The US-China AI Race: A New Cold War in Silicon

The US-China AI rivalry is the defining technology conflict of the twenty-first century. It is being fought not with missiles but with export controls, chip architectures, talent visas, and lines of code. Its outcome will determine whether the world’s most powerful AI systems are built under democratic governance with at least nominal transparency and safety commitments, or under authoritarian governance with state-directed objectives and no independent oversight.

This is not a metaphor or an exaggeration. It is the explicit framing used by senior officials in both governments. When the Biden administration imposed sweeping semiconductor export controls on China in October 2022, National Security Advisor Jake Sullivan described them as reflecting a new American strategy: rather than maintaining a relative advantage, the United States would seek to maintain “as large of a lead as possible” in foundational technologies. The message to China was unambiguous: the US intends to prevent China from achieving parity in AI capability, indefinitely.

China’s response has been equally unambiguous. President Xi Jinping has repeatedly designated AI as a “core national strategic technology” and directed that China must achieve self-sufficiency in semiconductors and AI foundations. The Made in China 2025 plan, the New Generation AI Development Plan, and successive Five-Year Plans have all elevated AI to the top of China’s technology priorities. The Chinese government has committed over $47 billion to its third-phase National Integrated Circuit Industry Investment Fund — commonly known as the “Big Fund III” — with the explicit goal of building a domestic semiconductor supply chain independent of American technology.

The stakes are civilizational. The nation that leads in AI will hold advantages across every domain of power: economic productivity, military capability, intelligence collection, scientific discovery, and the capacity to shape global norms and institutions. Both the US and China understand this. Neither is willing to be second.

---

The Export Control Regime: Three Rounds of Escalation

The US semiconductor export controls represent the most significant technology sanctions since the Cold War. They have been implemented in three major rounds, each more restrictive than the last, and they have reshaped the global semiconductor industry.

Round One: October 2022

The Bureau of Industry and Security (BIS) issued sweeping new rules restricting the export of advanced semiconductors, semiconductor manufacturing equipment, and related technologies to China. The key provisions:

Restriction	Detail
Advanced chips	Banned export of chips exceeding certain performance thresholds (targeting NVIDIA A100, H100-class GPUs)
Manufacturing equipment	Restricted export of equipment for fabricating chips at 14nm and below
US persons	Prohibited US citizens and permanent residents from supporting Chinese semiconductor manufacturing
Scope	Applied to any chip designed or manufactured using US technology, regardless of country of origin

The “US persons” restriction was particularly significant. It meant that American citizens working at Chinese chip companies had to choose between their jobs and their citizenship. Dozens of senior engineers and executives at Chinese semiconductor firms resigned within weeks of the rules taking effect.

Round Two: October 2023

The second round closed loopholes exploited after the first. China had been purchasing chips just below the restricted performance thresholds — NVIDIA’s A800 and H800, designed specifically to comply with the October 2022 rules while still providing substantial AI training capability. The October 2023 rules:

• Lowered performance thresholds to capture A800/H800-class chips
• Added new “performance density” metrics to prevent workarounds through chip clustering
• Expanded restrictions to cover additional countries suspected of serving as reexport channels (including Saudi Arabia and the UAE, before bilateral agreements were reached)
• Tightened end-use monitoring requirements
• Added additional Chinese entities to the Entity List

Round Three: 2024 Expansion

The third round addressed the reality that China was accessing restricted technology through intermediaries in Southeast Asia, the Middle East, and Central Asia. It also targeted the emerging threat of China developing its own AI chip architectures:

• Expanded geographic scope of controls
• Added restrictions on advanced packaging technologies (critical for chiplet-based designs that could allow China to achieve near-frontier performance without frontier fabrication)
• Tightened controls on EDA (electronic design automation) software
• Added new Chinese AI companies and chip designers to the Entity List
• Implemented new licensing requirements for cloud computing access to restricted chips

Effectiveness Assessment

The export controls have unquestionably slowed China’s AI hardware development. China’s most advanced domestically produced chips remain approximately two to three generations behind the frontier. Huawei’s Ascend 910B, China’s most capable AI training chip, delivers roughly 50-70% of the performance of NVIDIA’s H100 while consuming significantly more power and exhibiting lower reliability.

But the controls have not stopped China. They have redirected Chinese investment, accelerated domestic R&D, motivated workarounds, and created a thriving gray market for restricted semiconductors. They have also imposed costs on American companies — NVIDIA estimated it lost $5 billion in China revenue in the year following the initial controls — and created diplomatic friction with allies who resent American extraterritorial application of export restrictions.

