Another gold rush is underway. Recast as the AI tech boom. Yet, AI is power hungry, a voracious beast consuming vast amounts of electricity. And water [1]. And those gas-fired power plants [2] …
Generative AI data centers are expensive to build, expensive to operate, and tax the electric grid and our limited water resources (like the Colorado River).
Companies that invest in AI see profit. Electric utilities see profit from AI-induced demand for power. Local governments see profit from property taxes paid by data centers. Sort of a greed cycle .. what could go …
This LA Times article reminds us that there’s no “free lunch” – the seductive offer by tech and social media to entice AI-sipping customers.
The “free lunch” in the saying refers to the formerly common practice in American bars of offering a “free lunch” in order to entice drinking customers. – Wiki
• LA Times > “Power demands of AI data centers raise concerns over cost, blackouts” by Melody Petersen (8-31-2024) – Experts warn construction frenzy could delay state’s transition away from fossil fuels.
In Santa Clara — the heart of Silicon Valley — electric rates are rising as the municipal utility spends heavily on transmission lines and other infrastructure to accommodate the voracious power demand from more than 50 data centers, which now consume 60% of the city’s electricity.
While the benefits and risks of AI continue to be debated, one thing is clear: The technology is rapacious for power. Experts warn that the frenzy of data center construction could delay California’s transition away from fossil fuels and raise electric bills for everyone else. The data centers’ insatiable appetite for electricity, they say, also increases the risk of blackouts.
According to the International Energy Agency, a ChatGPT-powered search consumes 10 times the power as a search on Google without AI.
And because those new chips generate so much heat, more power and water is required to keep them cool.
By 2030, data centers could account for as much as 11% of U.S. power demand — up from 3% now, according to analysts at Goldman Sachs.
Notes
[1] This article uses infographics to visualize data center cooling footprints (water and electricity loads, which vary by state) for processing ChatGPT prompts.
Those transactional footprints beg the months of cooling required to first train chatbots. Estimates range from 700,000 to 22,000,000 liters of water to train some AI models.
Water replenishment rates by tech companies are unclear, while they pledge to make AI less thirsty.
References on methodology are included.
• Washington Post > POWER GRAB > “A bottle of water per email: the hidden environmental costs of using AI chatbots” by Pranshu Verma and Shelly Tan (September 18, 2024) – AI bots generate a lot of heat, and keeping their computer servers running exacts a toll.
While the exact burden is nearly impossible to quantify, The Washington Post worked with researchers at the University of California, Riverside to understand how much water and power OpenAI’s ChatGPT, using the GPT-4 language model released in March 2023, consumes to write the average 100-word email.
Even in ideal conditions, data centers are often among the heaviest users of water in the towns where they are located, environmental advocates said. But data centers with electrical cooling systems also are raising concerns by driving up residents’ power bills and taxing the electric grid.
An earlier article uses infographics as well to better understand what goes on in data centers.
• Washington Post > POWER GRAB > “Our digital lives need massive data centers. What goes on inside them?” by Antonio Olivo and William Neff (September 17, 2024) – We toured a facility in Northern Virginia to see how it works and to understand why water use and energy consumption are such a concern.
More than half a million people in Northern Virginia live in a neighborhood that’s less than a mile from a data center. That’s more than 1 in 5 residents.
There are four main types of data centers:
An “enterprise” data center serves the needs of the company that owns it. Think of a corporation that stores in-house information on its own computers.
Larger “hyperscale” data centers, owned by companies such as Amazon or Meta, have computer servers that cater solely to the company’s customers.
“Edge” data centers are smaller buildings in or near major population centers, where digital connectivity becomes almost instantaneous for, say, a passing driverless car.
Equinix is among the world’s largest owners of “colocation” data centers. Those facilities lease space to other businesses that hook up their servers to cables that belong to the data center company.
[2] As in plans for global data centers …
• Data Center Knowledge > “Oracle CloudWorld 2024: Embracing Multi-Cloud, Nuclear Energy, and Other Event Highlights” (September 12, 2024) – Alternative energy sources are crucial to the success of data centers globally.
