Alibaba Cloud — Model Studio

  Executive Briefing

Alibaba Cloud is the cloud and AI computing arm of Alibaba Group (NYSE: BABA), China's largest technology conglomerate, and the world's fourth-largest infrastructure-as-a-service provider by revenue. Within the AI infrastructure landscape, the company's most strategically significant offering is Model Studio — a one-stop platform for deploying, fine-tuning, and consuming AI models via OpenAI-compatible REST APIs, with the proprietary Qwen model family as its centerpiece. From a single endpoint, developers in Singapore, Frankfurt, Virginia, Hong Kong, or Beijing can call frontier-class LLMs, reasoning models, vision-language systems, code agents, and full-modality speech-and-video models — all on pay-as-you-go token pricing that is benchmarked aggressively against Western equivalents.¹

The company is led at the group level by CEO Eddie Wu (Wu Yongming), who has publicly framed AGI as Alibaba's primary corporate objective and chairs the AI Task Force responsible for all foundation-model investment. In April 2026, Alibaba executed a sweeping AI leadership restructure: internationally recognized computer-vision pioneer Li Feifei (Fei-Fei Li) was appointed CTO of Alibaba Cloud, responsible for cloud technology and AI cloud infrastructure, while Zhou Jingren (Jingren Zhou) — a former Microsoft Research veteran who had served as Alibaba Cloud CTO — was elevated to Chief AI Architect at the group level and given stewardship of the Tongyi large-model business unit, now operating as a standalone commercial entity.² Together they represent a deliberate pairing of global academic credibility and deep platform engineering.

The Qwen open-source model family has emerged as the dominant developer-acquisition strategy worldwide. By January 2026, Qwen derivative models on Hugging Face exceeded 200,000 — the first open-source LLM family to reach that milestone — with cumulative downloads surpassing 10 billion.³ More than half of all open-source model downloads globally are reported to trace back to the Qwen lineage. That ecosystem creates a compounding flywheel: developers who fine-tune on open weights naturally migrate to Model Studio's hosted inference when they need production throughput, prompt caching, or commercial SLAs. The consumer-facing Qwen App had reached a reported 234 million users by May 2026.⁴

What makes Alibaba Cloud unusual among hyperscalers is vertical depth. Unlike AWS Bedrock or Azure AI Foundry, which primarily serve third-party models, Alibaba controls the entire value chain: T-Head (Pingtouge) designs its own AI silicon; Apsara is a proprietary distributed OS; Model Studio and DashScope compose the inference and application layer; and the Qwen family is developed in-house with dedicated research infrastructure. The Zhenwu M890 chip unveiled in May 2026 — offering three times the performance of its predecessor, 144 GB on-chip memory, and native FP4 precision — signals a serious push toward silicon independence from Nvidia, a pressing concern given US export controls on advanced GPUs to China.⁵

  At a Glance

  Origins & Founding

The institution now known as Alibaba Cloud was formally founded on 10 September 2009 — Alibaba Group's tenth anniversary — by Dr. Wang Jian, who had been recruited from his position as Executive Vice President at Microsoft Research Asia by Jack Ma.⁸ The founding mandate was audacious in its specificity: break Alibaba's "IOE" dependency — the acronym for IBM servers, Oracle databases, and EMC storage that underpinned every major Chinese internet company of the era — and replace them with a proprietary distributed computing stack. Wang's team in Hangzhou wrote the first lines of the Apsara distributed operating system on 1 February 2009, months before the formal founding date, and the Apsara system continues to underpin Alibaba Cloud's global infrastructure to this day.⁸

The original thesis — commoditizing computing in the way electricity was commoditized in the early twentieth century — was simultaneously a cost-reduction strategy for Alibaba's own businesses and an external commercial proposition. Public cloud services opened to external customers by 2011, and the system proved itself under fire during the Singles' Day festival of 2010, when 2.4 billion page views flowed entirely across Alibaba Cloud infrastructure.

