San Francisco and Oakland Joined 300-Plus Cities Trying to Stop Data Centers This Week. By One Count, Only Three of Those Fights Have Actually Worked.

On October 16th, San Francisco Mayor Daniel Lurie signed a 45-day moratorium on new data centers, pausing a 9-megawatt Bayview-Hunters Point facility's plan to triple in size under a permit the city issued back in 2016. The same week, Oakland's city council passed its own version, 7-0. Supervisor Shamann Walton, who sponsored San Francisco's pause, summed up the reasoning in six words.

Until we know what that is, it needs to stop.Shamann Walton, San Francisco Board of Supervisors

Both cities joined a list that, by one count, now runs past 300 local governments nationwide -- a wave large enough that San Francisco, headquarters to OpenAI and Anthropic, is now restricting the industry it helped build. The irony wasn't lost on anyone covering the vote: a city whose economy runs on the AI boom just told its own data centers to wait. San Francisco's compute mostly sits in leased capacity in Virginia, Texas, and a dozen other states anyway, so the optics outran the practical effect in this one case. But the symbolism is real, and it's part of why this week's vote became national news instead of two more entries on a list nobody outside Bayview-Hunters Point was tracking.

That list is the actual story. A tracking effort from Heatmap Pro puts the value of U.S. data center projects canceled or shelved so far in 2026, mostly over local opposition, at roughly $260 billion. The Information Technology and Innovation Foundation, a Washington think tank, published its own estimate on October 16th putting potential job losses from that opposition in the millions. Numbers like that are why the backlash reads, at a glance, like an industry in real trouble. It's also, according to the research firm that actually went and checked the math project by project, a wave that's mostly not working.

What happens when someone checks the math

SemiAnalysis, an industry analysis firm whose numbers get cited by everyone from The Economist to utility investor decks, published research in mid-September mapping more than 300 local data center moratoriums, bans, and permit pauses against the firm's own tracked pipeline of hyperscale projects. It found 77 campuses, representing roughly 20 gigawatts of planned capacity, sitting inside a jurisdiction with some kind of restriction on the books. Then came the part most coverage of the backlash skips: checking each of those 77 projects against the actual rule -- what the ordinance covers, when it took effect, and whether the project had already secured its permits before the restriction passed.

The result, recirculated widely this week as San Francisco's and Oakland's votes put moratoriums back in the news: out of 77 campuses sitting inside a restricted jurisdiction, three are genuinely being delayed. Those three total about 1,525 megawatts. Add the capacity caught by New York's statewide executive order and the real number actually held up by the entire 300-plus-restriction wave comes to roughly 2.3 gigawatts -- against a pipeline SemiAnalysis projects will add 38 gigawatts of new data center capacity in 2027 alone, more than double what came online this year.

Why most bans miss

The gap between 300 restrictions and three real delays comes down to something mundane: most local moratoriums get written and passed fast, usually in response to one specific, already-controversial project, and they don't always end up reaching the project that prompted them. A restriction passed after a developer already has its permits in hand doesn't retroactively unwind them. A pause written to cover "new applications" doesn't touch a campus already under construction. A county ordinance aimed at one parcel can leave a data center a mile away, sitting in a different zoning jurisdiction, completely untouched. San Francisco's own ordinance is a case study in how narrow this can get: it pauses new approvals and the expansion of existing facilities, but it doesn't touch the handful of data centers the city has already permitted and that are running today, and nobody at the Board of Supervisors could say how many that actually is. A pause that can't count what it's pausing is exactly the kind of restriction SemiAnalysis's methodology would flag as real on paper and uncertain in practice.

SemiAnalysis's per-project check is, in miniature, the same exercise behind GridPermit's own Kill Test: a law on the books is a fact, but whether that law actually reaches a specific project is a separate fact, and the second one is the only one that changes what gets built. A county commission can pass a moratorium unanimously and put out a press release the same afternoon, and both of those things can be entirely true while the project the moratorium was aimed at keeps pouring concrete three miles away, grandfathered in under a permit issued before anyone on the current commission took office.

A rival count says the opposite

Not everyone reads the data this way. Data Center Watch, a research shop that's been tracking local opposition since 2023, puts the damage far higher -- tens of billions of dollars in projects it counts as blocked or delayed over its first two years of tracking, with a newer estimate running into the hundreds of billions for 2026 alone. The two firms aren't just reaching different conclusions from the same facts; they're counting different things. A look at Data Center Watch's own case list shows a real share of its "blocked" total comes from developers who withdrew one proposal and refiled a similar one nearby -- the same company, the same investment, a different parcel. SemiAnalysis doesn't count a relocated project as a loss. Data Center Watch does.

Both methodologies are defensible, and both produce a real, citable number. Neither one, on its own, tells you whether the specific data center a given county is fighting over this month is actually going to get built. That's the trap in treating either tracker's topline figure as a verdict: "300 bans" sounds like an industry under siege, and "2.3 gigawatts actually delayed" sounds like the backlash is mostly theater. The honest answer sits in between, and it only shows up when someone checks the specific project against the specific rule, the way SemiAnalysis did and the way Data Center Watch's critics say its own count sometimes skips.

What this means for San Francisco and Oakland

Which brings the story back to the Bay. San Francisco's 45-day pause can be extended to roughly 22 months if a Planning Department study, due by a hearing expected in November, finds real environmental impact -- the kind of agency-confirmed finding that would actually move this fight from a publicly indicated risk to something a developer has to price into a deal. That study hasn't been written yet. Oakland's version is even newer, with no study timeline reported at all. On the numbers above, both cities just joined a list where the overwhelming majority of entries never end up changing what gets built. Whether San Francisco and Oakland end up among the three restrictions that do, or the 297-plus that don't, depends on specifics neither city has nailed down yet: what the Planning Department's study actually finds, whether a permanent ban follows the 45 days, and whether the facilities these ordinances target had already secured what they needed before anyone voted.

I think that's the real lesson here, more than either side's topline number. A moratorium is a real political fact the day it passes. Whether it's a real commercial fact -- whether it kills a project, delays it, or never touches it at all -- is a separate question that only gets answered parcel by parcel, agency confirmation by agency confirmation, the same evidence-ladder climb GridPermit runs on every site regardless of which way that week's headline points. The industry's 38-gigawatt 2027 forecast and San Francisco's brand-new moratorium are both true at the same time. Betting a deal's viability on the headline instead of the paperwork is how a sponsor ends up surprised either way.