Michael De Chiara, senior partner at Zetlin & De Chiara looks at the AI revolution.
We are all aware of the newest revolution in computing known as artificial intelligence or Ai. We have heard stories of the virtually unlimited potential this computing revolution may bring to mankind, a sapient machine capable of human reasoning only quicker and with better memory that has a database that is always expanding. However, this burgeoning Ai revolution with its seeming unlimited potential is completely dependent on having the power to supply the massive data centers housing the vast array of computer processing chips, hardware, software and storage capability which is what really constitutes Ai.
So the Ai story is really a story about massive computer data capability and the data centers necessary to generate that data processing capability. Those massive data centers are the computing engines that comprise Ai. And like all engines, the data center computing engines run on power. Specifically, electric power.
So, some fundamental questions for Ai are what are the projected power needs for Ai and what sources of energy will fuel the Ai revolution? Where will this power come from? How will the U.S. compete with other countries, principally China, India, and Europe if we can’t power the massive and complex data centers that are required to provide Ai? The answer to the last question is simply that we must find and/or generate and/or produce whatever power is required. Failure is not an option.
In order for non-engineers to understand the magnitude of the anticipated power requirements for Ai, it is important to understand, in easily comprehensible terms, what electric power is and how much power is required in order to begin to appreciate the challenges and potential solutions to this critical problem.
So let’s start with some basic electric power related terminology. Electric power is described by terms such as watts, kilowatts, megawatts, gigawatts and terawatts. As most engineers know, a watt is the basic unit of electric power output or transmission. What is a watt of power? Without getting too technical, think of a 100 watt electric lightbulb. A 100 watt electric lightbulb uses 100 watts of power per second to stay on at full power.
That means 100 watts of electric energy is being transmitted to and used by that lightbulb every second. Now that we understand what a watt is (a measure of electric energy consumption or transmission per second) we can get a basic feel or understanding of energy demand required to power the Ai revolution by a few simple examples.
Example 1, a solar panel. A typical single solar panel that is found on the roof of a house, when it is operating at maximum efficiency in direct sunlight, generates about 440 watts of electric power per second. That would only be enough power to light four and a half 100 watt electric lightbulbs. Most homes and appliances in homes have requirements based on a kilowatt
(KWH) of electric power. A kilowatt is one thousand watts, a kilowatt hour (KWH) is a thousand watts of power being delivered for an hour, another basic energy measuring unit.
Example 2, a refrigerator. Another simple example of energy usage that we can all relate to is a refrigerator which typically requires 1.4 KWH of power per day. However, when we start talking about houses, then the overall energy needs to grow dramatically.
Example 3, a home. A typical home’s power usage is approximately 30 KWH per day.
Now when we start talking about multiple houses, the energy demand is measured not in kilowatts (thousands of watts), but in megawatts, millions of watts. Megawatt usage per hour is shortened to MWH. Now a typical home consumes about 1.0 megawatts of power every month.
Example 4, a small municipality. So a small municipality with 1,000 homes would require 1,000 megawatts of power per month.
Finally, when we start talking about the electric power of consumption of data centers which are necessary to provide the computing capability that actually makes up Ai or any other complex computer based information processing centers, we are talking in units of one thousand megawatts or a gigawatt (GW) which is a billion watts. The example of 1,000 homes per month using 1,000 watts of power (1 megawatt) becomes 1,000 megawatts per month or one gigawatt per month.
Example 5, the Hoover Dam. In terms we can hopefully relate to, the Hoover Dam produces roughly 2 GWs of power per second, assuming the water level of Lake Mead is at the appropriate level. Now, a “typical” Ai Data Center, assuming such a thing even exists, is projected to consume between 100 to over 500 MW of power per second. So, in terms we can now grasp, the newer larger projected Ai Data Centers, assuming they consume 500 megawatts of power, will each need twenty-five percent of the output of the Hoover Dam to operate continuously on a daily basis. So if we construct ten new Ai data centers, we would need 2.5 new Hoover Dams. Currently, projections for new Ai data centers run in the thousands. So you can see there is a dramatic need for more power. Thousands of new data centers would require thousands of new Hoover Dams.