The AI model-serving layer was seeded much later, with the internal beta of the Tongyi Qianwen (Qwen) large language model in April 2023 and its public release in September 2023 following Chinese regulatory clearance. The formal platform brand — Model Studio — debuted at a Singapore event in February 2024, unifying what had previously been accessed through the DashScope SDK dating to 2023.⁹ The T-Head chip unit (Pingtouge Semiconductor), which now provides the proprietary silicon for AI inference, was established in 2018 and shipped the Hanguang 800 AI inference chip in 2019.

  History & Timeline

    2009–2018: Building the foundation

Alibaba Cloud spent its first decade constructing the infrastructure plumbing that would later support AI workloads at scale. The company cleared the ISO 27001:2005 certification in 2012 — the first Chinese cloud provider to do so — merged with HiChina (domain registration and hosting) in 2013, and began international expansion in 2015 with a Silicon Valley data center. By 2016 it had appeared in the Gartner Magic Quadrant and established partnerships with SK Holdings and SoftBank. The company provided the full cloud infrastructure for the 2018 Asian Games, demonstrating its ability to handle large-scale real-time workloads. The T-Head (Pingtouge) semiconductor unit was established in 2018, setting the stage for proprietary silicon.

    2019–2022: Silicon, scale, and the AI build-up

The Hanguang 800 AI inference chip shipped in 2019, the first production silicon from the T-Head unit. In 2021 Alibaba unveiled the Yitian 710, an ARM-based server processor. By 2022 the company had expanded to 29 global regions and 91 availability zones, making it the largest IaaS provider in Asia Pacific. Internally, research teams had begun serious pre-training on what would become the Qwen foundation model.

    2023–2024: Qwen goes public; Model Studio launches

April 2023 saw the beta launch of Tongyi Qianwen, Alibaba's flagship LLM. Open-source weights for Qwen 7B were released in August 2023, and regulatory clearance enabled a public product launch in September.⁸ Open-source releases of Qwen 72B and 1.8B followed in December 2023. At the Apsara Conference in October 2023, Tongyi Qianwen 2.0 was announced alongside Model Studio upgrades. The branded Model Studio platform formally debuted in February 2024, including fine-tuning, evaluation, and OpenAI-compatible inference endpoints.⁹ In June 2024 the Qwen2 family launched, and in July 2024 Alibaba Cloud published research on a proprietary 15,000-GPU Ethernet interconnect, abandoning NVLink for large-cluster GPU fabric.¹⁰

    2025–present: The open-source flywheel and infrastructure surge

Qwen2.5 launched in September 2024 under Apache 2.0 licensing, dramatically lowering barriers to commercial use of the open weights. CEO Eddie Wu committed ¥380 billion in AI and cloud capital expenditure in January 2025.⁶ By April 2025, Qwen3 (Apache 2.0, 235 billion parameter MoE architecture, 36 trillion token training set) was released — one of the most capable open-source models available at launch.¹¹ US restrictions on Nvidia H20 exports to China (April 2025) added urgency to the proprietary silicon roadmap. The Apsara Conference in September 2025 announced expansion to eight new data-center regions and a partnership with NVIDIA on physical AI and humanoid robotics. By January 2026, Qwen derivative models on Hugging Face exceeded 200,000, the first open-source LLM family to do so.³ April 2026 brought the major leadership restructure placing Fei-Fei Li as Alibaba Cloud CTO and elevating Zhou Jingren to head the newly autonomous Tongyi business unit.² On 19 May 2026, Alibaba unveiled the Zhenwu M890 chip, the Qwen3.7-Max LLM, and the Panjiu AL128 Supernode server, the most significant infrastructure announcement in the company's AI history.⁵

  What They Offer — Products & Platform

Model Studio is Alibaba Cloud's branded AI model-serving and application-development platform, accessible via an OpenAI-compatible REST API, the DashScope Python, Java, Node.js, and Go SDKs, or the Qwen App consumer product. Activation is free; billing is pay-as-you-go per token with discounts for batch invocation and prompt caching. Five API regions serve global developers: Singapore, US Virginia, China Beijing, China Hong Kong, and Germany Frankfurt.¹

