are becoming much more expensive, while the group’s capital expenditure and free-cash-flow burden rise.

Both can be true. That is precisely why technological success does not resolve the investment case. The next proof is not another model release or demand headline. It is whether the profitable infrastructure layer can carry the application losses and capital demands of the full stack.

BoomRisk is a financial-risk monitoring framework, not an investment recommendation or market forecast. Adjusted EBITA and free cash flow are company-defined non-GAAP measures. Segment growth and announced capital allocation do not establish future returns.