Jul

15

Only some people agree, but the power industry believes there may be a demand-supply mismatch from AI data centers. Here are some summary views - from the American Nuclear Society's Nuclear Newswire (April 2024):

Major tech companies see artificial intelligence (AI) as something that will transform their industry, and there is a race to be first. When they look for clean, dependable power 24/7, nuclear clearly stands out as a good match. Constellation [the nation's largest nuclear utility] summarized it best in its recent forecast:
• AI and data center growth will drive power demand.
• Major tech companies are expected to invest $1 trillion in data centers over the next five years.
• In the next five years, consumers and businesses will generate twice as much data as all the data created over the past 10 years.
• AI data center racks could require seven times more power than traditional data center racks.
• Between now and 2030, domestic data center electricity consumption is expected to grow anywhere from 6.5 percent to 7.5 percent (335 terawatt-hours to 390 terawatt-hours).
• In its report, Data Centers 2024 Global Outlook, global real estate services company JLL has said that "AI is driving extreme scale for new developments with requirements now ranging from 300 megawatts (MW) to over 500 MW."

From the IEEE Spectrum (June 2024):

Scientists have predicted that by 2040, almost 50 percent of the world's electric power will be used in computing. What's more, this projection was made before the sudden explosion of generative AI.

From Data Center Dynamics (May 2024):

US utility Dominion expects to connect 15 more data centers to the grid in Virginia over the course of 2024, after connecting 15 facilities last year totaling almost a gigawatt of capacity [1 gigawatt = 1 nuclear plant]. In its most recent earnings presentation this week, the company said it had connected 94 data centers with more than 4GW of capacity in Northern Virginia since 2019. This included 15 data centers totaling 933MW in 2023, and 15 more are due to be connected in 2024. The company didn't include the capacity of those 15 facilities going live this year, and in the earnings call, CEO Robert Blue said he doesn't know how quickly they will ramp up to full capacity.

For those who think new nuclear power is the solution (2024), this is not a quote but a fact: The new Vogtle nuclear power plant took about 20 years to design and build, from concept to commercial operations. This recent construction schedule was set by an experienced nuclear utility that previously built access to transmission on a nuclear site they've owned for decades.

The critical metric is not the overall demand. Data centers' demand sits on the grid 24/7, so generating capacity must be available 24/7. While massive amounts of energy are already oversupplying some US power markets, most new sources originate from part-time wind, solar, and battery assets. Those part-time assets cannot serve the 24/7 load demanded by data centers. Therefore, the critical metric is the difference between the base supply and the constant load.

With growing 24/7 demand, a fleet of legacy power plants (natural gas, nuclear, coal) is needed to fill in the [significant] gaps left by part-time renewable energy sources. That fleet currently exists, but its overall capacity is declining. Retired plants (to the extent they can be summoned) and new generation will be needed.

However, any new base generation will experience poor capacity factors and difficult gross revenues. Both impair investors' revenues and erode their expected levelized cost of energy. Even if investors overcome profitability concerns, the time it takes to commercialize any new traditional generating asset exceeds the expected demand for new power (extreme example: Georgia Power).

These projections and concerns appear to contradict current trends. Demand has declined in the United States, Europe, and the United Kingdom. Current reporting suggests there could be too much supply, particularly in Europe. However, if projections described by ANS, IEEE, and utilities are correct, the opposite problem could be presented: insufficient supply. If supply becomes the issue as expected, scarcity curves will be taxed, unprofitable generating assets will become profitable, and residential, commercial, and industrial consumers will pay more. This issue is not limited to North America.

Humbert H. writes:

I was listening to an interview of some fund manager from Reno earlier today and he was talking about power shortage around where he lives due to AI server farms. He said they could be quickly and cheaply addressed with new gas powered plants, but due to the Biden administration now requiring all such plants to have complete carbon sequestration this stopped them from being a practical solution.

H. Humbert writes:

Increased the energy supply for data centers is the obvious and near-term brute-force solution. Of course (almost) everybody not in the tech industry assumes that the joule per bit per second for data centers can't be improved and hence producing more energy is the only solution using nuke. In fact Sam Altman said that too, what conventional thinking can possibly go wrong, right?

Zubin Al Genubi asks:

What would be a good way to invest in modern nuclear power? How about Bill Gates project?

Asindu Drileba adds:

I would suspect via buying Uranium ETFs? I first saw this conjecture from following the financier Lyn Alden.

Mark Zuckerberg of recent also mentioned in an interview that Energy and not Compute will be the number 1 bottle neck to AI progress.

H. Humbert responds:

The energy being the presumed AI investment proxy won't last in the long term. Increasing the energy supply is just an incremental engineering no-brainer approach to solve a longer term problem and the approach is not disruptive and it doesn't change the world.

Stefan Jovanovich offers:

Radiant Nuclear
Kaleidos: a Portable Nuclear Microreactor that Replaces Diesel Generators

Peter Penha writes:

A relevant interview on the Hidden Forces podcast with Brian Janous who was hired by Microsoft in 2011 to focus on energy (Google had just hired someone themselves as they thought the cloud might become something) - wound up as VP of Energy.

AI data centers need to be where they can individually draw the electricity of a city like Seattle (800 MWh) - so away from major urban areas - discusses the history of the grid from Sam Insull through to where we are going…also on the efficiency / consumption of AI chips - his view with AI is Jevons Paradox will apply and the more efficient the chips and the (new) grid gets the more consumers will demand.


Comments

Name

Email

Website

Speak your mind

Archives

Resources & Links

Search