From going to IOE to going to CUDA, why is Alibaba "free-range" Brother Pingtou at this time?

New Stand Pro

2026.11.20 10:32

· From overseas

Full text 5888 words

Text | New Stand Pro

On the evening of November 19, Alibaba CEO Wu Yongming revealed in an earnings call that Qianwen has trained next-generation models on Pingtouge chips, and Bailian has also provided model inference services based on domestic chips.

Three months ago, Alibaba announced a business adjustment: Cloud Intelligence Group and Pingtou Ge were integrated into "AI cloud and computing services." The two businesses are put into the same financial division, and Alibaba hopes that chips and cloud computing will cooperate more closely.

On September 11, Pingtou Ge secretly submitted a listing application to the Hong Kong Stock Exchange. Alibaba has just strengthened the synergy between the two businesses and is preparing to let Pingtou Ge accept independent valuation from the capital market.

According to the spin-off announcement, Pingtou Ge will still be a subsidiary of Alibaba after its listing. Alibaba hopes to retain the synergy between chips and cloud computing while allowing Pingtou Ge to have independent financing channels. This is significantly different from the arrangement prepared to completely spin off the Cloud Intelligence Group three years ago.

Pingtou Ge already has annual revenue of tens of billions of yuan. Compared with 2019, it no longer only makes AI inference chips. Server CPU, network interconnection, smart network card, storage controller, and the core chips needed in a data center of Alibaba Cloud, Pingtou Ge almost wants to make them himself.

Zhenwu has served more than 650 companies through Alibaba Cloud, and Pingtou Ge has also begun to cooperate with partners such as server manufacturers to provide hardware to the outside world. Now, it is preparing to go public independently.

According to Alibaba’s announcement, the independent listing is mainly to allow investors to independently evaluate the value of Pingtouge, so that it can directly enter the equity and debt markets in an independent capacity, and to make it more recognized by customers, suppliers and partners. However, how much external customers can be enlarged will have to be tested by the market after independent valuation.

After eight years of raising him, Brother Flathead finally wants to start his own business

When Hanguang 800 was first released, Alibaba had already thought about its business model.

In 2019, while introducing the computing performance of the chip, Zhang Jianfeng announced that the Hanguang 800 was specially customized for scenarios on Alibaba Cloud. "If it is market-oriented, more detailed optimization issues need to be considered." This sentence is modest in the context of a press conference, but borderline in the chip business: the first priority of this chip is to make Taobao's image recognition faster, rather than letting external customers queue up to place orders.

This is in line with Alibaba’s early technological tradition. Around 2009, Taobao's business scale grew rapidly, and traditional IT systems became increasingly difficult to support demand. Wang Jian promoted the Feitian distributed computing system, and Alibaba gradually replaced expensive minicomputers with general-purpose servers. On May 17, 2013, Alipay’s last IBM minicomputer went offline.

Since then, Alibaba has developed the computing power that originally solved its own problems into cloud services for other enterprises.

Brother Pingtou initially followed this path. When Alibaba creates a new business, there are usually two prerequisites: there are real workloads internally, and the performance reduction can be verified by itself. Hanguang 800 first serves Taobao image recognition, and Yitian first serves Alibaba Cloud general computing. This is the logic.

The conditions for release are more stringent, external commercialization must actually occur, parent cash flow can no longer be guaranteed indefinitely, and the capital market is willing to give multiples to independent tracks. At the time of "1+6+N" in 2023, Alibaba believed that "independence means liberation". Rookie submitted orders and Alibaba Cloud lined up, but later took them back. The reason is not that the assets are bad, but that once they leave the matrix, the synergistic value will be cut off - the value of Cainiao lies in the fulfillment network, and the value of the cloud lies in the continuity within the data and regulatory framework.

Pingtou Ge is the first asset where synergy does not constitute a prerequisite for survival: asset-light chip design, no heavy asset baggage, no license restrictions, external customers and external delivery have actually taken place.

