Introduction
Five Cents examines China’s AI Chip Push and the question behind Huawei’s new roadmap: can it build a usable AI-computing alternative without matching Nvidia chip for chip? We’ll look at its Ascend processors, giant cluster plans, manufacturing bottlenecks, and why export controls are reshaping the market. The place to start is Huawei’s shift from selling a processor to building the whole machine around it.
What is happening
Huawei’s latest strategy is not simply to launch another AI accelerator. Its planned Ascend nine-sixty chips are paired with Atlas systems. These combine processors, servers, networking, storage, management hardware, and software. The goal is a tightly integrated AI infrastructure stack, built largely around Chinese suppliers.
That matters because AI performance is no longer determined by one chip alone. Training a large model means coordinating thousands of accelerators, moving vast amounts of data between them, and keeping them busy rather than waiting on memory or network traffic. A weaker individual processor can still be useful if the cluster is available, efficient, well connected, and supported by workable software.
Huawei is betting on that system-level argument. It says its UnifiedBus interconnect can link extremely large numbers of processors. Its SuperPod and SuperCluster designs aim to make separate servers behave more like one computer. Those are company targets, not independently verified proof of performance. But the direction is clear: Huawei wants to compete on the entire data centre, not just on a benchmark chart.
The commercial logic is especially strong inside China. Export restrictions have made access to leading foreign AI chips less predictable, while Beijing is pushing for secure domestic supplies of strategic AI technology. Government procurement, state-linked customers, cloud providers, laboratories, and model developers can give domestic hardware a large testing ground.
The context
That market is valuable even if early products cost more or require extra engineering. Developers who port models to Huawei’s CANN software platform build experience outside Nvidia’s CUDA ecosystem. Each deployment creates more tools, more trained engineers, and more pressure for libraries and frameworks to support the local hardware.
But this is where ambition meets the hard part: manufacturing. Huawei can design an advanced accelerator, yet producing it in large, reliable volumes depends on foundry capacity, yields, equipment, memory, packaging, and power-hungry data-centre infrastructure. China has shown it can make advanced chips through difficult workarounds. But those methods can be slower and less economical than the manufacturing processes used by the leading global foundries.
The constraint is not whether a chip can be demonstrated. It is whether enough good chips can be made, on time and at a viable cost. High-bandwidth memory is another major obstacle. AI accelerators need huge amounts of very fast memory, and that memory is difficult to manufacture and package. Huawei has published ambitious specifications for future products, but domestic HBM capacity remains less mature than the supply available to leading international chip makers.
Recent pricing pressure in China suggests memory is becoming as important a bottleneck as the compute chip itself. Then there is software. Nvidia’s advantage is not only its hardware. Years of code, libraries, developer habits, and production tools have grown around CUDA. Huawei has expanded support for PyTorch and says more models are being trained natively on Ascend. That is meaningful progress, but compatibility on paper is different from smooth performance in a real production environment.
Porting workloads can require rewriting kernels
Porting workloads can require rewriting kernels, tuning memory use, and debugging tools that mature platforms have refined over many years. US export controls sit at the centre of this story. They restrict China’s access to advanced processors and key manufacturing technologies, raising the cost and difficulty of catching up.
Yet they also give domestic substitution a stronger political and commercial case. If foreign supply may be limited by licensing rules, Chinese firms have a reason to accept the compromises of local hardware. That does not mean the controls have failed. They can slow access to frontier computing and constrain manufacturing progress.
But they also encourage the creation of a more self-contained Chinese stack. The policy tension is real: limiting China’s most advanced capabilities may reduce short-term access, while continued sales of selected foreign chips could preserve some dependence on foreign ecosystems. For AI companies, the likely result is not a clean switch. Many will run mixed fleets.
Foreign accelerators may remain attractive for software maturity and broad compatibility where they are legally available. Huawei can appeal through domestic supply, policy support, and lower exposure to licensing uncertainty. Operating both, however, adds cost and complexity. The broader consequence is a more fragmented AI supply chain. The competition now includes memory, packaging, optical links, power, cooling, cloud infrastructure, and developer tools.
What comes next
It is becoming a contest between ecosystems rather than a race for the fastest single processor. So Huawei has not shown that China has overtaken Nvidia at the leading edge. It has shown a serious effort to make chip-level parity less decisive by building an alternative system around domestic demand and integrated infrastructure.
Its real test is execution: enough advanced chips, enough memory, software that developers can trust, and clusters that work at scale.
To continue, you can generate Five Cents Nvidia’s CUDA Moat or Export Controls and the Chip Supply Chain. And with that, you're up to speed in a few minutes.