Projecting the near term energy requirements for Ai, that is for the data centers that generate the computing capability necessary to support Ai, let’s start with the fact, as generally understood, existing data centers currently consume 1-2% of all global electricity and 4% of the total energy consumed in the United States. By 2030, that percentage is projected to grow to 3-5% of all electricity produced globally and 12-16% of all power consumed in the United States. Where will the energy to meet these new enormous energy requirements come from?
Currently, and for the foreseeable future, there are only a few discrete sources of power generation available.
They are: (1) gas and oil power plants with natural gas fired plants the predominate source; (2) nuclear power plants; (3) coal fired power plants; and (4) “Renewable” sources of power which are solar wind, hydro, and geothermals but predominately solar due to the low efficiency and reliability of wind power. In 2024, natural gas supplied roughly, 40-45% of the United States energy needs. Coal provided roughly 15% of U.S. energy needs, Nuclear energy provided roughly 20% of U.S. energy needs, while other renewables provided approximately another 20% of U.S. energy needs, with hydro power accounting for roughly 7%. However, the hydro energy component is essentially exhausted.
In terms of cost, while Renewable energy costs are falling, there are critical issues with their reliability since they depend on wind and sunlight and limited hydro capacity. Over the next decade, while a small portion of the new power demands may be met by Renewables, they will not address the dramatic increase in energy demand.
The obvious and likely solution to the short term(i.e. next decade) need for increased energy demand being driven by the Ai revolution will therefore be met by a combination of fossil fuel, principally natural gas and coal fueled powerplants and new nuclear energy power plants with some smaller contribution from Renewables, primarily solar. A typical natural gas supplied power plant produces roughly 300 to 800 MWs of electric power,
while a type of nuclear power plant produces about the same 300 to 800 MWs of electric power and a large coal power plant produces about 1,000 MW of electric power or a gigawatt of power. So you can see massive amounts of those three types of power sources, as well as reliable Renewable sources, principally solar, will be required to meet the power demands of Ai. Since the projected number of Ai related data centers over the next ten (10) years is projected to be in the thousands, the corresponding number of power plants necessary to support the Ai is likewise projected to be in the thousands. To quantify thousands of gigawatts of energy requirements the conversation now shifts to the term for, thousands of gigawatts of power or terawatts (TW).
In order to understand the magnitude of the terawatt of power, let’s go back to the Hoover Dam. As previously mentioned, since the Hoover Dam produces two (2) gigawatts of power, you need five hundred Hoover Dams to produce one (1) terawatt of power. And if the typical nuclear power plant produces, say 500 megawatts of power, you would need two thousand of those to produce a terawatt of power. As for coal plants, you would need one thousand large coal plants to produce a terawatt of power. This is where conventional technology, specifically the design and construction of both massive data centers and the design and construction of massive sources of a power generation enter the picture, and become the controlling element to the growth of Ai capability in the United States.
One caveat to this analysis is the always present hope that a new technology, or a breakthrough in existing technologies, will create a new, cheaper, more dependable, efficient and less risky energy alternative in the future. However, barring such a breakthrough, we will be designing, building, and operating a dramatic increase in natural gas, coal and nuclear power plants over the next decade. Further, those regions of the United States that recognize this and have the political and economic will to support and encourage this growth will become the Ai centers of the country and enjoy all the economic and social benefits that will naturally follow from that commitment to provide energy to support Ai. You can see that the topography of large areas of the country is now likely to change because of the energy demands of Ai. The construction of thousands of new power plants will dot the landscape across America. Also, the environmental impacts as well as the need for massive construction technology and labor, as well as engineering capability will be pushed beyond all current capabilities. Paradoxically, conventional American engineering and construction sectors are now, and for the foreseeable future will be the key drivers in creating our new Ai revolution.