The United States locked up 72 percent of new global AI computing capacity between 2022 and June 2026, leaving the rest of the world to fight over the scraps, TIME reported.
That concentration is the skeleton key to the bipolar AI race. While Washington and Beijing hoard the chips, the electricity, and the researchers — three-quarters of NeurIPS authors in 2025 worked for U.S. or Chinese employers, a Carnegie Endowment study found — everyone else is reduced to tenants on somebody else’s infrastructure.
The International Energy Agency estimates the two powers burned through 70 percent of global data-center electricity in 2024. For the Global South, the bill comes due in cultural erasure: 43 percent of Common Crawl is English, 0.2 percent is Hindi, and Quechua training data still reads like a colonial archive.
The Bottom Line
- U.S. and China control 72% and 10% of new AI compute respectively; 18% remains for 190+ other nations.
- Frontier models impose a “language tax” — Hindi and Odia cost more tokens per message than English.
- Sovereign models (Sarvam, Mistral, Jais) still lean on U.S./Chinese open weights, inheriting their biases.
The Anthropic Lockout That Woke the Middle Powers
In June, the U.S. government ordered Anthropic to block foreign nationals — including its own staff — from accessing the Fable 5 and Mythos 5 models. Anthropic took both models offline for all users rather than comply selectively.
The restrictions lifted weeks later, but the signal was received from Nairobi to New Delhi: access to frontier intelligence can be revoked by a single executive order. That episode, TIME reported, “invigorated the efforts of several middle powers to build sovereign models of their own.”
India backed Sarvam, Canada doubled down on Cohere, France championed Mistral, and the UAE’s G42 — chaired by national security adviser Sheikh Tahnoun bin Zayed Al Nahyan — launched Jais. Yet most sovereign models are fine-tunes of U.S. or Chinese open weights.
“Knowing how to speak a language is not the same as understanding the social norms of its speakers,” the TIME analysis said, citing research showing regional models fluent in local tongues but blind to local contexts.
Alignment by a Handful of Crowdworkers
The bias pipeline narrows further at alignment. A 2022 Anthropic study found 20 paid online workers supplied 80 percent of the preference data. Meta’s newer Community Alignment project, designed to capture cultural nuance, still drew participants from only two Global South countries and covered English and Hindi — ignoring hundreds of other Indian languages.
“We are desperate for productivity gains,” Nigeria’s digital economy minister Bosun Tijani told reporters when asked why Abuja declined to sign Finland’s UN proposal to keep AI “under human control.” “People talk about risks, but we don’t have time to wait.”
The Coalition Math
Washington’s Pax Silica initiative and Beijing’s new World Artificial Intelligence Cooperation Organization are pressing allies to pick sides. The alternative, Brookings’ Douglas Rediker argues, is a coalition of the unaligned: the 46-nation Global Partnership on Artificial Intelligence, the India AI Impact Summit in February, and the UN Development Programme’s compute build-out across Kenya, Malawi, Rwanda, South Africa, Togo, and Zambia.
Pooling chips, data, and evaluation benchmarks is the only way to escape the “spaceship for a school run” problem — deploying massive frontier models for tasks a small, localized model could handle.
Founded in 2018 in Abu Dhabi, G42 is chaired by the UAE’s national security adviser, Sheikh Tahnoun bin Zayed Al Nahyan, and led by founder and Group CEO Peng Xiao. The company became the sole registered shareholder of the messaging app ToTok, which was removed from Apple and Google stores in 2019 after The New York Times reported it was used for surveillance.
In 2020, state-owned Mubadala Investment Company took a stake, transferring Injazat and Khazna to G42. Silver Lake invested $800 million in 2021, and Microsoft committed $1.5 billion in April 2024, placing its president, Brad Smith, on G42’s board.
| Metric | United States | China | Rest of World |
|---|---|---|---|
| New AI Compute Share (2022–Jun 2026) | 72% | 10% | 18% |
| Data-Center Electricity (2024) | ~70% combined | 30% | |
| NeurIPS 2025 Author Affiliations | three-quarters | the rest | |
The next inflection point arrives when sovereign models face their first real benchmark against frontier systems in medicine, cybersecurity, and autonomous robotics. If they cannot keep pace, the countries betting on them fall behind in the very sectors AI promised to revolutionize. The billions living outside the bipolar axis are still waiting for a vote on what intelligence looks like in their languages, their hospitals, their schools.