Q&A: Can a centuries-old technology like hydropower meet AI’s rising energy demands?

Surging demand for the information and processing energy of synthetic intelligence is placing a hidden pressure on U.S. electrical grids.

Generative AI techniques like ChatGPT can every devour as a lot electrical energy in a single day as 180,000 typical U.S. households. The power required to coach these giant language AI fashions is much more staggering: Coaching GPT-4 required over 50 gigawatt-hours, about 0.02% of California’s annual electrical energy output and 50 instances greater than what was used for GPT-3.

As AI adoption grows, California’s largest utility, PG&E, forecasts demand may double by 2040.

Shon Hiatt, an affiliate professor on the USC Marshall College of Enterprise and director of the Enterprise of Power Transition initiative, sees hydropower as a promising resolution. This centuries-old, clear and renewable energy source is underutilized in the USA, says Hiatt, however may play an important position in powering the AI revolution—and will assist ease the stress on nationwide electrical grids.

Throughout Green Week, USC News spoke with Hiatt to discover how hydropower may assist meet AI’s rising power wants and help a extra sustainable future.

What challenges does the AI revolution current by way of power consumption?

U.S. electrical energy demand is projected to surge over the following 5 years, with development charges doubling from final 12 months’s estimates. This sudden enhance is pushed by three predominant elements: the rise of AI data centers, federally sponsored manufacturing crops and widespread electric vehicle adoption.

As a result of knowledge facilities require fixed, dependable energy, renewable sources like wind and photo voltaic cannot present with out large battery backup. In consequence, utilities might want to rely extra on pure fuel, coal and nuclear plants to satisfy the growing demand.

Trying forward, electrical energy demand for knowledge facilities is projected to extend by 13%-15% yearly by way of 2030. There may be not sufficient deliberate electrical energy era improvement to accommodate projected AI knowledge middle development.

That is driving Big Tech firms into the power sector, as seen in Amazon’s current $650 million deal to accumulate a Pennsylvania knowledge middle powered by an on-site 2.5 gigawatt nuclear plant. Because the demand for electrical energy continues to develop, it is clear that the power panorama is present process a major transformation.

How can hydropower assist tackle these challenges?

A comparatively fast resolution is repowering present hydropower crops and placing generators on present reservoirs.

Hydropower can present baseload power in contrast to wind and photo voltaic, that are intermittent because of clouds, climate, and so on.

We’re not constructing any new reservoirs (dams) on this nation. That doesn’t imply we can’t enhance hydropower from present reservoirs, nonetheless. The U.S. Division of Power estimates that as much as 10 gigawatts of power will be created by upgrading present powered services. That is one thing that may be accomplished inside months with capital.

Furthermore, lower than 3% of the greater than 90,000 reservoirs in the USA produce energy. Putting in generators and turbines on these reservoirs may present an extra 12 gigawatts of energy. Placing generators on present reservoirs may also be accomplished in a well timed method—in some states, a matter of months.

What are the benefits of investing in hydropower over different renewable power sources to help AI’s rising demand?

Each power supply has a tradeoff. Photo voltaic scales linearly and has thus the biggest land footprint of present energy sources. Wind can have an effect on birds and sea life. Nuclear and mixed cycle pure fuel have the smallest footprint for power output.

Run-of-the river hydropower has a lot decrease environmental affect than reservoir hydropower because it requires no reservoirs. Run of the river diverts some water from a watershed to energy a turbine downstream whereas the remainder of the water flows down the mountain.

The U.S. Division of Power estimates that the U.S. has 65 gigawatts of unexploited hydropower power that may come from ecologically pleasant run-of-the-river services. Nonetheless, improvement of run-of-the-river services can take years to develop because of authorities licensing and allowing limitations.

Frankly, it’s probably that the demand for electrical energy for AI data centers over the following 5 years can be met by new mixed cycle natural gas services. They’re fast to place up, require a small footprint, and the U.S. at the moment has a big abundance of low cost fuel. Small modular nuclear reactors are one other potential resolution. Nonetheless, they don’t seem to be prone to come on-line till 2030 on the earliest.