US and China Each Hold Advantages in Different Segments of the Global AI Race

The question of who is "winning" the AI race between the United States and China depends entirely on which race you mean. Across compute capacity, model performance, patent filings, talent pipelines, industrial deployment, and military applications, the two countries have developed fundamentally different strengths — and the lead changes depending on which metric you measure.

The 2025 AI Index Report from Stanford's Institute for Human-Centered AI found the US holds nine times China's AI compute capacity and seventeen times Europe's [1]. But China produced 23.2% of all global AI publications in 2023, more than any other country, and accounted for 69.7% of all AI patent grants that year [1]. These are not contradictory facts. They reflect two nations pursuing AI dominance through different strategies, with different advantages and different vulnerabilities.

The Chip Gap: Export Controls and China's Domestic Push

The semiconductor supply chain remains the most critical bottleneck in the AI competition. Since October 2022, when the Biden administration imposed sweeping export controls on advanced chips and chipmaking equipment, China's access to top-tier AI accelerators has been sharply curtailed [2].

The impact has been measurable but not total. SMIC, China's largest chip foundry, remains stuck at 7nm process technology because it cannot access extreme ultraviolet (EUV) lithography machines from the Dutch firm ASML [3]. U.S. Commerce Secretary Howard Lutnick testified that Huawei produced only 200,000 AI chips in 2025 — compared to Nvidia's 4.5 million [3]. The best U.S. AI chips are currently about five times more powerful than Huawei's best offerings, a gap that the Council on Foreign Relations projects could widen to seventeen times by 2027 [3].

Yet China has not stood still. Chinese chipmakers have found workarounds, using deep ultraviolet immersion lithography (DUVi) tools — which are not subject to export restrictions — enhanced with multi-patterning techniques to produce chips that approach the cutting edge [4]. SMIC's revenue rose 16% in 2025 to a record $9.3 billion, with projections of $11 billion in 2026 [5]. TrendForce projects that the domestic share of China's AI chip market will reach 50% in 2026, up from just 25% in 2024 [6].

Source: TrendForce / CSIS

Data as of Jan 1, 2026CSV

In December 2025, the Trump administration partially reversed course, allowing Nvidia to sell H200 chips — nearly six times more powerful than the export-compliant H20 — to approved Chinese firms in exchange for a 25% revenue stake [7]. This decision drew criticism from analysts who argued that exporting 3 million H200 chips to China would provide more computing power than China's domestic production could achieve until 2028-2029 [3].

Model Performance: Closing the Gap

On frontier AI models, the US maintains a lead — but it is narrowing faster than most analysts expected. American companies still produce the most capable large language models, with Chinese LLMs lagging by roughly seven months on average in benchmark performance [8]. The US holds the top positions in closed-source models, with OpenAI, Anthropic, and Google leading on reasoning and coding tasks.

But China's open-source strategy has proven formidable. In January 2025, DeepSeek released R1, an open-source reasoning model competitive with OpenAI's o1 at inference costs 90% lower [9]. Chinese open-source models — including DeepSeek and Alibaba's Qwen — now account for 30% of all global AI downloads, double the US share of 15.7% [9]. On the MMLU and HumanEval benchmarks, the gap between Chinese and American models has "essentially vanished" [9].

This matters because DeepSeek demonstrated that the American approach — massive capital, maximum compute, closed models — is not the only viable path [1]. Competitive models can be built with less compute if the engineering is clever enough.

Research Output: Volume vs. Influence

The research landscape tells a split story. China dominates in volume: it leads in total AI publications, AI patents, and increasingly in top-conference acceptances. At NeurIPS 2025, Chinese institutions surpassed the US as the top contributor country for the first time, with Tsinghua University alone accounting for 4.73% of accepted papers [10]. Predictions suggest China will also lead at ICLR 2026, marking its first time dominating a top AI conference [10].

Source: OpenAlex

Data as of Jan 1, 2026CSV

Source: Stanford HAI AI Index 2025

Data as of Apr 1, 2025CSV

In patents, China's lead is overwhelming. Between 2019 and 2023, Chinese AI patent grants grew from 42,075 to 85,451, while US grants declined from 16,204 to 12,212 [1]. But patent counts can be misleading — many Chinese patents are filed for defensive or incentive-driven reasons and may not reflect commercially significant innovations.