The platform delivers several distinct product tiers. At the flagship level, Qwen3-Max (context: 262K tokens, thinking and non-thinking modes) and Qwen3.7-Max (announced May 2026 with prompt caching) provide the highest-capability general-purpose inference. Qwen3.5-Plus and Qwen3.5-Flash offer 1M-token context windows at balanced and cost-optimized price points respectively. Qwen3.6-Plus, released on 2 April 2026, introduced an agentic coding and multimodal reasoning capability with compatibility for Claude Code and Cline agentic workflows.¹²

Specialist model tracks include: QwQ-Plus (reinforcement-learning-based reasoning); Qwen-Long (10 million-token context window for large-document analysis); Qwen-Coder with an integrated Coding Agent; Qwen-VL (vision-language understanding with OCR); QVQ (visual chain-of-thought reasoning); Qwen-Omni and Qwen-Omni-Realtime (full modality: text, image, audio, and video inputs; text or speech outputs); Qwen-Math (mathematical problem solving); and Qwen-OCR (document and handwriting extraction). Generative media models include the Wan series for text-to-video and image-to-video, and Z-Image for bilingual text-to-image generation. Speech synthesis and recognition, plus text and multimodal embedding, round out the catalog.¹

In the China Mainland (Beijing) region, Model Studio also distributes third-party models including DeepSeek, Kimi, MiniMax, and GLM, positioning the platform as a multi-model marketplace for domestic developers.

A notable enterprise developer offering is the AI Catalyst Program, providing selected startups with up to 2 billion free Model Studio tokens plus $120,000 in cloud credits. New users in the Singapore region receive approximately 70 million total free tokens (1 million per model across most Qwen proprietary models, 90-day validity) plus 1,650 seconds of video generation credit.

  Technology & Infrastructure

Alibaba Cloud's infrastructure strategy is defined by three parallel investments: proprietary AI silicon, a global data-center buildout, and a proprietary software stack — each designed to reduce dependence on external suppliers, particularly Nvidia, in the context of ongoing US export controls.

    Proprietary silicon — T-Head Zhenwu

The T-Head (Pingtouge) chip division, established in 2018, is Alibaba's internal semiconductor unit. Its latest AI accelerator, the Zhenwu M890 (unveiled 19 May 2026), delivers three times the performance of its predecessor (the Zhenwu 810E, 2024), with 144 GB on-chip memory, 800 GB/s inter-chip bandwidth, and native precision support from FP32 to FP4.⁵ The accompanying ICN Switch 1.0 networking chip provides 25.6 Tbps aggregate bandwidth, enabling congestion-free 64-accelerator clusters. As of May 2026, more than 560,000 Zhenwu units had shipped to over 400 customers across 20 industries.⁵

    Nvidia GPU fleet

Alibaba maintains a substantial Nvidia GPU fleet alongside its proprietary silicon. Reports indicate that Alibaba, ByteDance, and Tencent jointly received US approval to purchase up to approximately 400,000 H200 units; the exact Alibaba allocation is unconfirmed.¹³ Following the April 2025 H20 export restriction to China, limited export licenses reportedly resumed in mid-2025. In October 2025, Alibaba published a GPU pooling system claiming to reduce effective GPU utilization by 82% via shared pooling across workloads.¹⁴

    Servers and networking

The Panjiu AL128 Supernode (unveiled May 2026) integrates 128 AI accelerators with petabyte-per-second internal bandwidth, designed for burst inference on large agentic workloads.⁵ For large-scale GPU clusters, Alibaba Cloud has published research on a proprietary Ethernet-based 15,000-GPU interconnect, abandoning NVLink-style proprietary fabric in favor of a custom Ethernet architecture.¹⁰

    Data centers and global footprint

As of the Apsara Conference 2025, Alibaba Cloud operates 91 availability zones across 29 regions globally.¹⁵ New regions announced include the first Alibaba Cloud data centers in Brazil, France, and the Netherlands, with additional sites in Mexico, Japan, South Korea, Malaysia, and Dubai targeted within approximately one year of announcement.¹⁶ Service centers are also opening in Indonesia and Germany.

