On October 7, 2022, the Biden administration's Bureau of Industry and Security published 139 pages of new export controls. Almost overnight, American companies were barred from selling advanced semiconductor chips — and the equipment to make them — to Chinese entities without a license. The targets included Nvidia's A100 GPU, the workhorse of large-scale AI training. It was the most sweeping technology export restriction since the Cold War.
Within weeks, Chinese AI labs scrambled to stockpile existing inventory. Nvidia reported a surge in orders from the Asia-Pacific region in the days before the rules took effect. The race had entered a new, explicitly adversarial phase.
Modern AI systems — large language models, image recognizers, autonomous-vehicle perception stacks — require enormous parallel computation during training. Graphics processing units (GPUs), originally designed for video games, turned out to be ideally suited to this work. By 2020 Nvidia controlled roughly 80–90% of the data-center GPU market, and its H100 chip (released 2022) offered performance roughly six times faster than its predecessor for transformer-model training.
China's largest AI companies — Baidu, ByteDance, Alibaba, Tencent, and dozens of well-funded startups — had become deeply dependent on Nvidia hardware. China's domestic chip industry, led by SMIC (Semiconductor Manufacturing International Corporation), was still limited to 7-nanometer production at best, roughly two generations behind TSMC in Taiwan and Samsung in South Korea. Without access to leading-edge chips, Chinese AI labs faced a hard ceiling on model scale.
One year after the initial October 2022 controls, the Biden administration broadened restrictions again, closing loopholes that had allowed exports of slightly slower chips (the A800 and H800 versions Nvidia had designed specifically to comply with the first round of rules). The message was clear: the U.S. intended to maintain a generational lead in AI compute.
Beijing's answer was to accelerate domestic substitution. The "Made in China 2025" program, launched in 2015, had already designated semiconductors as a strategic priority. Post-2022, China's National Integrated Circuit Industry Investment Fund (the "Big Fund") injected hundreds of billions of yuan into chipmakers. Huawei's HiSilicon division intensified work on its Kirin and Ascend lines. In September 2023, Huawei surprised analysts by shipping the Mate 60 Pro smartphone running a 7-nm chip manufactured entirely within China — a significant, if partial, proof of concept for domestic capability.
On the AI software side, Chinese researchers continued to publish prolifically. Papers from Chinese institutions regularly topped citation counts at NeurIPS and ICLR. Moonshot AI, Zhipu AI, Baidu's ERNIE Bot, and Alibaba's Tongyi Qianwen all launched capable large language models in 2023, demonstrating that algorithmic progress could partially offset hardware disadvantages — at least for inference if not for cutting-edge training runs.
Hardware is only part of the story. AI capability depends equally on researchers and engineers. A 2022 MacroPolo report found that roughly 60% of the world's top AI researchers were working in the United States, but that China was producing the largest single cohort of AI PhDs globally. Notably, the majority of top-tier Chinese AI researchers were working at American universities or companies — a talent dynamic that made the competition simultaneously collaborative and adversarial.
The U.S. response included scrutiny of Chinese nationals working on sensitive AI programs and restrictions on visas in certain research areas. China responded by funding "Thousand Talents" style programs to attract researchers back, and by building elite AI institutes at Tsinghua, Peking University, and through the Chinese Academy of Sciences.
Both governments have framed AI leadership in explicitly national-security terms. The U.S. National Security Commission on AI (2021) called AI "the most powerful tool for competitive advantage in the 21st century" and recommended maintaining "at least a two-generation lead" in semiconductor manufacturing. China's 2017 New Generation AI Development Plan called for China to become the world's primary AI innovation center by 2030.
You're advising a technology policy think tank preparing a brief on AI chip competition between the U.S. and China. Use this AI analyst to explore the strategic logic of export controls, China's domestic alternatives, and the risks of technological decoupling.
Reporting by +972 Magazine and the Israeli newspaper Local Call revealed that the Israel Defense Forces had deployed an AI system called "Lavender" to generate targeting recommendations — identifying individuals suspected of being Hamas or Palestinian Islamic Jihad militants. A second system, "Where's Daddy?", was reportedly used to track targets to their homes. Investigations suggested that tens of thousands of individuals had been assigned AI-generated scores, with human review sometimes lasting only seconds per target.
The IDF disputed elements of the reporting but confirmed using AI to process intelligence data. The episode forced a global conversation about whether the speed and scale of AI-assisted targeting was compatible with international humanitarian law's requirements of distinction, proportionality, and precaution.
