In recent years, continuous iterations of Chinese chips and the successive commercialization of supernode clusters have alleviated the tight supply of domestic computing power.
As China's AI industry no longer needs to be anxious about the thirst for computing power, the next new problem emerges: how can the computing power already built truly enter enterprise businesses and transform into usable and easy-to-use AI capabilities?
As the reservoir gradually fills up, canals need to be built to deliver resources to where they are needed, and AI is no exception.
With abundant computing power, a mature and open software system is also needed to efficiently channel the computing power into model R&D and industrial applications.
Take DeepSeek as an example. As an advanced and open-source Chinese model, it is one of the top choices for many industries to adopt inclusive AI. On September 30, DeepSeek open-sourced its infrastructure components for the Ascend computing platform, including high-level language compilation tools, computing libraries, and distributed communication libraries, corresponding to the components it previously open-sourced for NVIDIA.
In the past, leveraging hardware performance advantages required complex operator development and cluster tuning; today, the optimization results verified by DeepSeek practices are released in an open-source manner. This means that a vast number of model researchers, enterprises, and individual developers can obtain a set of reusable and improvable underlying tools from Ascend foundational software represented by CANN when using Ascend computing power, without the need to reinvent the wheel.
This greatly lowers the threshold for model development and training, enabling more enterprises and developers to use advanced models and computing power at lower engineering costs.
From this, it is not difficult to see that foundational software is the critical channel between computing hardware and model applications. The open-source and open nature of foundational software directly determines whether AI capabilities can be smoothly used by more people.
After bidding farewell to the thirst for computing power, China's AI still needs a grand canal of foundational software to channel computing power and models into all walks of life, injecting vitality into the intelligent economy. So, how is the construction of this canal progressing?
Almost every monopolistic software field will eventually see the emergence of an open-source counterpart. Take operating systems, for example: Windows strictly controls its source code, and subsequently, Linux became completely open-source to the whole society.
The open-source and open nature of foundational software is a necessary means to break the single supply structure. In the AI era, an open software ecosystem is particularly needed. DeepSeek's recent open-sourcing of Ascend foundational components profoundly reflects this demand.
AI developers often sigh that "one day in AI is like a year in the human world." The iteration and update speed of models is extremely fast. Developers must keep up with model iterations while handling underlying adaptation and performance optimization. If every change in model structure requires writing operators from scratch, re-tuning, and waiting for the hardware platform to update supporting tools, R&D time will be heavily consumed in repetitive engineering work.
Therefore, a set of software tools that can keep up with model changes and unleash hardware performance is urgently needed by AI developers. A closed system can only wait for a centralized vendor to release a new version. In contrast, an open-source, open, and collaboratively driven technical ecosystem is obviously more adaptable to the development pace of the AI era.
On September 30, DeepSeek open-sourced a series of foundational components for the Ascend AI foundational software and hardware platform, among which the TileKernels project added Ascend backend support, enabling the same set of Python interfaces to run on overseas GPUs and Ascend NPUs. Developers do not need to force-fit their work for Ascend; they can efficiently run their original work on Ascend computing clusters, reducing engineering pressure on developers and decreasing reliance on a single supplier.
Model vendors are willing to open-source, but without an open underlying platform, developers still cannot use them effectively. Ascend software infrastructure, represented by CANN, plays this critical role: it provides support for underlying development and hardware adaptation, and offers foundational software capabilities for upper-layer AI frameworks and applications. Developers can conduct operator development, compilation, tuning, and execution within this system.
Of course, objectively speaking, open-source models and open software do not mean completely exempting hardware adaptation work. DeepSeek officials also mentioned that the current implementations of the open-source projects are targeted at specific software and hardware environments such as the Ascend 950 series products, and do not provide unconditional support for all Ascend models. In other words, in actual use, developers may still have some necessary engineering work to do.
This may be a weathervane. After China's AI infrastructure has achieved conditions such as leading models and abundant computing power, the importance of a complete and open foundational software system is becoming increasingly impossible to ignore.
Previously, many Chinese foundational software projects were also open-sourced to attract developers and users. However, a large number of open-source projects are in a "three-no" state of no maintenance, no community, and no applications. The day the code is uploaded is the end of its life.
If the open-source software ecosystem is not well developed, the related hardware and applications will naturally fail to grow. NVIDIA GPUs serve as a good comparison. The large-scale application of overseas chips is inseparable from the community resources, co-building habits, and developer stickiness accumulated over more than a decade, which serve as a moat.
Open-source and openness pave the riverbed for foundational software. But to keep the software ecosystem flowing endlessly, it is also necessary to attract developers to continuously join. How to make developers willing to come, stay, and continuously contribute in the community has been the most difficult threshold for domestic foundational software to cross over the past few decades.
Fortunately, the Agentic era is bringing positive changes.
First, the original development methods have changed, bringing new opportunities and variables for Agentic-based innovative software platforms.
Agents assisting in task completion are becoming a new paradigm adopted by an increasing number of developers, and past workflows are facing reshaping. Take operator development as an example. This was once a daunting task for many engineers, requiring an understanding of the model's computational logic, familiarity with the programming methods of specific hardware, and then sequentially completing code writing, compilation, testing, and performance tuning. Today, Agents are beginning to link these steps together. A software platform where Agents can be selected for the entire process means a simpler and smoother development experience, thereby attracting more AI-native developers.