    Software stack

The Apsara distributed operating system, written entirely in-house beginning in 2009, forms the foundation of Alibaba Cloud's infrastructure. The DashScope API layer provides the developer-facing inference gateway with OpenAI-compatible routing. The ModelScope platform hosts an open-source AI application community with approximately 23,000 AI apps, 95% built by individual developers.¹⁷ Agentic reinforcement learning (RL) training mechanisms underpin the Qwen model training pipeline.

  Model Catalog & Performance

Model Studio's catalog spans proprietary Qwen tiers, open-weight checkpoints, specialist models, and third-party models. The table below covers the primary API-accessible proprietary and open-weight models as of June 2026.

Cross-links to model cards: Qwen3-Max · QwQ-Plus · Qwen-VL · Qwen-Coder.

The open-source Qwen family published on Hugging Face (maintained by Qwen) has generated more than 200,000 derivative models — the largest open-source LLM derivative ecosystem globally as of January 2026 — with cumulative downloads exceeding 10 billion.³

  Pricing & Performance Position

Model Studio uses per-token, pay-as-you-go pricing with two systemic discounts: batch invocation (50% off both input and output tokens for asynchronous jobs) and prompt caching (up to 90% reduction on cache-hit reads; Qwen3.7-Max: $0.25/M cached vs $2.50/M fresh input).¹

Representative prices in the Singapore (international) region are as follows.¹

China Mainland (Beijing) pricing is reported to be 60–70% cheaper than Singapore rates, reflecting domestic market dynamics. The competitive positioning on price is explicit: Qwen2.5-72B at approximately $0.23/M tokens all-in is benchmarked publicly at roughly one-tenth the cost of GPT-4o.¹⁸ Morgan Stanley has identified Alibaba as a top pick on the basis of this full-stack pricing advantage.

  People & Leadership

Alibaba Cloud's AI leadership underwent a significant restructuring in April 2026, reflecting the elevation of Tongyi to a standalone commercial business unit rather than an internal capability.

Eddie Wu (Wu Yongming) serves as CEO of Alibaba Group and chairs the AI Task Force, which oversees all foundation-model investment and strategy. Wu has been public in framing AGI as Alibaba's primary corporate objective and has set explicit revenue targets for the Tongyi and Model Studio business lines.

Li Feifei (Fei-Fei Li) was appointed CTO of Alibaba Cloud in April 2026, taking responsibility for cloud technology strategy and AI cloud infrastructure.² Li is internationally recognized as a pioneer of modern computer vision — she co-created ImageNet and served as Director of the Stanford AI Lab before a tenure as Chief Scientist of AI and ML at Google Cloud. Her appointment signals deliberate credibility-building in Western developer and enterprise markets.

Zhou Jingren (Jingren Zhou) serves as Chief AI Architect at the group level and heads the Tongyi large-model business unit, which was elevated to a standalone structure in the same April 2026 restructuring.² Zhou previously served as Alibaba Cloud CTO and spent eleven years at Microsoft Research before joining Alibaba, giving him both research and enterprise-infrastructure depth.

Wu Zeming is the Group CTO, coordinating business technology platforms and the AI inference platform across Alibaba's operating entities. Wang Jian, the founder of Alibaba Cloud, has departed an executive role but remains an influential figure in the company's engineering culture. Fanyu Wu serves on the AI Foundation Model Task Force alongside Eddie Wu and Zhou Jingren.

  Position within Parent Org

Alibaba Cloud operates as the Cloud Intelligence Group segment within Alibaba Group Holding, a public company listed on both the NYSE (BABA) and Hong Kong Stock Exchange (9988). There are no external funding rounds or independent valuation events: the unit is entirely funded through Alibaba Group's balance sheet and capital markets access.

The financial trajectory is significant in scale and direction. In Q4 FY2026 (quarter ended March 31, 2026), the Cloud Intelligence Group reported revenue of RMB 41,626 million (~$6.04 billion), up 38% year-over-year, with AI-related products representing approximately 30% of cloud revenue and achieving eleven consecutive quarters of triple-digit year-over-year growth.⁷ The annualized AI model and application revenue run rate reached approximately $5.2 billion as of Q4 FY2026.⁷

CEO Eddie Wu has set explicit targets: annualized revenue from model and app services exceeding RMB 10 billion by Q2 FY2027, and exceeding RMB 30 billion by the end of FY2027 — with a five-year vision of surpassing $100 billion in combined AI and cloud revenue.⁷ These are forward targets and subject to revision.

