Generate or Perish: Preparing for the Disruptive Power Hunger of Artificial Intelligence

Subramaniam knows a thing or two about supply chains and keeping pace with change. FedEx is only 50-something years old, but it has grown from an American startup competitor to the U.S. Postal Service into a global giant with hundreds of airplanes, thousands of cargo vehicles and moving millions of packages every day.

And even he was caught off-guard by the rise of AI, the phenomenon which one of superchip maker NVIDIA’s top creative leaders will explain later.

For Subramanium, FedEx and any number of companies hoping to stay alive over the next few decades, AI is a reality which grows more real and demanding every day.

“I didn’t know the AI revolution was coming, but it did come,” the FedEx CEO pointed out.

Hype? Tell that to the meteorite of technological evolution coming to sweep away the old and reward the adaptable. Already, Schneider Electric CEO Olivier Blum and James Lee, partner at private equity giant Blackrock Global Infrastructure Partners (GIP) both pointed out during the Innovation Summit that a majority of commercial and industrial companies are utilizing generative AI in their operations.

Hyperscale activity is proof positive that generation is golden

And this AI adoption increases the pressure on the energy grid to a point where the old utility model, already under the stress of aging equipment without transformational upgrade, cannot keep up with the hundreds of GWs in new capacity which will be needed in the next decade.

“We firmly believe it is not a bubble,” said BlackRock GIP’s Lee. “There are strong fundamentals, first in the rapid acceleration of adoption. . . That adoption is driving real revenue.”

And it is driving hyperscalers such as Microsoft, Google, Amazon, Meta and Oracle to get off the energy sidelines and invest big in support for next-generation capacity both on the grid and possibly off-grid. These new power plants could be microgrids or co-located power plants, which the data industry sometimes calls energy parks, powered by natural gas and next-gen nuclear resources.

On-site power by any other name is still as relevant

None of these small modular reactor projects have happened yet, but they are moving closer to reality with the momentum of a well-positioned asteroid of change. The old utility model is not going to work in this new world, and survival of the fittest means flexibility and forward investment.

“They are not calling them microgrids but calling them energy parks,” noted Jana Gerber, North American Microgrid sector president for Schneider Electric, in an exclusive interview with Microgrid Knowledge at the Innovation Summit North America.

“The concept is very adaptable (of microgrids to provide prime and backup power for data centers and AI factories),” Gerber said. “We must shift the concept of DERs (distributed energy resources) so it’s not going to be just solar. . . We’re working on early stages of what projects are going to look like and what is the DER mix that will be required.”

For the future hyperexpansion of AI and cloud-based facilities in this Industrial Compute Age, that mix could be gas generators, with renewables and battery storage and eventually maybe even SMR nuclear. Everything is on the table.

To lay out the options, it might help take a step back and understand what AI is and how it is radically different in terms of data center computing. Who better to do that than Rev Lebaredian , vice president of Omniverse and Simulation Technology for leading processing unit maker NVIDIA.

Maybe you’ve heard of them? First company to reach a $5 trillion market valuation, dominant chip and graphic processing unit (GPU) player in the era of AI, and an overnight success story which took 30 years to develop.

Lebaredian was a visual effects maestro within Hollywood firms who joined NVIDIA and its founder, Jensen Huang, to create the same level of graphically fantastical worlds for gaming chips. That seemed good enough for a time, but Huang and his leadership team wanted to take the highest performing computing to solve major challenges in the physical world.

“Our goal was to build special computers to solve problems that were nearly impossible,” he recalled.

Game changer in deep learning and AI supercomputing

And eventually they produced the most powerful GPU possible, one so powerful there was no computing capacity to handle it. So they waited, and waited, about seven years until research at the University of Toronto achieved real breakthroughs in the potential of supercomputer-level deep learning, neural networks and natural language processing. It certainly didn’t start with ChatGPT.

“Finally someone figured out what to do with all that computing power,” Lebaredian noted. “This is going to change everything.”

To empower that change the technology requires AI factories, which are not the same thing as data centers. In Lebaredian’s illustration, a data center is essentially a warehouse or storehouse which can be accessed with relatively minor computing power.

An AI factory is just what its name implies—the power-hungry manufacturing and refining center where all of that data is distilled into something useful on a massive scale, such as gaining incredible insights into healthcare or industrial processing or facility operations and maintenance, so far beyond writing an easier school report that it becomes quaint to think of AI that way.

“You put into a factory all the raw materials and energy, and all the raw materials are reconfigured into it and out comes a refined production that is better than all of its parts,” he said. “With the factory you want to maximize the density as much as possible.”

Sitting on a volcano at the fault line of energy supply and demand

And that’s where a large part of the latest revolution in energy will need to be deployed. Currently, data computing racks consume energy capacity at about 140 kW per rack, but that is likely going to 1 MW per rack by 2027 and producing the possibility of 1-GW AI-enabled data computing centers by the end of decade. One gigawatt is equal to the average nuclear reactor unit.

The realization, shared not only by companies but also by the Trump Administration, political rivals and leaders throughout the world, is that you cannot lose the AI race. Winning that race, or keeping pace, will demand energy production at a growth rate the grid has not handled since its early days.

Mark Christie, former chairman of the U.S. Federal Energy Regulatory Administration said that electricity load has been lately growing five to six times faster than the commissioning of new generation capacity. This warp speed of demand over supply could produce a real human crisis in coming years, from energy unaffordability to access to reliable energy for the individual.

“We are sitting on a volcano,” Christie said during the second day Innovation Summit keynote at the Fountainebleau, quoting an old line by French writer Alexis de Tocqueville about his country on the brink of a revolution. He worries that the same thing could happen if America doesn’t get a handle on how to adequately power energy-hungry AI and a growing population.

“The political volcano could be from energy prices, and we’ve got to address that,” Christie added. “We cannot forget that what retail customers are paying is going to be a huge part of it. If we don't, the volcano could blow up and it could be bad.”

Enter microgrids and energy parks, or whatever the data industry decides to call them, to the rescue. Building out capacity at the edge of the grid or even off-grid could circumnavigate interconnection and regulatory delays, but even so the impetus is on the business, political and technological contributors to collaborate and figure out how to balance supply and demand so it doesn’t come down on us like space rock.

Utilities are certainly not ignorant of this, and they are growing capital expenditures by 60% or more in recent years.

“That investment is happening but that investment by itself is not enough,” BlackRock GIF’s Lee pointed out. “We need an all-of-the-above solution on baseload . . . Gas plants with carbon capture, next-gen geothermal and all forms of nuclear. We’ll need all of these to be able to meet the task at hand.”

The world as we know it may depend on it, and that is clearly not artificial hype.