Navigating the geopolitics of US–China AI competition | Sihao Huang

Strategic Context and Existential Risks

• A scenario where AI reaches human-level capability across domains by 2040 could trigger rapid acceleration in science and technology via AI-driven research.
• Global coordination is critical to prevent a scenario where a single nation (e.g., China or the US) deploys superintelligence without international consultation.
• A "first-mover" advantage could lead to existential conflict; for example, Chinese leadership might perceive US dominance via AI as an existential threat, potentially prompting pre-emptive military action.
• The "race dynamic" creates a prisoner's dilemma where nations feel compelled to deploy unsafe military AI to maintain battlefield parity, despite knowing the systems are unreliable.
• Potential pathways to catastrophic risk include not just frontier AGI, but smaller, specialized models capable of biological weapon design or sophisticated cyber warfare.
• Authoritarian regimes may use AI to eliminate the need for a "selectorate" (a ruling coalition), automating repression to reduce the number of people the regime must appease, thereby increasing the regime's stability and capacity for extreme policy.

Current State of Chinese AI Capabilities

• China's large language models (LLMs) are currently estimated to be 1.5 to 2 years behind the US frontier (e.g., GPT-4 level), though open-source models like DeepSeek V2 and Moonshot Kimi are competitive.
• Despite being behind in frontier general models, China leads in computer vision, with companies like SenseTime and Hikvision deploying advanced surveillance capable of tracking individuals via clothing and gait without facial recognition.
• The Chinese LLM market features intense competition among four to five major contenders (e.g., Alibaba, Zhipu, Baidu), resulting in a "price war" where inference costs have dropped by 90–99%.
• Chinese developers face a "compute overhang" from stockpiling American chips (A800, H800) prior to export controls, which allowed them to reach near-parity on previous generation models.
• While capable of innovation (e.g., Tsinghua students publishing Sora-like vision transformer papers), Chinese labs lack the massive GPU clusters and data infrastructure required to mass-produce frontier video generation models.

Semiconductor Constraints and Export Controls

• US export controls implemented in October 2022 and updated in 2023 restrict China's access to leading-edge chips (NVIDIA A100, H100) and Extreme Ultraviolet (EUV) lithography machines.
• China is currently unable to produce 3nm semiconductors domestically and remains stuck on 7nm and older nodes, significantly limiting the compute density required for future models (GPT-5/6).
• The "compute density" metric, introduced in 2023 controls, prevents China from buying even cut-down chips by focusing on performance-per-silicon-area rather than just bandwidth.
• Despite restrictions, China is importing "compliant" lower-performance chips (e.g., NVIDIA H20) for inference and smuggling restricted chips (A100/H100) valued at hundreds of millions of dollars.
• Domestic semiconductor production is plagued by low yields; while China can make 7nm chips, output is an order of magnitude lower than TSMC or Intel.
• China has invested over $100 billion via the "National Integrated Circuit Fund" to indigenize the supply chain, focusing on legacy nodes where it holds a comparative advantage.
• A major corruption scandal in 2023 involving the Wuhan government and a fake TSMC project resulted in the arrest of the Tsinghua Unit Group leader and a shift in strategy toward stricter supply chain verification.

AI Governance and Information Control

• China's AI regulation is significantly faster and more comprehensive than the US or UK, with the Cyber Administration of China (CAC) issuing enforceable rules on generative AI, deepfakes, and algorithms within months of their release.
• Primary regulatory motivation is information control and censorship; models must adhere to "socialist values" and cannot generate content perceived as sensitive or contrary to the state.
• Regulations for generative AI require user verification via National ID numbers, creating a database linking specific content generation to individual identities.
• While China regulates societal impact and information flow, there is a gap in enforcing safety during the training phase for models not yet publicly deployed.
• Chinese safety research is active but framed differently, focusing on "blue teaming" (internal security) and preventing content that violates party lines rather than global alignment or existential risk.
• Recent standards from the CAICT (Ministry of Industry and Information Technology) address long-term risks like deception and self-replication, aligning somewhat with global safety concerns.
• The Chinese government is currently integrating AI into its economic stimulus strategy, potentially shifting from infrastructure (bridges) to compute clusters, though fiscal constraints from local debt limit immediate scalability.

Future Scenarios and Cooperation

• China may eventually bypass compute limits through algorithmic breakthroughs (e.g., neuromorphic computing, brain-inspired AI) rather than simply scaling up hardware.
• A potential "AI winter" in the US due to diminishing returns on the scaling paradigm could allow China to gain ground if it has integrated AI more deeply into its economy and maintained funding through alternative methods.
• Near-term Western policy goals should include establishing an equivalent AI safety institute in China and harmonizing red-teaming standards to ensure mutual verification.
• Long-term goals must focus on creating continuous dialogue channels (Track 1.5/2) to manage volatile geopolitical relations and coordinate on emerging risks like agent governance or compute governance.
• Track 2 dialogues (non-governmental) have successfully produced commitments like the "Digital Declaration" and are crucial for building epistemic communities and testing policy ideas before formal adoption.
• Western policymakers should avoid viewing China as a unitary actor; its regulatory landscape involves complex, competing ministries (CAC, Ministry of Science & Technology, MIIT, NDRC) with overlapping mandates.
• China's industrial policy for semiconductors is characterized by a "complexity" approach where local governments compete and innovate (e.g., Shanghai vs. Guangdong) under national directives, similar to a machine learning evolution process.
• International cooperation is currently in a "honeymoon period" regarding shared fears of catastrophic risk, which should be leveraged to establish baseline safety standards before political interests diverge.