The Spring Festival Gala is just the beginning; the war for AI cloud has only just started.

When the live broadcast signal of the 2026 CCTV Spring Festival Gala lit up for the first second, it not only heralded the end of the old year but also marked the beginning of China's AI cloud era.

This was no ordinary technical support for a gala event. According to official disclosures, Volcano Engine was the exclusive AI cloud partner for the 2026 Spring Festival Gala, providing end-to-end technical support that covered core scenarios such as ultra-high-definition live streaming, real-time interactive scheduling, and massive concurrent user handling. This meant that the interaction requests of hundreds of millions of viewers and the smooth delivery of live broadcasts during the gala relied entirely on its AI cloud technology.

For Volcano Engine, this was a showcase of its capabilities to the entire industry. 'Only cloud services that can withstand the triple tests of traffic, stability, and security at the scale of the Spring Festival Gala are qualified to compete in the core battlefield of the AI era,' an industry insider told Xinmou.

Thousands of miles away, at Alibaba's Xixi Campus in Hangzhou, Alibaba Cloud Intelligence was equally busy. Despite not securing the exclusive AI cloud partnership for the Spring Festival Gala, they reserved ample computing power for the billions of AI calls made through QianWen App, Taobao, Alipay, and Fliggy during the Spring Festival period. Just a month earlier, according to LatePost, Alibaba was considering increasing its planned investment of 380 billion yuan over the next three years in AI infrastructure and cloud computing to 480 billion yuan.

As for Baidu Intelligent Cloud, the Spring Festival interactive features of Wenxin Yiyan, the cloud service scheduling for intelligent driving, and the AI computing power demands of Baidu Search all rested on their clusters. Shen Dou, Executive Vice President of Baidu Group and President of Baidu Intelligent Cloud Business Group, once stated, 'In the era of the intelligent economy, new infrastructure is essential for support. This is the AI-first cloud.'

Over the past two decades, Chinese internet giants have fought key battles in e-commerce, social media, short videos, and payments. But this time is different—previous wars were fought over traffic entry points; this time, it's about the underlying infrastructure of the AI era. If large models are the engines of the AI era, then AI cloud is the runway and airport that supports them. Microsoft has leveraged its deep integration with OpenAI through Azure to breach the defenses of AWS and Google in the global cloud market. Google has redefined cloud service standards in the AI era with its full-stack capabilities of Gemini + TPU + Google Cloud.

In China, the three giants with the most core large model capabilities—ByteDance, Alibaba, and Baidu—are waging a war in the AI cloud arena that will determine the landscape of the next decade. This is not a tentative (tànshì xìng, exploratory) battle in a localized field; losing it could mean losing entry tickets to the entire AI era.

01

The War Truly Begins After the Turning Point of Large Models

The real trigger for China's AI cloud war was not the launch of ChatGPT in 2023 but the full-scale eruption of the large model industry in 2025.

When the large model craze first emerged in 2023, cloud services were merely 'computing power suppliers' for large model companies—startup and tech giant teams procured GPU resources from cloud providers for training and inference, with a relationship more akin to that of a client and a contractor. The industry consensus at the time was that the core competition in large models lay in algorithms, with cloud services serving as mere supporting infrastructure.

But everything changed in 2025. The advent of DeepSeek-R1 directly shifted the competition in large models from laboratory benchmarks to real-world scenarios for C-end users. With Doubao surpassing 100 million daily active users, Yuanbao's user base growing tenfold, and QianWen App undergoing frequent iterations, large models transformed from a technical concept into products used by hundreds of millions of users daily.

This transformation brought about disruptive demands for cloud services. An industry veteran calculated that a large model product with over 100 million daily active users processes more than 50 trillion tokens daily, requiring tens of thousands of GPU cards for inference alone each day. Meanwhile, a full-scale pre-training of a large model with hundreds of billions of parameters demands thousands of cards running continuously for months, with computing power costs reaching hundreds of millions of yuan. More importantly, the upper limit of a large model's capabilities depends not only on algorithms but also on underlying computing power scheduling, network optimization, and inference acceleration—all of which are controlled by cloud providers.

The industry logic was completely reversed: previously, large models needed the cloud; now, the cloud must bind with large models. Cloud providers without large model capabilities can only serve as mere 'porters' of computing power without any premium. Conversely, large models without proprietary cloud services will always be constrained by computing power and unable to deploy their capabilities to a vast array of enterprise clients.

Microsoft Azure's success set an example for domestic giants. By the end of 2025, Azure far exceeded expectations, with AI clients contributing significantly to non-AI revenue, and a computing power shortage projected to persist into 2026. The key lay in its deep integration with OpenAI—from dedicated clusters for model training to specialized architectures for inference optimization and Copilot services for enterprise clients, cloud and large models were fully integrated.

