Data center power consumption surges, artificial intelligence threatens global power system

Editor's note: Earlier this year, Musk predicted that there would be power shortages in the coming years to meet the needs of artificial intelligence and electric vehicles. He warned that the simultaneous development of these technologies will put pressure on the existing power infrastructure, and by 2025, the power supply may not be enough to power the increasing number of artificial intelligence chips.

The New Yorker has also reported that it is estimated that OpenAI's ChatGPT consumes more than 500,000 kilowatt-hours of electricity every day to respond to the requests of about 200 million users every day. That's more than 17,000 times the average daily energy use of U.S. households.

Bloomberg just wrote an article saying that in many regions around the world, the power demand for artificial intelligence data centers has surged, and existing local power facilities can no longer meet their power needs. "Artificial intelligence has caused serious damage to the global power system."

The following is the full text translation of the Bloomberg report, compiled and translated by Ai Faner.

Original link: https://www.bloomberg.com/graphics/2024-ai-data-centers-power-grids/
Original author: Josh Saul, Leonardo Nicoletti, Saritha Rai, Dina Bass, Ian King and Jennifer Duggan

Like much of northern Virginia, Loudoun County was once known for its horse ranches and American Civil War sites. But over the past 15 years, many of the community's fields and forests have been cleared to build the data centers that form the backbone of our digital lives.

Now, the rise of artificial intelligence is driving demand for larger data centers, further changing the landscape and putting huge pressure on the region's energy grid.

On a sunny afternoon this spring, the newest data center was nearing completion. When completed, the 200,000-square-foot building will consume as much energy as the daily electricity consumption of 30,000 U.S. homes.

But first, it needs to get enough power…

▲ Databank data center in Virginia. Image source: Black Ridge

Data center provider DataBank, which owns the Virginia data center, is in desperate need of energy. An unnamed "big tech" customer leased an entire data center and was so eager to use the data center to obtain computing resources for artificial intelligence applications that it installed an installation in the building before DataBank planned to power it. Ok server.

"That's the problem with artificial intelligence. They need a lot of power, and once you get it, they want it immediately," said James Mathes, who manages part of DataBank's facilities. "Right now, artificial intelligence is like a blank check."

An almost overnight surge in power demand from data centers has outstripped existing power supplies in many parts of the world, according to interviews with data center operators, energy providers and technology executives. This dynamic has resulted in years-long wait times for businesses to connect to the grid, and growing concerns about power outages and rising prices for residents living in the most data center-dense markets.

▲ Inside Apple data center. Image source: Apple

Silicon Valley’s drive to develop artificial intelligence at all costs has led to a dramatic increase in electricity demand, which also threatens the entire country’s energy transition plan and the clean energy goals of trillions of technology companies. In some countries, including Saudi Arabia, Ireland and Malaysia, all the data centers they plan to build will require more energy to operate at full capacity than existing renewable energy supplies, according to a Bloomberg analysis of the latest data.

Sweden's data center power demand could double by the end of this decade and then double again by 2040, according to official estimates. In the UK, AI’s electricity consumption is expected to increase by 500% over the next decade. In the United States, data center electricity consumption is expected to increase from 3% of all electricity in 2022 to 8% in 2030. Goldman Sachs describes this as "power growth unprecedented in a generation."

There are more than 7,000 data centers built or in various stages of construction around the world, compared with only 3,600 in 2015.

If these data centers continue to operate, the total annual electricity consumption can reach 508 terawatt hours. This is greater than the total electricity generation in Italy or Australia in an entire year.

By 2034, global data center energy consumption is expected to reach 1,580 terawatt hours, equivalent to the entire energy consumption of India.

▲ Inside a Google data center. Image source: Google

These are just estimates, and there is still a lot of uncertainty about how the current AI craze will develop. There are also discrepancies between data center developers' forecasts of power demand and actual power generation.

While tech companies are quick to point out that even as data centers continue to expand, they account for less than 2% of global energy use, an April report from Goldman Sachs estimated that this number will rise by the end of the decade. The number may rise to 4%. Given that overall electricity demand has been nearly flat for years and even falling in some areas, any one percentage point increase has a huge impact.

John Ketchum, CEO of NextEra Energy in the United States, said that over the next two decades, U.S. electricity demand is expected to grow by 40%, while in the past two decades, U.S. electricity demand has only increased 9%. Data centers are the biggest reason for the surge in demand, Ketchum said, with electrification and manufacturing also being other factors. When asked why data centers were suddenly consuming so much power, his answer was straightforward: "Because of artificial intelligence." He cited the energy required to train models and the ability of artificial intelligence to draw conclusions from data it has never seen. Reasoning process. "This is 10 to 15 times the power consumption."

▲ The wind farm of NextEra Energy Company in the United States. Image source: NextEra Energy

Amazon, the largest cloud service provider in the United States, Microsoft and Alphabet's Google have all announced goals to run their data centers entirely on green energy – Amazon next year, Google and Microsoft by 2030. All three companies said they are looking into technological ways to reduce electricity use or balance grid demand more efficiently. This could include squeezing more efficiency from chips and servers, arranging equipment in better ways to reduce cooling needs, and shifting loads to different areas based on the availability of energy, especially green energy.

