China’s AI chip industry is no longer trying to build an Nvidia clone. Under sustained US export controls that block access to the most powerful general-purpose GPUs, the country’s largest technology companies are pivoting toward application-specific integrated circuits, custom chips designed to do one thing extremely well rather than handle any workload. The shift is creating a domestic semiconductor ecosystem that may end up architecturally distinct from the Nvidia-dominated model that powers AI in the West.

At the centre of this divergence is a design choice that export controls have accelerated. General-purpose GPUs, the kind Nvidia sells, are flexible and programmable, making them ideal for the fast-moving research phase of AI development where model architectures change constantly. ASICs sacrifice that flexibility for raw efficiency, delivering faster performance at lower power consumption for specific AI tasks. In a market where the best Nvidia hardware is unavailable, the economics of custom silicon become far more compelling.

Three paths to custom chips

Chinese companies are pursuing three distinct ASIC architectures. Huawei is betting on neural processing units through its Ascend series, including the widely deployed 910C and the upcoming Ascend 950. Cambricon Technologies is building domain-specific architectures with its Siyuan 590 and 690 series. Alibaba is taking a third route through its semiconductor unit T-Head, which launched the Zhenwu M890 parallel processing unit at its annual cloud computing summit last week, claiming three times the performance of its predecessor.

On the GPU side, Moore Threads leads the domestic effort. Founded in 2020 by Zhang Jianzhong, Nvidia’s former China executive, the company has dedicated itself to general-purpose chips like the MTT S5000 series. Biren Technology, Enflame, and Iluvatar CoreX are also competing in the space, but none has achieved the scale of the ASIC leaders.

A Morgan Stanley report published on 8 May made the market dynamics clear. Huawei is projected to capture 62% of China’s domestic AI accelerator market in 2026, followed by Cambricon at 14%. Among big tech firms building proprietary chips, Baidu and Alibaba are each expected to take roughly 5%. The ASIC heavyweights are winning on volume and momentum.

Performance is no longer the bottleneck

The performance gap between Chinese chips and Nvidia’s export-compliant hardware has narrowed significantly. Morgan Stanley data shows that Huawei’s Ascend 950 cards and Cambricon’s Siyuan 690 can outperform Nvidia’s H20, the most powerful chip Nvidia is currently permitted to sell to China, by 50 to 150% as measured in tokens per second.

Huawei expects AI chip revenue to reach roughly $12 billion in 2026, up from $7.5 billion in 2025. Nvidia’s share of the Chinese AI accelerator market has effectively collapsed to zero, a development that CEO Jensen Huang has described as a “horrible outcome” for the United States because it breaks the software dependency on Nvidia’s CUDA ecosystem that took two decades to build.

For China’s highly commercialised AI market, which focuses on deploying applications to hundreds of millions of users rather than conducting frontier research, the ASIC approach makes particular sense. Inference, the process of running a trained model at scale, rewards the kind of narrow optimisation that custom silicon provides. Training new models still benefits from GPU flexibility, but the revenue is in deployment.

The software stack problem

Hardware performance is only half the equation. The deeper challenge for China’s chip industry is breaking the lock-in created by Nvidia’s CUDA platform, the software layer that millions of AI developers worldwide use to write code for Nvidia hardware. CUDA’s network effects are enormous. Virtually every AI framework, every research paper, and every pre-trained model assumes CUDA compatibility.

Huawei is building CANN as its alternative, while Moore Threads has developed MUSA. DeepSeek has spent months rewriting its core code to work with Huawei’s CANN framework, moving away from the CUDA ecosystem. But semiconductor analyst Zhang Haijun notes that as AI models grow more complex, the boundaries between custom ASICs and flexible GPUs are “becoming increasingly blurry,” suggesting that the winning architecture may eventually combine elements of both.

Omdia chief analyst Su Lian Jye frames the choice practically: enterprises with robust AI engineering capabilities and a clear roadmap benefit from ASICs, while those running mixed workloads still lean toward general-purpose GPUs. For now, market momentum in China favours the specialist approach, partly by choice and partly because the general-purpose option from Nvidia remains either unavailable or restricted.

A structurally different ecosystem

The long-term consequence of this divergence may be more significant than the near-term performance benchmarks. If China’s AI industry standardises on a mix of Huawei NPUs, Alibaba PPUs, and Cambricon domain-specific chips, each running its own software stack, the result will be a fragmented but domestically self-sufficient ecosystem that operates on fundamentally different architectural assumptions from the West.

That fragmentation carries costs. Developers building for the Chinese market may need to support multiple hardware platforms simultaneously, increasing complexity. Cross-border AI collaboration becomes harder when the underlying compute stacks are incompatible. And the lack of a single dominant platform means no Chinese chip maker benefits from the kind of ecosystem lock-in that made Nvidia’s CUDA so powerful in the first place.

But the direction is set. US export controls intended to slow China’s AI progress have instead accelerated a structural redesign of its chip industry, pushing it toward custom silicon, domestic software stacks, and an architecture that no longer depends on American hardware. Whether that ecosystem can match the pace of innovation in the Nvidia-powered West is the defining question of the AI chip race.