Large data centers that power modern AI consume a colossal amount of electricity. Until recently, the discussion around this topic boiled down to one question: how to provide enough power to run it all. However, at the CERAWeek conference – one of the world's key forums in the energy industry – a different idea was voiced: what if AI infrastructures themselves could help the power grid, not just strain it?
From Consumer to Grid Participant
NVIDIA, in partnership with Emerald AI, has presented an approach where so-called “AI factories” – large-scale computing complexes designed specifically for artificial intelligence tasks – become flexible elements of the energy system. Simply put, they don't just consume electricity on a fixed schedule; they can adapt to the current state of the grid by reducing their load during peak times, helping maintain stability, and generally behaving like a “smart” participant in the energy market.
This changes the logic behind infrastructure design. Previously, you had to build with a surplus to handle peak consumption – more capacity, more reserves. But if the load becomes manageable and predictable, the need for such overbuilding is reduced. This is a win-win for everyone: infrastructure operators, energy companies, and ultimately, consumers.
Several major energy companies have already joined the collaboration: AES, Constellation, Invenergy, NextEra Energy, Nscale Energy & Power, and Vistra. They all plan to work on creating generating capacity that will be integrated with the new AI factory architecture, including through projects involving on-site local generation.
How Many “Thoughts” Fit into a Single Watt
There is one metric that has recently become more important than many others in the world of AI infrastructure: tokens per second per watt. A token is, roughly speaking, a unit of text that a model generates in response to a prompt. The more tokens a system can produce per unit of consumed energy, the more efficient it is.
This is not just a technical indicator – it represents real money and a real strain on the infrastructure. Companies that operate AI systems at an industrial scale pay huge sums for electricity. Improving energy efficiency directly lowers the cost of every single query.
Jensen Huang, founder and CEO of NVIDIA, put it bluntly in a recent interview:
"Power consumption is a problem, but it's not the only one. That's why we are working so hard on co-designing the entire system to improve the tokens-per-watt metric by orders of magnitude every year."
And judging by NVIDIA's track record, these are not just empty words. The company offers a comparison: since 2012 – with the release of the Kepler architecture – the number of tokens achievable within the same energy budget has grown by more than a millionfold. Yes, you read that right: a million times. This is the result of years of consistent improvements in chips, architecture, and software.
Huang describes modern AI infrastructure as a “five-layer cake”: energy, chips, infrastructure, models, and applications. Energy is at the base. Without it, nothing else works.
Robots on Solar Farms and Digital Twins of Nuclear Reactors
At the same CERAWeek conference, several companies from the NVIDIA ecosystem demonstrated exactly how AI is helping to build the energy sector of the future faster and more reliably.
Maximo, a solar robotics company that spun out of AES, announced the completion of a 100-megawatt solar installation at the Bellefield site. The installation was performed by autonomous, AI-controlled robots. This is significant because one of the primary obstacles to scaling up solar energy is the speed of construction: there is a shortage of skilled installers, and demand for new capacity is outpacing supply. Robotic installation helps to bridge this gap.
TerraPower, in collaboration with SoftServe, unveiled a digital twin for its next-generation nuclear power plants, specifically for the Natrium reactors. A digital twin, in short, is a virtual model of a physical object that allows design solutions to be tested beforehand without any real-world construction. According to the company, this approach shortens the design cycle from years to months. For the nuclear energy sector, where every approval stage takes years, this represents a major shift.
Adaptive Construction Solutions announced a national vocational training program – in partnership with NVIDIA – to train specialists who will build and maintain AI factories and their energy infrastructure. The shortage of skilled labor for the construction and operation of such facilities is a problem just as real as the shortage of power. The program is geared towards mass training for in-demand trade professions.
Digital Twins of Infrastructure: The New Design Standard
Meanwhile, at the conference, several major infrastructure equipment manufacturers described how they are rethinking data center design.
GE Vernova presented an approach where substations, power grids, and the AI factory's loads are simulated together – before physical construction begins. This allows them to identify potential bottlenecks in advance and mitigate risks associated with connecting to the grid.
Schneider Electric announced new validated designs compatible with the Vera Rubin architecture. The company simulates power, cooling, and control systems in a single environment, allowing operators to optimize parameters long before physical construction begins.
Vertiv focused on standardizing physical infrastructure with repeatable power and cooling modules that integrate into the overall architecture, enabling AI factories to be scaled up more quickly and with fewer risks.
The common idea behind all these announcements is the same: test everything in the digital realm first, then build. This approach reduces the cost of mistakes and accelerates the ramp-up to full operational capacity.
Why This Matters Beyond the AI Industry
The conversation around AI energy efficiency is often perceived as a purely corporate issue – a headache for the companies footing the bill for data center electricity. But in reality, the scope of the problem is much wider.
The growth of AI workloads is putting real pressure on power grids, especially in regions where many new facilities are being built. If these facilities consume energy in an uncontrolled manner, it impacts the grid's reliability for everyone else. An approach where AI factories become “smart” grid participants – able to react flexibly to its state – has the potential to mitigate this risk.
This is not a silver bullet. The construction of new generating capacity is still essential, and it is underway – though at a slower pace than demand is growing. But the fact that major industry players have started to view the power grid not just as a background resource, but as a full-fledged part of the system to be co-designed with the computing infrastructure, marks a noticeable shift in thinking.