But some tech leaders, such as OpenAI CEO Sam Altman, say adapting to this new situation will require an energy breakthrough — likely from nuclear power. Microsoft and Amazon are also betting on nuclear power, with Amazon recently buying a nuclear-powered data center in Pennsylvania and signaling its willingness to buy more. For now, the path forward remains unclear. Microsoft recently acknowledged that its push for artificial intelligence is jeopardizing its long-held goal of becoming carbon negative by 2030.

“We need terawatts of traditional green energy, whether it’s wind or solar, and it’s going to take a global effort,” said Amanda Peterson Corio, Google’s director of global data center energy. The scope of discussion is not limited to artificial intelligence needs. One terawatt is equivalent to the power generation of approximately 1,000 nuclear power plants. "Of course we want to decarbonize ourselves, but if we just decarbonize ourselves and not the entire grid, what's the point?"

▲ Apple has launched a series of carbon-neutral products. Image source: Apple

In today's data centers, you're likely to find thousands of Nvidia H100 graphics cards — the engine of the generative AI boom — each consuming up to 700 watts, nearly as much power as a 60-inch flat-screen TV. eight times. Data centers built to train AI models require even more power. For example, Microsoft has tens of thousands of Nvidia graphics cards connected in a data center used to develop OpenAI technology. These graphics cards are powered by networking and other types of chips, and coupled with the cooling equipment used in data centers, the power consumption is even greater.

Conventional wisdom in Silicon Valley holds that the amount of energy needed will only increase. Nvidia's latest graphics card, the B100, consumes almost twice as much power as the H100. Nvidia believes that companies will be able to do more with fewer chips, but Ian Buck, head of Nvidia's accelerated computing group, acknowledged that AI deployment is likely to increase. "People like to fill up their data centers," he said.

▲ NVIDIA H100 graphics card. Image source: NVIDIA

The development of artificial intelligence is also changing with each passing day, and people are keen to develop larger artificial intelligence models. The Microsoft supercomputer built in 2020 trained OpenAI's GPT-3, using 10,000 of the latest artificial intelligence chips at the time. According to a May speech by Microsoft Azure Chief Technology Officer Mark Russinovich, Microsoft's supercomputers used 14,400 top-of-the-line Nvidia H100 graphics cards in November 2023, and the next supercomputer is being designed Will be 30 times more powerful than now.

At the same time, Nvidia CEO Jensen Huang said that many countries will push to establish their own "independent property" artificial intelligence systems to stay competitive and process data locally. The global battle for AI supremacy is likely to hinge on which countries have enough data centers and power to support the technology. If so, Loudoun County is a microcosm of what's to come in the rest of the world — and the challenges other countries will face to stay competitive.

Data center capital of the world

Over the past five years, Dominion Energy Inc., which serves Loudoun County and is known as the "data center lane," has connected 94 data centers that consume a total of about four gigawatts of electricity. Now the company is fielding requests for data center campuses that will consume several gigawatts of electricity, enough to power hundreds of thousands of homes, with two or three data centers that combined could consume the equivalent of 2019 The total power consumption of all facilities connected during the year.

The surge in demand has created a backlog. Data center developers now have to wait longer to connect their projects to the grid. "It could be as fast as two years or as slow as four years, depending on what needs to be built," Dominion President Edward Baine said in an interview.

▲ Dominion Electric Company. Image source: FFXnow

Dominion Power is working to build the infrastructure to support this plan. New power lines suspended from giant metal towers and laid along roads and creeks supply power to these towering, windowless data centers. The company is building a large new wind farm and many solar farms near the coast, but coal and natural gas plants could also stay online longer.

In late 2022, Dominion Power submitted a previously unreported letter to regulators asking for permission to build new substations and transmission lines to meet "unprecedented" load growth. In the letter, Dominion Power said it experienced 18 load relief warnings that spring. These warnings are issued when the grid operator informs the company that load shedding may be required. "Load shedding" is a technical term for a controlled interruption of power to customers, which may include rolling blackouts.
"This was well beyond normal, safe operating procedures," Dominion Power told regulators.

▲ A DataBank data center in Ashburn, Virginia. Image source: Bloomberg

Virginia is not alone in its efforts to meet electricity demand. David Bloom, chairman of UK data center operator Kao Data, said that in west London, historically seen as a data center hub, new facilities will have to wait until 2030 to be connected to the grid. “We are now being pushed into new locations,” he said. Likewise, in southern Sweden, where the renewables market is strong, demand for grid connections is so great that companies may have to wait years. "We have a queue and you need to get your tickets," said Peter Hjalmar, regional manager for southern Sweden at Germany's E.ON SE power utility, according to a recent analysis by Morgan Stanley Reports show that many new artificial intelligence data centers across the United States are expected to consume 100 megawatts of electricity each, leading analysts to question: "How will all proposed data center projects be powered in a timely manner?" According to NextEra , with demand so high, major tech companies are bidding for data center space that can be powered at all times.

