On July 30, a research team published the AgenticCANN framework on arXiv, a knowledge-augmented agentic evolution system for the automated generation of Ascend C operators, the programming language for Huawei’s NPUs. The paper, identified as arXiv:2607.26661, outlined potentially profound strategic implications. It is the first time this task has been tackled, and the implication is direct: if large language models can autonomously program Chinese hardware, China’s dependence on NVIDIA software—the real moat of CUDA—could be nearing its end. This is not a chip war; it is a war for programming talent, and automation has just changed the rules.
The problem no one had solved: programming Ascend C
Optimizing Ascend C operators is critical for inference performance on Huawei’s NPUs, but it demands deep knowledge of the underlying hardware. The ecosystem around this language is young: it has less documentation, fewer examples, and a far smaller engineer base than CUDA, whose programming model boasts two decades of libraries, tools, and a global developer community. That asymmetry has been the Achilles’ heel of China’s technological sovereignty strategy: manufacturing advanced chips is one thing; programming them efficiently is another.
Large language models had already shown potential in automated CUDA kernel generation. But Ascend C poses unique challenges that, according to the paper’s own abstract, remained unexplored. AgenticCANN is not a minor adaptation of existing techniques: it is the first system to tackle automated operator generation for this language.
AgenticCANN: agentic evolution for a young ecosystem
The approach combines two elements. The first is agentic evolution: autonomous agents that iterate and improve solutions incrementally. The second is knowledge augmentation: the system learns from the specific structure of the hardware and from already-validated optimization patterns. In this way, deep knowledge of Huawei’s NPUs, which traditionally required years of experience, can be extracted, codified, and applied automatically.
From engineer scarcity to agent abundance
The geopolitical context is essential to gauge the scope of this advance. The United States has restricted the export of advanced chips to China and has also limited access to design tools and, de facto, to the building of a developer community around alternative ecosystems. NVIDIA has built its dominance not only on hardware but on CUDA: a software monopoly that turns any alternative into a risky bet for engineers. If a Chinese programmer must choose between CUDA, with its vast documentation and libraries, and Ascend C, with a nascent ecosystem, the rational short-term decision is often CUDA.
AgenticCANN directly attacks that barrier. If large language models can autonomously generate optimized Ascend C operators, the need for a large community of specialized developers drops drastically. This is not about eliminating engineers but multiplying their productivity: an agentic system can iterate multiple variants, evaluate their performance, and select the best ones in a fraction of the time a human team would need.
The missing piece on the digital sovereignty board
For years, the debate over China’s technological sovereignty has centered on chip manufacturing. But making a chip is only the first step: without a mature software ecosystem, hardware is a broken promise. The dependence on CUDA has been a silent but decisive dependency; even with domestic chips, the lack of established programming tools has limited their adoption in artificial intelligence applications.
The timing is no coincidence either. ByteDance projects annualized AI revenues, according to SCMP, positioning it at the forefront of China in monetizing this technology. Kimi K3, Moonshot AI’s model, is already available to enterprises through Fireworks AI on Microsoft Foundry, giving Azure customers a managed route to run it via OpenAI-compatible APIs. And the European Union has launched a call to establish up to seven AI gigafactories with public funding. In this scenario, the ability to program domestic hardware automatically is not a marginal advantage: it is the condition for Chinese AI not to depend on the infrastructure of its adversaries.
Automation as the new frontier of competition
AgenticCANN suggests that China is positioned to achieve something that seemed impossible: bypassing the software barrier. If automatic kernel generation works for Ascend C, China’s computing ecosystem could counter CUDA not by competing on the same terms but by changing the playing field. It will not need a developer community the size of NVIDIA’s; it will need better AI agents capable of programming domestic hardware.
Open questions remain. The paper is recent—published just hours ago—and has not yet passed scientific community scrutiny. No independent benchmarks are known that compare the performance of operators generated by AgenticCANN against manual optimizations. But the direction is unmistakable: the automation of hardware programming is the next frontier of global technological competition. While the West debates the regulation of AI models, China is using those same models to eliminate its greatest structural weakness. The question is no longer who makes the best chips, but who can program them without relying on armies of engineers. If this work is confirmed, the answer could shift the balance of technological power.