Geopolitical AI Races and Governance Gaps
The China-US AI competition shapes standards, export controls, and talent flows. AI in Taiwan, Ukraine, and Arctic scenarios. Smaller states caught in supply chain coercion and chip bans.
The DeepSeek Shock: China's Efficiency Breakthrough
DeepSeek-V3 and R1
On January 20, 2025, Chinese company DeepSeek released the R1 model, achieving competitive performance with U.S. frontier models at a fraction of the cost. DeepSeek-V3 was trained for $5.576 million in compute over 55 days on 2,048 NVIDIA H800 GPUs, totaling 2.79 million GPU hours. The R1 reinforcement learning phase cost an additional ~$294,000. By comparison, OpenAI's GPT-4 cost over $100 million to train. The disclosed costs exclude prior research, ablation experiments, and data preparation—but even accounting for these, the cost differential is staggering.
DeepSeek-R1 uses a Mixture of Experts architecture with 671 billion parameters and 37 billion activated per forward pass. On benchmarks, it dramatically outperformed GPT-4: 79.8% on AIME 2024 mathematics (vs. GPT-4's 9.3%), 97.3% on MATH-500 (vs. 74.6%), and a Codeforces Elo of 2,029 (vs. 759). The model was released under a permissive open-source license.
The NVIDIA Crash
On January 27, 2025, NVIDIA dropped approximately 17%, losing $589 billion in market capitalization in a single day—the largest single-day loss in stock market history. The Nasdaq 100 sank 3% and the S&P 500 dropped 1.5%. The crash was triggered by the realization that AI models may require far fewer chips than assumed, undermining the investment thesis for the entire AI hardware sector. DeepSeek demonstrated that U.S. export controls may be driving Chinese innovation rather than preventing it.
The Closing Performance Gap
The gap between the best Chinese and U.S. AI models shrank from 9.3% in 2024 to 1.7% in February 2025, suggesting parity is approaching. China generates four times as many STEM graduates annually as the United States. The efficiency-first approach forced by chip restrictions may produce models that are not just competitive but fundamentally more resource-efficient than American counterparts.
The Chip War: Export Controls and Evasion
Biden's AI Diffusion Rule
Published on January 15, 2025, days before Biden left office, the AI Diffusion Rule created a three-tier country framework. Tier 1 (unrestricted): the U.S. plus 18 allies including Japan, the Netherlands, Taiwan, and the UK. Tier 2 (quantity limits): approximately 150 countries, where Tier 1 companies must keep 75% of total AI compute in Tier 1 nations and no more than 7% in any single Tier 2 country. Tier 3 (restricted): approximately 24 countries including China, Russia, Iran, and North Korea. In May 2025, the Trump administration rescinded the rule before it took effect.
Trump Administration Reversals
On December 8, 2025, the Trump administration announced a policy allowing H200 and similar chips to approved Chinese customers. On January 13, 2026, the Bureau of Industry and Security revised its license review for NVIDIA H200 and AMD MI325X from "presumption of denial" to "case-by-case review"—a significant softening. The following day, Trump imposed a 25% tariff on advanced AI chips not destined for U.S. supply chains. In November 2025, the Commerce Department authorized export of 70,000 NVIDIA GB300 chips to the UAE and Saudi Arabia.
Smuggling Networks
In December 2025, Operation Gatekeeper dismantled a smuggling network that moved $160 million worth of NVIDIA H100/H200 GPUs to China between October 2024 and May 2025. Chips were relabeled with the fake brand "SANDKYAN" and classified as generic computer parts. Key defendant Alan Hao Hsu of Missouri City, Texas pleaded guilty in October 2025. Over $50 million in NVIDIA technology and cash was recovered. Chinese tech companies had ordered more than 2 million H200 chips for 2026, exceeding NVIDIA's inventory of 700,000 units—suggesting demand for smuggled or diverted chips vastly exceeds what enforcement can intercept.
China's Domestic Chip Development
Huawei Ascend Production
Huawei's 2025 production target for the Ascend 910C was approximately 300,000 units, doubling to 600,000 units in 2026. Total Ascend product line production for 2026 targets up to 1.6 million dies. The Ascend 910C delivers 60-70% of NVIDIA H100 inference performance and up to 320 TFLOPS FP16. Huawei's CloudMatrix 384 system (384x Ascend 910C via optical interconnect) claims 300 petaFLOPS BF16, reportedly surpassing NVIDIA's GB200 in certain configurations.
SMIC Manufacturing
SMIC, China's leading foundry, reached advanced-node capacity of 45,000 wafers/month by end of 2025, targeting 60,000 in 2026 and 80,000 in 2027. SMIC is producing significant volumes of Ascend dies on a 7nm-class (N+2) process without access to EUV lithography equipment. This represents a remarkable feat of engineering around export restrictions, though performance per watt remains behind TSMC's leading-edge nodes.
