The booming electricity demands of the nascent artificial intelligence gold rush boggle the mind and will likely hit ordinary ratepayers in the pocketbook if nothing is done to mitigate the downsides of this impending technological revolution.
According to a recent series of articles in The Seattle Times, co-published with ProPublica, businesses, farmers and residents in central Washington are already beginning to feel the pinch on their power supplies and likely facing substantial rate hikes next year. And that is happening before the full impact of artificial intelligence on information services, which some have suggested could increase data-center electricity demands tenfold.
These data centers — of which there are now over 100 in the state — require oodles of electricity for two major reasons: to power the racks of computer servers inside and to keep their sophisticated microprocessors from overheating. When thousands of these chips are sequestered in an enclosed space, the heat produced can halt their operation; it must be removed by power-guzzling cooling fans or liquids to keep them functioning.
The largest of these facilities, known as hyperscale data centers, draw over 100 megawatts (100 MW, or 100,000 kilowatts) — roughly 10% of the capacity of a large modern power plant. Already, there are several such centers in central Washington. Many other, smaller data centers cluster around Seattle and Tacoma.
All told, data centers in the Pacific Northwest currently require average electrical power of around 1,300 MW, according to energy analyst Massoud Jourabchi of the Greenway Research Group. That’s almost 6% of the average Pacific Northwest regional load, but it will grow rapidly during the next five years. In the local Seattle/Tacoma area, data centers already consume over 10% of available power.
Data-center operators like Microsoft need to secure cheap, reliable power they can use with minimal interruption. That’s why they located many of the larger centers in Grant and Douglas counties, close to Columbia River dams, and contracted with the local public utilities to supply hydropower 24 hours a day. But according to a Times article, the utilities are now having to purchase extra “unspecified power” from the grid to meet the needs of all their customers. Usually generated by out-of-state suppliers, often using fossil fuels, this power comes at a higher cost that must be passed on to its consumers.
One might fault these public utility managers for selling the crown jewels of their domains to voracious outsiders, although the lure of lucrative local tax payments probably influenced their thinking. But they likely did not anticipate the explosive growth of the data-center power needs, which have recently been doubling every four to five years.
A 2024 study by the Northwest Power and Conservation Council projects that NW regional data-center power requirements could double again by 2029. This estimate is extremely uncertain, however. Only an added 2,000 MW is needed for such “tech loads” in one low-end scenario, but 6,500 MW may be required in extreme cases. Much of this huge uncertainty arises from the unknown impact of artificial intelligence, which is only now being felt as Amazon, Google, Meta, Microsoft and others implement AI technologies in their data centers.
But one thing is obvious from this analysis: Data-center power requirements will dominate electrical load growth over the coming years, substantially more than the expected demand from electric vehicles. The next generation of graphics microprocessors soon to be installed in servers to enable AI applications (for example, Nvidia’s Blackwell chipset) will draw over a kilowatt apiece and require nearly as much power to cool them. Such demands exceed the average power needs of a typical Washington house, which will draw 1 to 2 kilowatts during normal weather conditions. (For a 30-day billing period, that translates to 720 to 1,440 kilowatt-hours of electrical consumption.)
Industry executives are struggling to accommodate these anticipated demands without requiring new fossil-fueled power plants, given their promises — and government mandates — to reduce greenhouse gas emissions. In Pennsylvania, for example, Microsoft just announced its intention to lease the full 835 MW output of a resurrected Three Mile Island nuclear reactor, enough to power 700,000 area homes. This nuclear power will be fed to the grid to supply its expanding fleet of AI-enabled data centers.
Elsewhere, utilities are delaying plans to retire nuclear and coal-fired power plants or converting the latter to natural gas in an attempt to meet anticipated data-center needs. Other utilities and companies are contemplating the possibility of building geothermal power plants, which might be an option for the Pacific Northwest, or hoping to commission small modular nuclear reactors.
But it will be difficult to bring these new power supplies online fast enough to meet the exploding data-center demands. Such a lapse could lead to more frequent near-term power shortages that will have to be met by purchasing costly grid power — as the Grant County public utility has done.
During the severe January 2024 cold snap, for example, Pacific Northwest utilities had to import an average 4,900 MW of power from systems in California and the Rocky Mountain states, at unit costs far above the baseline hydropower prices. Such extra charges feed into the costs per kilowatt-hour paid by ordinary ratepayers, currently growing by 5% to 6% annually in the Puget Sound area.
And as voracious data-center electricity demands exceed the limits of available supplies, the ruthless law of supply and demand will inevitably kick in and increase our rates still further. The only other option would be rolling blackouts, which nobody wants.
If not met by renewable energy sources in the coming years, these demands may also endanger the near-term goals of Washington’s 2019 Clean Energy Transformation Act. Hydropower is already tapped out, and the low-hanging fruit of energy efficiency has largely been picked. And highly variable solar and wind power require long-term storage to provide the steady 24-hour supplies favored by data centers.
Power shortages and consequent rate increases therefore look increasingly likely if artificial intelligence is allowed to proceed unchecked and unregulated. Realizing the full benefits of this technological revolution — which was recognized this year by the Nobel Prizes in physics and chemistry — will require electricity that does not yet exist and will be difficult if not impossible for utilities to supply. This threatened energy shortage may prove to be the Achilles’ heel of the promising AI industry.