Researchers from the University of Oxford's Oxford Internet Institute and Aalto University's Department of Computer Science have published a report on AI infrastructure, cloud providers, and accelerators.
Using publicly available data, they compiled a list of each provider's cloud regions as of November 2024.
The public cloud providers included in the study are Amazon Web Services, Google Cloud Platform, and Microsoft Azure in the United States, as well as Alibaba Cloud, Huawei Cloud, and Tencent Cloud in China.
In Europe, notable providers include Exoscale, Hetzner, and OVHcloud.
The study also identified AI accelerators from companies such as Nvidia, Google, Amazon, Microsoft, Alibaba, and Huawei.
Overall, the census discovered a total of 225 cloud regions, which are clusters of large interconnected data centers located in the same geographical area.
Out of these 225 cloud regions, 132 were found to be equipped with AI accelerators.
In South America, only Brazil hosts public cloud AI computing facilities, while in Africa, only South Africa has such infrastructure.
Public cloud AI computing is primarily concentrated in North America, Europe, the Middle East, and Asia.
The report introduces the concept of "compute sovereignty," highlighting its importance in discussions about AI governance among governments and industries.
The researchers explore key questions related to this concept:
(1) how much AI computing a country has within its borders
(2) the nationality of the companies that own the AI computing data centers, and
(3) the nationality of the vendors providing the AI accelerators used in these data centers.
They analyze these questions empirically through the lens of cloud computing infrastructure, focusing on nine leading public cloud providers that represent approximately 70% of the global market.
The findings suggest that "compute sovereignty" varies between countries, depending on the level of analysis applied.
Identifying the most relevant level of analysis is influenced by government policy goals and national contexts, leading to various policy trade-offs.
Policies designed to attract data centers can enhance the security of critical computing resources.
Still, they may also lead to increased energy, water, and land consumption, along with corresponding socioeconomic and environmental impacts.
Different regional and supply chain approaches come with their own sets of trade-offs.
In the future, could compute producers control access to a critical resource?
Is an A.I. divide forming as the biggest companies and richest countries scramble for GPUs, and much of the world is being left behind?