Powering AI data centres – challenges and opportunities

That’s more than the entire electricity consumption of Japan today. The energy requirements of AI data centres also differ from their non-AI counterparts. AI data centre developers and operators need to carefully consider their energy sourcing, and energy generators and suppliers will need to be ready to offer not only more energy, but the right sort of energy too. 

In this article we analyse some of the potential challenges and legal questions that come with powering AI data centres, alongside potential routes forward.

Growing energy demands of AI

Sam Altman, OpenAI’s CEO, has described powering AI as the “hardest part” of satisfying demand. Each search with OpenAI’s ChatGPT typically requires 2.9 Wh per request, almost 10 times the 0.3 Wh required for a Google Search.  Whilst the overall consumption numbers for AI data centres are currently unconfirmed, some estimates put the scale of consumption from the use of ChatGPT alone at 226 GWh annually. This is roughly equivalent to fully charging the entire current stock of US electric vehicles. There is, therefore, a substantial challenge for data centre operators in procuring this energy and a huge opportunity for energy generators in providing the same. 

As AI models develop in sophistication, the resulting increased computing power will necessitate rising electricity requirements for AI data centres, which will be compounded by the growing popularity of the associated AI products. These data centres need reliable, substantial, scalable, and cost-effective energy supply. The risk of price volatility and supply instability when obtaining energy on the open market is substantial for AI data centre operators who will need to find alternative solutions. There are therefore two principal considerations: the first is the power source and the second is how the purchase of electricity will be structured. 

Sourcing power

AI data centre operators will require the following from their energy supply:

1. Energy density: AI data centres have a higher energy demand than the standard data centres required for technologies like cloud computing for two reasons. Firstly, the graphics processing units (GPUs) required for AI have greater computing power and correspondingly require more electricity to operate. Secondly, this demand is compounded by the parallel energy requirements to cool these GPUs. This makes AI data centres more energy dense than their non-AI counterparts, even before we start looking at size, where hyperscale data centres are increasingly being used by AI developers and system owners. 
   
   
2. Reliability: AI data centres have a consistent energy demand with the priority being to prevent downtime as AI consumers want consistent and uninterrupted access to AI products. 
   
   
3. Cost effectiveness: Energy will be the largest operating cost for an AI data centre and energy sources must therefore be cost effective. 
   
   
4. Scalability: As the race towards more powerful AI systems gathers pace, AI data centres must be scalable and so must their energy sources. 
   
   
5. Sustainability: In addition to the above, AI data centre owners are likely to align their operations with sustainability and net-zero goals. This may be a way to distinguish AI businesses within the market. We have also begun to see governments mandate that data centres be powered from renewable sources, including China, who in 2023 announced their ‘Green Data Centre’ standard which requires data centres to be powered by 100% renewable energy by 2032. The EU has also implemented a number of efficiency standards which affect data centre owners, including RED3 which exist alongside independent standards such as the Climate Neutral Data Centre Pact. For further information please see our articles on EU and Singaporean data centre sustainability legislation. 

Given these specific requirements, there are a limited number of power sources for data centre owners to consider. Traditional oil and gas generation has been able to supply a large part of energy centre demand and is appropriate in many aspects – however, the increased regulatory and market movement towards sustainability makes the future for these sources for AI data centres less certain. Furthermore, whilst data centres owners in certain locations may look at resource specific sources such as geothermal or hydroelectric, on a global scale, it will increasingly be wind and solar powering AI data centres, with hydrogen and nuclear presenting themselves as potential disruptors.

Wind and solar

Due to the sustainability requirements outlined above, renewable sources, particularly wind and solar, will likely be at the forefront of powering AI data centres. They offer cost effective renewable power which aligns with corporate net-zero goals and can help future proof against potential regulation which could make utilising non-renewable sources such as oil and gas unviable. 

The primary issue with sourcing electricity from wind and solar on their own is intermittency. Solar and wind have low-generation periods due to the variation in the forces of nature from which these technologies derive power. However, the solution to this is the integration of energy storage, primarily battery energy storage (BESS). BESS allows excess energy being created during high generation periods to be stored and then released during low generation periods, as well as allowing low-cost energy to be stored for use during high-cost peak times. This will allow data centres to be supplied with a cost effective, reliable and steady baseload. Such BESS can be co-located with the generation asset, located at the data centre or at a separate site.  Hydrogen may also offer an alternative energy storage solution for data centres which can either be produced on-site from low-cost renewable energy at peak-generation times or purchased wholesale (though noting the current market for wholesale hydrogen is in its infancy). For more information on energy storage, please see our Unlocking Energy Storage report.

A secondary issue with wind and solar is the scalability which is crucial for AI data centre owners during the current period of AI growth. It is difficult for these types of generators to simply add more turbines or PV panels as land constraints create a physical barrier to expansion. Land permissions and grid connection amendments can also create delays which limits the agility of such projects.  AI data centre owners may therefore need to enter into further agreements down the line to purchase power from additional sources if their demand outstrips generator capacity.  Those contracts are not easy to navigate. They are long term, complex agreements, often with a financing angle which can result in data centre owners incurring costs and delays.

Despite these challenges, the wind and solar market continues to grow alongside the AI data centre boom with big tech remaining at the forefront of the wind and solar power purchase agreement (PPA) market as these companies look to power their AI and non-AI data centres. Amazon, Meta, Google and Microsoft continue to dominate and Amazon remains the largest corporate clean energy buyer in the world. These companies are also among the largest players in the AI market and will continue carrying AI data centre growth alongside renewable PPA growth. 

The nuclear option

Small modular reactors (SMRs) are being discussed as an innovative and potentially promising solution for powering AI data centres. It can be hard to visualise nuclear power stations built in a production line in contrast with traditional images of an enormous dome reactor core and cooling towers. In fact, smaller SMR models are likely to look like shipping containers, which could be stacked on top of one another. The modular design of these reactors, with capacities up to 400 MW (as opposed to traditional 1,000 MW nuclear reactors), allows for prefabrication, rapid on-site assembly and scalability, making them an attractive option for data centre designers.

Nuclear power offers several advantages for AI data centres. It provides carbon-free energy, which is increasingly important to AI data centre owners, as well as offering a reliable and large baseload power supply. This is crucial for the continuous operation of AI data centres, which, as discussed, have a higher energy density than non-AI data centres. Additionally, SMRs have a smaller footprint than solar or wind farms which makes them better suited for co-location at the data centre. Whereas a 300 MW solar farm would require between 1,200 and 2,100 acres of land, an SMR can sit on less than 100 acres. This means they can be more easily co-located and byass national grid systems, the benefits of which are explained below. They are also highly scalable given their modular nature which makes them an agile solution in a continually changing AI data centre market. 

However, the deployment of SMRs is not without challenges. The high potential for cost and time overruns during the construction phase pose significant financial risks, as do the regulatory and PR hurdles related to nuclear power. Moreover, past project failures have led to scepticism among investors, making it potentially difficult to secure third-party funding. For similar reasons, reactor vendors will not want to build on balance sheet and instead will try to sell the asset to the data centre developer which passes on the substantial risks. The regulatory landscape may be changing, however. The UK, for example, announced in February 2025, the removal of restrictions on designated nuclear sites, opening up the possibility for nuclear power across the nation with an eye to commercial SMRs being used to power energy intensive sites such as AI data centres. 

Despite these challenges, major technology companies are actively investing in nuclear power for their data centres. Google has partnered with Kairos Power to develop SMRs specifically for AI data centres while Amazon has signed an agreement with Energy Northwest to deploy SMRs in Washington state. Larger, non-SMR nuclear solutions are also being explored with Microsoft planning to renovate an existing nuclear power station in Pennsylvania to power its AI ambitions alongside a pledge to support a $30bn BlackRock fund investing in AI infrastructure. 

Nuclear fusion is another technology potentially relevant to AI data centres. Long awaited but so far out of reach for practical and commercial purposes, this technology could solve not just the issue of powering AI data centres, but of powering the entire clean energy transition. Again, the AI giants are leading the way in investment in this market with the Sam Altman backed start-up Helion closing a $425m Series F funding round alongside a deal with Microsoft to supply its operations (in particular its AI data centres) with nuclear fusion generated electricity by 2028 (the first such PPA of its type). 