Take Ascend's CANNBot as an example. Developers can describe the computational requirements of operators, and the tool assists in completing code generation, compilation, and testing, providing a reusable workflow for the entire process. The intelligent development capabilities of MindStudio also use agent technology to automatically identify precision anomalies and solve performance bottlenecks. Simple and easy-to-use Agents run through the entire development process, greatly reducing manual engineering efforts and creating conditions for CANN to attract AI-native developers.
Changes in development methods have won more existing users for China's foundational software. Then, after AI democratizes technology, it brings new incremental users to foundational software.
When Agents can assist in handling some engineering tasks, engineers and developers who were originally unfamiliar with this system also have the opportunity to participate more directly, continuously attracting external developers to the software ecosystem. Currently, we see that some Chinese foundational software has crossed the ecological tipping point. For example, Ascend CANN has entered normalized open-source community operations, with monthly active developers exceeding 5,200, open-source code reaching 37.23 million lines, and more than half of the active contributions coming from outside Ascend. The active participation and continuous contributions of external developers in the community indicate that the Ascend ecosystem is no longer supported by a single enterprise and truly possesses sustainable vitality.
However, it is necessary to maintain a bit of clarity. Currently, CANN ranks first in the activity ranking of open-source projects in China, but even with this scale, it is not considered large when placed in the magnitude of top global open-source communities. The next stage of China's foundational software has just begun, and there is still a long way to go.
As more developers are willing to continuously use, contribute to, and maintain it, boats are finally sailing endlessly on the canal of China's foundational software. But sailing is not the purpose; conducting commercial activities and creating economic benefits by leveraging the transport capacity of the water is.
The true value of foundational software cannot stop at development; it must ultimately lead to industry.
Computing power and models flow into all walks of life through foundational software, improving efficiency and creating value in real production and operations. Only then can technical capabilities be transformed into economic benefits, which in turn support the development of the software ecosystem.
Take supernodes as an example. The emergence of supernodes provides new support for the development of China's AI computing infrastructure and shows the industry the possibility of alleviating computing power supply pressure, becoming the mainstream form of computing infrastructure in the Agentic AI era. However, for enterprises to use supernodes effectively, it is absolutely not as simple as just spending money on computing power construction. The ease of use, compatibility, and development efficiency of the software platform will all affect the cost and efficiency of enterprises using computing power.
Over the past year, in addition to hardware breakthroughs such as Ascend supernodes, the exploration in foundational software in China's AI is equally worthy of attention. To lower the threshold for computing power and allow hardware capabilities to enter actual application scenarios more smoothly, Ascend has done a lot of work.
On the one hand, opening up as much as possible, the operator programming, operator libraries, communication libraries, and tools of CANN have all been fully open-sourced, and the Mind series application enablement suites have also been fully open-sourced. This allows developers to modify and adapt based on the open code, making the operators better fit their own model structures and computational requirements. After development is completed, they can also use supporting tools for performance tuning.
Through full opening, developers take more engineering work into their own hands, which not only reduces the cost of repetitive development and shortens the model adaptation cycle, but also improves resource utilization, allowing existing computing power to exert greater value.
On the other hand, Ascend CANN actively integrates into the mainstream development ecosystem, embraces cutting-edge industry technologies and community developers, and adapts to developers' habits rather than requiring others to adapt to it. For example, the Triton API is fully compatible with the community. Developers can continue to use the familiar Triton programming interfaces and development methods, reducing the workload of relearning toolchains to adapt to new hardware. An engineer accustomed to mainstream frameworks like PyTorch can run the code written in PyTorch on Ascend with almost no modifications.
It can be seen that supernodes provide a stronger computing foundation for China's AI, while foundational software is attempting to shorten the distance between computing power and industrial applications. Of course, software optimization still has its limits. To thoroughly solve the problems of difficult computing power for enterprises and high AI costs, it requires not only the continuous improvement of software but also continuous breakthroughs in hardware to provide more abundant computing power supply, as well as enterprises exploring more valuable AI application scenarios to put computing power and other investments to the best use.
Only when computing power supply, software capabilities, and industrial needs form a linkage can China's AI truly sail into the next leg of the intelligent economy.
Today, the capabilities of large models are continuously improving, enterprises' willingness to use AI and agents is unprecedented, and investments in computing infrastructure are also increasing. But these investments, stronger models, and more computing power do not necessarily bring business value. From this perspective, the open-source and open nature of foundational software is not just a technical choice, but also the only way to release the value of the AI industry.
When AI needs to enter more extensive application scenarios, open-source and open foundational software is needed to lower the usage threshold, reduce inefficient investments in reinventing the wheel, and accelerate the industrial implementation of models and computing power.
Today, the canal has been built. Ascend foundational software, represented by CANN, has become one of the important infrastructures supporting China's AI.
But do not rejoice too early. Just as a canal is not completely finished once built and requires continuous maintenance and reinforcement in the later stage, the vitality of foundational software comes from the ecosystem and cannot be separated from the source of vitality provided by developers and commercial users. China's AI foundational software cannot do without you, without everyone, to use it, pay attention to it, raise an issue in the community, or upload a piece of code. Flow together with the foundational software and sail into the AI era together.