Export controls, domestic models, and a bifurcating stack.
China’s AI race is a contest over models, compute, semiconductors, deployment and strategic autonomy, conducted under increasingly consequential US technology restrictions. Chinese firms including DeepSeek, Alibaba and Baidu are advancing models and adoption while facing constrained access to leading AI chips, making efficiency, domestic capacity and open-source distribution central competitive tools. The result is pressure toward more sovereign and potentially bifurcated AI stacks, with access to hardware, cloud capacity and models becoming geopolitical questions.
The US-China AI competition extends beyond headline model performance. It encompasses access to advanced chips and chipmaking technology, domestic semiconductor capacity, training and inference infrastructure, talent, model releases, deployment channels and the rules governing their use. This is a contest over the sovereign AI stack: the ability to control the dependencies required to build and operate AI systems at scale.
China’s national AI ambitions are therefore closely tied to industrial policy and security priorities. A countrywide push has combined government funding, a domestic chip focus and changes to the regulatory environment, while Chinese technology companies continue to work within domestic rules that include tighter regulation and censorship. These conditions shape both the direction of AI development and the kinds of products and deployments that can scale.
US export controls on advanced AI chips and related technology are a central organizing force in China’s AI development. The restrictions have squeezed Chinese AI companies’ access to Nvidia hardware and other high-end compute resources, while US rules have also targeted chips that fall beneath earlier technical thresholds and considered limits involving AI memory chips and equipment. Chinese institutions and companies had relied on chips from Nvidia and AMD before export limits tightened.
The restrictions are intended to slow access to frontier compute, but their effects are not one-directional. They may impede China’s AI-chip progress and create a bottleneck for future models, while also increasing incentives for Chinese rivals to develop alternatives, optimize limited hardware and alter model-development approaches. The competitive question is consequently not simply whether compute can be restricted, but how rapidly firms can substitute, conserve or obtain it through other channels.
DeepSeek has become an important example of adaptation under chip constraints. Its releases have been associated with efforts to produce strong models without the latest chips, and Chinese firms such as DeepSeek and 01.ai have pursued lower-cost approaches that use smaller training data sets. These strategies suggest that efficiency in training, model design and collaboration can matter alongside absolute access to the most advanced hardware.
Open-source releases are also a route to distribution and refinement. Alibaba, Baidu and DeepSeek have used open-source models to engage outside developers and global talent, potentially decentralizing development while reducing dependence on a purely domestic ecosystem. Broad adoption of models such as DeepSeek in Chinese services, including among automakers and smartphone sellers, illustrates how model competition can shift from creating a model to building applications and distribution around it.
China’s domestic semiconductor effort is a long-term response to exposure in the supply chain. GPUs have become a focal point of US-China technology competition, and the push for locally controlled chips is linked to the wider goal of securing cloud, inference and deployment capacity. Yet Chinese AI executives and other observers have pointed to limited resources and export curbs as persistent constraints on closing the gap with the US.
Physical export controls also face enforcement and routing challenges. Chinese AI developers have accessed advanced American chips through intermediaries, and Chinese companies have reportedly used Malaysian data centers by transporting hard drives containing training data to facilities with advanced Nvidia chips. Such practices reflect AI access arbitrage: restrictions on hardware movement can coexist with ways to obtain compute or model-development capacity through third countries, affiliates or remote infrastructure.
The competition is increasingly about who sets the terms of AI access, not just who develops the strongest model. US policy has explored rules that govern how many AI chips countries can obtain, while model capabilities from the United States and China can prompt policy shifts. As models become relevant to economic activity, critical institutions and national security, customer eligibility, hosting location and deployment geography can become matters of state policy.
Chinese open models may broaden China’s international technology reach, including through lightweight, lower-cost offerings and infrastructure associated with Huawei and ZTE in Africa. At the same time, the United States retains important strengths in AI chips, model quality and sales, even as Chinese firms expand their models and deployments. What to watch is whether export controls produce durable separation in chips, cloud and model ecosystems, or whether open-source distribution, alternative infrastructure and indirect compute access keep the global AI stack interconnected.
Grounded in the archive and knowledge graph. Browse all topic guides, the concept reference, or the posts.