As AI cements its place at the forefront of the new ‘industrial revolution’ and drives growth for nations across the world, a new form of geopolitical contestation is emerging – the battle over AI compute infrastructure.
A recent study by Vili Lehdonvirta, Bóxī Wú, and Zoe Hawkins from the Oxford Internet Institute, Aalto University, and the Australian National University, respectively, has revealed a stark divide in the global distribution of AI computing resources. This divide, categorised by the researchers as ‘Compute North,’ ‘Compute South,’ and ‘Compute Desert,’ could have profound future implications for AI governance and global technology power dynamics.
The geography of AI computing
AI development relies heavily on three core components: data, algorithms, and compute power. The latter, which includes advanced graphics processing units (GPUs) and AI chips, is becoming increasingly vital as the technology scales up. The research team behind the study broke the world into three distinct regions based on the availability and sophistication of public Cloud AI compute resources:
- Compute North: This category includes countries primarily in the Global North, such as the United States and several European and Asian nations. These countries possess significant AI compute capabilities, hosting the most advanced GPUs necessary for both developing (training) and deploying (inferencing) AI models. For example, the United States leads in providing access to Nvidia’s latest H100 and A100 GPUs, which are critical for cutting-edge AI development.
- Compute South: Predominantly composed of Global South countries, this category includes nations with AI compute resources geared more towards AI deployment rather than development. These regions often rely on older, less powerful GPUs, such as the Nvidia V100, suitable for running existing AI models but not for developing new, advanced ones.
- Compute Desert: This group includes countries with little to no public Cloud AI compute infrastructure. Many of these countries, often lower-income, depend on foreign-hosted compute resources to access AI capabilities, limiting their ability to independently influence AI governance.
The implications of a divide
The uneven distribution of AI compute infrastructure has several significant implications, most importantly in concerns over sovereignty, power dynamics, and heightening the existing global economic divide.
Countries with advanced AI compute resources, such as those in the Compute North, are in a much stronger position to shape global AI governance. The United States, in particular, holds a competitive edge due to its substantial compute infrastructure and strategic export controls that limit other countries’ access to cutting-edge AI technologies. These controls have effectively barred China from acquiring the latest AI GPUs, reinforcing the U.S.’s dominant position.
The concentration of AI compute power in the Global North also almost identically mirrors broader economic inequalities seen in the world presently. Countries in the Compute South and Compute Desert have limited capabilities to participate in the development of advanced AI technologies, which could perpetuate a new ‘compute divide’ that mirrors the digital divide of past decades.
Additionally, some countries are adopting strategies to enhance their AI compute capabilities as a matter of national policy. For instance, China is developing its own AI processing chips in response to U.S. export restrictions. In Europe, concerns over ‘digital sovereignty’ are driving demand for locally hosted AI compute resources, especially for sensitive or large-scale AI training tasks.
Why is there a divide?
There are several factors that contribute to the concentration of advanced AI computing infrastructure in regions such as the US, such as:
- Government policies and export controls: The U.S. has implemented stringent export controls on advanced GPUs, preventing countries like China from accessing the latest AI compute technologies. This strategy aims to maintain technological leadership and control over AI development.
- Market dynamics and demand: The initial demand for advanced AI development in the U.S. led Cloud providers to concentrate their most advanced compute resources there. This created a “path dependency,” where the early establishment of infrastructure in one region makes it less likely for newer regions to catch up quickly.
- Regulatory and political considerations: In regions like Europe, local regulations and policies around ‘digital sovereignty’ are creating incentives for the localisation of AI compute infrastructure. This is particularly relevant for training large AI models, which may involve sensitive data that cannot be transferred outside the region.
The future of AI compute governance
The researchers expect the current concentration of AI compute infrastructure to grow even further, potentially tipping global inequalities to a boiling point, especially if trends continue. However, there are some factors, such as diversification, improved regulation, and international partnerships, that could lead to a shift in this trend.
The emergence of new technologies and providers, such as non-Nvidia GPUs and custom AI chips, could alter the landscape of AI compute. For example, the European LUMI supercomputer, equipped with AMD GPUs, represents a shift towards more distributed compute infrastructures.
As AI governance increasingly relies on compute power, there could also be greater calls for transparency and regulatory oversight, especially for infrastructures managed by private entities. Countries like Norway have already begun implementing requirements for data centre operators to register and disclose their capabilities.
Meanwhile, countries in the Compute Desert may seek alternative ways to bolster their AI capabilities, such as through international partnerships or government-led initiatives to build local compute infrastructures.
