Our economy is both fueled and limited by energy sources. Before the Industrial Revolution, human strength and ability were supplemented by animals. We still retain a measure of power in terms of the ability of a horse. One horsepower (hp) is equivalent to about 750 W, which is about 5x what one person can deliver. 

The Industrial Revolution, which began in 1760, was distinguished by the introduction of energy sources that offered even more power than a horse. Initially, power was supplied using wind and water, until the switch was made to steam, wood, coal, and eventually, oil. 

However, the transfer of energy from the source to the action was through a mechanical linkage. This meant that the application was always in proximity to the engine. Mills were located next to water sources. Trains carried their fuel and steam engines with them. 

Electrification introduced a new era when the conversion of power from its raw source into an easily transportable form allowed the use of the energy remotely. Thomas Edison’s first DC power station on Pearl Street in New York City came online in 1882, delivering 100 kW. By 1896, Westinghouse’s Niagara Falls AC power station came online, delivering 37 MW of power. The need for electric power was driven by the light bulb and the streetcar. 

Fast forward to today and the power consumption of the electronics industry. What is fueling this growth is the introduction of artificial intelligence (AI)/machine learning processing in data centers. 

According to a report released in 2024 by Newmark¹, 5% of the data center applications in 2021 were for AI applications. By 2025, 30% of the applications in data centers will be for AI. Capex spending by the companies creating large data centers has a compound annual growth rate of 30%.

AI processing is notoriously power-hungry. The same Newmark report says the typical power consumption of a rack in a data center has been about 12 kW, while an AI processor rack consumes as much as 50 kW. 

For example, the NVIDIA H100 GPU has a peak power consumption of 700 W. This is 1 hp of power consumption for one AI processor module. If the rail voltage rail is 1.8 V, this is a current draw of 400 A. 

This is roughly equivalent to what an internal combustion car’s starter motor draws, which is why battery jumper cables are so thick. A Tesla electric motor only draws 200 A when it is accelerating. Other processor chips are in the same category, drawing 500 to 1000 A, but at lower rail voltages. 

The estimated sales volume of the H100 in 2024 is about 2 million units, or about 3.5 million units cumulatively since it was introduced. When  all of these processors are operating, the processors alone will consume 2.5 GW of power, not including the rest of the electronics and their cooling systems. A U.S. household uses about 1 kW of power on average. This is equivalent to the power consumption of 2.5 million households, or the entirety of Los Angeles. 

According to the International Energy Agency, in 2022, data centers worldwide consumed a total of 460 Twatt-hours of energy, accounting for 2% of global electricity usage.² This is, on average, 40 GW of electric power usage.  In the U.S., data centers accounted for about 6% of total power consumption. In Ireland, data centers accounted for 17% of the nation’s power usage in 2022. Power usage in data centers is expected to almost double by 2025. This means data centers could account for nearly 10% of the U.S. power consumption within the next few years.

According to the Newmark study, “Energy is the number one challenge for the data center market.” It is no wonder that Microsoft, Google, and Amazon have struck deals with operators and developers of nuclear power plants, according to the New York Times. Microsoft was reported to have made a deal to revive the Three Mile Island nuclear plant in Pennsylvania. Amazon and Google have stated that they are considering new, modular nuclear plants, located adjacent to their data centers. Okla, a developer of modular nuclear power plants, signed a 20-year purchase agreement with Wyoming Hyperscale to use Okla’s Aurora powerhouse to power new data center campuses. 

Just as important to adding to the power grid is reducing the demand for power by more efficient chip design and power delivery. For a 1000 A processor operating at 1 V, the resistive load created by the chip is 1 mΩ. For the IR losses to be less than 10% of the power consumption, the DC resistance in the power distribution path needs to be less than 100 µΩ. This means more and thicker copper layers and optimal design of the via field from the copper layers to the device being powered. This means measurement capability in the µΩ range to test the final designs. 

The Industrial Revolution began with harnessing power sources that amplified the ability of a person. With today’s generation of AI and network processors, one single chip requires 1 hp to operate. The need for speed will be the driving force behind energy production and efforts to make increasingly efficient use of available energy.

REFERENCES 

1. “2023 U.S. Data Center Market Overview & Market Clusters,” Newmark, January 2024, https://www.nmrk.com/insights/market-report/2023-u-s-data-center-market-overview-market-clusters. 
2. “Electricity 2024, Analysis and Forecast to 2026,” International Energy Agency, January 2024, https://iea.blob.core.windows.net/assets/6b2fd954-2017-408e-bf08-952fdd62118a/Electricity2024-Analysisandforecastto2026.pdf.