---

Huawei: The Sanctions Survivor

Huawei Technologies is the most important case study in the limits of export controls. Placed on the Entity List in 2019, subjected to the most restrictive technology sanctions ever applied to a single company, and cut off from Google’s Android ecosystem, TSMC’s fabrication services, and virtually every American semiconductor supplier, Huawei was widely expected to wither.

It did not. Huawei’s revenue declined sharply in 2021-2022 but rebounded. The company launched the Mate 60 Pro smartphone in August 2023 featuring the Kirin 9000s processor — a 7nm chip fabricated by China’s SMIC using a workaround process that stunned American analysts who believed China was incapable of producing chips at that node.

In AI, Huawei’s Ascend series chips have become the foundation of China’s domestic AI training infrastructure. The Ascend 910B is not as powerful as NVIDIA’s H100, but it is available in quantity, and Huawei has invested heavily in the CANN software framework to make it usable for AI training workloads. Chinese frontier labs, unable to legally purchase NVIDIA hardware, have increasingly adopted Ascend-based training clusters.

Huawei’s survival demonstrates a fundamental limitation of export controls: they can slow technological development, but they cannot stop a nation of 1.4 billion people with a $18 trillion economy and a government willing to spend whatever it takes. The controls have imposed a performance tax on China’s AI development — Chinese labs need more chips, more power, and more engineering effort to achieve results comparable to their American counterparts — but they have not imposed a performance ceiling.

---

China’s Frontier Labs: The Competition Intensifies

The Western narrative that China is “behind” in AI is increasingly inaccurate, and dangerously complacent. China has produced multiple frontier-capable AI laboratories, several of which have demonstrated capabilities that rival or exceed Western counterparts in specific domains.

The Major Players

Lab	Parent	Key Models	Strengths	Estimated Compute
Baidu	Baidu (public)	ERNIE series	Chinese-language NLP, search integration, enterprise deployment	Large domestic clusters
Alibaba Cloud	Alibaba Group	Qwen series	Open-weight models, multilingual capability, cloud integration	Massive cloud infrastructure
ByteDance	ByteDance	Doubao/Skylark	Recommendation systems, multimodal, integration with TikTok/Douyin	Enormous (TikTok scale)
01.AI	Independent (Yi-Series)	Yi series	Founded by Kai-Fu Lee, open-weight, efficient architectures	Moderate
DeepSeek	High-Flyer Capital	DeepSeek-V3, DeepSeek-R1	Reasoning capability, efficiency innovations, open-weight	Unknown (private hedge fund backing)
Zhipu AI	Tsinghua-affiliated	GLM series	Academic research, bilingual capability	Moderate
Moonshot AI	Independent	Kimi	Long-context capability, consumer applications	Growing

DeepSeek: The Warning Shot

DeepSeek deserves particular attention. Founded by Liang Wenfeng, the founder of quantitative hedge fund High-Flyer Capital, DeepSeek has produced models that have shocked the Western AI research community with their capability-to-cost ratio.

DeepSeek-V3, released in late 2024, demonstrated performance competitive with GPT-4 and Claude 3 Opus on multiple benchmarks while reportedly being trained at a fraction of the cost. DeepSeek-R1, a reasoning-focused model released in early 2025, achieved performance on mathematical and coding benchmarks that rivaled or exceeded OpenAI’s o1 model.

What made DeepSeek particularly alarming to American policymakers was its training efficiency. DeepSeek published papers describing training techniques — including multi-head latent attention and mixture-of-experts architectures — that allowed frontier-level performance using significantly less compute than American labs required. If China can achieve comparable AI capability with fewer and less advanced chips, the strategic logic of export controls is fundamentally undermined.

DeepSeek also complicates the narrative that AI requires Western-style venture capital and corporate structures. It was funded by a single individual’s private wealth, built by a relatively small team, and operates with minimal public-facing institutional structure. It demonstrated that the barriers to frontier AI development may be lower than assumed.

---

Talent Flows: The Brain Drain Battlefield

AI capability is ultimately a function of human talent, and the flow of AI talent between the US and China is one of the most consequential dynamics in the rivalry.

The United States remains the world’s dominant destination for top AI researchers. A 2023 study by MacroPolo found that approximately 60% of the world’s top-tier AI researchers (defined by publications at NeurIPS, ICML, and ICLR) were based in the US, and a significant share of those were Chinese-born researchers who had come to the US for graduate education and remained.