Where the US still leads is in research influence. American institutions contribute the most top-100-cited AI publications, and the most commercially deployed models continue to originate from US labs [1]. The quality-versus-quantity distinction remains real, even as it narrows.

The Talent Pipeline Under Strain

Perhaps the most consequential and least predictable variable is talent. Chinese-born researchers constitute a significant share of AI talent at US labs and universities. But that pipeline is under unprecedented strain.

Around 50 tenure-track scholars of Chinese descent left US universities for China in the first half of 2025 alone, based on a tally by Princeton University researchers, with more than 850 such departures since 2011 [11]. At least 85 rising and established scientists working in the US have joined Chinese institutions full-time since early 2025 [12]. The second Trump administration has enacted sweeping visa restrictions — revoking visas, pausing appointments, proposing fixed durations of stay, and eliminating Optional Practical Training (OPT) for graduates [13].

China has responded with targeted recruitment. Beijing launched the "K visa program" in October 2025 to attract returning technical talent, offering immigration pathways without requiring prior employment contracts [14]. Over 50% of DeepSeek's breakthrough researchers never left China for schooling or work, suggesting China's domestic training pipeline is increasingly self-sustaining [8].

The irony is sharp: the US may be pushing away the very researchers whose contributions underpin its AI lead. A $100,000 visa price tag, floated as a policy proposal, could "hurt the innovation and competitiveness of the U.S. industry," according to analysts cited by Time [8].

Follow the Money

The investment picture is lopsided in the US's favor — but the gap is less meaningful than it appears.

Source: Stanford HAI AI Index 2025

Data as of Apr 1, 2025CSV

US private AI investment reached $109.1 billion in 2024 and surged past $174.6 billion in 2025, driven by massive bets on frontier model companies [15]. China's private investment was $9.3 billion in 2024, rising to roughly $10 billion in 2025 [15]. The US held 97% of global generative AI deal value in the first half of 2025 [15].

But China's state-directed funding operates on a different logic. The government launched a $47.5 billion semiconductor fund in 2024 and added a separate $8.2 billion AI venture fund in 2025 [16]. Total Chinese government AI investment is projected to reach $56 billion in 2025 [15]. By comparison, US federal AI R&D spending was $3.3 billion for FY2025, though defense-related AI spending has surged to $4.3 billion [17].

The disparity in private capital reflects structural differences. US investment is concentrated in a handful of foundation model companies (OpenAI alone has raised over $20 billion) [8], while China spreads investment across a broader ecosystem of applied AI, robotics, and infrastructure. The question is whether massive frontier-model spending yields proportional returns — or whether China's more distributed approach proves more durable.

Military and Surveillance Applications

Both countries are aggressively integrating AI into military systems, but with different emphases. The Pentagon's Replicator Initiative aims to field thousands of low-cost, expendable autonomous drones [18]. The US leads in computer vision models for intelligence, surveillance, and reconnaissance (ISR), analyzing satellite and drone imagery at scale [18].

China's People's Liberation Army has prioritized AI-enabled unmanned systems, with open-source procurement records from 2023-2024 revealing extensive investment in intelligent unmanned vehicles, drone swarms, and AI-commanded electronic warfare systems [19]. Chinese defense scientists have developed experimental generative AI capable of commanding drones equipped with electronic warfare weapons [20]. The PLA has also invested in facial recognition, gait recognition, and deepfake-related technologies for base security and psychological warfare [19].

Does America's lead in foundational models translate to a meaningful defense advantage? The relationship is indirect. Foundation model capabilities feed into computer vision, natural language processing for intelligence analysis, and autonomous decision-making. But China's approach — deploying large quantities of cheaper, less sophisticated autonomous systems — may prove equally effective in many operational scenarios. The US Department of Defense has historically struggled to move AI from research to fielded systems quickly, a problem China's more centralized procurement system may mitigate [18].

In January 2025, President Trump revoked Executive Order 14110, which had imposed ethical guidance on AI applications including military uses, and replaced it with Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence" [21]. The signal was clear: deregulation in pursuit of dominance.

The Case for China's Applied AI Dominance

The steelman case for China rests on several structural advantages that frontier model benchmarks do not capture.

Scale of deployment: China had 2,027,200 industrial robots in operation as of 2024, more than five times the US figure of 393,700 [1]. Sixteen cities are testing robotaxi services, with Baidu operating approximately 400 vehicles across 1,160 square miles in Wuhan alone [9]. Chinese workers are more likely to use AI regularly for work (25%) than their American counterparts (15%) [1].

State coordination: China's New Generation AI Development Plan, issued in 2017, set explicit targets — AI industry parity by 2020, world-leading in some fields by 2025, and becoming the primary center for AI innovation by 2030 [22]. By the plan's own benchmarks, China has met or exceeded its 2025 targets in patents, publications, and several applied domains.

Regulatory tolerance: China allows faster real-world testing and deployment of AI systems in manufacturing, surveillance, and transportation than Western regulatory frameworks typically permit. This produces more real-world training data and faster iteration cycles.

Energy: China has generated more electricity than the US since 2010 and has begun construction on more nuclear power plants than the rest of the world combined [8]. Energy is, according to one expert, "the one where the U.S. is least competitive" among AI inputs [8].

The counterargument is that deployment at scale is only as good as the underlying models. If frontier model capabilities determine which AI applications are possible, the US's lead in foundational research creates options that China cannot yet replicate. But DeepSeek's success suggests that compute-efficient approaches can partially compensate for raw hardware disadvantages.

The Third Players: AI Non-Alignment

The binary US-China framing increasingly misses the strategies of third countries and regions. Middle powers are not simply choosing sides — they are pursuing what Chatham House calls "sovereign AI" strategies [23].

India is positioning itself as an alternative governance leader. US tech giants have pledged billions in AI investments in India, and New Delhi will host the Global AI Impact Summit in 2026 to advance its claim as a bridge between the two superpowers [24]. India's strategy focuses on building domestic AI capacity while maintaining relationships with both Washington and Beijing.

The Middle East is emerging as a significant player. Saudi Arabia and the UAE signed AI-focused partnerships with the US in 2025, but Gulf states also maintain technology relationships with Chinese firms [25]. Their sovereign wealth funds provide capital that both superpowers want.

The EU has increased AI defense investments and maintains the world's most comprehensive AI regulatory framework through the AI Act. Europe lacks a frontier model competitor but controls important parts of the supply chain — notably ASML's monopoly on EUV lithography [25].

Southeast Asia is hedging actively. Countries like Singapore, Vietnam, and Indonesia are accepting AI investments from both US and Chinese firms, building what Modern Diplomacy calls a "sovereign AI axis" that resists binary alignment [26].

Rather than choosing a side, these countries are pursuing four strategies identified by Chatham House: specializing in a particular supply chain segment, aligning with one superpower, sharing sovereignty with peer nations to amplify influence, or hedging by using AI from multiple sources [23].

The Semiconductor Self-Sufficiency Question

Beijing has set a target of 70% semiconductor equipment self-sufficiency in mature process technologies by 2027, with 80% overall semiconductor self-sufficiency by 2030 [27][28]. Industry executives have drawn up plans to meet these targets, and equipment localization is advancing: China's semiconductor equipment industry saw rapid growth in 2024-2025 [27].

But independent assessments are mixed. The 70% target is "certainly attainable in some but not all areas" according to industry analysts — achievable for mature nodes (28nm and above) but not for cutting-edge production [27]. SMIC has reported that some foreign-purchased tools could not be put to work because supporting equipment was unavailable, preventing production line formation [27].

If China does achieve broad semiconductor self-sufficiency by 2028-2030, the US loses its most powerful coercive tool: the ability to restrict chip access. Export controls work only as long as alternatives don't exist. But even optimistic Chinese projections acknowledge a multi-year gap in the most advanced nodes, and the performance gap between domestic and foreign chips continues to widen in absolute terms even as China closes the relative gap [3].

What the Scoreboard Actually Shows

The AI competition between the US and China is not a single race with a single finish line. It is a series of parallel competitions across research, hardware, deployment, talent, investment, and military applications. The US leads in frontier models, compute capacity, private investment, and research influence. China leads in patents, publications volume, industrial deployment, open-source model adoption, and applied AI at scale.

The more important question may not be which country leads overall, but which advantages prove more durable. The US's compute and capital advantages depend on maintaining semiconductor chokepoints and attracting global talent — both of which face growing challenges. China's deployment and scale advantages depend on continued access to capable models and chips — which export controls and its own domestic efforts are reshaping in real time.

What is clear is that neither country can be described as simply "winning" or "losing." They are playing different games, with different rules, and the outcome depends on which game turns out to matter more.