Alibaba Cloud

Alibaba Cloud, also known as Aliyun, is the cloud computing and artificial intelligence business of Alibaba Group. It sells public and non-public cloud services, operates data and model-development platforms, and commercializes Alibaba's Qwen family of large language models. Alibaba describes the business as a three-layer stack of infrastructure as a service, platform as a service, and model as a service.^[1]

Alibaba reports the business through its Cloud Intelligence Group. That reporting segment is broader than any single Alibaba Cloud product, so segment revenue should not be read as revenue from Qwen, Model Studio, or public cloud alone. For fiscal year 2026, ended March 31, 2026, Cloud Intelligence Group reported revenue of RMB158.132 billion, up 34 percent from the prior year, and adjusted EBITA of RMB14.265 billion.^[1]

In its fiscal 2026 annual report, Alibaba cited Gartner's April 2026 study to describe itself as the fourth-largest infrastructure-as-a-service provider worldwide and the largest in Asia Pacific by 2025 revenue. It cited IDC to describe itself as China's largest public-cloud provider by combined IaaS and PaaS revenue in 2025.^[1] These are rankings within specified revenue categories, not measures of all cloud software, total computing capacity, or model quality. The same filing said Cloud Intelligence Group offered computing services in 34 regions as of March 31, 2026.^[1]

Organization and business scope

Alibaba Cloud provides elastic compute, storage, networking, databases, big-data processing, security, cloud-native services, and model services. Public-cloud customers generally pay by consumption or subscription. Non-public cloud work can combine hardware, software licenses, installation, application development, and maintenance.^[1] The portfolio therefore spans conventional enterprise infrastructure as well as training and serving a foundation model or building an application with generative AI.

The distinction among Alibaba's AI products is important:

• Qwen is a family of models developed by Alibaba.
• Model Studio, called Bailian in China, is a managed service for invoking and adapting Qwen and third-party models.
• Platform for AI, or PAI, covers a broader model lifecycle, including data labeling, development environments, distributed training, evaluation, optimization, and deployment.
• ModelScope is an open model and dataset community. It can connect to Alibaba Cloud resources, but it is not the same product as Model Studio.

Alibaba Cloud is a business within Alibaba Group rather than a separately listed company. Alibaba announced a plan in May 2023 to distribute Cloud Intelligence Group to shareholders and prepare it for a separate listing.^[2] In November 2023, the company cancelled the full spin-off, saying expanded United States restrictions on advanced-computing chip exports created uncertainty and that the transaction might not enhance shareholder value.^[3]

History

Alibaba established its cloud business in September 2009. Its early work included the distributed platform that became the Apsara cloud operating system and the ODPS data-processing service, later renamed MaxCompute.^[4][5] Alibaba Cloud's current product description still identifies Apsara as the proprietary operating system underlying its cloud services.^[4]

Alibaba's MaxCompute history records ODPS entering stable operation in October 2010, reaching a single-cluster scale of up to 5,000 servers by August 2013, and beginning commercial service in July 2014.^[5] These milestones describe one data platform inside the wider cloud business. They should not be treated as counts of Alibaba Cloud's entire fleet.

International expansion accelerated in the middle of the next decade. In 2015, Alibaba opened a Singapore data center and designated Singapore as the headquarters for its international cloud operations.^[6] In January 2017, Alibaba and the International Olympic Committee announced a partnership through 2028 under which Alibaba became the IOC's official cloud-services and e-commerce-platform-services partner.^[7]

Alibaba created DAMO Academy in October 2017 and said it expected to invest more than US$15 billion in research and development over three years. DAMO expands to Academy for Discovery, Adventure, Momentum and Outlook.^[8] It was framed as a group-wide research program rather than a product line with a fixed set of commercial services.

Alibaba's T-Head chip unit introduced the Hanguang 800 inference processor in 2019 and the Arm-based Yitian 710 server processor in 2021.^[9][10] In November 2022, Alibaba Cloud launched ModelScope with more than 300 ready-to-deploy models at launch. The service covered vision, language, and audio tasks and supported online testing, fine-tuning, cloud deployment, and local deployment.^[11]

Alibaba Cloud introduced Tongyi Qianwen in April 2023 and opened the service to the public in September after China's first national rules for public generative-AI services took effect.^[12][13][14] The family expanded through releases including Qwen2 in 2024 and Qwen3 in 2025. The product history is best understood as a sequence of model families, not as one model whose parameter count and context window remain constant.