Ali Farhadi, CEO of the Allen Institute for AI, said expected data center growth may also encounter limitations in the transmission lines carrying power. "I don't think we can move that much power around the world, let alone generate it," he said.

Energy analysts say data centers will become more efficient over time, but they will also get larger. The average size of data center facilities worldwide is now 412,000 square feet, a nearly five-fold increase from 2010, according to market intelligence firm DC Byte.

Experts say a surge in data center demand, coupled with significant investments in new substations, transmission lines and other infrastructure by power companies like Dominion, also increases the likelihood that customers will see higher energy prices. The cost of some upgrades is typically distributed to electricity customers throughout the region, appearing as a line item on each person's monthly utility bill.

Goldman Sachs estimates that U.S. utilities will have to invest approximately $50 billion in new generation capacity to support data centers. "This will increase both wholesale energy prices and retail energy prices," said Patrick Finn, a power market analyst at energy consultancy Wood Mackenzie.

A Dominion Power representative said costs, including grid modifications, will be allocated among each category of customers, from residential to industrial, based on the actual cost of serving each type of customer. As a result, residential customers' share of transmission costs has declined in recent years, while data center shares have increased.

▲ Power transmission lines. Image source: Unsplash

There are also early signs of some rate increases in Ireland, another heavily saturated market. Ireland's mild climate has attracted numerous data centers from Microsoft, Amazon and other companies, and these facilities are expected to consume a third of Ireland's energy by 2026, up from just 18% in 2022.

Wholesale electricity prices in Ireland have been on average a third higher than elsewhere in Europe this year. Sarah Nolan, senior modeler at Cornwall Insight, said other factors, including Ireland's geography, also played a role, but growing demand for data centers also played a role. First, just before the artificial intelligence boom started, Ireland also took strong measures to restrict the construction of data centers.

To manage energy constraints, Ireland's state-owned electricity operator implemented a moratorium in Dublin in early 2022 and set out conditions for connecting new data centers to the grid, including prioritizing those that generate their own electricity. Donal Travers, director of technology, consumer and business services at IDA Ireland, which attracts foreign direct investment, said Ireland's state-owned electricity operator had a "comprehensive" plan to build the grid. . But he said he expected restrictions on new large grid connections to "probably last until around 2028."

Meanwhile, other regions are eager to open their doors.

▲ On March 20, 2024, a new data center is under construction in Gainesville, Virginia. Image source: Bloomberg

next virginia

When Rangu Salgame looks to Malaysia, he sees the next Virginia "taking shape". Johor, the southernmost state in Peninsular Malaysia, has a policy to expedite data center approvals. On top of that, it's also close to Singapore, long a hub for data centers, which has put a moratorium on new construction for several years to manage energy growth on the tiny island.

Once a sleepy fishing village, the outskirts of Johor Bahru are now vast construction sites. Microsoft and Amazon are investing in the region, as is Salgeim's company, Princeton Digital Group (PDG). At Sedenak Tech Park, about 40 miles south of downtown Johor Bahru, towering cranes dot the sky. PDG's new 150-megawatt data center occupies a corner of the campus opposite similar facilities from other providers.

"We completed the process from groundbreaking to production within 12 months," said Salgame, who expects to have 300 megawatts of capacity in Johor within two years. "Not all parts of the world can produce at this speed and scale."

Along with Malaysia, Texas has become one of the fastest-growing data center markets in the United States, thanks in part to its independence and the promise of a deregulated power grid that reduces wait times. According to DC Byte, Microsoft Vice President of Energy Bobby Hollis said that the data center in Texas can be connected to the grid within the one to two years required to build the data center.

Hollis said Texas is rich in solar energy resources and has some of the best wind energy resources in the world in the state's Panhandle Plains region. But hot weather and tight water supplies are prompting Microsoft and Google to try to wean their data center cooling equipment off water. However, the alternative method requires more energy, about 5% more energy on average, according to Microsoft.

While power appears to be abundant in Texas, it has its limitations. Hollis said supplies of solar panels and other equipment needed for clean energy are starting to tighten. The Electric Reliability Council of Texas also recently warned that it was underestimating demand in the San Antonio area, where Microsoft's massive data center campus is located, which could cause cascading blackouts across the state in the future.

▲ Electric Reliability Council of Texas. Image source: Wikipedia

Opposition to data centers is growing in Virginia. At a Prince William County Board of Supervisors meeting in March, frustrated residents spoke out against a zoning change that would have allowed data centers about twice as tall to be built on certain land. One woman told officials a data center was turning her quiet neighborhood into a "dystopian nightmare."

Rachel Ellis, a 48-year-old housewife who spoke remotely, said the change would mean more demands on an already strained power grid. "Continuing to approve data center construction without understanding where the power is going to come from is a reckless act of management," she said.

The Board of Supervisors voted after hearing from more than a dozen people who opposed the zoning change. Larger data centers are approved.

# Welcome to follow the official WeChat public account of Aifaner: Aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo