The U.S.-China artificial intelligence (AI) rivalry is no longer confined to semiconductor manufacturing. While American companies like NVIDIA dominate the hardware market, the battle for global AI leadership now hinges on software ecosystems, strategic partnerships, and the ability to create self-reinforcing technological advantages. This shift reflects a broader contest over control of the “AI stack”—the interdependent layers of hardware, software, and infrastructure that power machine learning models. As China’s Huawei and other domestic firms push to challenge U.S. dominance, the competition is evolving into a complex struggle for influence over both physical components and digital frameworks.
The AI Stack: Hardware as Foundation, Software as Leverage
The AI stack consists of two critical layers: hardware (semiconductors) and software (development tools and platforms). While semiconductors provide the computational power needed to run AI models, software determines how efficiently those models are developed, optimized, and deployed. NVIDIA’s dominance in the hardware market—accounting for up to 70% of global AI chip margins—has been reinforced by its CUDA platform, a proprietary software architecture that enables developers to harness GPU capabilities seamlessly. CUDA has become the de facto standard for AI development, creating network effects that lock users into NVIDIA’s ecosystem.
China’s Huawei, meanwhile, faces significant hurdles in closing the hardware gap but is making strides through state-backed initiatives. Its Ascend chips and CANN (Compute Architecture for Neural Networks) aim to replicate NVIDIA’s software advantages, while partnerships with domestic labs like DeepSeek are helping bridge the divide between U.S.-centric tools and Chinese alternatives. The battle for the AI stack thus extends beyond chip manufacturing to the creation of competing software ecosystems that shape the future of AI innovation.
NVIDIA’s Software Moat: A Two-Decade Advantage
NVIDIA’s CUDA platform is a cornerstone of its dominance. Launched in 2006, CUDA has evolved into an open standard for parallel computing, enabling developers to write code that runs efficiently on GPUs. Its success lies in three key factors: network effects, developer lock-in, and ecosystem integration.
Network effects mean that as more developers adopt CUDA, the platform becomes increasingly valuable. A growing user base attracts more libraries, tools, and frameworks (like PyTorch), which in turn attract further adoption. This self-reinforcing cycle makes it difficult for alternatives to gain traction. Developer lock-in is reinforced by NVIDIA’s zero-price strategy: CUDA is freely available, reducing barriers to entry while ensuring long-term dependency on its hardware.
The integration of CUDA with popular frameworks like PyTorch solidifies its position. Most AI developers write code in PyTorch, which relies on CUDA for execution. This creates a switching cost—developers would need to rewrite their code or abandon familiar tools to use competing platforms. As a result, NVIDIA’s margins remain exceptionally high, with chip profits reaching 70-80%, far exceeding those of rivals like AMD or Intel.
Huawei’s Strategy: Bridging the Gap Through Open Innovation
Huawei’s approach to challenging NVIDIA’s dominance is both strategic and incremental. The company has focused on two primary fronts: open-sourcing its software tools and leveraging existing frameworks to reduce adoption friction.
The open-sourcing of CANN in 2025 was a direct challenge to CUDA’s proprietary model. By making its toolkit freely available, Huawei aimed to attract developers and build a community around its Ascend chips. However, early feedback from developers highlights usability challenges: CANN is less intuitive than CUDA, and its ecosystem remains underdeveloped compared to the decades of maturity in NVIDIA’s platform.
To address these gaps, Huawei has also developed torch_npu, a PyTorch-compatible backend plugin that allows developers to run their existing code on Ascend hardware without rewriting it. This lowers the switching cost for developers accustomed to PyTorch and CUDA, making it easier to transition to Huawei’s stack. The strategy mirrors China’s broader approach to technological catch-up: open-access tools, state subsidies, and domestic market protection.
Collaborations with domestic AI labs like DeepSeek further bolster Huawei’s position. DeepSeek’s models, which are optimized for both NVIDIA and Ascend chips, serve as a bridge between the two ecosystems. By supporting open-weight models—publicly available, high-performance AI systems—Huawei is creating a pathway for developers to experiment with its hardware while maintaining compatibility with global standards.
The Strategic Implications of Software Ecosystems
The competition over software ecosystems has profound implications for global AI leadership. NVIDIA’s CUDA advantage is not just a technical edge—it represents economic and strategic power. Control over the development tools that shape AI innovation gives companies leverage over both developers and end-users. A developer trained in PyTorch and CUDA is unlikely to switch to an alternative platform without significant cost or risk, reinforcing NVIDIA’s position as the de facto standard.
For China, overcoming this software moat would mark a critical breakthrough. If Huawei can build a self-sustaining ecosystem around CANN and MindSpore, it could reduce reliance on U.S.-centric tools while expanding its domestic market. However, challenges remain: CUDA’s extensive library of pre-optimized code, the time required to mature CANN, and the inertia of established developer communities all pose obstacles. Whether Huawei can replicate NVIDIA’s network effects remains an open question.
Europe’s Absence in the AI Stack Race
While the U.S.-China rivalry dominates headlines, Europe’s role in the AI stack competition is more nuanced. The continent lacks both a global AI chip designer and a software platform comparable to CUDA or CANN. This absence limits Europe’s ability to influence the broader AI ecosystem, as hardware and software layers are deeply intertwined through co-design.
Without domestic chip design capabilities at scale, European developers face challenges in optimizing software platforms for their specific needs. The economic value of the AI stack—measured in market share, innovation, and control over global standards—accrues overwhelmingly to those who dominate both hardware and software. While Europe possesses critical inputs like advanced lithography equipment, it remains an input provider rather than a full contender in the race for AI leadership.
This dynamic raises questions about Europe’s strategic options. Targeted industrial policies could help close the compute gap, while open-source initiatives focused on niche areas—such as automotive AI or climate modeling—might allow Europe to carve out a unique role. However, achieving genuine sovereignty or reducing dependence requires a deliberate strategy that balances innovation with global collaboration.
Conclusion
The U.S.-China AI rivalry is no longer solely about semiconductors; it is a battle for control over the entire AI stack. NVIDIA’s CUDA platform has established an enduring advantage through network effects and developer lock-in, while Huawei’s efforts to replicate this model highlight the growing ambition of Chinese firms. The success of these strategies will determine whether China can close the gap in both hardware and software, reshaping the global landscape of AI innovation.
For readers, key areas to watch include Huawei’s progress in building a viable alternative to CUDA, Europe’s potential role in shaping a more balanced AI ecosystem, and the broader implications of stack control for global technological governance. As the competition intensifies, the outcome will not only define the next era of AI but also influence how innovation is shaped across borders.
Source
Read the original report: https://www.bruegel.org/analysis/stack-battles-us-china-artificial-intelligence-rivalry-moving-beyond-chips-alone
Leave a Reply