Photo credit: www.cnbc.com
The rapid expansion of artificial intelligence (AI) technology is driving an unprecedented increase in the construction of data centers, leading to a sharp rise in demand for electrical power to support and cool these vital facilities. This surge has raised significant questions about the United States’ capacity to provide sufficient electricity to support AI’s growth, as well as concerns about the ability of the aging electrical grid to manage the heightened load.
“If we don’t start considering this power challenge in a new way, we’ll fall short of realizing our aspirations in AI,” stated Dipti Vachani, the automotive sector lead at Arm. The company’s energy-efficient processors are favored by major tech companies such as Google, Microsoft, Oracle, and Amazon because they can cut energy consumption in data centers by up to 15%.
Nvidia’s latest offerings, including the Grace Blackwell chip, utilize Arm-based CPUs designed to perform generative AI tasks using 25 times less energy than previous models. “The approach of conserving energy necessitates a completely different design philosophy compared to simply maximizing performance,” Vachani added.
The concept of increasing computational efficiency—often summarized as “more work per watt”—is a promising solution to the challenges posed by AI’s growing energy demands. However, experts warn that such measures alone are not sufficient.
For instance, a single query through ChatGPT consumes nearly ten times the energy of a typical Google search, according to a Goldman Sachs report. Generating one image using AI technology can use as much energy as charging a smartphone.
This dilemma is not a new one; as early as 2019, it was estimated that training certain large language models could produce carbon emissions equivalent to the lifetime output of five gasoline-powered vehicles.
According to Google’s latest environmental report, greenhouse gas emissions associated with its operations surged nearly 50% from 2019 to 2023, partly due to the energy consumption of data centers. Microsoft also reported almost a 30% increase in emissions from 2020 to 2024, closely linked to its data center activities.
In Kansas City, high energy demands from Meta’s AI-focused data center have caused plans to retire a coal-fired plant to be reconsidered, highlighting the urgent need for reliable power sources.
Increasing Power Demands
Globally, there are more than 8,000 data centers, with the majority located in the United States. The Boston Consulting Group projects that demand for data centers will increase by 15% to 20% annually through 2030. By that year, data centers are expected to consume about 16% of all U.S. electricity, a substantial increase from just 2.5% prior to the launch of OpenAI’s ChatGPT in 2022. This surge in demand is comparable to the annual power consumption of approximately two-thirds of American households.
During a CNBC visit to a Silicon Valley data center, experts discussed the industry’s strategies to accommodate this rapid demand growth and secure adequate power.
“We anticipate that the energy demand from AI-specific applications will be on par with or potentially exceed the historical needs associated with cloud computing,” noted Jeff Tench, executive vice president of North America and Asia-Pacific at Vantage Data Centers. Many tech giants rely on businesses like Vantage to house their servers, which can require significant amounts of power. Vantage centers are equipped to use upwards of 64 megawatts of power—enough to power tens of thousands of homes.
According to Tench, many of these power allocations are secured by single customers who may occupy the entire data center space. As AI applications expand, these power requirements could escalate dramatically, theoretically reaching hundreds of megawatts.
Santa Clara, California, where the CNBC team visited Vantage, has historically been a hub for data centers due to its proximity to tech giants in need of resources. However, Tench indicated a “slowdown” in the market due to a lack of available power from local utilities. In response, Vantage is constructing new campuses in Ohio, Texas, and Georgia.
“The industry is actively searching for locations with nearby renewable energy sources, such as wind or solar,” Tench explained. He noted that some facilities are transitioning from coal-fired power generation to natural gas, while others are exploring opportunities to utilize energy from nuclear plants.
Vantage is also expanding a campus near Phoenix, Arizona, that will increase its capacity to 176 megawatts.
Resilience of the Power Grid
The aging electrical grid frequently struggles to manage power loads, even in regions where generation capacity is adequate. A major challenge is transporting energy from production sites to consumption points, with extensive upgrades to transmission lines being a costly and protracted solution.
“This process can incur massive costs that ultimately get passed on to utility customers,” explained Shaolei Ren, an associate professor at the University of California, Riverside.
A recent $5.2 billion project aimed at enhancing power transmission in Virginia—dubbed “data center alley”—met with opposition from local consumers wary of rising utility bills.
Another promising strategy involves deploying predictive software to address vulnerabilities in the grid, particularly at the transformer level. “Every unit of electricity generated must pass through a transformer,” noted VIE Technologies CEO Rahul Chaturvedi, who highlighted that the average transformer in the U.S. is about 38 years old, making them a common source of outages. Unfortunately, replacing them can be an expensive and slow process. VIE Technologies has developed small sensors that can be attached to transformers to anticipate failures and optimize load distribution.
Chaturvedi reported that his company’s business has tripled since the introduction of ChatGPT and is expected to double or triple again next year.
Cooling Solutions for Data Centers
To maintain optimal operating conditions for generative AI, data centers are anticipated to require between 4.2 billion and 6.6 billion cubic meters of water by 2027 for cooling purposes, as detailed in Ren’s research. This volume eclipses the total annual water withdrawal of several countries, including half of the U.K.
“There is significant concern regarding AI’s energy consumption; however, we could tackle this if we embrace nuclear energy more decisively,” argued Tom Ferguson, managing partner at Burnt Island Ventures. He highlighted that water resources pose the most pressing limitation for future AI developments. Ren’s research indicates that processing just 10 to 50 queries on ChatGPT can consume as much water as a standard 16-ounce bottle.
While traditional cooling methods typically rely on water, the Vantage data center in Santa Clara employs large air-conditioning units that do not require water intake. Other innovations include liquid cooling solutions directly applied to chips.
“For many facilities, implementing these methods necessitates extensive retrofitting. However, at Vantage, we integrated a design six years ago that allows us to utilize cold water loops on the data hall floor,” Tench stated.
In response to growing energy demands, companies like Apple, Samsung, and Qualcomm are promoting on-device AI solutions that aim to reduce the reliance on power-intensive cloud computing resources.
Ultimately, Tench emphasized that the achievable levels of AI deployment will correspond closely with the energy available to support data centers, which may be less than current aspirations. Nonetheless, he noted that numerous stakeholders are actively working to address the existing limitations on power supply.
Source
www.cnbc.com