Military applications of AI span a wide spectrum, from logistics optimization to fully autonomous lethal systems. It is useful to distinguish between:
AI-enabled decision support: Systems that process intelligence, surveillance, and reconnaissance (ISR) data to present recommendations to human operators. The U.S. military's Project Maven (2017–present), which uses computer vision to analyze drone footage, is a canonical example. Google's engineers famously resigned in protest when the company's involvement became public in 2018.
Human-on-the-loop systems: Weapons that can act autonomously but with a human capable of overriding them. Raytheon's Phalanx close-in weapon system, which automatically engages incoming missiles, has operated in this mode since the 1980s. More recent examples include Israel's Iron Dome and the HARPY loitering munition.
Fully autonomous lethal systems (LAWS): Weapons that select and engage targets without human intervention. No state officially acknowledges deploying these, but credible reports from the 2020 Nagorno-Karabakh conflict described Turkish-made Kargu-2 drones operating in an autonomous mode against human targets — a finding noted in a UN Panel of Experts report.
Launched in 2017, Project Maven contracted Google to apply computer vision to analyze video feeds from surveillance drones flying over conflict zones. It generated over 3,600 signatures per month, drastically reducing manual analyst workload. After internal protests and a public letter signed by more than 4,000 Google employees, Google declined to renew its Maven contract in 2018. The DoD subsequently moved the project to Palantir. Maven continues to operate and has expanded significantly in scope.
Strategists worry that AI integration into nuclear command-and-control systems could compress decision timelines in ways that increase the risk of accidental war. A 2020 RAND Corporation study, "Stabilizing the Military Balance in the Taiwan Strait," and subsequent work by the Nuclear Threat Initiative identified several pathways to instability:
Sensor-to-shooter compression: AI systems that detect a launch and automatically queue a retaliatory strike could reduce the time for human verification from minutes to seconds, increasing false-alarm risk. The 1983 Soviet Petrov incident — where a duty officer correctly identified a satellite malfunction as a false alert rather than a U.S. first strike — is the archetypal near-miss that a fully automated system might have missed.
Deception vulnerability: Adversarial inputs (spoofed sensor data, jamming) could cause AI targeting systems to misidentify civilian or neutral objects as threats, with catastrophic consequences.
Opacity of intent: When both sides deploy AI systems that make rapid, opaque decisions, the risk of escalatory spirals increases because neither side can read the other's intentions in real time.
Since 2014, the United Nations' Convention on Certain Conventional Weapons (CCW) has hosted discussions on "lethal autonomous weapons systems." As of 2024, no binding treaty exists. The U.S., Russia, and China have each resisted mandatory prohibitions, while smaller states and NGOs like the Campaign to Stop Killer Robots advocate a preemptive ban. The debate mirrors the earlier struggle over landmines and cluster munitions — with AI adding new layers of technical complexity.
Russia's 2022 invasion of Ukraine became the first large-scale peer conflict in which AI-enabled drone warfare played a central role. Ukraine deployed commercially sourced quadcopters running computer-vision models to identify and engage Russian armor. Ukrainian startup Saker developed AI-powered targeting software that could be installed on consumer drones. Russia deployed Shahed-136 loitering munitions (of Iranian design) in mass saturation attacks against Ukrainian infrastructure.
The conflict demonstrated that AI-enabled autonomous capabilities are no longer the exclusive province of great powers — a $500 drone running open-source vision models can be effective on a modern battlefield. This democratization of military AI has profound implications for future conflicts and arms control.
You're preparing testimony for a Senate Armed Services Committee hearing on autonomous weapons systems. Use this advisor to stress-test arguments for and against lethal autonomous weapons, explore international humanitarian law constraints, and think through escalation risks in AI-enabled conflict.
On March 13, 2024, the European Parliament voted 523 to 46 to pass the EU Artificial Intelligence Act — the world's first comprehensive legal framework governing AI systems. After three years of drafting, revision, and a dramatic last-minute scramble to address generative AI (which had not existed when the legislation was first proposed in 2021), Europe had done what it did for privacy with GDPR: written the rules first and bet that the world would follow.
Whether that bet would pay off was far from obvious. American tech companies had already begun lobbying European governments to soften implementation. Chinese officials said nothing publicly about the Act, but Beijing had been quietly building its own AI governance architecture for two years.
The EU AI Act organizes AI systems into four risk tiers:
Unacceptable risk (banned): Social scoring systems of the kind deployed by Chinese local governments, real-time biometric surveillance in public spaces (with narrow exceptions), subliminal manipulation, and AI that exploits vulnerable groups.