This creates a paradox for US policymakers. Chinese-born researchers are critical to American AI leadership. But China’s intelligence services have documented programs to recruit overseas Chinese scientists, and several high-profile espionage cases have involved Chinese nationals working in US technology companies. The result is a climate of suspicion that risks driving away exactly the talent the US most needs.

The Department of Justice’s “China Initiative,” launched in 2018 and wound down in 2022, prosecuted researchers at US universities for undisclosed ties to Chinese institutions. While some cases involved genuine security concerns, others appeared to be motivated by ethnicity rather than evidence, and the initiative had a documented chilling effect on Chinese-born researchers’ willingness to work in the US.

China has invested heavily in “reverse brain drain” programs, offering generous compensation packages, research funding, and prestigious titles to attract overseas Chinese researchers back home. The Thousand Talents Program and its successors have recruited thousands of scientists across all disciplines, including AI. The effectiveness of these programs is debated — many participants maintain dual affiliations rather than relocating permanently — but they have unquestionably strengthened China’s domestic research base.

The talent picture is further complicated by the rapid growth of AI education in China. Chinese universities now graduate more STEM PhDs annually than American universities, and the quality of the top Chinese AI programs (Tsinghua, Peking University, Zhejiang University, Shanghai Jiao Tong University) has improved dramatically. A growing share of frontier AI research is produced by researchers who were trained entirely in China, never having worked or studied in the West.

---

Taiwan: The World’s Most Dangerous Chokepoint

Taiwan is the single most dangerous vulnerability in the global AI supply chain. Taiwan Semiconductor Manufacturing Company (TSMC) fabricates over 90% of the world’s most advanced semiconductors — the chips that power AI training, smartphones, advanced military systems, and virtually every other frontier technology application. There is no substitute. There is no backup. If TSMC’s fabrication facilities in Taiwan were damaged, destroyed, or seized, the global AI industry would effectively halt.

This reality creates a strategic dilemma of extraordinary complexity:

For the United States: Taiwan’s semiconductor capability is an American strategic asset, because TSMC fabricates the chips that power American AI and military systems. But Taiwan’s geographic proximity to China and the growing capability of the People’s Liberation Army mean that this strategic asset is located in the one place most vulnerable to America’s primary rival. The US is attempting to mitigate this through the CHIPS and Science Act, which provides $52.7 billion in subsidies for domestic semiconductor manufacturing, including TSMC’s new fab complex in Arizona. But the Arizona fabs will not reach full production until 2028-2029 at the earliest and will initially produce chips at least one generation behind TSMC’s Taiwan facilities.

For China: Taiwan’s semiconductor capability is both the greatest prize and the greatest deterrent. Seizing Taiwan would give China control of the world’s most advanced chip fabrication capability. But a military action against Taiwan would almost certainly damage or destroy that capability — TSMC’s fabs are extraordinarily fragile, and the company’s engineers could sabotage their own equipment rather than allow it to fall into Chinese hands. Moreover, any attack on Taiwan would trigger catastrophic economic consequences for China itself, whose technology industry depends on TSMC-fabricated chips as much as any country’s.

For Taiwan: Semiconductor manufacturing is Taiwan’s “silicon shield” — the theory that its indispensability to the global technology supply chain deters Chinese military action. But this deterrent is eroding as both the US and China invest in alternative fabrication capacity, and as the strategic value of controlling AI hardware increases to the point where a rational actor might calculate that the benefits of seizing Taiwan outweigh the costs.

---

Military AI Applications

Both the US and China are integrating AI into military systems at an accelerating pace, and neither side has full visibility into the other’s capabilities or doctrine.

The United States has invested billions in military AI through programs including Project Maven (AI-assisted intelligence analysis), the Joint All-Domain Command and Control (JADC2) architecture (AI-enabled coordination across military branches), and numerous classified programs in autonomous weapons, cyber operations, and electronic warfare. The Pentagon’s Replicator initiative, announced in 2023, aims to field thousands of AI-enabled autonomous systems within 18-24 months.

China’s military AI investments are less visible but no less ambitious. The People’s Liberation Army has published doctrine emphasizing “intelligentized warfare” — a concept that envisions AI as transformative across every domain of military operations, from logistics and intelligence to autonomous combat systems. China has demonstrated autonomous drone swarms, AI-enabled naval warfare systems, and cyber capabilities that US officials have described as peer-level.

The risk of miscalculation is severe. If both sides deploy AI systems that operate faster than human decision-making, the escalation dynamics of a crisis — particularly over Taiwan — could outpace the ability of political leaders to control. This is the AI equivalent of the nuclear hair-trigger problem, and it has no solution currently on offer from either side.