The capital investment underpinning this growth is substantial. Alibaba has confirmed a ¥380 billion (~$52.7 billion) commitment to AI and cloud infrastructure over three fiscal years from FY2025.⁶ Reports have surfaced of a potential increase to $69 billion over the same period, though this remains unconfirmed by the company.⁶ The human cost of this investment is visible in cash flow: FY2026 non-GAAP net income fell 62% and free cash flow turned to a RMB 46.6 billion outflow as capital expenditure ramped aggressively.

  Customers & Partnerships

Enterprise deployments span a broad range of industries, illustrating the breadth of Model Studio's addressable market. AstraZeneca China built an adverse event reporting tool using Qwen and Model Studio, reporting 95% accuracy and a claimed 300% improvement in processing efficiency. China Eastern Airlines was the first commercial Qwen App partner, deploying a natural-language flight booking interface. Consumer and retail adopters include KFC China, Luckin Coffee, Mixue, and Shiseido.¹⁷

International adoption is growing. GladCube in Japan uses Alibaba Cloud visual generation models for digital marketing. Japanese AI lab FLUX co-developed FLUX-Japanese-Qwen 32B, an open-source Japanese-language LLM built on the Qwen architecture. 0G Foundation provides on-chain autonomous agent access to Qwen through an integration with Model Studio. Developers using Claude Code and Cline can access Qwen3.6-Plus through its agentic coding APIs.¹²

Platform and marketplace integrations include Dify Enterprise on the Alibaba Cloud Global Marketplace and an NVIDIA Physical AI partnership — integrating the NVIDIA humanoid robotics and physical AI stack on the Platform for AI within Alibaba Cloud.¹⁵

The developer community is the largest quantitative indicator of ecosystem health. The ModelScope platform hosts approximately 23,000 AI apps; the Qwen family on Hugging Face has generated more than 200,000 derivative open-source models (a global first for open-source LLMs) with cumulative downloads exceeding 10 billion.³ The consumer Qwen App is reported to have reached 234 million users by May 2026.⁴

  Competitive Position

Alibaba Cloud holds an estimated 7.7% global IaaS market share (up from 7.2% in 2024, per Gartner's April 2026 data), ranking it fourth globally behind AWS, Microsoft Azure, and Google Cloud.¹⁹ In Asia Pacific, it is the largest IaaS provider by revenue, with a reported 22.5% regional share (up from 20.8%), and it dominates domestically in China with approximately 36% domestic cloud share.¹⁹

On the model-serving layer, the competitive dynamic differs from the infrastructure layer. Against AWS Bedrock, Azure AI Foundry, and Google Cloud Vertex AI, Alibaba Cloud can credibly argue that it is the only hyperscaler that also produces globally competitive open-source foundation models at scale — making the platform, the models, and the silicon all proprietary rather than licensed from third parties. The Qwen open-weight ecosystem's reach — accounting for a reported majority of global open-source LLM downloads — creates a developer moat that pure-play inference APIs cannot replicate.

Domestically, Tencent Cloud and Huawei Cloud are the principal rivals for enterprise AI cloud workloads, with Huawei occupying a particularly strong position in inference hardware given its Ascend chip line. Omdia's 2025 "GenAI Cloud Titans in Asia and Oceania" report highlighted Alibaba's full-stack generative AI solutions and developer-friendly open-source approach as key differentiators.²⁰

The principal risk factors in the competitive position are geopolitical: US export controls on advanced chips constrain access to Nvidia H100 and H200 GPUs in China, creating both a ceiling on available training compute and urgency behind the Zhenwu silicon program. The proprietary-silicon trajectory (Zhenwu M890, 560K+ units shipped) indicates that Alibaba is betting on achieving silicon independence within the next few years.

