Who controls frontier AI? Safety, ethics and the growing US–China battle for AI supremacy

Artificial intelligence has entered another consequential and dangerous phase. The debate is no longer simply about who can build the most powerful AI model. Increasingly, it is about who controls those models, who can obtain them, and what happens when increasingly powerful AI capabilities become freely available around the world.

At the centre of this debate is the concept of open-source and open-weight artificial intelligence. The distinction is important. In traditional open-source software, developers generally have access to the source code and the freedom to inspect, modify and redistribute it. The Open Source Initiative argues that genuinely open-source AI should similarly provide sufficient access to the components and information necessary to study, modify and reproduce the system.

Open-weight AI is narrower. Here, developers release the numerical parameters, or “weights”, that a trained neural network has learned. These weights can allow organisations and individuals to download, modify and operate sophisticated AI models on their own infrastructure.

OpenAI itself now offers open-weight models through its gpt-oss family under the permissive Apache 2.0 licence. These models can operate on infrastructure controlled by the user rather than through ChatGPT or OpenAI’s hosted API.

This openness has enormous benefits. This means that universities can conduct research without depending entirely upon major technology companies. Start-ups can innovate at considerably lower cost. Countries across Africa and the developing world can adapt models to local languages and circumstances. Organisations can deploy AI locally where privacy or data sovereignty prevents sensitive information from leaving their infrastructure.

But openness creates a difficult question: What happens when tomorrow’s frontier AI becomes powerful enough to cause serious harm?

It comes with a proliferation problem because once the weights of a powerful AI model are released publicly, retrieving them is fundamentally different from withdrawing access to a cloud service.

Copies can spread across thousands of computers, countries and networks.

The same AI capability that helps a scientist discover medicines could potentially assist another person in designing biological threats. A cybersecurity model capable of identifying vulnerabilities could also help malicious actors exploit them. Models could potentially be modified to remove safety restrictions and facilitate fraud, sophisticated cyberattacks, disinformation or autonomous harmful activities.

“This openness has enormous benefits. Which means that universities can conduct research without depending entirely upon major technology companies. Start-ups can innovate at considerably lower cost.”

The 2026 International AI Safety Report highlights precisely why rapidly improving general-purpose AI capabilities require stronger risk-management mechanisms. Leading laboratories are consequently developing increasingly elaborate frontier-safety frameworks.

Anthropic, for example, operates a Responsible Scaling Policy under which stronger safeguards are introduced as model capabilities and associated risks increase.

The fundamental ethical dilemma, therefore, becomes:

How do we democratise AI without democratising catastrophic capability?

The competition between America and China in the AI power race is not helping matters, as the rest of the world appears to be onlookers in this AI race.

This question becomes considerably more complicated because AI is now intertwined with geopolitics.

The United States possesses OpenAI, Anthropic, Google, xAI and many of the world’s most advanced semiconductor and cloud-computing companies.

China has responded aggressively through companies including DeepSeek, Alibaba and Moonshot AI, increasingly using open-weight models as part of its rapidly expanding AI ecosystem. And the technological gap is narrowing remarkably quickly.

Stanford University’s 2026 AI Index reports that the performance gap between the leading American and Chinese models has effectively closed. As of March 2026, the leading US model was ahead by only about 2.7 percent, with American and Chinese models repeatedly exchanging positions near the top of global rankings.

Recent developments have made the geopolitical environment even more troubling. Reports this July indicate that disagreements over Chinese frontier models, advanced semiconductors, model distillation and open-weight releases are threatening emerging US–China cooperation on AI safety.

Humanity cannot afford an AI equivalent of the nuclear arms race in which every safety precaution becomes interpreted as slowing oneself down while helping one’s competitor. AI supremacy will increasingly depend upon four strategic resources:

Compute. Chips. Data. Energy. Talent.

Electricity is rapidly becoming another strategic component of AI infrastructure as enormous data centres require unprecedented power generation and grid capacity.

The danger is therefore not merely an AI race. It is becoming an AI power race, technologically, economically, militarily and literally electrically.

The solution to this AI power race cannot simply be to close every AI model. That would concentrate one of humanity’s most transformative technologies within a handful of corporations and powerful countries. Nor can the answer be unrestricted proliferation.

In the short term, frontier developers should adopt mandatory capability testing, independent red-teaming, cybersecurity standards, model-weight protection and transparent reporting of serious AI incidents. Open-weight releases should undergo specific evaluations before publication, particularly for biological, chemical, cyber and autonomous capabilities.

In the medium term, governments should establish internationally compatible thresholds for frontier AI. Models exceeding specified capability or compute thresholds should face enhanced evaluation and governance requirements before unrestricted release. Cloud providers and semiconductor companies should also participate in monitoring extraordinarily large training runs while protecting legitimate research and commercial confidentiality.

The long-term solution must ultimately be international.

The United States and China must recognise that AI safety represents an area of mutual survival rather than geopolitical concession. There should eventually be an international frontier-AI governance architecture involving the US, China, the European Union, the United Kingdom, Africa and other major regions, supported by common evaluation standards, incident-notification mechanisms and scientific cooperation.

Africa must have a seat at that table. We cannot remain merely consumers of whichever technological ecosystem Washington or Beijing eventually dominates.

Most importantly, increasingly autonomous AI must remain subject to meaningful human responsibility.

I have consistently advocated what I call Responsible Human-in-the-Loop—RHITL. The more powerful and autonomous machines become, the greater, rather than smaller, human accountability must become.

Artificial intelligence should remain open enough to distribute opportunity and competitive enough to stimulate innovation but governed strongly enough to protect humanity. Because the greatest danger may ultimately not be that America wins the AI race or that China wins it.

It is that humanity becomes so obsessed with winning the race that we forget to ask where the race is taking us.

Sonny Iroche is the founder and CEO of GenAI Learning Concepts Ltd, an artificial intelligence strategist, and a member of Nigeria’s National Artificial Intelligence Strategy Committee. He also serves on Thematic Working Groups in the Africa AI Council of Smart Africa and the UNESCO Technical Working Group on Nigeria’s AI Readiness Assessment Methodology. He is a graduate of the University of Oxford’s postgraduate programme in Artificial Intelligence for Business.

Join BusinessDay whatsapp Channel, to stay up to date

Open In Whatsapp