Policy
US-China AI Dominance Leaves Rest of World Behind
The bipolar concentration of computing power, research talent, and language support in artificial intelligence is forcing middle powers to pursue sovereign models—but major gaps remain.
Key takeaways
- The US received 72% of new global AI computing capacity between 2022 and June 2026, China 10%, leaving just 18% for all other countries combined.
- Hindi, spoken by over 500 million people, accounts for only 0.2% of Common Crawl training data, while English represents 43%.
- Middle powers including India, Canada, France, and the UAE are developing sovereign AI models, but many inherit biases from US or Chinese base models.
- A June 2026 US government order temporarily blocked all users from accessing certain Anthropic models, exposing how quickly access can be cut off.
- Countries outside the US-China duopoly should form coalitions, pool computing resources, and develop shared testing frameworks to gain influence over AI governance.
The artificial intelligence landscape has crystallized into a two-power race, with the United States and China commanding the infrastructure, talent, and resources that will shape the technology's future. This concentration is leaving the rest of the world in a precarious position—dependent on AI systems they have little power to influence or control.
Between 2022 and June 2026, the U.S. received approximately 72% of the world's new AI computing capacity, while China captured 10%, according to estimates by Epoch AI. That leaves just 18% for all other countries combined. The imbalance extends to energy consumption, with U.S. and Chinese data centers accounting for roughly 70% of global electricity used by such facilities in 2024, per International Energy Agency data.
The language and representation gap
The dominance goes beyond hardware. A Carnegie Endowment study of over 10,000 researchers who published at NeurIPS, a premier AI conference, in 2025 found that three-quarters were affiliated with U.S. or Chinese institutions.
This concentration has profound implications for the roughly 3.4 billion people who speak languages poorly supported by frontier AI models. A recent Common Crawl snapshot—a web archive widely used for training language models—revealed that approximately 43% of pages were in English, while Hindi, spoken by more than half a billion people, accounted for just 0.2%.
Research by Aditya Vashistha, who leads the Cornell Global AI Initiative, has documented how AI assistants cause users in India to adopt American writing styles and lose cultural nuance. The economic impact is tangible: processing the same content in Hindi or Odia can require more tokens than in English, effectively imposing a "language tax" on developers building applications in those languages.
Why it matters
As AI systems enter schools, hospitals, and government offices worldwide, decisions about safety standards, values embedded in models, and language support are being made almost entirely in Washington and Beijing. A June incident underscored this vulnerability: the U.S. government ordered Anthropic to block foreign nationals from accessing certain models, temporarily taking them offline for all users. The restrictions were later lifted, but the episode demonstrated how quickly access can be severed by a single government's decision.
The sovereign model response
Several middle powers have launched efforts to develop sovereign AI models: Sarvam in India, Cohere in Canada, Mistral in France, and Jais in the United Arab Emirates. Governments in India and Canada are moving to invest directly in domestic AI firms, including through equity stakes. In the UAE, Jais emerged from a partnership involving Inception, a subsidiary of G42, chaired by a member of Abu Dhabi's ruling family.
Yet significant questions remain about these models' capabilities and cultural responsiveness. Many sovereign models are built by adapting open-weight models from the U.S. or China and can inherit biases from those systems. Research shows some regional models can write fluently in local languages while misunderstanding the social norms and contexts of their speakers.
The challenge intensifies as frontier models drive advances in mathematics, chemistry, medicine, and cybersecurity. Countries relying on sovereign models that cannot keep pace risk falling behind in critical scientific and security domains.
Coalition strategies
Vashistha argues that countries seeking greater influence over their AI future should form coalitions of their own. The Global Partnership on Artificial Intelligence, which includes 46 countries and the European Union, offers one foundation. In February, India hosted the first major government AI summit in the Global South, while Finland proposed measures at the UN General Assembly to keep AI "under human control."
Yet agreement remains elusive. Twenty-two countries signed Finland's proposal, but others including Nigeria and India did not. Nigeria's minister of digital economy reportedly stated: "We are desperate for productivity gains. People talk about risks, but we don't have time to wait."
Vashistha recommends that countries lacking infrastructure to build frontier models should pool resources for shared regional computing facilities, coordinate on data sharing and training initiatives, and collaborate on testing frameworks that account for local languages and institutional realities. He also notes that smaller, task-specific models often suffice for everyday applications—using frontier models for routine tasks is "like taking your kids to school in a spaceship."
These details were first reported by TIME in an essay by Aditya Vashistha, assistant professor at Cornell's Ann S. Bowers College of Computing and Information Science.