In February 2025, Alibaba announced that it would invest at least RMB380 billion, then about US$53 billion, in cloud and AI infrastructure over the following three years. The company said the amount exceeded its cloud and AI spending over the previous decade.^[15] In September 2025, Alibaba Cloud announced plans for its first regions in Brazil, France, and the Netherlands, plus additional capacity in Mexico, Japan, South Korea, Malaysia, and Dubai.^[16] By the end of fiscal 2026, the company reported 34 operating regions, so that later figure should be used instead of the 29-region count in the September announcement.^[1]

Cloud infrastructure

Alibaba Cloud's infrastructure layer includes virtual machines, object and block storage, virtual networking, content delivery, databases, data warehouses, Kubernetes services, and security products. A region is a geographic area containing one or more zones. Zones within a region are designed as separate failure domains, while product availability, prices, endpoints, and features can differ by region.^[17] The reported count of 34 regions therefore does not mean every Alibaba Cloud service is available in every location.

Apsara coordinates Alibaba's proprietary cloud stack, while services such as MaxCompute expose higher-level data processing. The company's origins inside Alibaba's commerce operations gave it workloads in transaction processing, search, recommendation, logistics, and large-scale analytics. Alibaba later offered the resulting infrastructure to external customers.^[1][5]

T-Head adds custom silicon to that stack. Alibaba described Hanguang 800 as an inference neural processor used internally for product search, translation, recommendation, advertising, and customer service. In a company-run ResNet-50 inference test, Alibaba reported 78,563 images per second and 500 images per second per watt.^[9] Those figures are vendor results from one workload, not a general comparison with all accelerators.

Yitian 710 is a 5 nm, 128-core Armv9 server processor with eight DDR5 memory channels and 96 PCIe 5.0 lanes. Alibaba announced it for its own Panjiu servers and data-center workloads in 2021.^[10] That announcement also described Intel, AMD, Arm, and Nvidia as infrastructure partners, which makes the in-house processors part of a heterogeneous supply strategy rather than evidence that Alibaba had replaced outside vendors.

By fiscal 2026, Alibaba reported that T-Head had put a proprietary GPU into production at scale for training, fine-tuning, and inference.^[1] In May 2026 it publicly named the Zhenwu M890, describing 144 GB of on-chip memory, 800 GB/s of inter-chip bandwidth, support for numeric formats from FP32 to FP4, and integration with its Panjiu AL128 system.^[18] These specifications and shipment figures come from Alibaba; independent workload results are needed for cross-vendor performance conclusions.

AI models and platforms

Qwen

Qwen began as text base and chat models, then expanded into code, mathematics, vision-language, audio, image, and video systems. The first technical report described base models and instruction-aligned chat models at multiple sizes, along with specialized code and mathematics variants.^[19] The family is central to Alibaba Cloud's model-as-a-service strategy, but Qwen is not synonymous with the whole cloud business.

Alibaba released Qwen3 in April 2025 as six dense models from 0.6 billion to 32 billion parameters and two mixture-of-experts models: a 30-billion-parameter model with 3 billion active parameters and a 235-billion-parameter model with 22 billion active parameters.^[20] The release combined optional reasoning and direct-response modes. Model weights were distributed through GitHub, ModelScope, and Hugging Face, and the technical report describes the initial Qwen3 family as Apache 2.0 licensed.^[20][21]

Open availability applies to particular releases, not automatically to every model carrying the Qwen name. Alibaba Cloud's July 10, 2026 documentation lists Qwen3.7-Max as a proprietary hosted model in Model Studio.^[22] On July 20, Alibaba announced Qwen3.8-Max-Preview on its Token Plan, Qoder, and QoderWork services and said open weights would follow.^[31] A preview announcement and a future release commitment are not the same as already published weights. This mix of downloadable, preview, and managed-only models is more precise than describing all Qwen models as open source. It also makes download comparisons with another family such as Llama sensitive to the date, repository, and definition of a derivative model.

