The battle for the AI stack: Nvidia versus Huawei, with Europe watching from the sidelines

The US-China technology rivalry is often described as a 'chip war', but that framing is increasingly out of date. The real contest covers the full artificial intelligence stack – not just the hardware that runs AI models, but the software that makes that hardware usable. On both fronts, the competition has narrowed to two companies: US-based Nvidia and China-based Huawei. Europe, despite its indispensable position upstream in the manufacturing process, particularly in lithography, is not a contestant, but rather a spectator.

On hardware, the United States remains ahead. Nvidia accounts for roughly half of the world's installed AI chip stock and as much as two thirds of installed computing capacity, and its chips remain the best performing on the market. Huawei is the most credible challenger, but its most advanced Ascend processors still lag on memory bandwidth, developer tools and power efficiency. Huawei compensates for this by packing processors together, resulting in an estimated two to four times more power usage for equivalent processing output.

Yet the performance gap matters less than it appears. US export controls on Nvidia's chips, combined with Beijing's subsidies, procurement preferences and cheap energy for firms using Huawei's chips, have created a captive market in China large enough to sustain Huawei through its immature phase. Huawei's Ascend chips already power a large share of China's data centres, which has allowed Huawei's AI chip revenue to rise sharply this year – revenue that is reinvested back into R&D. Nvidia, for its part, has largely conceded the Chinese market to Huawei. China is buying time: pressing Washington for access to the most advanced US chips while nurturing its domestic alternative. The Trump administration's transactional approach to export controls has made this two-track strategy easier, not harder.

If the story ended with hardware, Nvidia's dominant position wouldn't be so clear, at least not in China and potentially elsewhere, too, as Huawei starts exporting its AI hardware. What makes Nvidia's advantage durable is CUDA, the proprietary software platform that has become the de facto standard for AI development. CUDA generates powerful network effects: because Nvidia dominates the installed chip base, developers optimise for CUDA-compatible hardware, which enlarges the developer community and the library of pre-optimised code, which in turn makes alternatives even harder to adopt. The result is extraordinary pricing power, with margins on chips of 70%-80%.

Huawei's response to this advantage mirrors the playbook China has used for hardware: subsidise adoption and steer a generation of captive domestic developers towards China's own software layer for AI. More specifically, in August 2025 Huawei open-sourced key components of its CANN toolkit – a direct strike at CUDA's proprietary model – and it has recruited Chinese AI labs and universities to join an Ascend developer community. Just as important is 'torch_npu', a plugin that lets standard PyTorch code, until now effectively tied to CUDA, run on Ascend processors, removing the single biggest switching cost for developers. The collaboration with DeepSeek, whose open-weight models are engineered to run on both Nvidia and Huawei hardware, widens the addressable market further.

None of this guarantees success. Developers still complain that CANN is buggy and less user-friendly than CUDA, whose two-decade stock of optimised libraries is enormous. But the direction of travel is unmistakable: none of these building blocks existed in meaningful form two years ago. The software gap, like the hardware gap, is no longer static.

Where does this leave Europe? Nowhere comfortable. Europe has real assets: ASML's monopoly on extreme ultraviolet lithography, IMEC's research leadership, and world-class AI science – but no designer of AI accelerators of global significance – Arm, in any case UK-based, licenses CPU architectures rather than accelerator designs – and no software ecosystem analogous to CUDA or CANN. Because hardware and software are co-designed, the absence of one makes the other nearly impossible to build. Europe is, in effect, an indispensable input supplier to a race in which it does not itself compete. The economic value in the AI stack accrues to those who control chip design and the software platform, not to component suppliers.

What can Europe do? First, learn from China's trajectory rather than dismiss it: China started from a position of dependency but is closing the gap through a deliberate combination of diplomacy, creating captive demand and sustained subsidies. This does not mean precisely copying China's playbook on freeing itself from US dependence on AI; but Europe needs to realise that over-dependence on the US or China for AI is increasingly problematic and nurture domestic alternatives.

Second, Airbus offers an applicable lesson on a different front. Any attempt from Europe to create its own stack must be done with the spirit of comparative advantage among European countries. Airbus became a European success story not because of a political deal over which country got which factory, but because the division of labour reflected genuine comparative advantage and the governing consortium subordinated politics to commercial logic. The analogy is imperfect – Airbus competed in a duopoly with slow product cycles and no software lock-in, neither of which holds for AI compute – but that is precisely why the governance lesson matters more, not less. A European-level effort to build a more independent AI stack needs to take into account that Europe does have relevant companies it can leverage, such as ASML, IMEC, Infineon, STMicroelectronics and Zeiss. In addition, Europe needs to continue pushing to control its own regulatory space.

Third, Europe needs to build demand through coordinated public procurement for AI compute. EuroHPC, the AI factories and InvestAI already fund supply-side capacity, but they do not aggregate demand, and procurement remains fragmented at member-state level. The first areas where this can happen are the public administration, healthcare, defence and/or research.

While none of this delivers a European Nvidia in the near term, AI is too important not to try. Building Airbus took twenty years; but the alternative is a future in which Europe's compute infrastructure and the terms under which European firms build AI models is subordinated permanently to Washington or Beijing. The global technology contest has evolved from who designs the best chips to who controls the full stack. Europe should at least give it a real try.

ZhōngHuá Mundus is a newsletter by Bruegel, bringing you monthly analysis of China in the world, as seen from Europe.

Sign up for the newsletter