Where Does Your AI Actually Run?
When executives talk about artificial intelligence, they still reach for software language: apps, copilots, agents. But the best recent policy research suggests that this vocabulary is no longer enough. The Bank for International Settlements describes AI as depending on a five-layer supply chain: hardware, cloud infrastructure, training data, foundation models and applications. The OECD likewise frames AI as a general-purpose technology underpinned by physical infrastructure, especially advanced chips and computing resources, and warns that these markets are susceptible to competition problems.
Photo: Google DeepMind
That distinction matters because the economics differ across layers. The BIS finds that upstream infrastructure markets are shaped by high fixed costs, economies of scale and scope, switching costs, network effects and consumer inertia, all of which create barriers to entry and favour larger organisations. In global cloud computing, the BIS reports market shares of 31 per cent for Amazon Web Services, 24 per cent for Microsoft Azure and 11 per cent for Google Cloud. In infrastructure-as-a-service, the segment most relevant to AI deployment, those three providers together accounted for almost 74 per cent of the global market in 2023.
The hardware layer is no less strategic. The BIS cites estimates that Nvidia’s share of the market for data-centre GPUs exceeds 90 per cent. It also notes that Taiwan Semiconductor Manufacturing Company produces more than 60 per cent of semiconductors globally and more than 90 per cent of the most advanced chips. The paper argues that the combination of concentration, switching costs, bundling and vertical integration can shape competition far beyond the semiconductor market itself.
This is why it is misleading to talk about “the AI market” as if it were one thing. The BIS explicitly distinguishes between highly concentrated upstream layers and more contestable downstream ones. Applications may multiply quickly and compete aggressively, while the infrastructure beneath them remains controlled by a relatively small group of actors. That does not mean AI will become a single monopoly. It does mean that the underlying routes through which AI is trained, deployed and scaled are not evenly distributed.
The investment data points in the same direction. UN Trade and Development reports that data centres accounted for more than one fifth of global greenfield project values in 2025, with announced foreign direct investment exceeding $270 billion. The value of newly announced semiconductor projects rose 35 per cent in the same year. These flows were concentrated in a handful of host countries, led by France, the United States and the Republic of Korea, while only a few emerging markets appeared among the top recipients. UNCTAD warns that without action, global investment risks becoming increasingly concentrated in a small number of sectors and regions, limiting its broader development impact.
Power makes this story even more concrete. BloombergNEF’s latest outlook projects that US data-centre electricity demand could reach 106 gigawatts by 2035, up 36 per cent from its previous forecast. Taken together, the evidence suggests that AI expansion is not only a software story. It is a story about chips, data centres, grids, land, financing and location.
This changes the strategic question. The issue for boards and governments is no longer whether to adopt AI. It is where the underlying compute sits, how concentrated the critical layers are, and how much operational dependence is being built into the organisation by choices about vendors, regions and model ecosystems. The OECD argues that effective competition in AI infrastructure will depend on authorities monitoring these markets closely and using a balanced mix of enforcement and advocacy tools.
AI strategy, in other words, is no longer only product strategy or workforce strategy. It is also infrastructure strategy. The leaders who understand that early will be better placed to manage risk, preserve bargaining power and avoid building their future on a route they do not control.