January 20, 2025, marked a notable date in the unfolding AI competition between the United States and China. In the U.S. capital, Donald Trump assumed the presidency, announcing later his administration’s ambitious AI strategy, "Winning the Race," framing the technological contest as a decisive factor with profound implications for humanity's future.
Simultaneously, in Hangzhou, China, DeepSeek, a relatively obscure firm, unveiled its R1 AI model. Industry observers described this launch as a "Sputnik moment" signaling China's emerging strength in AI development.
The nature of the AI race remains subject to varied interpretations. According to AI policy analyst Lennart Heim, competing goals might encompass deploying AI solutions within economies, advancing robotic systems, or striving toward artificial general intelligence with human-like capabilities. Despite these complexities, multiple metrics indicate the United States currently retains a leading position in AI progress, though Heim cautions that a comprehensive assessment remains hindered by the absence of certain crucial data.
Central to this leadership is access to compute power, the computational resources essential for training AI models. Daniel Kokotajlo, executive director of the AI Futures Project, identifies compute as arguably the foremost driver behind recent advancements. This dependency presents significant challenges for Chinese AI enterprises.
Since 2022, Biden administration policies have restricted exports of advanced chip manufacturing equipment to China, escalating in 2023 to bans on the chips themselves. These measures curtail Chinese access to the sophisticated hardware needed for AI workloads. DeepSeek’s CEO, Liang Wenfeng, emphasized that while capital availability is sufficient, the primary obstacle consists of these export restrictions.
However, policy shifts under the Trump administration in January introduced export regulations potentially permitting Chinese firms to acquire approximately 890,000 Nvidia H200 AI chips. This volume surpasses projections for China's domestic chip production by 2026 more than twofold, as highlighted in a report by the Center for a New American Security. Janet Egan, a co-author of the report, remarks that such access could notably enhance China’s AI capabilities, effectively equipping its key strategic competitor.
Nonetheless, initial Chinese customs responses reportedly blocked imports of the H200 chips, raising questions about whether these resources will reach AI developers within China. Analyst Chris Miller suggests that China has incentives to appear as though it is restricting chip imports to promote domestic technology adoption and negotiate diplomatic positioning with Washington.
The success of DeepSeek’s R1 under such constraints illustrates the potential impact of skilled teams operating with limited resources. Stanford researchers observed that over half of the contributors to this breakthrough had not studied or worked outside China, challenging assumptions about an inherent U.S. edge in AI talent.
Data from the NeurIPS conference indicates that China has outproduced the U.S. in top AI researchers. Although many Chinese experts eventually relocate to the U.S., the proportion remaining in China more than doubled from 2019 to 2022. Immigration policies imposing costly visa fees may impede U.S. innovation and competitiveness, according to Temple University business professor Subodha Kumar.
Energy availability represents another critical input in AI training's intensive power demands. While American AI companies actively seek partnerships with energy providers, China boasts a long-term advantage, having produced more energy than the U.S. since 2010. Miller notes that energy is where the U.S. lags most in competitiveness regarding AI inputs. Should China overcome its chip supply bottlenecks through eased export controls or increased domestic chip production, its substantial energy resources could prove decisive.
Currently, American dominance in AI chip technology and a greater share of leading researchers have enabled deployment of some of the world's most advanced large language models (LLMs). According to Epoch AI, Chinese LLMs generally trail American counterparts by approximately seven months. Additionally, Chinese models often utilize a process called "distillation," training on outputs from more capable models to enhance performance, as Heim explains. For example, DeepSeek's R1 reportedly self-identified as ChatGPT, OpenAI's language model, during testing.
In terms of commercial success, revenue derived from AI products provides an indicator of market adoption. Alibaba, the developer of the Qwen family of models, publicly reports financial figures due to its public listing, though AI model development is ancillary to its dominant Cloud Intelligence business. In 2024, Alibaba Cloud reported annualized revenues of around $22 billion.
Comparatively, American AI startups, founded later and focused solely on AI development, are rapidly closing the gap. OpenAI’s Chief Financial Officer Sarah Friar revealed that the company exceeded $20 billion in revenue by November 2024, marking a milestone in monetization amid the competitive landscape.
This complex interaction of technological capabilities, supply chain regulations, talent migration, energy endowments, and revenue growth illustrates the multifaceted nature of the U.S.–China AI race. While the United States presently maintains leadership in key areas, uncertainties remain regarding how emerging policies and domestic developments in China might alter this balance.