Nuclear Power in the AI Age
Ellison made one of Oracle’s biggest headlines of the week with his Monday earnings call announcement the company would invest in three small nuclear reactors to power a data center with over 1 GW of AI capacity. Although his keynote address at CloudWorld didn’t provide any additional details, Ellison said that while Oracle now runs over 162 data centers globally, they could soon be operating over 1,000 facilities.
“Alternative energy sources are crucial to their success,” Kevin Sullivan, a principal at Pricewaterhouse Cooper, told Data Center Knowledge. “Nuclear, on a personal level, makes me a little nervous. Other alternative energy sources might be better options. But all options need to be available.”
As reported in this article (and elsewhere), here is yet another portent of the thirst for electric power by the boom in artificial intelligence. A novel idea? More similar deals to come? More move fast and let things break (again), eh.
• Washington Post > “Microsoft deal would reopen Three Mile Island nuclear plant to power AI” by Evan Halper (September 20, 2024) – Supposedly shuttered for good (decommissioned) in 2019, the Pennsylvania plant would come back online by 2028 if approved by regulators.
Pennsylvania’s dormant Three Mile Island nuclear plant [site of the 1979 partial reactor meltdown] would be brought back to life to feed the voracious energy needs of Microsoft under an unprecedented deal announced Friday in which the tech giant would buy 100 percent of its power for 20 years. … the energy equivalent it takes to power 800,000 homes, or 835 megawatts.
The four-year restart plan would cost Constellation [plant owner Constellation Energy] about $1.6 billion, he said, and is dependent on federal [public] subsidies in the form of [undisclosed] tax breaks earmarked for nuclear power in the 2022 Inflation Reduction Act.
Constellation will also need to clear steep regulatory hurdles, including intensive safety inspections from the federal Nuclear Regulatory Commission, which has never before authorized the reopening of a plant. The deal also raises thorny questions about the federal tax breaks, as the energy from the plant would all be produced for a single private company rather than a utility serving entire communities.
“It doesn’t address the core issues that are making the current practice of AI unsustainable by definition,” she [Sasha Luccioni, the top climate executive at sustainable AI start-up Hugging Face] said of the deal. “Instead of monopolizing decommissioned nuclear power plants, we should be focusing on integrating sustainability into AI.”
Like if it’s so great, why the AI promotional blitz? Maybe because companies are spending $billions and $billions; and if no one uses it (like in The Simpsons episode S7 E6, if no one looks at the giant advertising statues come to life) … Maybe because of “duels for supremacy” (remember “Tastes Great. Less Filling”).
So, star-studded commercials … in the Olympics, helping-your-kid scenarios, corporate pitches, smartphone gotta-have’s, …
• Washington Post > “Here’s how much tech companies are spending to tell you AI is amazing” by Shira Ovide (August 13, 2024) – If AI is supposed to be so amazing, why are tech companies trying so hard to persuade you it’s great?
Notes
[1] The Simpsons episode S7 E6 “Treehouse of Horror VI” segment “Attack of the 50-Foot Eyesores” – when advertising statues come to life and terrorize the town, there’s only one remedy (as Lisa discovers), “Just don’t look.” (This episode first aired on the Fox network in the United States on October 29, 1995.)
Here’s an article illustrating “the ‘volatility’ of data center development.”
Who will make the required long-term financial commitments? And do data centers really create a lot of local jobs? [1]
• Washington Post > POWER GRAB > “Tech giants fight plan to make them pay more for electric grid upgrades” by Caroline O’Donovan (September 13, 2024) – If the power company spends big on new infrastructure but the power demand it was built to serve [the data center boom] doesn’t materialize, other customers — including business and residential payers — will be stuck with the bill, the utility said.
Notes
[1] This article discusses data centers and jobs. It has lots of photos and some video. Some stories of workers’ paths to certification as data center technicians. HVAC technicians. The scope of demand for data center workers.
• Washington Post > POWER GRAB > “Data centers are everywhere. What it’s like to work in one.” by Danielle Abril (September 17, 2024) – These seemingly invisible workers are critical to keeping our connected world humming.
Nuclear Power in the AI Age – coming to an area near you?
10-16-2024 Small nuclear reactor projects
See also:
• BBC > News > “Google turns to nuclear to power AI data centres” by João da Silva, Business reporter (Oct 14, 2024) – The agreement with Kairos Power aims to bring the first nuclear reactor online by 2030.