In September 2018, Alibaba integrated the chip R&D teams of Zhongtian Micro and Damo Academy to establish Pingtou Ge. In addition to the Hanguang 800 mentioned above, the Yitian 710 will be released in 2021. The latter uses Arm architecture, is designed for cloud data centers, and undertakes general computing tasks. As of February this year, Pingtouge has delivered more than 470,000 chips, and its annual revenue has reached 10 billion yuan.

The automotive industry was an early entry area. In May of this year, Pingtou Ge revealed that the smart driving industry has deployed more than 130,000 computing cards, and its customers include more than 30 car companies and smart driving companies such as Changan, GAC, BYD, and Xpeng. They require a lot of computing power to train perception, prediction and end-to-end driving models. For car companies, directly purchasing servers and building computer rooms is not the only option. They can also rent Zhenwu computing power on Alibaba Cloud on demand.

The usage of financial institutions is different. Shanghai Pudong Development Bank introduced Zhenwu into the computing cluster to process documents and forms with large models. Large model companies require larger-scale computing resources. Pingtou Ge not only has customers, but also has increasingly complex usage scenarios.

The consolidation in June created an information black box. After merging Pingtou Ge into "AI Cloud and Computing Services", the market "saw" the existence of this asset for the first time in the official caliber of Alibaba's financial report, but could not see its separate revenue structure, separate profits, and separate costs. The merger method allows Pingtou Ge's tens of billions of revenue to be folded into a business segment with a growth rate of 45%, which not only proves "presence", but also temporarily hides "volume".

The core of the difference is the quality of the word "customer".

Wu Yongming revealed in March this year that among Alibaba Cloud's public cloud and hybrid cloud products, more than 60% of Pingtouge chips are being used by external commercial customers. The 60% here refers to the chips deployed on the cloud serving external customers. It does not mean that 60% of Pingtouge's revenue comes from direct sales of chips outside the group.

A car company rents Alibaba Cloud computing power and pays for cloud services. Pingtou Ge provides chips for Alibaba Cloud, and revenue may be generated through internal procurement and settlement within the group. How much of the annual revenue of tens of billions of yuan comes from the group's internal supply, and how much comes from the chip and server sales outside the group. Public information has not yet provided a complete breakdown.

The market has obviously not reached a consensus on how Pingtou Ge should be priced. An investment institutional person gave a valuation range of US$20 billion to US$30 billion, while JPMorgan Chase’s previous rough estimate was US$25 billion to US$62 billion. The disagreement itself is a signal: some people see the prototype of "China's NVIDIA", while others see a chip department that has not yet fully escaped the walls of Alibaba Cloud.

Same orchard, different harvest

Pingtouge chose to go public at this time and caught up with the changes in the focus of domestic AI chip competition.

In the past few years, domestic AI chips have entered the procurement lists of operators, government and enterprises from the computer rooms of major Internet manufacturers. In addition to computing performance, companies discuss bidding, equipment delivery and subsequent maintenance.

In October, IDC released a set of data. In the first half of 2026, in China's data center AI acceleration chip market, Nvidia's shipment share has dropped to 34%, while Huawei Ascend has reached 28%, shipping 572,000 units in the first half of the year. Pingtouge, Cambrian and Kunlun Core occupy approximately 7%, 5% and 4% of the market share respectively.

Export controls have freed up the market for domestic chips, but insufficient production capacity has set a ceiling for Pingtou Ge’s shipments. Wu Yongming admitted that the proportion of self-developed chips deployed within Alibaba Cloud is "still not large", mainly because of limited production capacity.

But in addition to shipment share, continued adoption by external customers also depends on software migration, system stability and delivery capabilities.

The CUDA ecosystem that NVIDIA has established over the years has its advantages in the accumulation of compilers, software tools and developers. Changing to a domestic chip often means that the algorithm team must re-adapt the software, and the operation and maintenance team must verify the stability of the server and network; the unit price of computing power is cheap, but it does not necessarily mean that the overall cost is lower.

Huawei has invested in this area very early. Focusing on Ascend chips, Huawei continues to build the CANN software platform and enters the data center market through complete machines, clusters, and super-node products. Chip, interconnect, software and system delivery are brought together so customers don't have to solve all adaptation issues from scratch.

There are different ways to sell chips: one is to hand the chips and servers directly to customers, and the other is to deploy the chips in the cloud and sell computing power to customers.

The former method is inseparable from server manufacturers, system integrators and operator channels. For banks, government agencies or large enterprises, procurement often requires bidding, testing, acceptance and other links. Just because a chip can pass the performance test does not mean that the server equipped with it can successfully enter the client room.

Kunlun Core has accumulated some experience on this road.

In 2025, in China Mobile's centralized procurement of artificial intelligence general computing equipment, servers equipped with Kunlun cores won 70%, 70% and 100% of the three "CUDA-like ecology" bid packages respectively, involving billion-yuan orders. In June this year, Qi Wei, vice president of R&D at Kunlun Core, revealed that the proportion of business from the external market has exceeded the proportion of internal supply to Baidu.

Another change occurred on the revenue side. After going public in 2020, Cambrian once faced pressure from high R&D investment and revenue scale. In the first three quarters of 2026, it achieved revenue of 9.526 billion yuan and net profit attributable to the parent company of 3.715 billion yuan. Although high R&D investment, inventory and cash flow will still affect the quality of operations, the rapid growth of revenue and profits has provided the market with independently measurable operating results.

Pingtouge’s advantage is that Alibaba Cloud provides it with ready-made customers, data centers and actual workloads.

In May of this year, Wu Yongming said that in the future, with the expansion of Pingtouge chip production capacity, Alibaba may directly sell AI servers to computing power centers and data center service providers, and jointly build computing power centers with partners. Compared with just deploying chips within Alibaba Cloud, this means that Pingtou Ge began to strive for more opportunities to directly sell hardware.

The tests to be faced on this road are also very specific. Kunlun Core has centralized procurement orders from operators, and Cambrian has independent financial reports. Pingtou Ge needs to gradually establish his own sales, channels, delivery and after-sales service systems to obtain more direct purchase orders from third-party customers.

Production capacity is also affecting the pace of competition.

Wu Yongming once admitted that the proportion of self-developed chips deployed within Alibaba Cloud is still low. The important reason is that domestic chip production capacity is limited. Pingtou Ge adopts a division of labor model in chip design and manufacturing, relying on external supply chains for wafer manufacturing, advanced packaging and other links. When production capacity is tight, it must not only meet Alibaba Cloud's internal needs, but also arrange deliveries for customers outside the group.

Even if the capital market is willing to pay for growth expectations, chip companies' revenue ultimately depends on actual delivery.

However, Pingtou does not have to completely copy the sales paths of Kunlun Core and Cambrian. Alibaba Cloud has mature enterprise customer channels. In the future, it can help Pingtou Ge expand hardware sales and continue to absorb a large amount of chip production capacity through cloud services. Compared with competing for customers alone from the beginning, this is Pingtou Ge's unique advantage.

The question is whether Pingtou can gradually transform the customer relationships that Alibaba Cloud has established into a hardware business that he can operate in the long term.

Brother Pingtou’s strongest trump card is still in the hands of Alibaba Cloud

In 2003, notebook computers became increasingly popular, and consumers began to pursue thinness, lightness, battery life, and wireless networking. Intel launched Centrino, which combines the Pentium M processor, 855 series chipset and wireless network components into a platform for unified design, optimization and verification.

As of September of that year, there were 92 notebook models based on Centrino technology. Intel no longer only sells CPUs, but also begins to participate in defining the overall machine experience, and uses computer manufacturers to bring the entire technology platform to the market.

More than 20 years later, Brother Pingtou is also trying to shift from a single chip to a computing platform. At the Yunqi Conference in September this year, Pingtou Ge released a new generation of Zhenwu V900. According to the official parameters, its single-chip performance is three times that of the previous generation M890.

Also unveiled with the V900 is a more complete system. The new generation of Panjiu super node server is planned to be launched in the first quarter of 2027. In addition to the AI ​​computing chip, the system also integrates the ICN Switch interconnect chip, Panmai smart network card and Zhenyue storage main control chip. The different chip products that Pingtou had accumulated in the previous eight years began to be put into the same computing system.

Alibaba Cloud has enough computing tasks that can be used to test the performance of the entire chip system. Today, large-scale models increasingly rely on multi-card parallel computing. Agent further increases the requirements for CPU scheduling, database access, tool execution, and storage reading and writing. No matter how powerful a GPU is, it cannot solve all performance bottlenecks in the data center.

Google chose to develop TPU for its own workloads, and Amazon optimized AWS cloud services through self-developed chips such as Graviton and Trainium. Alibaba also hopes to use its advantages of owning chips, models and cloud platforms to reduce coordination costs between products from different manufacturers.

This full-stack capability has obvious value within Alibaba. The model team can optimize the algorithm for the chip, the chip team can adjust the design according to the model load, and the cloud platform is responsible for scheduling and deployment. For training, inference, and large-scale Agent applications, software and hardware collaboration can reduce part of the adaptation and operation and maintenance costs.

But as a chip company, Pingtou Ge cannot win the market solely by relying on this kind of internal collaboration.

On April 3, the Zhenwu Wanka intelligent computing cluster jointly built by China Telecom Guangdong Company and Alibaba Cloud went online in Shaoguan, Guangdong. A month later, Jiechuang Intelligence confirmed in investor exchanges that it had cooperated with Pingtouge to use its own brand servers, Pingtouge chips and Evergreen Cloud applications to develop AI all-in-one machines for central state-owned enterprises, government affairs and other customers.

Pingtou Ge already has cases of external landing, but there is still a long way to go between "walking into other people's computer rooms" and "independent hosting". Alibaba Cloud, operators and system integrators still play an important role in platform construction, deployment or service delivery.

Can an enterprise without the long-term on-site support of Alibaba Cloud's technical team be able to deploy, maintain, and upgrade this system with the same ease?

Software is one of the hurdles.

In July this year, Pingtou Ge open sourced the T-Head SAIL software stack and claimed that it has been adapted to more than 260 mainstream training and inference frameworks. The average adaptation time for new versions of mainstream inference frameworks is less than 7 days. It hopes that engineers who originally used tools such as PyTorch and vLLM will not have to rewrite a lot of code for Zhenwu.

However, Alibaba Cloud's public SDK compatibility test also left some specific details. Among them, the pass rate of vLLM 0.9.1 version in a set of unit tests was 74.52%. In comparison, the pass rate of vLLM 0.8.3 version in the same document is 94.65%, and there are obvious differences in the adaptation performance of different versions.

Another SDK description also mentioned that when SGLang runs some Qwen3 models, there are problems such as reduced high concurrency throughput and poor performance of specific configurations, and gives suggestions for switching calculation backends and adjusting configurations.

For customers who are preparing to purchase computing power on a large scale, these details are precisely the factors that determine the cost of deployment.

The software adaptation problem faced by Brother Pingtou is also being solved by other domestic chip manufacturers. In September this year, Huawei disclosed that Ascend has covered more than 90 mainstream third-party communities such as PyTorch, Triton, and vLLM, and external developers accounted for 61% of the CANN open source community. On November 18, Cambrian announced the completion of Day 0 inference adaptation of SenseNova Loop-9B, and opened the inference link of the new model on the MLU platform through vllm_mlu.

The software ecology of domestic chips is rapidly making up for the mistakes. The competition has moved from supporting multiple frameworks to how quickly new models can run through after release, whether performance is stable, and how much engineering costs customers need to pay for migration.