  Notable Events

Several moments warrant specific attention beyond the main narrative.

In October 2016, ITHome's CEO publicly migrated away from Alibaba Cloud citing service overselling and outages; Alibaba Cloud issued a public apology, an early test of its enterprise reliability positioning.⁸

In April 2025, US authorities restricted exports of Nvidia H20 chips to China, affecting Alibaba alongside ByteDance and Tencent. Limited export licenses reportedly resumed in mid-2025, but the restriction accelerated investment in the Zhenwu silicon program.¹³

The January 2026 milestone of 200,000 derivative models on Hugging Face — the first open-source LLM family to reach that threshold — was a significant public marker of the Qwen ecosystem's reach relative to Meta's Llama family, which had been the prior benchmark for open-source LLM adoption.³

The April 2026 leadership restructuring, combining Fei-Fei Li's appointment as Alibaba Cloud CTO with the elevation of the Tongyi unit, represented the most significant organizational signal that Alibaba intends to compete for global AI platform market share rather than restricting its ambitions to the Chinese market.²

The May 2026 simultaneous announcement of the Zhenwu M890 chip, Qwen3.7-Max, and Panjiu AL128 Supernode marked the first time Alibaba had coordinated a silicon, model, and server announcement at the same event — signaling maturation toward a vertically integrated AI platform narrative analogous to what Apple does in consumer hardware.⁵

  Outlook & Roadmap

Alibaba is in the midst of the most aggressive strategic repositioning in the company's history, betting that AI and cloud will displace e-commerce as the group's primary growth engine. The signals are consistent and multi-layered.

On the model side, the Tongyi Lab's elevation to a standalone commercial business unit with Zhou Jingren as its head gives the Qwen program dedicated P&L accountability and an explicit mandate to monetize foundation models as core commercial products rather than internal infrastructure. The pipeline — Qwen3.7-Max (May 2026), continued open-source releases, ongoing RLHF and reasoning research — indicates a quarterly cadence of model updates for the foreseeable future.

On the infrastructure side, the Zhenwu M890 chip and the Panjiu AL128 Supernode represent a clear roadmap toward silicon independence from Nvidia. Over 560,000 Zhenwu units have shipped as of May 2026; the target trajectory is plainly to bring a meaningful share of Alibaba Cloud's own inference workloads onto proprietary silicon, reducing exposure to US export control risk while lowering unit inference costs.

Global data-center expansion — with first-ever footprints in Brazil, France, and the Netherlands, and expansions in Mexico, Japan, South Korea, Malaysia, and Dubai — directly targets developer markets where open-source Qwen has built an organic install base but where Alibaba previously lacked local serving infrastructure.¹⁶ The appointment of Fei-Fei Li as Alibaba Cloud CTO can be read in this context: her international standing provides enterprise credibility in markets where Alibaba's brand is less established.

The revenue targets are explicit. CEO Eddie Wu's stated goals — RMB 10 billion in annualized model and app revenue by Q2 FY2027 and RMB 30 billion by year-end FY2027 — imply continued high-double-digit growth from the current run rate and require sustained enterprise customer acquisition outside of China.⁷ The five-year $100 billion AI and cloud revenue aspiration is a public commitment around which the entire capital program is organized.

The primary uncertainties are geopolitical (the trajectory of US chip export controls and any retaliatory measures affecting Alibaba's non-China business), competitive (whether Meta, Mistral, or DeepSeek erode the open-source Qwen distribution advantage), and financial (whether the current investment cycle translates to durable revenue before free cash flow pressure intensifies). All revenue targets cited are forward guidance and subject to revision.

