Local Consent Is the Unpinned Dependency in Your Compute Roadmap

On November 10, more than a hundred people packed Boise City Hall to talk about buildings full of computers they will never see, and everyone who testified backed a pause, many of them asking for permanent limits. The council approved a six-month moratorium on large data centers that took effect immediately. Nobody in that room was thinking about the region variable in your Terraform, yet hearings like this one now sit upstream of cloud capacity, right next to chips and power, because the money behind new campuses increasingly depends on how they go. For the lender's view, this breakdown of how local consent became a credit variable in data center finance lays out the mechanics clearly. This post looks at the same shift from our side of the stack: what changed in 2026, how a county vote turns into an API error, and what engineers can do about a dependency nobody can pin. A region is a building, and a building is a loan We talk about cloud capacity as if it were weather: some days the GPUs are there, some days they are not. In practice, capacity is a construction schedule. A campus needs land, zoning, permits, a utility agreement and a tenant, and lenders release money in stages, with the developer proving at every draw that the project still meets the terms of the loan. In August, senior bankers told Reuters that community opposition now factors into how they judge these loans, and that they lean toward states friendlier to data centers. Karen Fang, who runs infrastructure and sustainable finance at Bank of America, defined project readiness as permits and approvals plus "the community support from the people who are going to live around it." That one line compresses the whole problem. Consent feeds the cost of capital, the cost of capital shapes how fast new halls get built, and build speed is the constraint hyperscalers keep citing to investors. Sundar Pichai said in April that Google was "compute constrained in the near term" and that its cloud revenue would have been higher with more supply. Microsoft has told investors for several quarters that Azure demand is running ahead of the capacity it can bring online. Synergy Research Group still expects total US data center capacity to double within three years, but its chief analyst, John Dinsdale, said in July that limited power and rising local concerns "are crimping many new plans for data centers." 2026 turned consent into a moving target The scale changed first. Data Center Watch, which tracks these fights, counted at least 75 US projects worth about $130 billion blocked or delayed in the first quarter alone, roughly as much as in all of 2025, while the number of organized opposition groups more than doubled to 833 across 49 states. By October, S&P Global reported that local governments in 40 states had an active ban or moratorium in place. Then the rules started changing under projects that were already in flight. In Git terms, 2026 was the year branch protection changed in the middle of the sprint. In early August, Texas, home to the country's largest construction pipeline, effectively paused its merge queue: Governor Greg Abbott ordered an audit of every data center moving through ERCOT's interconnection process, and the grid operator stopped approving energization for large data center loads until the checks are finished. The audit plan even includes a community impact review, with reports due in December. In July, New York put a one-year pause on large new facilities. And on November 3, Janesville, Wisconsin, added a required reviewer: any project there above $450 million, including a proposed $8 billion, 800-megawatt campus on the old GM site, now needs a popular vote. The midterms made the trend hard to wave away. The Associated Press put it bluntly: data centers took a beating in the midterm elections, with about three-quarters of voters in the AP Voter Poll opposing a new data center in their own community and local restrictions passing in 44 of the 45 referendums that had finished counting. Campaigns spent more than $380 million on TV ads that mentioned data centers, and Morgan Stanley warned clients that the results could split the country into places that welcome new facilities and places that do not. From a county vote to a stack trace Most of this reaches developers indirectly, so it helps to follow one chain from end to end. Project Jupiter is a 2.45-gigawatt campus planned for Doña Ana County, New Mexico, with Oracle as the tenant and capacity meant for OpenAI's Stargate program. It has drawn water concerns and regulatory setbacks, including New Mexico officials' rejection of a proposed natural gas pipeline extension in July. In September, Oracle sent the developer, a unit of Blue Owl Capital, a force majeure notice that preserves its right to keep paying the lower development-stage rent for longer if power is not ready for the planned 2028 opening. Oracle says the project remains on schedule, and Blue Owl says the financial commitments have not changed. Still, look at what happened: a fight over how the campus gets its power became a clause in a lease, and that clause now decides who absorbs a delay. That clause is an exception handler written in legal language, and the largest tenants negotiate theirs long before the first rack arrives. Everyone else meets the same failure as an unstructured error: InsufficientInstanceCapacity on AWS, ZONE_RESOURCE_POOL_EXHAUSTED on Google Cloud, AllocationFailed on Azure, or a polite email saying your reservation start date has moved. None of those messages mentions a planning commission, but more and more of them begin at one. The lever that belongs to engineers Residents rarely object to computation as such. They object to a load that is huge, constant and opaque, and that may trigger grid upgrades their own power bills help pay for. Software cannot settle every argument about water or noise, but it can change the shape of the load, and utilities increasingly treat that shape as a reason to connect a facility sooner. Google, for example, has written 1 GW of demand response into its long-term utility contracts with Indiana Michigan Power, TVA, Entergy Arkansas, Minnesota Power and DTE Energy, delivered by limiting or shifting machine learning workloads when the grid is stressed. In a March demonstration in the UK, EPRI, Emerald AI, National Grid and Nebius cut the power draw of AI compute by up to 40% with sub-minute response, while service levels for high-priority jobs held throughout. None of that works unless the workloads cooperate. A training job that checkpoints every few minutes can yield to the grid; one that checkpoints once a night cannot. A pipeline built from idempotent steps can pause and resume; one that charges a customer twice on retry cannot. The cloud already runs drills for this, and most of us skip them: AWS gives Spot instances a two-minute interruption notice, and Google Cloud gives Spot VMs about 30 seconds. Code that treats those signals as routine is what turns flexible load from a slide in a pitch deck into an operating fact. A checklist for a dependency you can't pin You cannot vote in every county that hosts your workloads, but you can stop assuming the capacity will simply be there. Five habits help more than any forecast: Make the region a variable, not a constant. Parameterize region and zone in your infrastructure code, keep container images, model weights and secrets replicated, and rehearse a deploy into a second region on a schedule, the way you rehearse restoring backups. Treat capacity errors as a normal state. Keep an ordered list of fallbacks across instance types, zones and regions, put reservations (On-Demand Capacity Reservations or Capacity Blocks on AWS, reservations on Google Cloud and Azure) only on paths that truly cannot wait, and let everything else queue instead of paging someone at 3 a.m. Ask your provider the questions a lender asks. Before you commit to future GPU capacity from a colocation provider or a neocloud, ask whether the site is fully entitled, whether any approval can still be appealed or put to a vote, whether power is contracted or only queued, and what the contract does if delivery slips: credits, substitution at another site, or the right to walk away. Separate work that needs a place from work that only needs a GPU. Latency-sensitive inference has to stay near users. Training, evaluation and batch jobs can follow capacity wherever it appears, as long as your data pipeline lets them move. Watch the boring feeds. If your roadmap depends on one provider's expansion in one county, subscribe to that county's planning agenda and the state utility commission's docket, and put moratorium expiry dates on the same calendar as your certificate renewals. The cloud has neighbors now It is tempting to file all of this under someone else's job: politics for politicians, credit for bankers. But the chain runs straight through our work. Consent shapes financing, financing shapes when capacity appears, and capacity shapes what we can ship and what it costs. The old joke says the cloud is just someone else's computer. In 2026 it is also someone else's town, and the town gets a vote. Teams that plan for that, by keeping systems portable and heavy jobs interruptible, will spend less time staring at capacity errors and more time shipping.