Middle Powers Can Avoid a Race to the Bottom on AI Data Centres

Summary

• Risk of a bidding war: Countries competing to host AI data centres are at risk of offering excessive concessions to AI companies.
• Cause for optimism: Governments have previously coordinated on minimum corporate tax and manufacturing subsidies – cooperation on data centres should be easier.
• A floor and a ceiling: Middle powers should agree on baseline transparency and security requirements for AI companies, and a cap on subsidies and tax concessions. 
• How to get there: Countries like Australia, Canada, France, and the UK could start by comparing notes on live negotiations, and over time work towards durable rules on data centre concessions.

 

Frontier AI needs infrastructure. The amount of compute used to train frontier models has more than quadrupled each year. Inference compute demand is rising around tenfold a year. To keep pace, the world needs more data centres. 

Most frontier AI compute is concentrated in the United States and China. But frontier AI companies want to build elsewhere to reduce geographic risks and avoid bottlenecks in US electricity generation and grid connections. Several ‘middle powers’ – here defined as countries beyond the US and China that can host significant compute capacity – are already lining up deals. France and Australia have announced major new AI data centre projects. OpenAI is backing a major cluster in the UAE. Canada is funding domestic AI data centres.

Governments want these projects on their soil for financial and strategic reasons. The UK government’s Data Centre Covenant, published in October, cited sovereignty, economic resilience, financial benefits, and securing access to AI models as key motivations for the country’s AI data centre build-out. 

However, as more governments seek to attract the same infrastructure, they may undermine one another’s ability to get good deals. 

The US state of Georgia estimated that its data centre tax breaks cost US$474.2 million in revenue in 2025, even though many of the data centres likely would have been built there anyway. Other jurisdictions, like the UK, have started to offer significant electricity discounts to attract data centre investment. In August, the World Bank warned that developing countries are also risking a race to the bottom, highlighting the decade-plus tax holidays that India and Uzbekistan are each offering to data centre operators.

We’ve seen this before 

Consider the global minimum tax (GMT) on corporations. Because companies can shift their profits to wherever tax is lowest, jurisdictions began undercutting each other's corporate tax rates to attract that profit. In response, some 140 jurisdictions agreed – through the OECD/G20 Inclusive Framework – on a 15% minimum effective tax rate for large multinationals. 

However, the GMT required years of negotiation and complex rule-making, necessary to address the mobility of the corporate tax base. The data centre problem is more tractable. One cannot wire transfer a data centre to the Bahamas. 

A closer precedent comes from European competition for large manufacturing plants. EU member states have long competed to attract factories, especially automotive plants, through tax subsidies. EU rules constrained those offers. This reduced the scope for firms to play governments against one another for ever-larger subsidies.

Semiconductors provide another recent example. The US and EU both want to attract new chip-fabrication plants. Concerned that these subsidy programs could turn into a dramatic transatlantic bidding war, both sides committed to avoiding a ‘race to the bottom’ in semiconductor supply chains and created senior consultation channels. The US and Japan agreed to share information on semiconductor support measures.

Policy implications

Middle powers can avoid a race to the bottom by agreeing to limit certain types of concessions. There are two questions to consider.

1. What should middle powers agree on?

Middle powers could try to preserve competition that makes them genuinely better places to build, while constraining competition that merely transfers public value to the firm or weakens public-interest safeguards.

For example, governments could agree to:

• Impose baseline transparency and security requirements on the AI companies leasing the data centres, such as mandates around incident reporting or disclosure of large training runs.
• Place limits on direct subsidies and tax concessions for data centre developers, where they do not serve the public interest. 

Countries can still compete on grid connection times – a priority for AI companies – as well as clean energy, workforce quality, and safety guarantees. The distinction between these advantages and direct financial incentives will not always be clear, but avoiding obvious subsidies is a useful starting point for coordination.

The data centre principles set out by Canada and Australia, or the conditions set out by the UK’s Data Centre Covenant, could provide a useful model for shared baseline requirements. They call for minimising water use and protecting other consumers from higher electricity costs, among other things.

2. How should middle powers coordinate?

Middle powers already negotiating with frontier AI companies are incentivised to move quickly, before another jurisdiction produces competitive offers and reduces their bargaining power. But in the medium-to-long term, middle powers would be collectively better placed if they coordinate now. 

Coordination mechanisms should ideally: 

• Cover enough potential hosts to constrain race-to-the-bottom dynamics.
• Operate quickly enough to influence negotiations happening now.
• Provide a durable framework for future negotiations.

On coverage, universal coordination is not needed. Only a handful of nations have the space, energy, and power connectivity required for gigawatt-scale build-outs. Canada, France, Australia, the UK, India, and the Gulf states are among the main competitors. 

Ideally, any such arrangement would involve all of these players. However, if the Gulf states or India declined, coordination among Western nations would still make sense. AI companies want to host their infrastructure in nations with legal predictability, less risk of US export controls, and lower exposure to regional conflict. A group of Western nations coordinating on subsidies can still be internationally competitive, even without Indian or Gulf-state buy-in. 

On speed, the semiconductor example shows that diplomatic mechanisms can enable rapid coordination. Governments should share information about what frontier firms are offering and requesting in data centre negotiations. If a company tells one government that another country has offered much better terms, officials should be able to ask each other whether that claim is true.

On durability, middle powers should begin developing longer-term arrangements to manage incentives. The OECD already supports similar arrangements for other areas, including minimum interest rates for aircraft export finance and limits on repayment periods for ship export finance. For data centres, participants could agree to cap total public support as a share of project cost. National governments could coordinate with state and provincial authorities to apply these rules.

Middle powers still have time to secure the benefits of data centres without a costly bidding war. But they must start coordinating now.