The Bank for International Settlements published a paper analysing the artificial intelligence supply chain across five layers, hardware, cloud infrastructure, training data, foundation models and AI applications. It finds substantial concentration in the hardware and cloud layers and describes how big technology companies are extending their footprint across the full stack through vertical integration, bundling and exclusive partnerships, with implications for consumer choice, innovation, operational resilience, cyber security and financial stability. On hardware, the paper notes that Nvidia accounts for more than 90% of data centre GPU revenues, supported by an exclusive software ecosystem around CUDA, and highlights dependencies such as advanced semiconductor production concentration. In cloud infrastructure, it reports global market shares of 31% for Amazon Web Services, 24% for Microsoft Azure and 11% for Google Cloud Platform, and finds the infrastructure as a service segment even more concentrated, citing switching costs, egress fees, lack of interoperability and ecosystem tying as reinforcing factors. While training data, foundation models and AI applications may be more contestable, the analysis points to potential “winner takes all” dynamics, including data scarcity pressures and dominance by a handful of foundation model providers, with GPT-4 estimated at 69% of 2023 generative AI revenue and ChatGPT at 60% of 2024 chatbot monthly visits; big tech firms are also described as major financiers and integrators, accounting for 33% of capital raised by AI firms in 2023 and nearly 67% for generative AI firms, alongside exclusivity arrangements such as OpenAI’s use of Azure and Anthropic’s use of Amazon infrastructure and chips. The paper notes that several competition authorities are gathering evidence and pursuing inquiries into AI market conduct, referencing a joint statement by the European Commission, the UK Competition and Markets Authority, the US Department of Justice and the US Federal Trade Commission, as well as investigations into major AI partnerships and Nvidia’s proposed acquisition of Run:ai. It also outlines policy options under discussion, including data-sharing measures, public training datasets, non-discrimination access requirements for foundation models, multi-cloud strategies and common application programming interface standards to reduce switching costs.

Original source View in tracker