Canada’s AtkinsRéalis has partnered with American technology manufacturer Nvidia to develop a range of nuclear powered, large-scale AI factories.
AtkinsRéalis claims the new framework will allow collaboration with developers and the implementation of AI infrastructure by deploying Nvidia technologies in the delivery process.
AtkinsRéalis’ aims to use its CANDU portfolio alongside Nvidia’s Omniverse DSX Blueprint for AI computing workloads to meet rising AI infrastructure demand as it outpaces the current available power supply.
GCR spoke with Sam Stephens, the head of digital for AtkinsRéalis’ nuclear arm.
What are your short and long-term goals for the project?
Most new global AI infrastructure capacity is currently being built in the US, driven by large‑scale investments from hyperscalers such as Meta, Microsoft, Amazon Web Services (AWS) and Google, plus an emerging class of new neocloud companies who are focused on AI computing power. This momentum is expected to continue as utilities increasingly align with hyperscale power demand.
As an AEC member in Nvidia’s Partner Network, we support the build-out of AI infrastructure by working with data centre developers who want to de-risk their projects with our engineering, design and project management services.
Over the next two years we aim to continue to transform the way we deliver infrastructure through the adoption of AI and Accelerated Compute and, in the longer term, work with those developers to ensure power supplies are reliable and sustainable.
We see nuclear energy providing secure energy supplies for the AI infrastructure. AI helps manage and operate the energy supplies to optimise performance, and can help transform delivery processes to provide that energy faster and with greater certainty.
How will nuclear power be integrated on site and how will it affect the design and layout of the development when compared to a conventional data centre build?
Integrating nuclear power on site changes the scale and layout of an AI campus compared with a conventional data‑centre build, but it also enables more resilient, efficient and sustainable design.
Nvidia’s DSX reference blueprint for a 1GW AI factory indicates that accommodating on‑site power generation and related infrastructure could increase the overall land footprint by approximately 20%-30%.
A nuclear power plant would be connected to the grid in order to provide essential services and support to local communities and infrastructure, not just an AI factory. So, there are opportunities to share grid infrastructure with the data centre.
With the latest data centre designs relying upon closed loop water cooling, as opposed to air cooling, we also see opportunities to use waste heat from the power plant with absorption chillers to increase sustainability and reduce operating costs.
Co‑locating AI data centres with nuclear power increases site scale, but it enables a more integrated, efficient and sustainable infrastructure model, where power, cooling and grid connectivity are designed together, not separately.
How did you come to the decision to use nuclear power, as opposed to a renewable source?
AtkinsRéalis supports a diverse energy mix and we see renewables already scaling to meet our future needs. However, we see nuclear playing a vital role, particularly as data centres demand secure, always on, 24/7 power. We need low carbon generation that is proven at scale and nuclear energy is one of the only power sources capable of meeting those needs globally.
Already, data centre campuses are being developed in the US with a mix of energy sources, including power purchase agreements with existing nuclear power plants. CANDU reactor technology is uniquely placed to support these needs with online refuelling and the use of unenriched uranium from stable regions of the world.
How scalable do you believe the technology is and when do you think the first project of this kind will be completed?
We see this as highly scalable.
AtkinsRéalis holds the exclusive license to CANDU® Reactor technology with 740MWe and 1000MWe designs and also supports a wide range of fission and fusion technology vendors, such as GE Vernova, RR SMR, EDF Energy, in the development and deployment of their solutions.
Projects of this kind are already underway, such as Energy North West’s deployment of X-Energy’s small modular reactor with AWS with AtkinsRéalis supporting as owner’s engineer
We expect an acceleration of the commissioning of these projects as operators look to secure longer term power supplies.
Do you think nuclear AI factories will become the model for data centre development?
Yes, we expect nuclear power to become the preferred approach as hyperscalers and neoclouds seek to avoid volatile fossil fuel prices and seek long term cost certainty, with further support from investors to meet ESG and CSR targets.
How will the use of digitisation help the project?
Nvidia and AtkinsRéalis are collaborating to identify ways that AI and Accelerated Compute can reduce timelines for new nuclear permitting, design and construction. We see digitisation as vital across the project and asset lifecycle.
For example, by developing digital twins on Nvidia’s Omniverse platform, integrating OpenUSD 3D data from all our design and analysis tools, we see opportunities to simulate and optimise plant designs from a construction, operation, maintenance and decommissioning perspective. AI will help our designers reuse information across projects and simulate environmental conditions to reduce risk to as low as reasonably practicable.
We are also applying AI to better predict project outcomes, critical to Nvidia and the overall ecosystem as speed to market is one of their primary concerns. AI can better predict risks of delivery and help teams to manage schedules, an area that we expect to further enhance with the help of Nvidia technologies such as agentic AI.
To scale nuclear power quickly enough we will need new ways of working, and digitisation promises to provide that breakthrough.
How do you envisage the future of AI infrastructure?
While the majority of AI infrastructure investment has been in the US to date, particularly to develop and train AI models, we expect that build out will be more geographically widespread.
This will be driven by a need to have infrastructure close to the people accessing these services, requiring ‘inference’ from the pre-trained models, but also with a shift towards ‘sovereign AI’ where countries seek to protect their information.
There is growing talk around Agentic AI being the next big wave, for example with openclaw being heralded as the next ChatGPT. This will drive more demand and more growth for AI infrastructure.
Physical AI is also gathering pace, where AI starts to interact with the environment around us, either through data collection through sensors or through operations conducted by autonomous machines, such as robotaxis or humanoids.
AI infrastructure will become as ubiquitous as the mobile phone or broadband infrastructure, with small and large data centres being developed as we have seen masts and exchanges and with AI chips amongst us in everyday life.
