Alibaba’s 133% Cloud EBITA Surge Validates In-House Chips—Even as the Stock Sells Off
Alibaba reported June-quarter 2026 earnings on August 20, and the most consequential number sat buried inside a freshly created reporting segment. AI Cloud and Compute Services—a new unit combining the old Cloud Intelligence Group with T-Head, Alibaba's semiconductor arm—posted adjusted EBITA of RMB5.63 billion on RMB48.44 billion of revenue. That is a margin of roughly 11.6%, up approximately 440 basis points from a year earlier, when the same operations earned 7.2%. Revenue rose 45%. EBITA rose 133%.
Alibaba ADRs fell about 5% by midday.
The Consolidated Pain Behind the Segment Gain
The divergence between segment strength and share-price weakness has a straightforward explanation. Group revenue grew only 9%. Adjusted EBITA dropped 30%. Net income collapsed roughly 75%. Free cash flow was negative RMB44.7 billion. Quarterly capex surged 75% to RMB67.7 billion—Alibaba said the cost of deploying some new servers had roughly doubled year-on-year because of component scarcity.
The other new segment tells the rest of the story. AI Labs and Applications, housing Qwen model development and consumer AI products, generated RMB3.34 billion of revenue against an adjusted EBITA loss of RMB13.86 billion. That single-segment loss is 2.46 times the entire positive EBITA produced by AI Cloud and Compute. Building the whole AI stack remains extraordinarily expensive even when one layer of it is working.
Zhenwu Reaches Commercial Scale
At the Zhenwu M890 chip launch in May, Alibaba said more than 560,000 Zhenwu-family accelerators had shipped to over 400 external customers. By the June-quarter release, that count exceeded 650 across more than 20 industries. On August 12, a 64-accelerator M890 supernode went into commercial service in Inner Mongolia, serving mixture-of-experts models with up to 10 trillion parameters. Kimi K3 and Qwen3.8-Max already run on it.
The corporate reorganization formalized this direction: merging T-Head into the cloud segment treats chip, server, network, cloud and model as one economic stack. Capital allocation matched the rhetoric. Share repurchases totaled just US$162 million—capex outspent buybacks by a factor exceeding 60—and a WSJ-reported memo disclosed that Alibaba agreed to sell Lingxi Games to Trustar Capital for over US$1.5 billion. Peripheral assets are being liquidated to fund compute.
Sanctions Created a Hybrid, and the Hybrid Favors Alibaba
A Financial Times report on August 19 revealed that Beijing has begun allowing limited Nvidia H200 imports, with ByteDance and Tencent each receiving approximately 10,000 chips. Washington had been willing to approve larger volumes; China itself is restricting deployment to cultivate domestic accelerator adoption.
The architecture emerging across Chinese hyperscalers is therefore heterogeneous: scarce Nvidia GPUs reserved for frontier training where CUDA and raw throughput justify the premium, domestic silicon like Zhenwu absorbing high-volume inference and production workloads where availability and cost-per-token matter more. Alibaba can monetize customers on whichever side they land. Those wanting Nvidia rent it through Alibaba Cloud. Those whose workloads run economically on Zhenwu consume Alibaba's own silicon—and Alibaba captures more of the stack. A standalone GPU-rental cloud generally lacks that second option.
Distribution Is the Actual Moat
This is where the earnings data converge into the specific insight that most coverage has missed. The defensibility of Alibaba's custom silicon program comes less from transistor benchmarks than from the fact that Alibaba can make its chip the default backend behind an existing cloud API. A Chinese accelerator startup must persuade a customer to buy unfamiliar hardware, retool a datacenter, rewrite CUDA kernels and accept adoption risk. Alibaba slots M890 underneath Model Studio, Qwen deployments or an existing cloud tenancy, and the customer never makes a "buy domestic silicon" procurement decision at all.
Alibaba then amortizes chip R&D and depreciation across internal commerce workloads (Taobao, Amap), Qwen inference, and hundreds of external cloud tenants. An independent compute lessor gets one monetization path. Alibaba gets several.
If AI Cloud and Compute margins keep expanding toward the mid-teens while capex remains elevated, U.S. export controls will have accomplished something Washington probably did not intend: accelerating the construction of a vertically integrated Chinese semiconductor-cloud platform whose competitive advantage compounds through distribution, workload density and a captive customer base that adopted the silicon without ever choosing to.
not investment advice