After Microsoft, that is financing the restart of an idle reactor, Amazon and Google just made the headlines by announcing that they both will bet on a very promising technology, small modular reactors or SMRs. The former (which also acquired earlier this year a data center powered by an adjacent nuclear power plant) will partner with Dominion Energy to explore this technology while the latter inked a deal with Kairos Power in which the nuclear startup will provide seven SMRs to power Google’s future (AI) data centers by 2035, with the first one scheduled for 2030.
Google and Amazon’s backing legitimates the SMR technology that is still in its infancy. Indeed, there has been very few SMR operational rollouts and these rollouts have been limited to China and Russia. That said, there is a large number of US startups including Nuscale, TerraPower (backed by Bill Gates) and Oklo (backed by Sam Altman) chasing this new nuclear eldorado and the US Nuclear Regulatory Commission has recently approved Kairos’ plans for a 35MW demonstration reactor, suggesting the US is advancing on the topic.
SMRs are expected to handle the challenges (poor safety, delays, and cost overruns…) that have plagued nuclear reactors over the last decades and are considered a next-generation energy source due to their modular construction, higher geographic flexibility, reduced costs (15% to 40% lower capex) and better safety (thanks to their distinct cooling technologies).
Because large-scale nuclear power plants take an average of six to twelve years to construct, and the permission process can take much longer, SMRs’ scalability and faster deployment (only three to five years of construction), along with ease-of-installation and use (assembled on-site), provide a comparative advantage.
Over the past 20 years, the power consumption of data centers has relentlessly increased as cloud computing, servicing always-on smartphones or providing video streams (Youtube, Netflix…), has become the norm for billions of users. So far, hyperscalers were able to deliver these online services with electric power delivered the traditional way by simply hooking the data centers to the grid.
But, since the advent of the now 2 years old Generative AI wave, the cards have been completely reshuffled for the whole IT industry but also for the energy and utilities sectors. The upcoming buildout of hundreds of AI-enabled data centers will trigger an electric power supply imbalance not seen in decades.
In fact, the next generation data centers, designed for AI training/inference, will have a significantly increased rack density (hence more chips) and require liquid cooling. These upcoming computing powerhouses will easily consume 500MW, with some of them even reaching the Gigawatt scale as clusters of more than 100’000 GPUs could be built in a not-too-distant future.
As the utilities/energy sectors will not be able to cope with the rapid rollout of these AI data centers, Google and Amazon’s strategy to rely on their own nuclear power source based on SMRs makes full sense, at least on paper. Some technological issues still need to be resolved, the availability of uranium is clearly not guaranteed (already in under-supply) while the regulatory setup is currently inexistent (allowing a private company to run its own nuclear reactors will be a tricky subject for any government, to say the least).
GenAI’s disruption is not only challenging the production of electric energy and its transport (the grid is also in need of a significant upgrade) but some segments of the industrials and materials sectors as well.
For example, some “old and low tech” equipment, like transformers, are currently in shortage. Transformers are used to convert (AC to DC) and to lower the electric power entering a data center to the adequate tension for all the electronic components forming the computing stack (CPUS, GPUS, memory, networking…). A specific type of steel, called Grain Oriented Electrical Steel, is used for the transformers’ core. This specific material is currently in shortage due to the heavy demand stemming from data center constructions.
Shortages across many industries are, currently, the price to pay to build the electric infrastructure which will form the base layer for upcoming technologies like autonomous transports, smart cities, Metaverse and many others.
Investing in the specific segments across the energy, utilities, industrials and materials sectors linked to the GenAI wave is an elegant way to be exposed to this revolution while benefitting from a lower volatility level due to the longer cycles inherent to these businesses.
