Alibaba has introduced a new internal concept: “Tongyunge,” shorthand for Tongyi Lab, Alibaba Cloud and T-Head. It refers to the integrated development of large models, cloud computing and chips, a trio Alibaba calls its AI golden triangle.

According to our understanding, “Tongyunge” was a name proposed by Alibaba founder Jack Ma during an April 2025 discussion with teams from Alibaba’s technology businesses. Late last year, Ma also named Ant Group’s health AI assistant “Afu.”

Alibaba Group CEO Eddie Wu also took part in the discussion. He said “cloud + AI + chips” would be the most important triangular support for Alibaba’s future technology strategy. Over the next decade, the biggest growth and uncertainty in cloud computing will come from AI-driven change. AI models that tightly coordinate software and hardware will become essential for the next generation of cloud computing companies.

Ma stressed internally that Tongyunge’s full-stack AI capabilities are both Alibaba’s advantage and its responsibility. Combining the three businesses so they help and support one another, he said, “will bring unprecedented opportunities to the world.”

He did not dwell much on specific businesses or technical road maps, instead focusing on vision. That is also how he usually speaks inside Alibaba: setting the tone.

Ma said: “Tongyunge’s mission is to let every person and every company participate in the AI era. Alibaba’s high technology is not only about pursuing the stars and the sea. It must also protect everyday life, bring change to the lives of ordinary people, and allow every ordinary person to live with dignity. We hope to bring the world into a kind era of high technology.”

Alibaba sees itself as one of the few Chinese technology companies with full-stack AI capabilities. One industry source said AI competition is not limited to consumer applications but is a comprehensive contest. Alibaba’s layout resembles Google’s in China: it has self-developed chips, cloud services, large models, AI applications and a business ecosystem.

On January 29, Alibaba disclosed for the first time its self-developed high-end AI chip, the “Zhenwu 810E,” also known as Alibaba’s PPU, or Parallel Processing Unit.

According to people familiar with the matter, cumulative shipments of T-Head’s Zhenwu PPU have reached the hundreds of thousands, surpassing Cambricon and putting it in the first tier among domestic GPU vendors. A week ago, Alibaba decided to support a future independent listing for its chip company T-Head.

Alibaba is trying to build its own AI narrative. It has pushed its Qwen large model into open source, bet on a broad ecosystem and opened new growth space for its cloud business. Alibaba is already benefiting: cloud revenue is expected to grow 30% to 40% in a single quarter, and almost every rise in its share price has been tied to AI.

The introduction of “Tongyunge” as an overall concept represents Alibaba’s AI infrastructure system. “Before, we were playing scattered cards, breaking up a straight flush and playing it one card at a time. What people can see now is a full straight flush being played,” said a person close to Alibaba’s core management.

Tongyunge: Alibaba Wants to Become New Infrastructure for the AI Era

In MiniMax’s newly filed prospectus, the company said its cap on prepaid computing power purchases from Alibaba Cloud over the next three years, from 2026 to 2028, will total $375 million. Including purchases made in recent years, MiniMax’s cumulative spending on Alibaba Cloud computing power will approach or even exceed Alibaba’s equity investment in the four-year-old large-model startup. That points to a large market for providing model capabilities and training services on the cloud, and renting computing power to other customers and startups. In the AI era, cloud has become a good business.

Many large companies have recognized this. In 2023, ByteDance founder Zhang Yiming said the operating-system-level opportunity of the current era is AI plus computing.

“AI will accelerate customers’ move to the cloud,” Liu Weiguang, senior vice president of Alibaba Cloud Intelligence Group and president of its public cloud business, previously told us. The momentum comes not only from GPU usage or calls to large-model APIs.

Cloud is the starting point of Tongyunge.

Before 2009, Alibaba had nearly bought up all of China’s IBM minicomputers, and Oracle databases were costing it hundreds of millions of dollars a year. It had to build its own cloud. In 2008, Wang Jian joined Alibaba and pushed the company to move away from IOE, shorthand for IBM minicomputers, Oracle databases and EMC storage. A year later, Alibaba Cloud was formally established.

A veteran Alibaba Cloud employee who has worked there for more than 10 years recalled that Alibaba identified two key points about cloud very early on. First, the combination of cloud software and hardware was critical, because cost reduction and efficiency gains improved the price-performance ratio of services. Second, one cloud needed to support multiple chips, and the company had to solve its dependence on any single chip by developing its own. “Creating T-Head in 2018 was an inevitable choice in the development of the cloud,” the person said.

He described the relationship among the Qwen large model, Alibaba Cloud and T-Head chips within Alibaba’s technology strategy: if AI applications are flowers and grass, cloud is the soil, the large model is the air above them, and chips are the moisture in the soil. All are elements AI applications cannot do without.

Qwen, Alibaba Cloud and T-Head chips are not three parallel technology lines. Alibaba has bound them into one AI infrastructure system, a choice meant to reduce uncertainty and improve system controllability.

T-Head’s role is to provide a portion of computing power resources in Alibaba Cloud’s real AI computing scenarios that can be directionally optimized and predictably supplied. This computing power may not be the strongest in general-purpose performance, but it can offer an important differentiated supplement in cost, energy consumption or specific workloads.