Platform for AI

PAI is Alibaba Cloud's end-to-end machine learning platform. Its documented modules include iTAG for labeling, Data Science Workshop for notebook and IDE development, Distributed Learning Cluster for training jobs, Designer for visual workflows, and Elastic Algorithm Service for deploying models as endpoints.^[23] PAI also connects specialized Lingjun compute, general cloud resources, and big-data engines to its training and serving tools.^[24]

The platform supports major frameworks including TensorFlow, PyTorch, Megatron, and DeepSpeed.^[23] Model Gallery provides no-code paths to deploy, evaluate, or adapt open-weight models. Current documentation includes both full-parameter and efficient supervised fine-tuning, including low-rank adaptation for supported models.^[25] Exact model, region, accelerator, and billing support changes over time, so a PAI capability should be checked against the current product documentation before deployment.

PAI also exposes tools for established computer vision and natural language processing workloads. Its documentation names ResNet-50 and Transformer-based models as examples for inference optimization and lists applications including recommendation, financial risk management, autonomous driving, and scientific computing.^[23][24] These are platform capabilities and solution categories, not proof that one model or configuration performs equally well in each field.

Model Studio

Model Studio is the managed model-service layer. As of July 2026, Alibaba documented Qwen and third-party models, OpenAI-compatible APIs, and multimodal services for text, images, audio, and video.^[22] Alibaba documented Model Studio availability in Singapore, Virginia, Beijing, Hong Kong, Tokyo, and Frankfurt, with separate endpoints, keys, model availability, features, and pricing by region.^[22]

Model Studio supports on-demand inference and, for selected models and regions, fine-tuning, evaluation, knowledge retrieval, and dedicated deployment. Alibaba's documentation distinguishes full-parameter training, efficient training, continual pretraining, and preference optimization.^[25] This managed workflow differs from downloading a weight file from ModelScope and running it on self-managed infrastructure.

ModelScope

ModelScope is an open model community and software ecosystem. At launch in 2022 it contained more than 300 models, including more than 150 that Alibaba described as state of the art for their tasks.^[11] Those are launch figures, not current inventory. The durable point is the service model: contributors publish models and datasets, users can test or adapt them, and artifacts can be deployed on Alibaba Cloud, another cloud, or local hardware.^[11]

ModelScope also distributes Alibaba models, but it is not restricted to Qwen and is not merely a download mirror. Conversely, placement on ModelScope does not by itself establish that every artifact has the same license, safety evaluation, support policy, or production readiness. Those details must be checked on the individual model card.

Research and applications

DAMO Academy has supported research in areas such as deep learning, language, vision, medical imaging, chips, and distributed systems. Its relationship with Alibaba Cloud varies by project: some work becomes a cloud service or internal system, while other work is published research.

PANDA, a pancreatic-lesion detection system involving DAMO researchers and clinical collaborators, is a well-documented example. A 2023 Nature Medicine paper reported training on 3,208 patients from one center, external testing across nine centers, and retrospective real-world evaluation on 20,530 consecutive patients.^[26] In the combined real-world evaluation, the paper reported 92.9 percent sensitivity and 99.9 percent specificity for lesion detection.^[26] Those numbers apply to the study's non-contrast CT cohorts and evaluation design. They do not establish a general cancer-screening accuracy, replace clinical diagnosis, or show performance in every population.

Alibaba Cloud has also supplied infrastructure for sports broadcasting and event operations. The IOC partnership runs through 2028, and Beijing 2022 moved core games systems to Alibaba Cloud. Olympic Broadcasting Services used cloud delivery for live signals, while the Olympic Channel used Alibaba infrastructure for distribution.^[7][27] These deployments demonstrate large event workloads, although they do not independently benchmark Alibaba Cloud against other providers.

PAI and earlier "ET Brain" programs have been presented for industrial inspection, recommendation, public services, and urban planning. A 2017 Alibaba agreement with Macau, for example, proposed work in tourism, health, government services, and transportation.^[28] Alibaba's current PAI documentation also lists finance-related risk management among its solution areas.^[24] Announced use cases should be separated from peer-reviewed outcome evidence; an integration or pilot does not on its own prove cost savings, safety, or social benefit.

Financial performance and market position

Cloud Intelligence Group's fiscal 2026 revenue was RMB158.132 billion, compared with RMB118.028 billion in fiscal 2025. Revenue excluding Alibaba-consolidated subsidiaries grew 33 percent, which Alibaba attributed mainly to public-cloud growth and adoption of AI-related products.^[1] Adjusted EBITA rose 35 percent to RMB14.265 billion.^[1]

For the March 2026 quarter, segment revenue was RMB41.626 billion. External revenue grew 40 percent year over year. Alibaba said AI-related product revenue had grown at a triple-digit rate for eleven consecutive quarters, represented 30 percent of external cloud revenue, and had an annualized value above RMB35.8 billion.^[29] The annualized figure is a run rate, not recognized annual revenue, and the 30 percent denominator is external revenue rather than total segment revenue.