Power purchase structuring

Once a power source has been identified, data centre owners have decisions to make regarding the structure of its purchase. A PPA allows corporate energy consumers to purchase power directly, and on a long-term basis, from renewable and nuclear energy generators. There are a number of structures available but all present long term price surety against a potentially volatile wholesale market leading to cheaper energy and easier financial planning.

Private wire PPAs

Where data centre owners have adequate resources and would like maximum control over their own power generation, a private wire PPA may be the most suitable option. 

This is available where the generation asset and the data centre are co-located or adjacent. The key benefits are exclusivity and a bypass of distribution and generation regulation which avoids costs associated with connection to public grid systems, thereby bringing down the PPA costs even further.  

By bypassing national grid systems, AI data centre developers can also bypass lengthy grid connection lead times which can often act as a roadblock for projects.

This is the key structure for nuclear powered AI data centres but is also commonly seen with renewables and in particular rooftop solar.

Physical PPAs

Physical PPAs (depending on regulatory requirements also referred to as sleeved PPAs) fit into two subcategories, pay-as-produced and pay-as-nominated, where the data centre owner either purchases all power produced or a set amount respectively and gets it delivered physically. A sleeved PPA involves an offtaker signing a PPA with an electricity generator in parallel with a separate agreement with a licenced intermediary who transports the energy from the generation site to the offtaker’s site and tops-up if needed. 

This structure allows an AI data centre owner to purchase power from a generation asset without the constraints of co-location which opens up possibilities to purchase from generation assets such as offshore wind and floating solar where land constraints may be less of an issue. 

Synthetic PPAs

Also known as a virtual or financial PPA, this structure does not require the physical delivery of power. Instead, generators sell power to a utility company in a conventional utility PPA who then supply electricity to the end-user. However, the end-user may make a separate agreement with the generator for a contract for difference (or alternative derivative) under which the parties agree to settle the difference between market price and an agreed ‘strike price’. 

This is often accompanied by the purchase of guarantee of origin to allow companies to offset their emissions. Such a structure is also utilised for cross-border PPAs. Note that the derivative contract can be considered a regulated financial instrument and so further regulatory advice, and subsequent authorisations and compliance may be needed.

Hybrid PPAs

This is less a separate structure but a subset of PPAs with integrated energy storage; usually as a separate cost. This is well suited for AI data centre wind and solar deals where BESS and other energy storage systems are crucial to delivering the high-baseload requirements.

Bird & Bird has considerable experience in advising on these structures. For more information, please visit our PPA Hub. 

The future of AI data centres

AI data centre owners have a number of considerations to make when looking to source power for their operations. A suitable power source must first be considered before a determination of the best way of delivering that power to the data centre asset.

It is also important to note that the next few years of data centre construction may be dictated by political trends, particularly in the US where according to one source, there are over ten times the number of data centres as in Germany, which, according to the same source, has the second most data centres worldwide. The early days of the Donald Trump administration look to be characterised by reduced federal support for renewable projects and a bolstering of the oil and gas industries. Furthermore, in January, Trump announced a $500bn private sector investment to support US AI infrastructure which looks to spur AI data centre construction. Therefore, in the US, which is a global hub for AI data centres, we may see developers moving away from renewables and towards nuclear and gas if these policies swing the balance of cost-effectiveness enough to make these the most viable energy sources. 

However, this is undercut by the demands of tech investors and consumers as well as the goals of major tech companies with both Google and Microsoft pledging 100% clean energy for their data centres by 2030. 

AI itself is also a potential solution to the issues its data centres are creating. Through predictive analytics and machine learning, AI can dynamically adjust cooling systems, manage server loads, and optimise power usage, reducing unnecessary energy consumption. Additionally, AI can predict maintenance needs and optimise hardware performance, leading to reduced downtime and more efficient energy use. By continuously analysing data centre operations, AI can identify patterns and implement energy-saving measures, ultimately leading to more sustainable and cost-effective data centre management.

How we can help

Our team works across borders and draws on the knowledge and experience of more than 70 international lawyers across key practices like AI, real estate and energy. For more information, please contact: Dr. Matthias Lang, Marco Nicolai, or the wider team here. 

With thanks to Josh Gallichan for his contribution to the article.