The United States and China together control roughly 90% of global AI computing power, attract between 70% and 80% of global AI investment, and host the largest concentrations of AI research talent in the world. For the remaining countries — the middle powers, the emerging economies, the Global South, this concentration of capability is not an abstraction. It is a structural condition that is already shaping which governments can access frontier AI, on whose terms, and at what cost.
That was the central finding of a panel discussion hosted by Rest of World during New York Tech Week, titled “The Great AI Divide.” The conversation brought together Sam Winter-Levy, a fellow at the Carnegie Endowment for International Peace; Aditya Vashistha, an assistant professor at Cornell University who leads the Cornell Global AI Initiative; and Peter Micek, general counsel and United Nations policy manager at digital rights group Access Now. Their conclusions were sobering: the geopolitical dynamics surrounding AI access are hardening in ways that will be difficult to reverse.
Managed Access and the Rationing Problem
Winter-Levy identified three converging trends that are making the pessimistic scenario, where a small number of wealthy countries and companies control who gets access to advanced AI, increasingly likely. The first is the shift by leading US developers toward managed-access models. Anthropic’s Mythos model, for example, is being rolled out initially to small groups of companies it selects, leaving other countries waiting to be chosen. The second trend is compute scarcity: demand for frontier AI is outstripping the available infrastructure, forcing companies to ration access. The third is the growing assertiveness of both the US and Chinese governments in determining who can use these systems.
“That puts most of the Global South in an uncomfortable position,” Winter-Levy said. “And the middle powers remain exposed to the disruptions that AI could cause, even if they don’t necessarily capture the benefit.” In other words, countries that are not building frontier AI are still subject to its economic and social disruptions, such as job displacement, information ecosystem changes, dependency on foreign platforms, without the compensating gains that come from being at the frontier.
This dynamic is not hypothetical. As EastFrontier has previously reported, China’s open-source models have been actively expanding their footprint in the Global South, partly because they offer an alternative to US platforms that carry their own access conditions. The Rest of World panel suggests that this competition for influence through AI access is now a defining feature of the geopolitical landscape.
Whose Values Are in the Model?
Vashistha raised a different but equally consequential dimension of the divide: the cultural and linguistic biases embedded in AI systems that were designed primarily for Western, educated, industrialized, rich, and democratic societies — what he called “WEIRD” societies, representing roughly fourteen to fifteen percent of the world’s population.
“What about the 85% and how they are represented in the AI systems?” he asked. The problem goes beyond early-generation failures like stereotyped image generation. It extends to preference datasets that are overwhelmingly in English, safety benchmarks that do not account for ableism despite one billion people worldwide living with disabilities, and models that continue to encode religious, linguistic, and identity biases that are “not what they were a couple of years ago, but are still deeply rooted in these models.”
For countries in the Global South that are increasingly relying on AI for healthcare, education, and public services, these biases are not merely an equity concern, they are a practical failure mode. A medical AI trained primarily on Western clinical data may perform poorly on patient populations with different genetic profiles, disease burdens, or healthcare-seeking behaviors. An AI-powered educational tool that does not support local languages excludes the communities it is supposed to serve.
The Leverage Question
The panel’s most strategically pointed discussion concerned what leverage middle powers actually have in negotiating access to frontier AI. Winter-Levy’s framework was transactional: countries that control something the great powers need, critical minerals, energy, semiconductor supply chain components, battlefield data, or large pools of local-language training data have a stronger hand to play.
“The Netherlands, Taiwan, Japan, and South Korea play key roles in the chip supply chain. That gives them quite a lot of leverage,” he said. “India obviously has huge amounts of data, even Ukraine has huge amounts of data from the battlefield that American companies are desperate for.” The implication is that AI access will increasingly be negotiated as part of broader economic and geopolitical bargains, rather than simply purchased on the open market.
Micek offered a concrete illustration of how this leverage can be applied in reverse. RightsCon, the world’s largest conference on the future of the internet, was cancelled this year after the Chinese government pressured the Zambian government to demand full moderation of online and offline panels, deny access to Taiwanese participants, and enforce adherence to the One China policy. Zambian authorities complied. “We have direct experience with being used as a ping-pong ball between great powers,” Micek said.
The Open-Source Escape Valve
One thread of cautious optimism ran through the panel: the possibility that capable open-source models could serve as an escape valve for countries that cannot or will not align with either the US or Chinese AI ecosystems. If good-enough open-source models can handle most practical use cases — and the rapid improvement of models like DeepSeek, Qwen, and Kimi suggests the gap with proprietary frontier models is narrowing — then the leverage of the great powers over AI access is significantly reduced.
“It might be the case that good enough open-source models that aren’t quite at the front end will work for most use cases,” Winter-Levy said. “And if that’s true, then the world is much less bleak for a lot of middle powers.” The caveat is significant: for the highest-stakes applications, national security, frontier scientific research, advanced cybersecurity, the gap between a good-enough model and the best available model may be decisive.
As EastFrontier has previously reported, China’s “good-enough” chips and models are already powering AI deployments across the Global South, offering a practical alternative to US platforms at a lower cost and with fewer political conditions attached. Whether that alternative is sufficient for the use cases that matter most — and whether it comes with its own set of dependencies and vulnerabilities — is the question that will define the next chapter of the global AI divide.