Google followed a similar path, using its self-developed TPU chips to support the training and inference of its Gemini large model and delivering a full-stack capability of 'chips + models + cloud' to enterprise clients through Google Cloud. According to Gartner data, Google Cloud's global market share rose to 13% in 2025, outpacing AWS in growth, with core increment (zēngliàng, incremental) gains also coming from AI-related services.

Against this backdrop, ByteDance, Alibaba, and Baidu—three domestic giants each possessing full-stack capabilities in 'chips-large models-cloud services'—made AI cloud their core battlefield in 2026. The three chose vastly different paths, much like they did when large models first emerged three years ago:

Alibaba bet on a full-stack closed loop from the outset, treating its 'TongYunGe' strategy—an integration of Tongyi Lab, Alibaba Cloud, and T-Head Semiconductor—as its core competitive edge, attempting to replicate Google's full-stack story.

Baidu was the earliest domestic player to layout (bùjú, deploy) AI-native cloud, using a combination of the PaddlePaddle deep learning framework, the Wenxin large model, and the self-developed Kunlun chip to deeply cultivate scenarios in government, enterprise, industrial, and intelligent transportation fields, following a route of 'technological precipitate + scenario landing.'

ByteDance's Volcano Engine started the latest but with the strongest momentum, leveraging extreme internal scenario validation at ByteDance—Doubao's daily processing of 63 trillion tokens, a growth rate exceeding 200% in half a year, and the traffic tests of Douyin's 800 million daily active users. Volcano Engine aimed to replicate the capabilities honed to perfection internally at ByteDance for external use.

The market landscape was being rewritten accordingly. Before 2025, the domestic cloud market was led by Alibaba Cloud, followed closely by Huawei Cloud and Tencent Cloud, with Baidu Intelligent Cloud and Volcano Engine ranking fourth and fifth. However, after 2025, AI cloud became the sole growth engine. Cloud providers without large model capabilities saw their growth rates slow, while the three with core large models rapidly reshaped the market landscape.

Among them, Volcano Engine's growth was the most astonishing. According to a report by international market research firm Omdia, as of October 2025, Volcano Engine held a 15% market share in the global enterprise-grade MaaS (Model as a Service) market, indicating its mature capabilities in serving B-end application scenarios.

It should be noted that MaaS is not the entirety of the AI cloud concept, with its revenue accounting for only a small portion of the entire public cloud market. However, this does not prevent Alibaba from refocusing its attention on AI cloud. In Liu Weiguang's view, Senior Vice President of Alibaba Cloud Intelligence Group and President of the Public Cloud Business Unit, even though MaaS has significant growth potential, the key lies in winning the entire new market of AI cloud, establishing a full-stack AI cloud capability that integrates software and hardware, and enabling enterprises to leverage stronger AI models at lower costs. This is the decisive factor in competition.

02 The Frontline Battle: Full-Stack Capabilities, Customer Acquisition, and Computing Power Arms Race

The first major clash among the three giants was the computing power arms race.

For AI cloud, computing power is the most critical ammunition. According to previous reports by LatePost, Alibaba is one of the companies with the largest GPU reserves in China, having procured tens of thousands of high-end GPUs from overseas in 2025 alone. It even purchased large quantities of consumer-grade graphics cards like the RTX 4090 to build inference clusters and supplement inference throughput.

Alibaba's confidence stems from its self-developed chip layout (bùjú, layout). T-Head Semiconductor's self-developed Zhenwu 810E AI chip has become one of the main chips in China's newly added AI computing power market, securing its first external major client order in 2025. In January 2026, sources close to the market revealed that Alibaba decided to support T-Head's future independent listing, aiming to further strengthen its chip capabilities and provide underlying support for Alibaba Cloud's full-stack strategy.

ByteDance's Volcano Engine follows a route of 'internal refinement and external output.' ByteDance's internal computing power demands serve as the best training ground for Volcano Engine—Doubao's daily processing of 63 trillion tokens, equivalent to handling billions of user requests daily, imposes far higher requirements on computing power scheduling, inference optimization, and disaster recovery backup than those of most enterprise clients. Additionally, securing the AI cloud partnership for the 2026 Spring Festival Gala was essentially a demonstration to the entire industry: Volcano Engine's capabilities can withstand the most extreme traffic tests in China.

More noteworthy is ByteDance's global layout . According to relevant media reports, ByteDance will fully accelerate the globalization of its AI business in 2026, with Southeast Asia as the core focus, and Volcano Engine serving as the infrastructure for ByteDance's AI globalization. In other words, Volcano Engine must not only support the growth of Doubao's overseas version, Dola, but also serve a large number of Chinese enterprises going global, following the path Microsoft Azure took to rise in the global market.

