Mainland China's cloud infrastructure market reached $14.7 billion in Q4 2025 according to numbers being attributed to research firms tracking the sector โ Omdia and Canalys are the standard sources for this data and have been publishing quarterly trackers throughout the year. The trajectory is unambiguous: $11.6B in Q1, $12.4B in Q2, $13.4B in Q3, and $14.7B in Q4. That is roughly 27% year-over-year growth, accelerating across the year, in a market that observers wrote off in 2023 when the chip export restrictions first hit. Alibaba Cloud is widening its lead โ Omdia put its share at 36% in a recent quarter, up from 33% earlier in the year โ with Huawei Cloud around 18% and Tencent Cloud around 10%. The remainder is split among Baidu, ByteDance's Volcano Engine, and a long tail of regional providers.
The driver everyone is naming is AI agents, and the named deployments are concrete enough to take seriously. Alibaba launched Wukong in March, an enterprise-facing agent for document editing, meeting transcription, and workflow automation, and the agent's growth on Aliyun is showing up in the cloud spend numbers because each Wukong session burns inference. Tencent's bigger move is integrating an AI agent directly into WeChat as a mini-app surface, with reach to 1.4 billion monthly users โ the closest thing the world has to a single-platform agent rollout at consumer scale. Baidu, ByteDance, and Zhipu have all raised inference token prices in April, which is what you do when demand outstrips supply rather than when you are trying to attract users. The market signal there is that compute capacity is binding and these companies are choosing margin over land-grab.
For Western builders, the relevant reading is not "Chinese clouds are growing fast." It is that the open-weights model ecosystem that the rest of the world is consuming โ Qwen, DeepSeek, Yi, Kimi โ is being trained and served on this infrastructure. When Alibaba Cloud gains share, the Qwen family gets more compute to train against. When Tencent monetizes WeChat agents at 1.4 billion-user scale, the data flywheel that feeds Hunyuan and other in-house models gets denser. The chip restrictions made the buildout slower and more expensive but did not stop it; the Chinese hyperscalers are running on Huawei Ascend, Cambricon, and a mix of stockpiled and gray-market Nvidia hardware, with software stacks that have been hardened by two years of forced workarounds. The gap to US hyperscaler training capacity is real but not what the export-control debate often assumes.
The honest caveats on the $14.7 billion number are worth flagging. The data ultimately comes from research-firm panels and industry contacts, not audited financials. Different firms count "cloud infrastructure" differently โ Canalys, Omdia, and IDC use slightly different methodologies and produce slightly different numbers each quarter. Year-over-year comparisons are reliable; absolute level claims should be read with a grain of salt. And the AI agent attribution, while plausible and corroborated by the named product launches, is hard to disentangle from baseline cloud growth for traditional workloads, currency effects, and a Chinese economy that has been weak in non-tech sectors. The headline number is real, the AI driver is real, and the strategic implications for the global AI buildout are real โ but builders making decisions on this should track the underlying research-firm reports rather than the press summaries that compress them.