Alibaba's Gartner and IDC rankings are useful measures of commercial position, but their scope matters. Gartner's cited table covers IaaS vendor revenue in United States dollars. IDC's cited result covers Chinese public-cloud IaaS and PaaS revenue.^[1] Neither ranking supports the baseline's exact worldwide, Asia-Pacific, or China share percentages, so those figures are omitted.

The RMB380 billion investment plan signals substantial capacity expansion, but it is a forward commitment announced by the company rather than a completed spend.^[15] Likewise, data-center openings announced in 2025 were plans. The fiscal 2026 filing's 34 operating regions is the more appropriate point-in-time measure for current infrastructure.^[1][16]

Regulation and geopolitical constraints

China's Interim Measures for the Management of Generative Artificial Intelligence Services took effect on August 15, 2023. They apply to public-facing services and set obligations involving information security, personal data, and service governance.^[13] Alibaba opened Tongyi Qianwen to the public the following month after regulatory clearance was reported.^[14] Enterprise-internal research and a public chatbot can therefore face different compliance pathways.

United States export controls are another material constraint. In October 2023, the Bureau of Industry and Security broadened controls on advanced-computing chips and semiconductor-manufacturing equipment for China and other destinations of concern.^[30] Alibaba explicitly linked the expanded restrictions to its November 2023 decision not to pursue a full Cloud Intelligence Group spin-off.^[3] Its fiscal 2026 filing continued to warn that export controls could limit access to technologies, components, and computing capacity needed for cloud and AI services.^[1]

Cloud customers also face ordinary architecture and governance choices: location, data residency, cross-border transfer, identity controls, encryption, model retention, and product-specific service terms. A provider's global region count or certification list does not determine whether a particular workload complies with the laws and contracts that apply to it.

Interpreting product claims

• Model names and availability change quickly. A model family can include both downloadable and hosted-only releases.
• Benchmark results are configuration-specific. Alibaba's Hanguang, Qwen, and Zhenwu announcements contain vendor measurements that should not be generalized beyond their disclosed tests.
• A region count is not a universal availability count. Each service has its own regional matrix.
• Segment figures are not product figures. Cloud Intelligence Group revenue cannot be assigned to Qwen, PAI, or Model Studio without a separate disclosure.
• Market rankings depend on category definitions. IaaS revenue, public-cloud IaaS plus PaaS revenue, and total cloud revenue are different measures.
• Announced investments, partnerships, and pilots describe strategy or deployment, not necessarily audited expenditure or independently measured outcomes.

