Two Nations Control 72% of New AI Computing Power
The U.S. and China dominate the infrastructure and data behind frontier models, forcing other countries to seek sovereign alternatives.
By Noor Graham ·

Between 2022 and June 2026, the United States acquired approximately 72% of the world's new artificial intelligence computing capacity. China captured 10% of that total. The remaining 18% was divided among every other nation on Earth. This concentration of hardware defines the current power dynamic in the AI sector.
The disparity extends to energy consumption. In 2024, data centers in the U.S. and China used about 70% of the electricity consumed by data centers globally. The rest of the world accounted for the remaining 30%. A study by the Carnegie Endowment analyzed more than 10,000 researchers who published at the NeurIPS conference in 2025. It found that three-quarters were affiliated with employers in the U.S. or China.
Language representation in training data mirrors these physical gaps. Common Crawl, a web archive frequently used to train language models, contains roughly 43% of its pages in English. Only 0.2% are in Hindi. That language is spoken by more than half a billion people. Researchers noted that training data for Quechua, spoken by about 8 million people in South America, often relies on texts from the era of Spanish colonization.
Cost structures also reflect these imbalances. AI companies typically charge users per token. Messages in Hindi and Odia often require more tokens than the same messages in English. Developers building applications in these languages face a higher processing cost for identical content. This financial penalty exists despite newer systems narrowing the gap.
The vulnerability of non-dominant nations became clear in June 2026. The U.S. government ordered Anthropic to block foreign nationals from accessing its Fable 5 and Mythos 5 models. The company removed both models from service for all users. Restrictions were lifted later that month, but the event demonstrated how quickly a single government can disrupt global access to AI tools.
Middle powers are responding by building their own models. India has developed Sarvam, Canada has Cohere, France has Mistral, and the United Arab Emirates has Jais. In the UAE, Jais was created by Inception, a subsidiary of G42. G42 is chaired by Sheikh Tahnoun bin Zayed Al Nahyan, the country's national security adviser. India and Canada are also taking equity stakes in domestic AI firms to protect national interests.
Many of these sovereign models adapt open-weight systems from the U.S. or China. They can inherit biases from their predecessors. Some regional models write fluently in local languages but misinterpret the social norms of their speakers. Training data and alignment processes often fail to capture these nuances. In a 2022 Anthropic study, roughly 20 paid online workers provided about 80% of the alignment feedback. Meta's Community Alignment project draws participants from only two Global South countries.
Safety testing remains limited. Existing evaluations cover too few languages and often assess isolated responses rather than multi-turn conversations. The U.S. and China are now organizing competing coalitions. In July, Beijing established the World Artificial Intelligence Cooperation Organization. Washington leads the Pax Silica initiative and plans to urge allies to choose sides.
The Global Partnership on Artificial Intelligence includes 46 countries and the European Union. In February, the India AI Impact Summit in New Delhi marked the first major government-hosted AI summit in the Global South. In September, Finland proposed measures to manage AI development pace at the United Nations General Assembly. Twenty-two countries signed the proposal to keep AI under human control. Nigeria and India did not sign.
Nigeria's minister of digital economy, Bosun Tijani, stated that the country is desperate for productivity gains and does not have time to wait. The United Nations Development Programme is expanding AI computing infrastructure in Kenya, Malawi, Rwanda, South Africa, Togo, and Zambia. This work aims to give these nations the shared resources needed to build and adapt AI systems independently.






