Artificial intelligence (AI) is increasingly considered a strategic technology that can provide significant economic, social, and military advantages. However, many types of advanced AI research require not just massive amounts of data, but also massive amounts of computing power. To make progress, researchers rely on hyperscale cloud computing infrastructures, which can provide stupendous amounts of compute in a scalable, on-demand fashion.
The ownership of these infrastructures is highly concentrated: Amazon, Microsoft, and Google together hold an estimated 70% of the global cloud infrastructure market, which includes the market for GPU compute used in AI research. The infrastructure is moreover extremely geographically concentrated. For example, Amazon’s P4d GPU instances are housed in giant data centres situated in just seven locations in the world: Virginia, Ohio, Oregon, Seoul, Tokyo, Frankfurt, and Dublin. In most countries AI researchers thus have to ship their data abroad to train a model. Governments are struggling to finance viable public alternatives. Efforts to develop distributed training infrastructures of sufficient scale remain nascent.
This project seeks to map the political geography of this computational infrastructure that underlies advanced AI by answering the following research questions:
- Which companies own hyperscale cloud computing infrastructure?
- How is it distributed around the world?
- How are governments attempting to shape its distribution?