China’s AI has posted a run of breakthroughs this year, from DeepSeek’s DeepSeek-V4 series to Moonshot’s Kimi K3, a 3-trillion-parameter open-source model, and Zhipu’s GLM-5.3, whose coding ability approaches top closed models. Behind the models sits compute. Despite US limits on advanced AI chips and semiconductor equipment, domestic compute such as Huawei’s Ascend has broken through the “usable” threshold into “good” and “scaled commercial” use, and the localisation rate of compute infrastructure is climbing fast.
At July’s World Artificial Intelligence Conference, Huawei showed the Ascend 950 supernode in real hardware for the first time, signalling a shift from single-card comparison to system-level breakout that offsets the gap in advanced process nodes. As AI spreads across industries, sovereign compute has become central to data, industrial and national security. Observer.com spoke with Fang Xingdong, a Zhejiang University professor and author of “The Rise of Ascend,” on how Ascend found its way out.
Fang says Ascend’s challenge was unique. Huawei’s past work in communications was “catching up” on tracks others opened. Ascend started differently: when Huawei decided to build AI chips in 2016, the industry was still early, and most AI chips served narrow scenes like cameras and autonomous driving. Huawei started almost in step with global peers. The harder challenge was the strategic uncertainty of true innovation from zero to one, yet under US-China rivalry Ascend became a core driver of Huawei’s resurgence and possibly a landmark of China’s high-tech rise.
The most damaging challenge came from outside. After TSMC stopped foundry work under US pressure, Ascend’s evolution was interrupted and faced a “generation gap” crisis. Many assumed Ascend had lost its future. It ultimately overcame the manufacturing ban and talent drain. The newly released Ascend 950 series showed stunning performance in testing, marking Huawei’s move from reliance on the global advanced supply chain toward full-chain localisation.
Fang argues China’s hardware gap was never mainly technical but about industrial playbook. Early Ascend 910 performance was not worse than foreign peers, but China had not built an ecosystem on its own root technology, dominated by the US system. Building that ecosystem from zero is an unprecedented challenge needing both corporate initiative and state industrial policy, as seen in 5G’s “moderately forward” infrastructure strategy.
On manufacturing, expanding semiconductor capacity takes time, and US limits on ASML EUV lithography choke advanced process. Fang is candid that China remains in catch-up mode and is not yet able to overtake the US overall, but the gap will not widen and can narrow over five to ten years as the paradigm moves from “rootless” to “rooted.” In April Huawei enabled full Ascend supernode support for DeepSeek-V4, and achieved 0day adaptation for Kimi K3 in July.
Fang also highlights “Tao’s Law,” announced by Huawei in May, which targets reducing the time constant tau through logic folding, full-stack synergy and system reconstruction, a pragmatic route as Moore’s Law stalls. On the open versus closed divide, he argues America’s closed model funnels trillions of dollars into a few giants, while China’s open route can reach far more of the world’s 8 billion plus people, and should become the mainstream global AI path.
Editor’s note: This is an adapted translation of the original Sohu report. It has been trimmed and restructured for readability for an international business audience.