Qwen’s advantage does not come entirely from the model itself. Its deep integration with Alibaba Cloud’s computing power scheduling, inference deployment and enterprise delivery systems means the coordination among model, computing power and cloud services shows up more in engineering efficiency and delivery capability than in any single technical metric.

Alibaba Cloud reported revenue of 39.824 billion yuan in the third quarter of 2025, up 34% year on year. Revenue from AI-related products has posted triple-digit growth for nine consecutive quarters, and the market expects Alibaba’s full-year revenue in fiscal 2026, from April 1, 2025 to March 31, 2026, to potentially reach 150 billion yuan.

According to our understanding, Alibaba is considering raising its planned investment in AI infrastructure and cloud computing over the next three years from 380 billion yuan to 480 billion yuan.

T-Head’s PPU: Serving Alibaba First, While Pushing Into External Markets

In early 2025, commercialization of T-Head’s self-developed AI chip, the Zhenwu 810E, entered a scaled phase. On one hand, it began carrying inference computing power demand on Alibaba Cloud. On the other, it launched small-scale resale in March 2025, marking its shift from internal use to broader external commercial operations.

The Zhenwu PPU has served more than 400 customers, including State Grid, the Chinese Academy of Sciences, Xpeng Motors and Sina Weibo. Most of those customers used Alibaba Cloud’s computing power services, which are backed by T-Head chips.

For T-Head, Alibaba Cloud is currently its most important customer.

We understand that T-Head is also actively expanding into external markets. In 2025, it secured orders from two major outside customers: Xpeng Motors and BYD each ordered more than 10,000 PPUs. In 2026, intelligent driving, embodied intelligence, and AI training and inference will all be key expansion areas for T-Head’s PPU.

With the explosion of generative AI, GPUs have become the most critical infrastructure in the global technology industry. NVIDIA has climbed to the top of global technology companies by market value, while domestic chip companies such as Cambricon, Moore Threads and Biren Technology have also set new records in the capital markets.

Six years ago, GPUs were still in their infancy in China. T-Head, then less than two years old, began developing the Zhenwu PPU series in 2020. China’s four domestic general-purpose GPU startups, Biren Technology, Moore Threads, MetaX and Iluvatar CoreX, were all founded around 2020.

T-Head’s main chip portfolio currently includes:

Hanguang 800, its first chip launched in 2019, is an AI inference chip using a self-developed architecture that can process 78,000 images per second. The chip has gradually been applied to Taobao’s main search scenarios during Singles’ Day.

At the same time, T-Head also began developing a more difficult general-purpose CPU chip. At the 2021 Apsara Conference, T-Head released Alibaba’s first general-purpose server chip, Yitian 710, which is now widely used through Alibaba Cloud in areas including video encoding and decoding, high-performance computing and gaming.

The Zhenwu 810E is an integrated AI acceleration chip for training and inference, applicable to AI training, AI inference and autonomous driving. Alibaba has deployed Zhenwu PPUs at scale for training and inference of the Qwen large model.

The Zhenyue 510 SSD controller chip has shipped more than 500,000 units, mainly covering low-latency, high-concurrency and high-reliability scenarios such as cloud computing data centers, high-performance databases, distributed storage and high-frequency financial trading.

Inside Alibaba, use of T-Head’s AI chips is mainly divided into two scenarios: AI inference computing power services provided externally by Alibaba Cloud, and computing power used by Alibaba Group’s internal businesses. For inference, T-Head chips are used first, while training flexibly combines T-Head and third-party chips depending on model size.

According to our understanding, Alibaba’s computing power supply in the first half of 2026 will mainly come from two parts: existing inventory chips suited to inference scenarios, used to meet basic computing power demand, and T-Head’s Zhenwu series chips, which will become the main support for Alibaba Cloud’s inference computing power demand.

One industry source said the performance of T-Head’s Zhenwu 810E under some typical workloads has entered the first tier of domestic computing power chips and can be compared with Huawei’s Ascend 910 series. In specific inference or constrained computing power scenarios, its overall performance can exceed NVIDIA’s A800 and approach NVIDIA’s H20. But in core metrics such as general-purpose computing power scale and memory bandwidth, it still has a clear generational gap with NVIDIA’s H100 and H200.

The industry generally believes that, against a backdrop of constrained computing power supply and persistent external uncertainty, increasing the actual use of domestically developed chips is not only about gradually improving performance. It is also a choice shaped by real constraints.

In terms of development paths, NVIDIA has built a highly closed but extremely sticky developer system through a model based on hardware first-mover advantage and the CUDA ecosystem. Huawei’s Ascend, by contrast, has used full-stack self-development and phased openness to build high barriers in highly customized scenarios such as government cloud and smart cities.

T-Head, by comparison, is trying to focus on system-level coordination among chips, cloud and models. By deeply binding itself with Alibaba Cloud and the Qwen large model, it can validate demand in real business scenarios and push rapid iteration of chips and software.

The future competition among domestic GPUs will hinge on who can build an ecosystem that is easy to use, general-purpose and developer-friendly. Only then will market substitution for NVIDIA become possible.