Last month Sup. Connie Chan, who’s running for Nancy Pelosi’s congressional seat, released a campaign ad excoriating San Francisco’s AI boom. “AI data centers are driving up our utility costs, displacing our community, and jeopardizing our environment,” she said, standing in front of a public library.
Chan’s not the only one raging against the machines. If elected, she’ll support a moratorium on new AI data centers penned by Congresswoman Alexandria Ocasio-Cortez (D-NY).
Chan’s opponent in the November race, state Sen. Scott Wiener, is calling for environmental regulations on data centers, better land use planning, expanded unemployment benefits to gird against mass job losses, and more. But he stops short of supporting a ban.
Data centers are now a bipartisan bête noire, according to a recent Gallup poll. More than 70 percent of Americans surveyed oppose or strongly oppose new data centers in their communities. Democrats are most likely to oppose such projects, but a majority of Republicans and independents also agree.
The U.S. has an estimated 3,000 data centers, and some 1,500 new ones are in development, driven by the computing demands of AI. Some of these new structures would cover more than a square mile.
That April estimate is already outdated. It also doesn’t include several proposals from a single Southern California developer — including one on the Cow Palace grounds, along San Francisco’s southern border. The way it’s described, it likely wouldn’t be affected by the ban Chan supports.
Turns out, though, that the city is already home to several centers larger than new ones that would fall afoul of the proposed ban. But most people probably wouldn’t recognize them as they walk right past.
Data in your backyard
It’s tough to estimate how many data centers are in San Francisco.
The Frisc found listings for half a dozen sites advertising server access for rent right now in locations like 274 Brannan Street and 200 Paul Avenue. Bay Area Air Quality Management District data confirm that diesel generators of the sort commonly used to back up data centers are running in these buildings.
That’s likely just the start. For example, this map says SF has 12 data centers, with clusters along the Embarcadero and in the Bayview, although it’s unclear if some of those facilities have closed.
Then again, others might not even be mapped because they’re small enough to go uncounted, says UC Riverside professor Shaolei Ren, who studies data infrastructure development.
Older centers, says Ren, provide “from a few megawatts up to tens of megawatts, whereas new data centers easily use hundreds.”
(One megawatt is enough to power hundreds of homes for a year; places with mild climates like San Francisco can stretch megawatts farther than places with extreme heating and cooling needs.)
These smaller centers have been an uncontroversial piece of internet and telecom infrastructure for decades. “There’s one right down the street from my office in Oakland,” political consultant Jim Ross tells The Frisc. “The only reason I know it’s a data center is I know somebody who did the [permits] for it.”
What Ross and Ren are talking about, and what Santa Ana, Calif.-based Global Stack is proposing at the Cow Palace, are far different than the sprawling proposals that have municipalities, states, and a majority of voters up in arms.
In fact, Global Stack doesn’t call its Cow Palace proposal a “data center.” In comments to The Frisc, CEO Dan Kang called it “emergency response infrastructure” meant to make SF more resilient — for example, to help first responders stay online with communications and data when the grid fails during a blackout or disaster.
The proposed site also includes a helipad and parking garage, and Global Stack wants to do something similar in county fairgrounds across the state. Folks in Napa, Sonoma, and Solano have already said no.
10 megawatts
When asked if the city government has signed a contract, Global Stack spokesperson Amanda Haig said the installation proposal is “still under early evaluation” with the Cow Palace. The company wants to approach “first responders, hospitals and public safety organizations” as customers. There are no agreements or contracts in place.
Plans call for a 10-megawatt facility. That would seem too small to run afoul of the Ocasio-Cortez moratorium on new AI data centers, which would apply to centers with maximum capacities above 20 megawatts.
But Chan spokesperson Ian Krager interprets the ban to apply as well to data centers of any size if used for AI. The bill’s exact language: “used for the development or operation of artificial intelligence models at scale.”
If you’re in San Francisco trying to watch a Netflix movie, and it’s located on a server in New York, it’s not going to be a great experience.