High risk: AI used in critical infrastructure, education, employment, essential services, law enforcement, border control, and the administration of justice. These systems must undergo conformity assessment, maintain technical documentation, enable human oversight, and register in an EU database.
Limited risk: Chatbots and deepfakes must carry transparency disclosures. Users must be informed they are interacting with AI.
Minimal risk: Spam filters, AI in video games — no requirements beyond existing law.
The Act also introduced rules specifically for general-purpose AI models (such as GPT-4 and Claude), requiring transparency about training data, adversarial testing, and cybersecurity compliance. Models above a compute threshold of 10²⁵ FLOPs face additional "systemic risk" obligations.
When the EU enacted GDPR in 2018, many companies — including major U.S. tech firms — found it easiest to apply its standards globally rather than maintain separate data practices for European users. This "Brussels Effect" (a term coined by Columbia Law professor Anu Bradford) refers to the EU's ability to set global standards through its market size. AI Act proponents believe the same dynamic will apply to AI governance — giving the EU's rules an outsized global footprint.
The U.S. approach has been more fragmented. President Biden's Executive Order on AI (October 30, 2023) — the most detailed U.S. federal AI policy action to date — directed agencies to develop safety standards, required developers of the most powerful AI systems to share safety test results with the government, and invoked the Defense Production Act to mandate reporting of large training runs.
But EOs are not law, and the incoming Trump administration in January 2025 revoked Biden's AI EO on its first day, replacing it with an order emphasizing "American AI dominance" and reducing regulatory friction. This represented the starkest example yet of how AI policy can reverse with elections — and why industry has been reluctant to make long-term compliance investments based on executive action alone.
The U.S. also relied heavily on voluntary commitments: in July 2023, seven leading AI companies (including OpenAI, Google, Microsoft, Meta, Amazon, Anthropic, and Inflection) signed White House commitments to watermarking AI-generated content, sharing safety information with governments, and investing in cybersecurity. These commitments were non-binding and unverifiable.
China has moved faster than most observers expected on narrow AI governance while simultaneously deploying AI tools that Western frameworks would classify as high-risk. Between 2021 and 2023, China's Cyberspace Administration issued a series of regulations:
Algorithm Recommendation Regulations (2022): Required disclosure when algorithmic systems were being used to shape content feeds, banned certain manipulative practices, and established opt-out rights.
Deep Synthesis Regulations (2022): Required labeling of deepfakes, prohibited use of AI-generated faces or voices to spread disinformation, and mandated real-name registration for services.
Generative AI Regulations (2023): Required operators of generative AI services to conduct security assessments, ensure training data legality, and prevent outputs that undermine "socialist core values" — a clause with obvious political dimensions.
Notably, China's rules apply primarily to public-facing services — not to government or military AI applications. This asymmetry is mirrored, to varying degrees, in U.S. and EU rules, where national security carve-outs are similarly broad.
| Dimension | EU AI Act | U.S. (2023 EO) | China (2021–23) |
|---|---|---|---|
| Legal form | Binding regulation | Executive order (non-binding law) | Binding regulations |
| Scope | Comprehensive, risk-tiered | Safety reporting, agency guidance | Sector-specific (algorithms, deepfakes, GenAI) |
| Enforcement | Fines up to €35M or 7% global revenue | Agency rulemaking; no direct penalty | Fines; service suspension; criminal liability |
| Generative AI | General-purpose model rules added 2023 | Voluntary commitments; reporting for large models | Security assessment required; values alignment |
| National security | Broad carve-out | Defense and intel exempted | Government/military largely exempt |
A persistent challenge across all three regimes: AI development cycles are faster than legislative cycles. GPT-4 was publicly released in March 2023; the EU AI Act, which had to be amended to address it, took another full year to pass. By the time the Act enters full force (2026–2027), AI capabilities may have advanced by another generation. Most legal scholars working on AI governance acknowledge this lag as the fundamental structural problem in AI regulation.
You're helping a multinational technology company map its AI products against the EU AI Act, U.S. executive guidance, and China's regulations. Use this advisor to determine risk classifications, identify compliance gaps, and think through the strategic implications of regulatory divergence between jurisdictions.
On November 1–2, 2023, the UK hosted the first AI Safety Summit at Bletchley Park — the wartime codebreaking facility that gave birth to the modern computer. Twenty-eight countries signed the "Bletchley Declaration," agreeing that frontier AI posed potentially catastrophic risks and that international cooperation was needed. The signatories included the United States, China, the EU, and — notably — most of the world's major economies.
