Alibaba’s June Quarter Shows the Full-Stack AI Flywheel Compounding in Public

Alibaba’s filed June-quarter results — real IR numbers, not a press leak — are the clearest public evidence yet that owning silicon, models, cloud, and applications simultaneously compounds. The cloud unit grew five times faster than the group. Its profit grew three times faster than its revenue. The group’s bottom line fell 75%.

What Happened

Alibaba’s June-quarter earnings filing — a real audited print, submitted to regulators — is worth separating from the private-lab figures circulating this week. OpenAI and Anthropic revenue estimates have appeared across multiple outlets in the past ten days, but those numbers conflict, are not filed with any securities regulator, and should not be read as a comparable quarter. Alibaba’s numbers are IR. That distinction matters when you are trying to reason about where the AI value chain is actually monetizing.

The headline: group revenue grew 9% year over year to RMB 269.0 billion. Cloud Intelligence Group’s external revenue grew 45% — a 22-quarter high — with AI-related product revenue posting its twelfth consecutive quarter of triple-digit year-over-year growth. Cloud adjusted EBITA rose 133% to RMB 5.6 billion, at an 11.6% margin. Model and application annual recurring revenue crossed CNY 16 billion. Those are the numbers that tell the stack story.

Hold the hedges upfront, because they belong at the front. Group net income fell roughly 75% year over year. Capital expenditure rose 75% to RMB 67.7 billion. Free cash flow swung to a RMB 44.7 billion outflow. The AI Labs and Applications segment ran an adjusted EBITA loss of RMB 13.9 billion. This is spend-ahead-of-return, not a clean profit story. The 45% cloud growth is also partly a mix effect: management explicitly noted they were proactively scaling back low-margin business, which flatters the growth rate relative to pure demand. The chip and supply-chain claims — the Zhenwu M890 processor, 650-plus cloud customers, 100-day hyperscale data-center delivery, compute demand outstripping supply — are management statements from the earnings call, not independently audited figures. The ~$7.3 billion AI run-rate is management’s own RMB-to-USD conversion. And the full-stack / vertical-integration-end-state framing developed below is our structural analysis; Alibaba reported a quarter, it did not narrate the industry.

The Structural Read

Four frames. The first three are reinforcing; the fourth is the constraint the other three have to clear.

1 — The Full-Stack Flywheel

Alibaba’s architecture runs from silicon to application without a seam. T-Head’s Zhenwu chips now serve more than 650 cloud customers; the new Zhenwu M890 SuperNode is designed to run inference for models above two trillion parameters, which is not a hypothetical — Qwen 3.8 Max, open-weighted last week at 2.4 trillion parameters, runs on it, as does Kimi K3. Above the silicon sits the cloud layer, which management is rebuilding explicitly as an agentic cloud. Above that sit the models — Qwen 3.8-27B, the 2.4-trillion-parameter Qwen 3.8 Max, more than three billion cumulative downloads, more than 300,000 derivative models built on the base weights. Above that sit the applications, including the newly launched QwenWork for enterprise. The CEO’s framing of the business — paraphrased from the earnings call — is that growing customer demand at any layer of this stack converts directly into commercial opportunity for the layers beneath it. That is not a marketing claim; it is the mechanism the 133% EBITA growth is proving out. Each new enterprise that runs Qwen needs cloud compute. Each cloud customer that scales pulls on T-Head silicon. The demand signal flows downward and the margin accrues at the infrastructure layer.

2 — The China Answer to the Supply Wall

Alibaba is not waiting for export controls to ease. It is building around the constraint. Its own silicon removes the Nvidia dependency at the inference layer — at least for models it controls. Management stated on the call that compute demand will continue to outstrip supply, and the strategic response is vertical ownership of that compute rather than reliance on an external vendor subject to U.S. export policy. The claimed 100-day build time for hyperscale data centers, if accurate, is a supply-chain advantage that compounds the chip ownership: it means Alibaba can convert capital into deployed capacity faster than competitors waiting for allocated GPU shipments. This is the same logic explored in Beyond Nvidia’s Moat — the moat around GPU supply is real, but it creates the incentive for vertically integrated players to route around it entirely. The chip and DC-build claims are management statements, not audited, and should be held as disclosed rather than verified. But the strategic direction is legible in the capex line regardless: RMB 67.7 billion of investment spending does not leave many alternative explanations.

3 — Open Weights as Distribution

Qwen 3.8 Max at 2.4 trillion parameters is free. More than three billion cumulative downloads and 300,000-plus derivative models are the distribution asset that makes it free. The logic is not charity and it is not a race-to-zero — it is the same loss-leader-to-platform mechanism running through the entire AI value chain, as developed in The AI Value Chain. A 2.4-trillion-parameter model that developers build on, fine-tune, and deploy in production creates a gravitational pull toward the cloud infrastructure underneath it. The 35% share of cloud external revenue now attributable to AI-related products is the commercial echo of three billion downloads. Open weights are the top-of-funnel; the cloud EBITA is the conversion.

4 — The Constraint: Group Economics Have Not Turned

The cloud unit’s margin is improving. The group’s economics are deteriorating. Net income down roughly 75% year over year, a RMB 44.7 billion free-cash-flow outflow, and a RMB 13.9 billion operating loss in the AI Labs and Applications segment are the cost of building the stack before the returns are in. The bet is that the application layer losses today are buying the customer base that will monetize through the cloud layer tomorrow. That is a coherent theory. It is also the theory every major AI spender is currently running, and not all of them will be right. Alibaba’s advantage is that it has more layers already in production than most — it is not building the stack from scratch, it is accelerating one that was already generating revenue. But the gap between cloud-unit profitability and group-level economics is the central risk: if the application layer does not convert at the projected rate, the RMB 67.7 billion capex cycle will have funded a cloud for competitors to use.

The Baidu data point from three days earlier rhymes with the same dynamic. Baidu’s GPU cloud grew 283% year over year and its AI Cloud Infrastructure segment reached RMB 7.3 billion at a 50% growth rate — while group revenue fell 4%. The barbell is becoming a sector pattern in China: an AI-infrastructure engine running at multiples of group growth while the legacy business drags the headline number down. Both companies are making the same bet, at roughly the same moment, with different stack completeness.

The broader framing — that Alibaba represents the vertical-integration end-state the rest of the industry is reaching toward from one end — is our structural read, not the company’s claim. Nvidia is climbing up from hardware toward an agentic harness, as covered in the NVIDIA AVO analysis. Anthropic is hiring toward the silicon layer, as the Amir Salek piece maps out. The direction of travel is consistent across players; the distance each has to travel varies. Alibaba is not at the end of the road, but it is currently the furthest along in terms of filed, auditable numbers. That distinction is worth preserving precisely because so much of the current AI revenue narrative is running on unverifiable private-lab figures.

Three Implications