Baidu Intelligent Cloud, on the other hand, was the earliest domestic player to layout (bùjú, deploy) AI computing power. Baidu established the Institute of Deep Learning in 2013, becoming the first Chinese internet company to elevate deep learning to a core technological innovation. In 2016, it launched PaddlePaddle, China's first deep learning platform. In 2018, it announced the self-development of China's first full-function AI chip for the cloud, 'Kunlun.' In 2019, it released the world's first knowledge-enhanced large model with hundreds of billions of parameters. By 2025, Baidu Intelligent Cloud had lit up China's first fully self-developed 30,000-card Kunlun chip cluster, capable of simultaneously supporting the training of multiple large models with hundreds of billions of parameters, marking an important milestone in the construction of domestic AI computing power clusters.

Baidu's core advantage lies in its deep integration with scenarios. The AI capabilities of Baidu Intelligent Cloud are deeply embedded in all of Baidu's core businesses: the AI computing power demands of search, user services of Wenxin Yiyan, the cloud-edge-end coordination of intelligent driving, and the city-level implementation of ACE Intelligent Transportation are all supported by Baidu Intelligent Cloud. Especially in the fields of government, enterprise, industrial internet, and intelligent transportation, Baidu Intelligent Cloud has secured many city-level intelligent transportation projects in China, with the stable computing power demands from these projects forming its fundamental base.

Behind the computing power arms race lies the competition for customers. Since 2025, the AI-related computing power demands of domestic internet companies, startups, and government and enterprise clients have exploded, becoming the core battleground for the three giants.

Alibaba Cloud's advantage lies in China's most comprehensive cloud service ecosystem and over a decade of accumulated enterprise clients. Volcano Engine's breakthrough lies in internet clients and overseas clients. ByteDance itself is one of China's largest internet companies, with a far deeper understanding of internet industry demands than other cloud providers. Baidu Intelligent Cloud, meanwhile, firmly holds onto the fundamental base of government, enterprise, and industrial sectors.

03 The Invisible Battlefield: Organizational Synergy and Departmental Rivalry

The AI cloud war among the giants is not just about computing power and customers but also about a company's internal organizational synergy capabilities. The deep integration of cloud and large models essentially requires breaking down internal departmental barriers within a company to achieve end-to-end collaboration from chips, infrastructure, models, to products—precisely the most challenging issue for large companies to resolve.

Alibaba clearly recognizes this issue. Alibaba formally proposed the 'TongYunGe' trinity strategy, fully integrating Tongyi Lab, Alibaba Cloud, and T-Head Semiconductor. The underlying logic is that previously, Tongyi Lab belonged to DAMO Academy, T-Head was an independent chip company, and Alibaba Cloud was an independent business group, with certain departmental barriers among them, often leading to fragmented efforts and difficulty in forming synergies.

After the implementation of the 'TongYunGe' strategy, technical roadmaps were largely unified, and resources were allocated centrally. For example, the training of Tongyi large models, from chip selection and cluster construction to training frameworks and inference optimization, was jointly designed by T-Head, Alibaba Cloud, and Tongyi Lab throughout the process, breaking down previous departmental barriers.

ByteDance's Volcano Engine follows a different synergy route. ByteDance's large model team, the Seed department, has been deeply integrated with Volcano Engine from the outset—all computing power services for Doubao are provided by Volcano Engine, with joint design and refinement in every detail from the underlying architecture of model training to inference optimization.

In February 2025, Wu Yonghui, former Vice President of Google DeepMind, joined ByteDance to lead the theoretical research of large models in the Seed department. His first move was to break down barriers between model departments and teams, enabling data sharing across all links and teams, with the most crucial aspect being integration with Volcano Engine's infrastructure team.

Baidu Intelligent Cloud was the earliest domestic player to achieve 'cloud-model integration.' Its technological R&D system underwent significant reforms, with the technical platform department (TPG) responsible for Wenxin large model R&D split into a basic model R&D department and an applied model R&D department, both reporting directly to Li Yanhong. This adjustment marked Baidu's shift from 'basic R&D priority' to 'application-driven,' pushing backbone (gǔgàn, backbone) talents more adept at technological implementation to the frontlines and accelerating the transition of large models from laboratories to industrial scenarios.

An internal source told Xinmou that Baidu's advantage lies in being an AI-native company from the outset. PaddlePaddle, Wenxin, and Intelligent Cloud were all essentially built around AI implementation, without much historical baggage. The difference is that 'Alibaba Cloud started with the cloud and then moved into large models; we started with AI technology and then built the cloud, so the synergy between cloud and models is naturally smoother.'

