Amazon and Microsoft are draining Seattle’s power grid to train the machines they are using to sack Seattle’s workforce. The city council voted unanimously on Tuesday to impose a one-year moratorium on new data-center construction. It also carved a twenty-megawatt expansion loophole for existing facilities that serve what lawmakers call a civic purpose—a distinction that matters only up to the point where a health-facility server room and an LLM training cluster are drawing from the same distribution substation.
The five proposed facilities would have consumed up to a third of Seattle’s current electricity demand, the kind of step-change that grid engineers notice immediately and that local ratepayers notice when their next bill arrives. Seattle City Light’s 2025 integrated resource plan had already flagged that projected load growth from large commercial customers could force the utility to acquire new generation capacity years ahead of schedule. This is the moment when the policy debate stops being about “AI governance” in the Beltway sense and becomes about rate cases, transmission-interconnection queues, and whether a family in Rainier Valley will see its electricity bill rise because two of the world’s largest corporations have decided that the future runs on a speculative technology.
An AI training run at scale is not a steady-state utility load. It is a bursty, GPU-dense draw that peaks unpredictably and then idles, requiring grid operators to maintain spinning reserve—natural-gas peaker plants burning fuel at roughly 40% thermal efficiency while spinning unloaded, emitting carbon for zero computational output until the next compute spike. A megawatt is a unit of power; it does not care whether it is spinning up an MRI or cooling a rack of H100s. This is the AI bezzle, the gravity-defying interval in which the capital expense of training ever-larger models looks like investment, but the revenue never materializes, and the cost falls on the grid.
You will have noticed the contradiction. Amazon and Microsoft have cut thousands of local jobs over the past year—a mix of post-pandemic restructuring and the capital reallocation toward AI that the companies’ own earnings calls tout—while committing a combined $390 billion to AI infrastructure this year alone. The money is not being spent on labor. It is being spent on GPUs, on specialized concrete pours, on power purchase agreements with nuclear operators, on the physical plant of what Cory Doctorow calls the AI bubble, a technology that, as he notes, “grows less profitable every day, and bosses have to force it on workers.” The force in this case is not just the layoff notice. It is the electricity bill. The same capital that fires the employee also bids up the price of keeping the lights on in the apartment the former employee can no longer afford.
The workers noticed. Amazon Employees for Climate Justice and 350 Seattle turned the contradiction into a letter-writing campaign that put nearly a hundred thousand emails in front of lawmakers, and the organizer Ben Jones put the logic bluntly: “AI is synonymous with people losing their jobs.” The mayor, Katie Wilson, told reporters that public pressure “supports and helps to spur on elected officials to do things that they already want to do,” a formulation so candid it borders on disarming. The city wanted to act. The workers gave it cover. The result is a moratorium that passed with unanimous consent.
The hyperscalers’ engineering is genuinely efficient: their newest facilities run at power-usage effectiveness below 1.1, meaning nearly every watt entering the building goes to compute rather than overhead. That is a real achievement. It is also, in the context of a moratorium with a twenty-megawatt back door, beside the point. Efficiency gains do not lower aggregate demand when the number and size of the facilities grow without constraint; they lower the per-unit cost, which makes it cheaper to build more of them—the Jevons paradox in a rack. The expansion amendment guarantees that even as per-server efficiency improves, aggregate load will still climb within city limits.
And the amendment itself is the place where policy meets physical plant. You cannot partition a distribution substation without building a second one, a thing anyone who has stood in a transformer bay knows. Because cloud providers do not expose per-tenant power metering at the rack level, the city cannot audit whether those extra megawatts are feeding emergency dispatch or a foundation-model training cluster. The carve-out that looks tidy on a legislative docket collapses under a tradesman’s question about what the copper is actually carrying.
The moratorium is in effect for one year, and there is a public consultation open at the municipal planning department while it holds. The consultation is the moment to push for an interoperability mandate on the cloud tenancies that host municipal APIs, forcing the hyperscalers to open the interfaces so that civic applications can be decoupled from the speculative compute load, and to pass a rate structure that prices peaker-plant standby time at the true marginal cost of carbon and copper. The moratorium exists because workers organized around twin grievances—jobs and power—that are not separate grievances at all. They are the same extraction machine, viewed from two sides. When a city council votes unanimously to pause it, even with a carve-out, that is not a policy failure. It is a proof of concept: the people who built the cloud can still refuse to be eaten by it. There is a legislative session in January. The deadline matters. The work is to be done.