The HBM Bottleneck
China's existing HBM (High Bandwidth Memory) stockpile was expected to be depleted by end of 2025. Domestic producer CXMT projected 2.2 million HBM stacks in 2026, but this supports only 250,000-400,000 Ascend 910C packages—less than Huawei needs for its production targets. HBM remains the critical bottleneck in China's AI chip independence. Huawei's product roadmap includes the Ascend 950PR (Q1 2026) and Ascend 950DT (Q4 2026), targeting higher memory bandwidth to address this constraint.
The Middle East: Petrodollars Meet Compute
UAE: $100 Billion in AI
The UAE directed $100 billion toward AI initiatives (March 2024), managed by MGX (created by Mubadala and G42). The UAE-US AI Campus in Abu Dhabi is a 26 square kilometer site with 5 gigawatt capacity when complete, with an initial 200 MW cluster by 2026. Partners include G42, OpenAI, Oracle, Cisco, NVIDIA, and SoftBank. The first phase will deploy 100,000 NVIDIA chips (Grace Hopper systems). Microsoft invested $1.5 billion in G42 equity with $4.6 billion in data center capex and $7.9 billion more planned from 2026-2029. The Technology Innovation Institute released the Falcon Arabic 7B and Falcon-H1 language models.
Saudi Arabia: The Transcendence Project
Saudi Arabia launched the "Transcendence Project"—a $100 billion AI investment program. HUMAIN, a full-stack AI company launched in May 2025, is building the ALLAM multilingual Arabic model. An xAI partnership will build a 500 MW data center with NVIDIA GB300 GPUs (announced November 19, 2025). Saudi Aramco is acquiring a significant minority stake in HUMAIN. Google Cloud and the Public Investment Fund agreed to a $10 billion partnership for a global AI hub. NVIDIA and SDAIA will deploy up to 5,000 Blackwell GPUs for a sovereign AI factory. The NEOM DataVolt AI Factory is a $5 billion investment for a 1.5-gigawatt AI factory in Oxagon, with Phase 1 (300 MW) operational by 2028, powered entirely by renewable energy. Regional data center capacity is expected to triple from 1 GW to 3.3 GW by 2030.
European Sovereign AI: Regulation and Investment
The Paris AI Action Summit
Held February 10-11, 2025, private investors pledged nearly EUR 110 billion in AI investment in France. The UAE committed EUR 30-50 billion for a large data center campus. Canadian firm Brookfield Corporation pledged EUR 20 billion. A $400 million endowment was established for the Current AI Foundation (public goods: datasets, open-source tools). The EU pledged EUR 150 billion through the AI Champions initiative, and the European Commission's InvestAI initiative added EUR 50 billion, for a combined total of approximately $320 billion. 60 countries signed the final communique—notably the U.S. and UK declined to endorse it.
European AI Champions
Mistral AI (France) raised EUR 1.7 billion at a EUR 11.7 billion valuation (~$13.8 billion), led by ASML, with participation from Andreessen Horowitz, NVIDIA, and Bpifrance. Mistral deployed 18,000 NVIDIA Grace Blackwell Superchips in a 40 MW data center. Aleph Alpha (Germany) raised $500 million in 2023 and pivoted to the Pharia "generative AI operating system" in early 2025, positioning for regulated industries under the EU AI Act. The EuroHPC AI Factories Network selected 19 sites across Europe, with EUR 10 billion in total EU investment (2021-2027). JUPITER in Germany became Europe's first exascale supercomputer. Total European AI funding reached EUR 4.6 billion in the first 10 months of 2025.
India and the Global South
India's AI Mission
India's IndiaAI Mission received Union Cabinet approval with a budget of Rs 10,372 crore (~$1.24 billion), with over $1 billion specifically for compute capacity, sovereign datasets, and frontier models. IndiaAI Mission 2.0, unveiled by Minister Ashwini Vaishnaw, pivots from infrastructure-building to research and adoption. Key infrastructure includes 38,000 GPUs, 600 AI Data Labs, a $1.1 billion state-backed VC fund for AI startups, and $18 billion approved for domestic semiconductor projects. BharatGen, India's indigenous LLM led by IIT Bombay, is a 17 billion parameter multilingual foundational model with Rs 900 crore in funding. Google pledged $15 billion for foundational AI infrastructure in India, and G42 (UAE) is deploying an 8 exaflop supercomputer for sovereign AI infrastructure.
Africa's AI Adoption
The AU Continental AI Strategy Phase I (2025-2026) focuses on governance frameworks and national strategies. 49 African countries endorsed the Africa Declaration on AI. At least 8 countries have adopted national AI strategies, with 5 more drafting them. The AI 10 Billion Initiative targets up to $10 billion for AI entrepreneurship and infrastructure. The data center market reached $1.26 billion in 2024, projected to reach $3.06 billion by 2030. Microsoft and G42 are building a $1 billion geothermal-powered data center in Kenya (100 MW by 2026). Microsoft committed ZAR 5.4 billion in South Africa and plans to train 1 million South Africans in AI by 2026. 230 million digital jobs are expected in Sub-Saharan Africa by 2030.
