When we think of “the cloud,” we often imagine data floating invisibly in the air. But the reality is much more concrete. The cloud resides in giant buildings called data centers, packed with powerful, energy-hungry computer chips. These chips, especially Graphics Processing Units (GPUs), are necessary to build and run powerful chatbots like ChatGPT, and have become critical infrastructure in the world of AI.
As AI becomes more capable, so too does the geopolitical importance of high-performance chips, and where in the world they're located matters. The U.S. and China are competing for stockpiles, and Washington has used sanctions to block Beijing from buying cutting-edge chips. But despite the stakes, there's surprisingly little public data about where exactly the world's AI chips are located.
A new peer-reviewed paper, shared exclusively with TIME ahead of publication, aims to fill that gap. “We tried to find out where AI is,” says Vili Lehdonvirta, the paper's lead author and a professor at the University of Oxford's Internet Institute. Their findings were grim: GPUs are concentrated in just 30 countries around the world, with the United States and China far ahead. Most of the world is in what the authors call “computing deserts,” areas with no rentable GPUs at all.
The findings have big implications not only for next-generation geopolitical competition, but also for AI governance — that is, which governments have the power to regulate how AI is built and deployed. “If the actual infrastructure that runs the AI or on which the AI is trained is on your territory, you can enforce compliance,” says Redonvirta, who is also a professor of technology policy at Aalto University. Countries that don't have jurisdiction over AI infrastructure have fewer legislative options and will be subordinate to a world shaped by other countries, he argues. “This has implications for which countries shape AI development and the norms around what is good, safe, and beneficial AI,” says Boksi Wu, one of the paper's authors.
The paper maps the physical locations of “public cloud GPU computing” – essentially GPU clusters available for rent through the cloud businesses of major tech companies. But the study has significant limitations. For example, it doesn't count GPUs owned by governments or privately owned by tech companies for their sole use. It also doesn't take into account the non-GPU types of chips that are increasingly being used to train and run advanced AI. Finally, it counts the number of computing “regions” (or groups of data centers containing those chips) that cloud companies offer in each country, rather than individual chips.
Read more: How “friend-shoring” made Southeast Asia a linchpin of the AI revolution
That's not for lack of trying: “The quantity of GPUs, and especially how they're distributed across (cloud) provider regions, remains highly confidential,” the paper notes. Despite the paper's limitations, the study is the closest to the latest public estimates of where the world's most advanced AI chips are located, and the authors argue it's a good indicator of the elusive bigger picture.
According to the paper, the United States and China have the largest number of public GPU clusters in the world. China surpasses the United States in the total number of GPU-enabled regions, but the most advanced GPUs are concentrated in the United States. The United States has eight “regions” where you can rent H100 GPUs, the kind that the U.S. government has sanctioned China for. China has none. This doesn't mean there are no H100s in China, it just means that cloud companies don't have H100 GPUs in China. The New York Times reported in August that a black market for regulated chips was booming in China, citing intelligence agencies and vendors who said millions of dollars' worth of chips had been smuggled into China despite the sanctions.
The authors argue that the world can be divided into three categories: the “computing north,” where the most advanced chips are located; the “computing south,” where there are some older chips that are suitable for running AI systems but not for training; and “computing deserts,” where there are no chips available for rent at all. These terms overlap to some extent with the vague notions of “global north” and “global south” used by some development economists, but Redonvirta says they are simply analogies to draw attention to a “global divide” in AI computing.
Wu said the risk of having too many chips concentrated in wealthy countries is that countries in the global South could become dependent on AI developed in the global North, without having a say in how it works.
This “reflects existing patterns of global inequality across the so-called Global North and Global South,” Wu said, threatening to “entrench the economic, political and technological power of Compute North countries and influence the institutions that shape AI research and development in Compute South countries.”