Mark Specht of the Union of Concerned Scientists
That would still seem to absolve Global Stack, if the use of its equipment ultimately aligns with its marketing pitch.
In follow-up comments, Krager wasn’t willing to concede that point. He said media reports call Global Stack an AI company and “if they’re not pushing back on the reporting then that’s their problem.” (In at least a couple of Global Stack’s presentations to county officials, it has said its centers would provide “world-class AI capabilities and blackout-proof operations.”)
Whether or not the Cow Palace site is associated with AI, its size makes “the development or operation of artificial intelligence models at scale” highly unlikely. The tech industry isn’t building new data centers the size of six football fields on a lark; they’re doing it because smaller centers can’t keep up with their data demands.
Outgrowing the farm
The first U.S. data center was built in 1945, but until recently many Americans knew these facilities by another name: server farm.
A server farm, with hundreds or thousands of machines storing data or programs, becomes a data center when it adds generators for uninterrupted power and a cooling system. But the definition also depends on who you ask and, critically, how big it is.
Whatever they’re called, modern life requires them fairly close by.
“If you’re [in San Francisco] trying to watch a Netflix movie, and it’s located on a server in New York, it’s not going to be a great experience,” says Mark Specht, a manager at the Union of Concerned Scientists (UCS), a science education nonprofit.
But many new centers don’t host Netflix movies, Specht says; they power generative AI models. Training AI software involves running thousands of high-performance chips simultaneously for months on end, consuming an enormous volume of electricity. A proposed center in Ohio would be the largest in the world and sprawl across more than 1,000 acres (1.6 square miles).
It’s also why people are worried about their effect on power grids, water use — one factor the Ocasio-Cortez bill cites is the use of liquid to cool the equipment — and noise and air pollution from massive backup generators. AI’s threat to decimate jobs is another fear, and perhaps the greatest one.
Hundreds of U.S. communities have enacted bans. Oregon lawmakers proposed a three-year ban on new projects in that state.
UCS estimates that California will add enough new centers over the next decade to consume the electricity needed to power some 3 million homes.
The politics of data
Like the tech company shuttles that spurred the “Google Bus” protests of the 2010s, the centers have become a public symbol of tech privilege and excess, but shorn of nuance.
There is room for nuance, however, as the Oakland-based environmental group The Utility Reform Network acknowledges. TURN is backing several state bills to regulate data center expansion, including extra company taxes to help reinforce the public electric grid.
TURN spokesperson Lee Trotman says sites already in SF aren’t on their radar: “We don’t have much to say about them other than they use less energy and water than AI data centers.”
Such distinctions haven’t come through in Chan’s broadsides or in the explanations of her position that her staff have provided. Leaving the nuance aside is to her advantage, say political observers.
“The details of any specific proposed new data center don’t really matter,” says San Francisco State University political scientist Jason McDaniel. “For most of the public, opinion on the data center issue is going to mostly be determined by overall feelings towards AI.”
“It’s a good issue for Chan because it helps define her” in the eyes of voters, says consultant Jim Ross, whereas “anything that reinforces the idea that he sides with big technology companies is a problem for Wiener.”
Chan and Pelosi, who has thrown her weight behind the Richmond District supervisor, pounced after Wiener’s campaign earlier this month posted a satirical “ConnieChan.AI” chatbot to lampoon her political record.
Wiener soon shut down the bot and apologized, saying the stunt “missed the mark.”
Wiener’s campaign did not respond to requests for comment on this story.
The chatbot highlighted one of the main differences between Chan and Wiener. As a supervisor, Chan has fought against looser rules that encourage new housing, while Wiener has been a main driver of those rules.
A data center ban bolsters Chan, McDaniel says. Her stances on housing have put her in a NIMBY minority, as evidenced by the June election results, when (for example) a one-issue, anti-housing candidate got steamrolled in the city’s wealthiest district by a 70-30 vote.
But data centers are one issue where Chan can feel confident saying “not in my backyard” (or in many others’ backyards), says McDaniel, even if those centers, in one form or another, have been there for years.
