Published on Oct. 08, 2026

DeepSeek has open-sourced a full set of infrastructure components for Huawei’s Ascend platform. The package includes Ascend support for TileLang, compute libraries and a distributed communication library, corresponding to components DeepSeek previously released for Nvidia hardware, reported the Chinese news agency Cailainshe citing other reports on September 30, 2026 (in Chinese).

TileLang provides a higher-level language for writing operators, the computational building blocks of AI models, while handling hardware-specific instructions underneath. DeepSeek says every TileLang operator it uses in training now has a high-performance Ascend implementation. The “high-performance” claim could not be independently verified.

This release of open-source infrastructure for Huawei’s Ascend platform addresses a practical obstacle to adopting Chinese AI chips, namely the cost of moving software away from Nvidia’s ecosystem.

Huawei supported the development work, including joint optimization of computation and communication in a supernode built around 128 Ascend 950 chips. The results were released through Huawei’s CANN software community almost simultaneously.

The collaboration therefore extends beyond translating code for a different processor to coordinating how many chips work together.

TileLang’s development history helps explain the approach. Developed mainly by a team at Peking University’s School of Computer Science, it was open-sourced in January 2025 and used in DeepSeek-V3.2-Exp from September 2025.
According to Chinese media reports, it now carries most operators used in training the DeepSeek V4 series and has adaptation partnerships with Huawei Ascend, Moore Threads, Sophgo and MetaX.

At Huawei Connect 2025, a team member demonstrated a FlashAttention operator written in 80 lines of code rather than more than 500. That example illustrates code simplification, not a benchmark for an entire model.

Industry insider Wang Minjian told the newspaper that migration costs and accumulated software expertise remained central obstacles for domestic chips.
He described Nvidia’s ecosystem as offering the experience of “write once, run anywhere” and argued that a hardware-neutral operator language could let developers target different chips through different compiler back ends.

The release establishes a concrete route toward that portability, but the supplied research contains no independent benchmark. A new software stack also requires substantial work to become mature.

Nevertheless, the Chinese semiconductor industry has taken another step to throw off its CUDA shackles since Nvidia has become collateral damage in U.S. President Donald Trump’s “chip war” against China.

Numerous high-voltage lines create serious risks. Workers on the front line of grid operation and maintenance are especially exposed. China Southern Power Grid and Unitree have now signed a strategic agreement to jointly develop robots, reported the economic daily Diyi Caijing on September 30, 2026 (in Chinese).

The robots will soon perform some of the industry’s most dangerous tasks, such as removing ice from high-voltage lines in winter. They will also be trained to perform recurring tasks such as inspections, equipment checks, and temperature measurements.

“From the point of view of industrial development, the value of the power industry is not just that it is ‘suited to using robots,’” the report quoted Wang Yongli, deputy director of the Energy Internet Research Center at North China Electric Power University, as saying.

“It has all three core conditions for the commercial use of embodied AI at once. High-value tasks, an environment for continuous training, and the ability to replicate at scale,” he said. “It may not be the industry that embodied AI can enter most easily, but it may well be one of the industries best placed to turn ‘can be demonstrated’ into ‘can produce over the long term.’

In April this year State Grid issued an internal development plan for embodied AI for 2026. It plans to buy 8,500 units of embodied AI equipment for a total of about 6.8 billion yuan, about US$1.00 billion. The planned purchases comprise 5,000 quadruped robots, 500 humanoid robots for live-line work, and 3,000 dual-arm inspection robots.

With the demand of China Southern Power Grid and other energy companies added, the industry expects investment in embodied AI in the power sector to exceed 10 billion yuan (about US$1.5 billion) in 2026.

China's grid operators are spurred on by the Ministry of Industry and Information Technology (MIIT) in Beijing and the state assets regulator, which jointly asked every state-owned enterprise in the country to “choose at least 10 scenarios for on-site training of humanoid robots“, Diyi Caijing reported.

For China's young robot makers, deals such as this one with a huge state-owned entity mean steady revenue and real-world scenarios to train their robots, two resources they urgently need to grow.

The pockets on your jeans may soon be sewn on by a robot. It is one example of how manufacturers in Ningbo, a port city southwest of Hangzhou, are seeking ways to upgrade traditional industries with embodied AI.

More than one billion pieces of clothing are produced in Ningbo every year, which is home to major textile and garment companies such as Youngor, Firs, and Romon.

Attaching pockets to jeans and shirts is a typical use case for a boring and highly repetitive task that robots can handle consistently. Their performance is also improving.

“The accuracy in placing the cut pieces is held within 2 millimeters, and this fine work is now done at scale,” Mou Ruining, an engineer at the Zhejiang Humanoid Robot Innovation Center, was quoted in the Economic Daily (Jingji Ribao) on October 1, 2026 (in Chinese).

Traditional industries such as clothing manufacturing, home appliances, and petrochemicals still account for approximately 55 percent of Ningbo’s industrial base. With its massive deep-water port, the city offers a real-world test environment for innovative robotics startups in nearby Hangzhou and Shanghai.

With a history of producing molds and other parts for vehicles, companies in Ningbo are also pivoting to become suppliers of joints, electric drives and other core components for humanoid and industrial robots.

“This year our company has built 20,000 sets of joint modules for humanoid robots in series production. They are used widely in handling and household services,” the Economic Daily quoted Wu Gaohong, general manager of Ningbo Nashi Electric Drive Robot, as saying.

At six in the morning in the container yard at Ningbo-Zhoushan Port, a quadruped robot makes its first inspection round of the day. It will not bark at you, but can automatically generate its inspection report and upload it to the dispatch center.

Taken together, tens of thousands of use cases like this all over China amount to a silent paradigm shift. It is no longer automatically true that the production of labor-intensive, low-cost manufactured goods must migrate away from countries where birth rates are falling and wage costs are rising rapidly. Ningbo and China, have begun to demonstrate the opposite.

The Chinese Academy of Sciences has launched Panshi ScienceOne, a platform designed to connect research data, models, software tools, large scientific facilities and autonomous laboratories. The academy’s Computer Network Information Center released the platform and an accompanying “Spark Program” on September 30, 2026, Renmin Wang, the People’s Daily website, reported that day (in Chinese).

Panshi aims to support the full research cycle, from forming hypotheses and reviewing literature to simulation, experiments, data analysis and the generation of new findings. It is also developing evaluation systems for trustworthy AI across five academic fields and a traceability mechanism for scientific resources.

Nearly 100 research institutes, universities and technology companies have joined the ecosystem. Through the Spark Program, participants are expected to share computing capacity, data and tool chains while developing talent and moving research results toward practical applications.

The initiative is relevant because China is treating AI for Science as shared infrastructure rather than a collection of isolated laboratory tools. Connecting scarce national facilities with reusable data and models could reduce duplicated work and help research-intensive industries translate discoveries into engineering applications.