National data centre projects are consolidating America’s AI lead

The writer is the author of ‘The New Geography of Innovation’ and winner of the Financial Times/McKinsey Bracken Bower Prize

The geography of AI is largely simple: the US and China control close to 90 per cent of all compute, 90 per cent of all investments and 90 per cent of all notable foundation models. But there is one leg of the AI race which still looks genuinely multi-polar: data centres.

There is an obvious reason for this. It takes enormous resources and years of effort to build national champions like OpenAI and Nvidia. But any country with enough land, capital and electricity can build a data centre. This is why countries that barely show up to the starting line in other parts of AI seem like outsized contenders here. The United Arab Emirates is planning a massive five-gigawatt AI campus, the largest concentration of AI infrastructure outside the US. India has assembled 45,000 GPUs as part of its national AI programme and is subsidising their domestic use. The EU, Saudi Arabia and Japan have all mobilised billions of dollars to host AI gigafactories.

When it comes to turning electricity into intelligence, the map starts to look less uneven. According to the International Energy Agency, in 2024 the US accounted for 45 per cent of global data centre electricity consumption, China 25 per cent and Europe 15 per cent.

Governments clearly believe they are acquiring more than just server capacity. Brussels says that these investments will advance Europe’s “technological autonomy”, while India has described its acquisitions as part of its effort to build “technological sovereignty across the AI value chain”. Nvidia is openly pitching itself as a partner to governments looking to build sovereign AI.

However, data centres do not constitute sovereign AI capability. Countries may host facilities but dispersion of hardware does not equal decentralisation of power. Instead, the rush to host them is giving the US an opportunity to turn its lead in advanced chips into a system of geopolitical alignment before China can offer a competitive alternative.

The UAE provides the clearest preview of this. On July 10, the US commerce department upgraded its status under US export rules, placing it alongside Nato allies, and approved two state-linked companies, G42 and Core42, to receive advanced equipment without individual licences.

But it is revealing what this preferential treatment means in practice. Access is provisional. The companies keep the status for only 270 days unless they “become US companies”. The UAE has to remove Chinese equipment from its systems. Peng Xiao, G42’s CEO, has said that the decision to cut ties “was not easy” and the company was in effect forced to choose sides.

Data centres do not guarantee strategic autonomy so much as offer a strategic bargain: access to frontier compute in return for geopolitical alignment. Abu Dhabi brings vast capital and defence ties. Not many countries have that much sway in Washington. Others will have to wait in line for US compute.

Compute infrastructure is one of the last remaining corners where the US enjoys a clear, largely uncontested lead in AI. Chinese models have nearly erased the performance gap with American systems. But while China has made enormous progress in research, applications and algorithmic efficiency, it remains weakest at the hardware layer — GPUs. More importantly, China lacks a globally accepted substitute for the Nvidia-centred software and hardware ecosystem that has become the default architecture of modern AI.

Brazil recently announced two major national AI computing projects: one involving China’s Huawei and iFlytek and the other with Nvidia, suggesting competition may be arriving. Huawei’s Ascend chips are improving, but it is still far from exporting Ascend-based gigafactories that buyers consider interchangeable with the US stack. Once a country pours billions into scaling compute organised around Nvidia hardware, it’s hard to displace that installed base. The US doesn’t need its GPU advantage to last for ever, only long enough to make it too expensive for everyone building on top of it to switch sides.

The precedent is the railway architecture of the British empire. Tracks spread across the map, but they built a commercial system centred on Britain. The AI build-out follows a similar logic: data centres sit on sovereign territory, but the network they form is American.

Host countries are getting operational capacity but not strategic autonomy. They won’t necessarily emerge as independent AI powers, though they may become important locations inside an American system. Autonomy will remain elusive. What appears to be decentralised sovereign AI capability may end up centralising US geopolitical influence.