---

  References

1. Alibaba Cloud Model Studio — What is Model Studio
2. Alibaba Cloud Model Studio — Model Pricing
3. Alibaba Cloud Model Studio — Models Catalog
4. Qwen — Wikipedia
5. Alibaba Cloud — Wikipedia
6. Alibaba Group FY2026 Earnings Materials
7. Alibaba Cloud debuts Model Studio — Computer Weekly
8. Alibaba unveils Qwen3.6-Plus — Alibaba Cloud Blog
9. Alibaba Cloud international expansion — Alibaba Cloud Blog
10. Alibaba maintains Asia-Pacific leadership — Alibaba Cloud Blog
11. Alibaba CEO on super AI cloud vision — TechNode
12. Alibaba $52.7 bn AI drive — The Register
13. Alibaba data centers in eight locations — Data Center Dynamics
14. Alibaba AI leadership reshuffle / Fei-Fei Li — PanDaily
15. Alibaba Cloud AI revenue and investment strategy — Market Chameleon
16. Alibaba Cloud +38% AI cloud revenue — Washington Times
17. Alibaba Zhenwu M890 chip — The Next Web
18. Alibaba AI structure revamp — Caixin Global

  References

1. Model Studio capabilities, API regions, pricing, and free tier — Alibaba Cloud Model Studio docs; pricing page. ↩ ↩² ↩³ ↩⁴ ↩⁵

2. Fei-Fei Li appointed Alibaba Cloud CTO; Zhou Jingren elevated; Tongyi as standalone unit — PanDaily, April 2026; Caixin Global, April 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Qwen derivative models exceeding 200,000 on Hugging Face and 10 billion cumulative downloads — Wikipedia: Qwen. ↩ ↩² ↩³ ↩⁴ ↩⁵

4. Qwen App 234 million users by May 2026 — reported figure, not independently verified; sourced from Alibaba Group earnings commentary. ↩ ↩²

5. Zhenwu M890 chip, Qwen3.7-Max, and Panjiu AL128 Supernode unveiled May 19, 2026 — The Next Web. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

6. ¥380 billion ($52.7 bn) capex commitment; potential increase to $69 bn reported but unconfirmed — The Register, September 2025; Alibaba Group investor materials. ↩ ↩² ↩³ ↩⁴

7. FY2026 Q4 Cloud Intelligence Group revenue RMB 41,626 M, +38% YoY; AI products ~30% of cloud revenue; 11 consecutive quarters of triple-digit YoY AI growth — Alibaba Group FY2026 earnings; Market Chameleon analysis. ↩ ↩² ↩³ ↩⁴ ↩⁵

8. Alibaba Cloud founding history, Apsara OS origins, Wang Jian — Wikipedia: Alibaba Cloud. ↩ ↩² ↩³ ↩⁴

9. Model Studio debut at Singapore event, February 2024 — Computer Weekly. ↩ ↩²

10. 15,000-GPU Ethernet interconnect research, July 2024 — Wikipedia: Alibaba Cloud. ↩ ↩²

11. Qwen3 released April 28, 2025; Apache 2.0, 235B MoE, 36T token training set — Wikipedia: Qwen. ↩

12. Qwen3.6-Plus released April 2, 2026 — Alibaba Cloud Blog. ↩ ↩² ↩³

13. H200 purchase approval (shared across Alibaba, ByteDance, Tencent); H20 restriction April 2025; limited licenses resumed mid-2025 — reported figures, exact Alibaba allocation unconfirmed. ↩ ↩²

14. GPU pooling system reducing utilization by 82%, October 2025 — The Register. ↩

15. 91 availability zones across 29 regions; NVIDIA Physical AI partnership — Apsara Conference 2025 announcements; The Register. ↩ ↩²

16. New data-center regions in Brazil, France, Netherlands, Mexico, Japan, South Korea, Malaysia, Dubai — Alibaba Cloud Blog; Data Center Dynamics. ↩ ↩²

17. ModelScope community, enterprise customer examples — Alibaba Cloud Blog. ↩ ↩²

18. Qwen2.5-72B ~$0.23/M tokens, ~1/10th GPT-4o price — comparative pricing as published by Alibaba Cloud; methodology varies and prices change frequently. ↩ ↩²

19. Global IaaS market share 7.7%; APAC 22.5% — Gartner April 2026 data as reported by Alibaba Cloud Blog; methodology varies by analyst. ↩ ↩²

20. Omdia 2025 "GenAI Cloud Titans in Asia and Oceania" report — cited in Alibaba Cloud marketing materials; primary report not directly verified. ↩