For comprehensive analysis of military AI, see: AI at War: The Military Applications of Inhuman Intelligence.

---

Data Advantages: Two Different Kinds of Dominance

The US and China each have data advantages, but they are fundamentally different in character.

China’s data advantage is population scale and government access. With 1.4 billion people, the world’s largest internet user base, and a government that faces no legal barriers to accessing private sector data, China can generate training datasets of a scale and comprehensiveness that no democratic society can match. Facial recognition systems trained on the Chinese population have access to billions of images. Language models trained on WeChat conversations have access to the communication patterns of a billion people. Autonomous vehicle systems in China benefit from road testing data generated across a vast and diverse geography.

America’s data advantage is internet dominance and quality. The English-language internet — the largest, most diverse, and most information-rich corpus of text ever created — is predominantly hosted on American platforms, indexed by American search engines, and stored on American cloud infrastructure. The vast majority of the scientific literature, technical documentation, and code repositories that are critical for AI training are in English. Models trained on this data have demonstrated advantages in reasoning, coding, and general knowledge that Chinese models have struggled to match.

The data competition is also a privacy competition. China’s willingness to use personal data without meaningful consent gives its AI systems certain practical advantages, particularly in surveillance and social control applications. The US and Europe’s (nominal) commitment to data privacy creates constraints on training data that Chinese labs do not face. Whether those constraints ultimately strengthen or weaken Western AI — by building public trust and preventing misuse, or by handicapping model development — is one of the central unresolved questions in AI policy.

---

The Decoupling Trajectory

The US-China technology relationship is decoupling, and the trajectory points toward further separation, not reconciliation. But decoupling is not binary. It is proceeding unevenly across different technology layers:

Layer	Decoupling Status	Notes
Hardware (advanced chips)	Extensive	Export controls have largely severed advanced chip flows
Hardware (mature chips)	Limited	China dominates mature node production, interdependence remains
Cloud computing	Progressing	US cloud providers face increasing restrictions on serving Chinese AI labs
AI models	Partial	Open-source models flow freely; proprietary models increasingly restricted
Research publications	Minimal	Researchers still publish in shared venues, but collaboration declining
Talent	Partial	Visa restrictions and espionage concerns reducing flows
Data	Extensive	Firewall, data localization, and divergent regulatory regimes create separation
Standards	Progressing	Competing standards bodies, divergent technical norms

Full decoupling is probably impossible and certainly undesirable. The US and Chinese technology ecosystems are deeply intertwined at the level of supply chains, research communities, and technical standards. But the trend is unmistakably toward greater separation, and the remaining connections are becoming increasingly politicized and securitized.

The most likely outcome is not two completely separate technology ecosystems but two partially overlapping ones with a growing “no man’s land” between them. Countries and companies in this middle ground — including the Gulf states, India, Southeast Asian nations, and many European firms — will face increasing pressure to choose sides. For analysis of how the Gulf states are navigating this dynamic, see: Gulf States AI: The $100 Billion Desert Bet.

---

What INHUMAIN.AI Is Watching

The US-China AI race is accelerating, and several developments in 2026 will shape its trajectory:

DeepSeek’s next generation. If DeepSeek or other Chinese labs continue to demonstrate frontier capability at fraction-of-Western-cost, the strategic logic of chip export controls collapses. American policymakers are watching this closely.

TSMC Arizona progress. The timeline and capability of TSMC’s Arizona fabs will determine how quickly the US can reduce its dependence on Taiwan-based fabrication. Delays or capability gaps have national security implications.

Talent policy. Whether the US can maintain its dominance in AI talent while managing legitimate security concerns about Chinese espionage will depend on immigration policy decisions that have so far been made on political rather than strategic grounds.

Open-source proliferation. Meta’s Llama, DeepSeek’s models, and others are freely available worldwide. The proliferation of capable open-source models may render hardware restrictions less relevant — or may create new categories of risk that neither government has adequately considered.

Gulf states as swing players. Saudi Arabia and the UAE are courted by both the US and China as AI partners. Their choices about technology alignment — which cloud platforms to use, which chip architectures to adopt, whose AI models to deploy — will influence the shape of the global AI order. See: Gulf States AI.

This is a rivalry with no natural equilibrium and no clear end state. The US and China are locked in a technology competition that neither can win decisively and neither is willing to lose. The consequences of that dynamic — for AI safety, for global stability, and for the billions of people who will live with the AI systems both nations are building — are the subject of INHUMAIN.AI’s ongoing coverage.

For the broader geopolitical context, see: AI Geopolitics: Who Controls Inhuman Intelligence Controls the Century.