Nuclear power for AI data centres

Following Google’s recent announcement that it is investing in nuclear technology to support its US data centres, nuclear power, particularly in the form of small modular reactors (SMRs), is emerging as a potentially promising solution to meet the growing energy demands of AI.

What are SMRs?

SMRs encompass nuclear reactors up to 300 MW (compared to roughly 1,000 MW conventional reactors), which can be prefabricated and relatively quickly constructed on site. This makes them ideal for use by tech companies who are racing to meet rising energy demands without the long lead in times which come with other energy sources. 

Currently, only China and Russia have operational SMRs but other nations, including the UK and US are developing their own designs. The US Nuclear Regulatory Commission certified its first SMR design for use in 2023, although this project was scrapped in the same year due to rising costs – something which will be addressed later in this article.

Growing energy demands of AI

Sam Altman, the CEO of OpenAI has described powering AI as the “hardest part” in 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 ChatGPT consumption alone at 226 GWh annually. This is roughly equivalent to charging the entire stock of US electric vehicles.

As AI models increase in sophistication and popularity, this energy demand will continue to grow as hyperscaler data centres require increasing computing power. These data centres will need reliable, substantial, and cost-effective energy supply. The risk of price volatility and supply instability when obtaining energy on the open market is potentially colossal for data centre owners and so tech companies need to find alternative solutions.

The nuclear option

Whilst many major tech firms are looking to renewables to power these data centres, nuclear presents an alternative and potentially more suited option.

Compared to single site wind or solar generation, nuclear is highly reliable. Moreover, its flat output curve, compared to the periodic daylight curve of a solar farm for example, makes it well suited for data centres which require large baseloads at all times of day. The modular nature of SMRs also make them highly scalable as further units can be added when data centres are expanded or rebuilt.

Furthermore, despite environmental safety concerns, which are being addressed through advanced features such as passive safety systems, nuclear does hold the benefit of being carbon free. This is crucial for tech companies who will want to ensure compliance with national and international emissions and sustainability standards as well as their own net-zero goals.

SMRs also require a relatively small footprint in comparison with solar or wind farms which makes them easier to accommodate on data centre sites. For example, a 300 MW solar farm would require between 1,200 and 2,100 acres of land. Meanwhile, an SMR can sit on less than 100 aces. This greatly reduces the site size which reduces land costs and opens the availability of alternative site locations.

Potential issues

Given the potentially catastrophic costs of nuclear accidents, the area is highly regulated which brings with it cost issues and time delays. 

Following the US Nuclear Regulatory Commission’s approval of NuScale’s innovative SMR design in 2023, the outlook for SMRs was highly positive. However, less than a year later, the project was scrapped, with the estimated project cost running up to £9.3bn in 2023 – up from the initial £3bn when it was announced in 2018. This financial strain was too much to make the project viable and ultimately cost is going to be a major issue for SMRs because ensuring bankability is often the main factor in getting energy generation projects off the ground.

Nuclear projects typically run well over their estimated costs and timescales. See France’s Flamanville-3 which is four times over the cost estimate and almost 12 years over the initial development time. This makes financing a huge problem as investors will be increasingly wary of nuclear projects. National government interest may also be waning with the US Energy Department under the last three administrations providing over $600m in funding for SMRs without a single operational plant to date.

SMRs are also inherently more expensive per unit of output in comparison with their conventional counterparts as the development costs are not proportional to the size of the project. This potentially makes them even less bankable, although guaranteed connection to a data centre does guarantee project revenue for the generator lifecycle.

Bankability may, however, be less of an issue for Big Tech, who have the funds and bargaining power to negotiate favourable contract terms with SMR developers which may well be giving them the confidence to gamble on a currently unproven technology.

Notable deals and developments

Most recently, in October 2025, Google signed an agreement with Kairos Power to produce a number of SMRs to power its AI data centres. The first reactor is expected to be operational this decade (pending regulatory approval) with more to follow by 2035. Meanwhile, Amazon have signed an agreement with Energy Northwest, a consortium of state public utilities, to develop a series of SMRs in Washington state.

Outside of SMRs, Big Tech are also investing in conventional nuclear power stations to provide their growing energy needs. Microsoft have plans underway to renovate an existing nuclear power station in Pennsylvania whilst Amazon have recently purchased 960 MW nuclear powered data centre from Talen Energy.

A nuclear future?

Whilst SMRs present an exciting prospect to deliver clean energy around the clock to AI data centres, it is crucial not to ignore the issues of regulation and cost which have so far severely limited their development. Whilst significant investment may lead to breakthroughs, it remains to be seen whether SMRs hold the key to AI hyperscaler growth against the background of more bankable options. 

For more information, please contact Josh Gallichan and Kate Deniston.