FROM THE HILL

The following is the written testimony of Mark Mills, Executive Director of the National Center for Energy Analytics, presented to the U.S. Senate Committee on Energy and Natural Resources. The Senate hearing on May 21st is worth watching.

**To give a brief update on my running progress (refer to previous posts), I've adopted Hal Higdon's training program following several recommendations. This program has introduced a new perspective on training, particularly the extended weekend runs. The summer heat poses distinct challenges. Recently, due to a packed schedule, I managed to fit in my long run at midday. With the intense sun and a strong wind offering little respite, maintaining a consistent heart rate was challenging. This run served as a reminder of the critical role of proper nutrition and hydration. It also reinforced the significance of mental strength in running. Keeping a positive mindset is a form of energy, and adverse running conditions present an excellent chance to develop mental resilience.

Testimony of Mark P. Mills

Executive Director, National Center for Energy Analytics

Distinguished Senior Fellow, Texas Public Policy Foundation

Before the

U.S. Senate Committee on Energy and Natural Resources

Regarding “Opportunities, Risks, and Challenges Associated with Growth in Demand for Electric Power in the United States.”

May 21, 2024

<> SUMMARY: ORAL REMARKS <>

Good morning. Thank you, Chairman Manchin and Ranking Member Barrasso, for the opportunity to testify.

Senators, if I may make a personal observation as a long-time observer of, and episodic witness before this great Committee, it is music to my ears to hear in the opening remarks and testimony from other witnesses so much bipartisan convergence on this critical issue. I began my career as a young physicist working in a semiconductor factory; this issue and the restoration of America’s manufacturing is very close to home.

We are, as you observed in your opening remarks, at one of those rare pivots in history. What we’re witnessing is the emergence of a significant, and ostensibly unanticipated, new vector for electricity demand. As you know, I’m referring specifically to the emergence of useful artificial intelligence (AI), especially combined with the next phase of expansion for the cloud computing infrastructure.

While there is both debate and outright guesswork about future specific uses and abuses for AI, whether for civilian or security purposes, no one doubts the AI boom is here.

Utility planners and energy pundits have rediscovered a basic truth: The arrival of new ways to boost the economy illustrates a long-standing correlation, a veritable iron-law, that links growth, with rising energy use, especially electricity use.

We have a good idea why we’re now seeing such surprising leaps in near-term forecasts for electricity demand. Put in monetary instead of physics or engineering terms, every $1 billion spent on datacenters leads to over $60 million a year in electricity purchase. And last year, before the AI acceleration kicked into high gear, capital spending in the U.S. on datacenters was running at about $100 billion a year.

Now the addition of AI-enabled hardware is accelerating both the rate at which datacenters are being ordered and the energy use per datacenter. A $1 billion spent on new AI-infused datacenter likely drives some $200 million in energy purchases a year. For context, a $1 billion spent on new EVs, or $1 billion worth of new chip factories, generates only about $20 million in annual electricity demand.

That AI has a voracious energy appetite is not news to the technical community. Myriad studies have pointed out that the so-called training of an AI tool—equivalent to building not flying an aircraft—can use as much electricity as a Tesla driven anywhere from 300 thousand to four million miles, depending on the application. And once built, AI consumes energy to operate, as much as ten-fold greater than the training. And there are, for all practical purposes, nearly unlimited numbers of potential applications for training and using AI.

Of course, AI will become far more energy efficient. The latest AI chips are already 100-fold better than a half-dozen years ago, and current trends point to another 100-fold gain by 2030. But efficiency won’t solve the so-called problem of rising electricity demands. It will do the opposite, just as it did in the rise of the first cloud era.

Improving energy efficiency lowers costs, which is what makes possible the proliferation of the benefits of a technology or service, a reality especially true in digital domains. Operating at the computing efficiency circa 1980, one smartphone today would use as much electricity as a skyscraper, and one datacenter would use as much power as the entire U.S. grid. But because of efficiency gains, the world has billions of smartphones and thousands of datacenters. Consequently, today’s global cloud infrastructure already rivals the energy used by global aviation, and that’s pre-AI. We can count on the pattern of the recent past happening again with AI.

Electricity demand coming from the expansion of the cloud will continue to overshadow the combined impact of new EVs and chip factories.

But whether the AI expansion is in fact fully realized in the U.S. will depend on the extent to which the electricity is available, when it’s needed, and at tolerable prices. Given the scales of electricity demand from the cloud and AI, especially when added to the demands from reshoring manufacturing and promoting EVs, the nation’s electricity producers will need full access to all options. Electric power planning now needs to focus on additions to, not transitions away from existing electricity production.

The U.S. is the world’s cloud leader today, much as the U.S. was the dominant aluminum producer 25 years ago. But producing $1 billion of aluminum requires uses some $400 million in electricity which, today, happens mainly on low-cost coal-fired grids in China, which has become the world’s dominant supplier while the U.S. now has a single digit percentage share of world production. Aluminum and silicon are different materials, but the energy implications of the two domains offer, I believe, a relevant lesson. Thank you.