However, both ByteDance, Alibaba, and Baidu must confront internal resource allocation issues. Especially since 2025, the incubation and cultivation of new businesses have demanded substantial computing power and technological resources, inevitably leading to resource competition with existing business fundamentals. This situation severely tests the strategic decisions and judgments of those at the helm.

04 Rebuilding Infrastructure: Will AI Cloud Make the World More Open or More Closed?

In 2006, Amazon launched AWS, ushering in the era of cloud computing. The industry's vision at the time was that cloud computing would turn computing power into a public infrastructure akin to water and electricity, enabling anyone or any company to access powerful computing capabilities at extremely low costs and breaking down technological barriers erected by giants.

Two decades later, cloud computing has indeed transformed the entire tech industry, but it has not broken down barriers. Instead, it has further entrenched the advantages of giants. Today, over 80% of the global cloud market is controlled by AWS, Microsoft, and Google. In China, the top five cloud providers account for over 90% of the market share.

The advent of AI cloud is pushing this trend to the extreme.

The founder of an AI startup did the math: to stay in the game in the AI era, at least 100,000 high-end GPU cards are needed. Estimating conservatively at an annualized cost of $10,000 per card, that translates to an investment of about $1 billion a year—and that's just the basic investment in computing power, not including the costs of talent, technology, and scenario implementation. For startups, this threshold is almost impossible to overcome.

The more core barrier lies in data and scenarios. Among all the data in the world, only a tiny fraction is publicly accessible via cloud services. The rest is either offline cold data or held by tech giants. ByteDance has massive content and interaction data from Douyin and Doubao, Alibaba has transaction data from e-commerce and payments, and Baidu has scenario data from search and intelligent transportation. This data is the core fuel for training large models and optimizing AI cloud services, which startups simply cannot access.

The deep integration of cloud and large models is forming a new closed loop. For example, if a customer uses Alibaba Cloud's services, they are more inclined to use Tongyi Qianwen's large model; once they use Tongyi Qianwen's model, it becomes difficult to leave Alibaba Cloud's computing support. The same logic applies to ByteDance and Baidu—this closed loop will make the barriers for giants higher and higher (higher and higher), while leaving startups with less and less room to survive.

The U.S. market has already validated this trend. Three years after ChatGPT ignited the AI wave, hundreds of AI startups have been established, funded, and explored new products globally. However, to date, the vast majority of market share remains in the hands of two giants: Microsoft and Google. Microsoft, through its integration of Azure and OpenAI, has captured a large number of enterprise customers; Google, with its full-stack AI capabilities, has defended its core business. Startups find it extremely difficult to carve out sufficient survival space within the giants' closed loops.

On the other hand, AI cloud is also breaking down some old barriers. Alibaba Cloud's Tongyi Qianwen and Baidu's Wenxin large models have both open-sourced their core versions, allowing any company or individual to develop their own applications based on these open models, thereby lowering the barrier to AI adoption to a certain extent. Volcano Engine has also opened up many of ByteDance's internally refined AI capabilities to small and medium-sized enterprises through cloud services, enabling them to gain the same AI capabilities as ByteDance without having to build from scratch.

By the end of 2025, Volcano Engine released a preview version of the Doubao mobile assistant, deeply embedding the Doubao large model into the mobile system, bypassing various apps and allowing users to complete operations that originally required repeated clicks through voice commands. Almost simultaneously, the Qianwen App announced its integration with Taobao, Alipay, Fliggy, and Gaode, with a "Task Assistant" feature capable of completing complex, multi-step tasks on behalf of users. Baidu's Wenxin Yiyan has also been fully integrated into all of Baidu's ecosystem products, including search, maps, and cloud storage.

The giants are all trying to create a unified gateway for the AI era, and AI cloud serves as the underlying support for this gateway. When users can complete all operations and access all information primarily through a single dialog box or gateway, it may seem on the surface that users have more freedom of choice. However, the control points that determine what they can see, what they can do, and the order in which various services appear are actually fewer.

We may now be building another Tower of Babel. In legend, humans attempted to build a tower to heaven using a single language, but the project ultimately collapsed due to the fragmentation of languages. Today, the giants are using AI to try to create a unified, omniscient gateway, constructing a Tower of Babel for the AI era.

The war for AI cloud has only just begun. By 2026, this war will enter a white-hot stage, with all three giants investing more resources and talent to compete for the underlying infrastructure of the AI era. This is not a quick battle but a protracted war; the final outcome will depend not only on the amount of computing power or the density of talent but also on who can truly open up AI capabilities to more people and who can genuinely create new scenarios and values for the AI era.

This article is an original piece from Xinmou.

— END —