But critics noted what the summit was not: most of Africa, Latin America, and South and Southeast Asia were absent or marginally represented. The AI safety conversation, like the AI development conversation, remained concentrated in a handful of wealthy nations. The diplomacy of AI was already mirroring the diplomacy of climate — where the countries most vulnerable to consequences had the least influence over decisions.
The UK AI Safety Summit was the first in a series of international AI governance convenings. South Korea hosted a follow-up summit in May 2024, and France took the chair for a third summit in February 2025. The process has established a modest but real international AI safety architecture, including the creation of national AI Safety Institutes (first in the UK and U.S., then others) and work on shared evaluation standards for frontier AI models.
Crucially, China signed the Bletchley Declaration — a significant diplomatic achievement given U.S.–China tensions. This demonstrated that even adversarial powers can find common ground on existential-scale AI risks, though implementation and enforcement mechanisms remain largely absent.
The UK's AI Safety Institute (AISI), launched in November 2023, became the first government body explicitly tasked with evaluating advanced AI systems before and after deployment. By mid-2024, the U.S. had established its own AISI within NIST, and the EU, Japan, South Korea, Canada, and Australia were building comparable bodies. These institutes share evaluation methodologies and red-team findings — a thin but real layer of international AI safety coordination.
For most of the world's countries, the AI landscape looks very different from the frontier race. The question is not "which AI model to train" but "whose AI to use" — and the geopolitical stakes of that choice are substantial.
Infrastructure dependency: Cloud computing — the substrate on which most AI runs — is dominated by Amazon Web Services, Microsoft Azure, and Google Cloud (U.S.) and Alibaba Cloud and Huawei Cloud (China). African governments adopting AI-powered public services are, in most cases, routing sensitive citizen data through servers physically located in Europe or the U.S. The African Union's Data Policy Framework (2022) identified this as a sovereignty concern.
China's AI diplomacy: Through the Digital Silk Road component of the Belt and Road Initiative, China has deployed AI-powered surveillance infrastructure — facial recognition cameras, smart city systems, social management platforms — across dozens of countries including Zimbabwe, Ecuador, Pakistan, and Serbia. These systems, frequently built by Huawei and Hikvision, often come with subsidized financing and capacity-training packages that create long-term vendor lock-in and data-sharing arrangements with Chinese entities.
Language and cultural gaps: The majority of large language models are trained primarily on English-language data. A 2023 study by Masakhane, an African NLP research network, found that of the 2,000+ languages spoken in Africa, fewer than 30 had meaningful AI tool support. This capability gap means that even where AI is deployed, its benefits are unevenly distributed.
Geopolitical competition is producing what some analysts call "techno-blocs" — groupings of countries that align on AI standards, data flows, and chip supply chains along broadly political lines.
The U.S.-led bloc includes the EU, UK, Japan, South Korea, Australia, Canada, and increasingly India. The "Chip 4 Alliance" (U.S., Japan, South Korea, Taiwan) has worked to coordinate semiconductor export policy. The Quad (U.S., India, Japan, Australia) has established AI working groups. The G7's "Hiroshima AI Process" produced voluntary AI governance principles in 2023.
The Chinese-led alternative operates through different mechanisms — less formal multilateralism and more bilateral technology agreements. China has used standards bodies (particularly ISO/IEC and ITU) aggressively to advance its preferred AI governance norms, often proposing standards that prioritize state oversight and social stability over privacy and individual rights.
The majority of the world's nations remain unaligned or strategically ambiguous — simultaneously purchasing Huawei infrastructure, hosting U.S. cloud services, and attending both Western and Chinese AI governance forums. India is perhaps the most important swing state: the world's most populous country, with significant AI engineering capacity, is deliberately maintaining strategic ambiguity to maximize leverage over both blocs.
Technical standards — the arcane specifications governing how AI systems interoperate, what safety tests they must pass, and how their outputs are labeled — are an underappreciated dimension of AI geopolitics. The ITU (International Telecommunication Union), ISO, and IEEE all have active AI standards working groups. China has been documented sending larger delegations and submitting more standards proposals than any other country. Standards, once adopted, have decades-long lock-in effects on technology development — and they are almost entirely invisible to public debate.
You're advising the foreign minister of a mid-sized developing economy navigating AI geopolitics. Your country is being courted by both U.S.-aligned cloud providers and Chinese Digital Silk Road programs. Use this strategist to think through the trade-offs of alignment, data sovereignty, and AI capacity-building for nations outside the great-power competition.