BRICS AI Cooperation
The 17th BRICS Summit (July 6, 2025, Rio de Janeiro) produced a historic declaration with 126 commitments spanning governance, finance, health, climate, and AI, plus a separate Leaders' Declaration on Global Governance of AI. The New Development Bank launched a $5 billion digital sovereignty fund for AI infrastructure. The BRICS+AI Alliance, launched December 2024, includes 20+ tech companies from Russia, China, India, Brazil, Iran, and the UAE. Brazil allocated $4.2 billion for its "AI for All" program (2025-2028).
Taiwan: The Most Dangerous Bottleneck
TSMC's Market Dominance
TSMC produces nearly 90% of the world's most advanced chips. Its critical facilities are located 110 miles from mainland China. For companies like NVIDIA, TSMC is the only viable fabrication option. The entire global AI revolution depends on a three-company, three-country chain: NVIDIA (chip design, California), ASML (lithography equipment, Netherlands), and TSMC (manufacturing, Taiwan). A disruption at any point in this chain would halt global AI development.
Conflict Vulnerability
Research indicates Taiwan is especially vulnerable to a Chinese quarantine before 2027. Shortages of critical raw materials, chemicals, and natural gas would emerge within weeks of a blockade. Stockpiling would not fully protect against prolonged disruption. The Ukraine war has become a testing ground for AI military systems, and Western companies use Ukraine to test military AI in combat conditions—capabilities that could be deployed in a Taiwan scenario.
TSMC's Arizona Diversification
TSMC's Arizona expansion represents the largest foreign direct investment in a U.S. greenfield project in history, totaling $165 billion. It received $6.6 billion in CHIPS Act direct funding. Fab 21 Phase 1 began producing 4nm chips in Q4 2024. Phase 2 targets 3nm production in 2027. Phase 3, breaking ground in April 2025, will produce chips using N2 and A16 processes by end of the decade. The vision is a "Gigafab" cluster of 6 fabs, 2 advanced packaging facilities, and 1 R&D center. But even at full capacity, Arizona will produce only a fraction of TSMC's Taiwan output.
AI Espionage and Technology Theft
Google Trade Secrets
Linwei "Leon" Ding, a former Google engineer, was convicted in January 2026 on 7 counts of economic espionage and 7 counts of trade secret theft. He stole over 2,000 confidential Google documents on TPUs, GPUs, and SmartNIC network cards. He began exfiltrating data on May 21, 2022, copying Google source files to Apple Notes, converting to PDF, and uploading to personal Google Cloud. By mid-2022 he was in talks to become CTO of a Chinese startup; by early 2023 he founded Shanghai Zhisuan Technologies Co.
Samsung Semiconductor Espionage
10 people were indicted on December 23, 2025, including a Samsung executive, for stealing Samsung's 10nm DRAM technology (developed over ~5 years at 1.6 trillion won cost). The beneficiary was Changxin Memory Technologies (CXMT), China's leading DRAM maker, which achieved mass production of 10nm DRAM in 2023—years ahead of projections. Estimated Samsung losses: 5 trillion won in revenue in 2025 alone, with total national damage in "tens of trillions of won." An SK Hynix engineer was arrested at an airport attempting to leak HBM technology to China. Another SK Hynix engineer received an 18-month prison sentence for smuggling documents to sell to Huawei.
International Governance: Why It Fails
No Binding Framework
No international body has meaningful enforcement power over AI development. The Council of Europe Framework Convention on AI entered into force on November 1, 2025, but includes exemptions for private sector regulation and national security. The UN Global Dialogue on AI Governance launched in September 2025 with 100+ countries participating, but produces only voluntary commitments. China has proposed creating a new international AI cooperation body headquartered in Shanghai. The UK AI Security Institute and similar national bodies coordinate but cannot enforce international standards.
Why Governance Fails
Major powers resist binding constraints that might limit their competitive advantage. AI development moves faster than diplomatic processes. Definitions of risk and responsible development vary across nations. Enforcement would require intrusive inspections of proprietary technology—something no major AI power would accept. Nations adopting Chinese surveillance technology become locked into Chinese ecosystems, while those adopting U.S. technologies face similar dependencies, creating lasting structural relationships that constrain foreign policy options.
Smaller States Under Pressure
Smaller nations face pressure to join U.S.-led chip export controls against China or risk losing access to advanced semiconductors themselves. Countries requiring data localization face pressure from both U.S. companies seeking market access and Chinese companies offering "sovereign" AI solutions. This creates competing spheres of technological influence where the AI race shapes not just technology policy but geopolitical alignment for decades to come.