See also

• AI in education

References

1. ^Alibaba Group Holding Limited, "Annual Report on Form 20-F for the fiscal year ended March 31, 2026," June 2026. sec.gov/...baba-20260331
2. ^Alibaba Group, "Alibaba Says Cloud To Spin-Off; Freshippo & Cainiao To Pursue IPOs," May 18, 2023. alibabagroup.com/...document-1596257223028244480
3. ^Alibaba Group, "Alibaba's First Dividend and Solid Q2 Results; New Management Team Commits to Investing for Growth," November 16, 2023. alibabagroup.com/...document-1663636438236790784
4. ^Alibaba Group, "Alibaba Cloud," accessed July 28, 2026. home.alibabagroup.com/...esses-1747835448811585536
5. ^Alibaba Cloud, "MaxCompute: History," updated July 10, 2026. alibabacloud.com/...maxcompute-development-history
6. ^Alibaba Cloud, "Aliyun to Open Singapore Cloud Data Center to Serve Southeast Asia," August 19, 2015. alibabacloud.com/...data-center-to-serve-southeast
7. ^Alibaba Group and International Olympic Committee, "International Olympic Committee and Alibaba Group Launch Historic Long-Term Partnership as Alibaba Becomes Worldwide Olympic Partner Through 2028," January 19, 2017. home.alibabagroup.com/...ument-1491589952163545088
8. ^Alibaba Group, "Alibaba Launches Global Research Program for Cutting-edge Technology Development," October 11, 2017. alibabagroup.com/...document-1491841175609409536
9. ^Alibaba Group, "Alibaba Unveils AI Chip to Enhance Cloud Computing Power," September 25, 2019. alibabagroup.com/...document-1491206538574954496
10. ^Alibaba Group, "Alibaba Cloud Unveils New Server Chips to Optimize Cloud Computing Services," October 19, 2021. alibabagroup.com/...document-1491124270925873152
11. ^Alibaba Cloud, "Alibaba Cloud Launches ModelScope Platform and New Solutions to Lower the Threshold for Materializing Business Innovation," November 3, 2022. alibabacloud.com/...es-modelscope-platform-and-new
12. ^Alibaba Group, "Alibaba Cloud Unveils New AI Model to Support Enterprises' Intelligence Transformation," April 11, 2023. alibabagroup.com/...document-1582482069362049024
13. ^State Council of the People's Republic of China, "China moves to support generative AI, regulate applications," July 13, 2023. english.www.gov.cn/...t_WS64aff5b3c6d0868f4e8ddc01
14. ^Reuters, "Alibaba opens AI model Tongyi Qianwen to the public," September 13, 2023. marketscreener.com/...nwen-to-the-public--44828307
15. ^Alibaba Group, "Alibaba to Invest RMB380 billion in AI and Cloud Infrastructure Over Next Three Years," February 24, 2025. alibabagroup.com/...document-1830678592242057216
16. ^Alibaba Cloud, "Alibaba Cloud Announces International Expansion Plans to Power the Next-Generation AI Innovations," September 24, 2025. alibabacloud.com/...602561
17. ^Alibaba Cloud, "Regions and zones," updated July 1, 2026. alibabacloud.com/...regions-and-zones
18. ^Alibaba Group, "Alibaba Unveils New AI Chip, Flagship Model, and Rebuilt Cloud Stack for Agentic Era," May 20, 2026. alibabagroup.com/...document-1994119844504535040
19. ^Jinze Bai, Shuai Bai, Yunfei Chu, et al., "Qwen Technical Report," arXiv:2309.16609, 2023. arxiv.org/...2309.16609
20. ^Alibaba Group, "Alibaba Introduces Qwen3, Setting New Benchmark in Open-Source AI with Hybrid Reasoning," April 29, 2025. home.alibabagroup.com/...ument-1853940226976645120
21. ^An Yang, Anfeng Li, Baosong Yang, et al., "Qwen3 Technical Report," arXiv:2505.09388, 2025. arxiv.org/...2505.09388
22. ^Alibaba Cloud, "What is Alibaba Cloud Model Studio," updated July 10, 2026. alibabacloud.com/...what-is-model-studio
23. ^Alibaba Cloud, "What is Platform for AI (PAI)," updated June 23, 2026. alibabacloud.com/...chine-learning-platform-for-ai
24. ^Alibaba Cloud, "Platform for AI: Product architecture," updated June 11, 2026. alibabacloud.com/...service-architecture
25. ^Alibaba Cloud, "Introduction to model fine-tuning," updated July 14, 2026. alibabacloud.com/...model-training-overview
26. ^Kai Cao, Yingda Xia, Jiawen Yao, et al., "Large-scale pancreatic cancer detection via non-contrast CT and deep learning," Nature Medicine 29, 3033-3043, 2023. nature.com/...s41591-023-02640-w
27. ^Alibaba Cloud, "First Olympic Winter Games to Host Its Core Systems on Alibaba Cloud," February 22, 2022. alibabacloud.com/...-core-systems-on-alibaba-cloud
28. ^Alibaba Cloud, "Alibaba Group Forms Partnership with Macau SAR Government to Drive Smart City Development," August 4, 2017. alibabacloud.com/...p-with-macau-sar-government-to
29. ^Alibaba Group, "Alibaba's Cloud Revenue Growth Accelerates to 40% as AI Strategy Delivers," May 13, 2026. alibabagroup.com/...document-1991364841188622336
30. ^United States Department of Commerce, Bureau of Industry and Security, "Commerce Strengthens Restrictions on Advanced Computing Semiconductors, Semiconductor Manufacturing Equipment, and Supercomputing Items to Countries of Concern," October 17, 2023. bis.gov/...s-semiconductor-manufacturing-equipment
31. ^Alibaba Group, "Alibaba Cloud Unveils Agent-Native Innovations at WAIC 2026," July 20, 2026. alibabagroup.com/...document-2016703577908576256