The same model can be jointly supported by the chip, software and cloud platform teams within Alibaba Cloud. In the external computer room, customers also need to confirm whether the operating system version, inference framework, network configuration, and monitoring tools match, and whether new problems will be introduced after the upgrade.

Centrino was able to move into a large number of computer brands back then by solving these complex problems as early as possible. If Pingtou Ge wants to turn an internal computing system into an industry product, he also needs to establish similar capabilities.

At least other customers can use them in a long-term, stable and economical way without having to equip Alibaba Cloud's technical team.

write at the end

On the eve of the 2018 Yunqi Conference, Jack Ma made a phone call to Zhang Jianfeng. Alibaba is about to set up its own chip company. According to the company's tradition, the team originally prepared a lot of animal names, which imply smallness and agility.

Jack Ma finally chose "Brother Flathead". This kind of honey badger that lives in Africa is not very big, but it is famous for its courage to challenge animals much larger than itself. There is a saying circulating on the Internet to describe it: take life and death lightly, and do it if you don't accept it. After the name was confirmed, many people in the team even went to get a haircut the next day.

But the cruelty of the chip business is that Brother Pingtou can be optimistic about life and death, but customers cannot ignore the stability of the production system. A chip shutdown may affect banking services, car model training, or the normal operation of thousands of computing cards.

Whether Pingtou Ge can succeed in this step depends not only on its own chip roadmap, but also on what Alibaba is testing with Pingtou Ge.

What it is testing is a new template for capital allocation: retaining controlling rights to defend Xinyun Synergy, while pushing hard technology assets to the open market to accept independent pricing and independent performance assessment.

If the market accepts this structure, Alibaba will have a reusable template for other hard technology assets with potential for external commercialization. If valuations and liquidity fall short of expectations, the path to spin-off will be narrowed again.

Alibaba Cloud is still the biggest source of confidence for Brother Pingtou, but the capital market still depends on performance.

Respect for the original work, if you are qualified, you should be appreciated.

According to the real-name requirement of the Cybersecurity Law, please bind your mobile phone number before posting a comment.

Express

10:46

Five departments issued the 2026 version of the "Guidance Catalog" for marine industry investment and financing

FTSE China A50 Index futures decline widens to 2%

10:43

The development of my country's marine medicine and functional products industry is accelerating, with a scale of nearly 100 billion

10:40

The robot concept made some changes in the market, and Clai Electromechanical hit the daily limit.

10:33

Spot silver's intraday decline expanded to 3%, at $41.56 per ounce

10:28

The Shanghai Composite Index fell more than 2%, and nearly 5,300 stocks fell in Shanghai, Shenzhen and Beijing.

10:22

The concept of physical AI is partially recovering, and Fantuoshu has increased by more than 10%.

The agricultural sector rebounded in shock, and Wanxiang Denong hit the daily limit.

10:14

Xi Jinping meets with Australian Prime Minister Albanese

10:09

PCB concept rose again during the session, Ultrasonic Electronics and Aohong Electronics hit their daily limit

10:08

Hang Seng Technology Index's decline widened to 4%

SK hynix considers converting its Kioxia-related convertible bonds into stocks

10:07

The concept of ice and snow economy is partially active, and the hot and cold icebergs hit the daily limit

10:03

Alibaba International Station’s orders and buyers in Europe have increased by 70% in the past three years, and the London Summit launched a small and medium-sized enterprise entrepreneurial package

10:01

Jiyuan Technology completed tens of millions of yuan in angel round financing

09:57

The ChiNext Index fell more than 3%, with nearly 5,200 stocks falling.

09:56

May Day Vision launches A-share listing guidance work

09:55

China launches first smart cockpit AI grading evaluation system

09:54

Ministry of Finance successfully issues US dollar sovereign bonds in Hong Kong

09:44

NetEase Hong Kong shares fell more than 18%

扫描下载App