$cd ~/log
LOG.013

2026-11-07

AI Notes: The New Gate for Compute Is Permission to Build

In the US midterm elections on November 3, there were 41 local ballot measures related to data centers across 10 states, 21 of them in Ohio. By the counts from Ballotpedia and CNBC, every single one except Harwood, North Dakota, came out against data centers. In CNBC's voter survey, 56% of people said data centers were a key or major factor in how they voted.

The governor's races went the same way. Amy Acton, Ohio's newly elected governor, supports a "conditional pause": data centers must bear 100% of their utility costs and be built only with union labor. Gina Hinojosa, the newly elected governor of Texas, has promised to halt new data center construction.

Two days later, OpenAI published a pledge: for the data centers it owns and operates, it will pay the full cost of power, publish its construction plans, seek community input, and stop signing nondisclosure agreements with public officials. Meta quickly said it had made a similar pledge back in March.

For the past two years, the list of constraints on AI infrastructure has been chips, power and capital. This week the list gained one more item: whether the community will let you build. It changes the industry in three ways: who pays the costs, how long things take, and what already-built capacity is worth.

The ballots change who pays

At the core of the opposition to data centers is an external cost. Huge amounts of new electricity demand drive up grid investment and power prices, and much of that used to be spread across the bills of residents and small businesses. Add tax breaks, water use and noise, and the local community bore a large share of the cost, while the gains went mostly to model companies and cloud providers far away.

Meta president Dina Powell McCormick put it bluntly at The Wall Street Journal's tech conference: data centers are what AI looks like in physical form, so people can say "I don't want it built here," and behind that is how people feel about AI itself.

The pledges from OpenAI and Meta essentially take that external cost back onto their own books: paying their own power bills, funding grid upgrades, directing tax surpluses toward local teachers and first responders, and hiring local workers where possible. The bill hasn't disappeared; it has just changed hands. Construction and operating costs per megawatt will go up, and the union labor condition proposed in Ohio would push up building costs further. The hidden subsidy data centers enjoyed in site selection over the past few years is being clawed back at the ballot box.

Time costs more than money

Money problems can still be solved with financing. Time problems are harder.

According to the International Energy Agency, connecting a 100-megawatt project to the grid in Northern Virginia takes about seven years. An executive at Mitsubishi Power told Bloomberg last month that gas turbine deliveries are already booked out to 2031. So some data centers have started converting engines from retired airliners into gas generator sets, because they can be up and running in 24 months, at the price of lower efficiency and higher emissions. xAI installed eight of these units in Mississippi to power Colossus 2 in Memphis. Last month Microsoft pulled out of Nscale's project in Loughton, England, because the local grid won't be able to supply enough power until the early-to-mid 2030s.

Now a layer of political approval time is stacked on top of the grid queue.

Time is expensive because chips go out of date. This week there were reports that Microsoft did not renew a compute contract it signed with Oracle in 2023 and is gradually shutting down the related servers, because they run older Nvidia chips that no longer suit its current workloads. If a data center gets power two years late, the chips meant for it may already be a generation behind. The value of compute depends on two things: which generation of chips it uses, and when it can come online. Every quarter of approval delay pushes the second back by a quarter.

When OpenAI's chief global affairs officer Chris Lehane introduced the company's Ohio project, he said the computing equipment in a data center has to be replaced every three to five years, so it brings lasting jobs to the area. Seen from another angle, the same fact is a depreciation clock: the equipment starts losing value the day it is bought, whether or not the building has its permits.

Data centers already powered are getting pricier

When new capacity has to wait for permits, capacity that is already built and already powered becomes more valuable.

The amended IPO filing Anthropic submitted on November 2 offers a concrete example. Third-quarter revenue was $16.96 billion, up 47% from the previous quarter, yet operating profit swung from $974 million in the second quarter to a loss of $567 million. The filing attributes this mainly to $5 billion of compute bought on short-term contracts, which is "significantly more expensive" than multi-year contracts. Its multi-year compute is usually locked in 12 to 18 months ahead; once demand runs past that, the only option is to fill the gap in the short-term market. The filing also warns that some compute suppliers rely on third-party financing to build capacity, and if that financing gets more expensive or dries up, the capacity in their contracts could be delayed, repriced, reduced, or never delivered at all.

Meta is on the other side. Powell McCormick said Meta gets calls every day from people wanting to rent the compute it has already built.

This is the most direct consequence of the permitting constraint: the pace of new supply is no longer set only by demand and capital but also by approvals, so swings in short-term demand fall more heavily on the price of existing capacity. Whoever locked in long-term capacity early comes out ahead; whoever relies on short-term purchases to cover shortfalls will see gross margins squeezed.

Where the value lands

Put these things together, and the scarcest thing in AI infrastructure is shifting from GPUs to a different set of assets: land that already has permits, grid connections already in the queue, equipment that can generate its own power, and data centers already running.

The way builders compete will change too. It used to be about who could get chips and who could raise money. From here, it is also about who can get a county to accept them: how much of the power bill they will pay themselves, how tax revenue gets shared, and how many jobs the community gets. These terms will be written into contracts, and into the cost of every megawatt.

The newly elected governors and members of Congress don't take office until January, and no new law has taken effect yet. But a project often takes three to five years to go from site selection to power-on, so this week's ballots shape the capacity curve from 2028 onward.

$ share --x Share on X ↗

© 2026 CYBERCAESAR · cybercaesar.bond