I spent a good chunk of this week reading about data centres, which is not what I expected to be doing in the week of the US midterms. But here’s the thing: on 3 November voters in 10 states decided 41 local ballot measures about data centres, and by the next day CNBC’s count had them siding against the data centres in every one except a single measure in Harwood, North Dakota. In Ashville, Ohio, more than 81% voted no on keeping the village’s deal for two data centres. That’s not a close call, that’s a landslide.

Most of the headlines read this as “Americans reject AI”. I don’t think that’s quite right, and I think the more interesting part of the story is what happened next, because it lands right on the desk of anyone running an engineering team that builds with AI.

People Voted Against Buildings, Not Models

Gallup’s poll on the subject found that 71% of Americans oppose an AI data centre in their area, more than the 53% who oppose a nuclear power plant nearby. (Let that sink in for a second: a server farm is now less welcome next door than a nuclear reactor. The nuclear industry must be quietly delighted.) When the people who opposed them were asked why, half pointed to resource use, with water and energy named most often, and about one in five raised local costs such as higher utility bills. Hardly anyone was objecting to what the models actually do.

So people didn’t vote on chatbots. They voted on the one part of AI that has a postcode, a planning committee and a line on the electricity bill. And bear in mind, that part is exactly where every prompt your team sends ends up.

OpenAI’s Answer: Pay Up, Open Up and Rename

OpenAI’s response came in two parts. On 2 November it announced its Grid Protectors Program, under which its future campuses agree with the local utility to cut their power use when the grid is under strain. The pilot is its planned 3.2 GW campus in Effingham County, Georgia, where Georgia Power can cut up to 1,000 MW at busy times: roughly a third of the site. A few days later came a broader framework: OpenAI says it will pay the full cost of electricity and grid infrastructure for its projects, prioritise low-water cooling, publish its electricity and water use, and stop relying on nondisclosure agreements with local officials.

Those are sensible commitments, and they go straight at what people actually complained about, so credit where it’s due. And then, in the same breath, OpenAI told Business Insider it would like everyone to call these buildings “compute campuses”. Computerworld counted the new name four times in one announcement, against twice for “data centers”.

Having lived in London for years, I can’t help thinking of Windscale, which became Sellafield in 1981, more than two decades after its famous fire. New sign on the gate, same site behind it. (And “campus”? I’ll believe it when I see the freshers’ week.) Every engineer has seen the software version of this too: the legacy monolith that got renamed “the platform” during a reorg and kept falling over every Tuesday, exactly as before. Renaming the thing never fixes the thing. Publishing the numbers and dropping the NDAs might, which is why those are the bits I’d watch, in actual contracts rather than press releases.

What This Means for Engineering Teams

Here’s the part I think most of the coverage misses. Read the Grid Protectors Program again with an engineer’s eye: the compute your agents depend on is now, by contract, interruptible. Up to a third of a flagship campus can be dimmed on a hot afternoon when everybody’s air conditioning is on. OpenAI hasn’t said which workloads give way first (training runs, batch jobs or the API you call from production), and I don’t expect it to spell that out any time soon.

On top of that, somebody pays for those power lines now, and it won’t be the “campus”. Those costs end up in the price of tokens sooner or later. I’m convinced per-seat SaaS pricing collapses into usage pricing once one person runs ten agents; the flip side is that usage pricing passes the electricity bill straight through to you. Your AI spend just acquired a weather forecast.

And yes, the irony isn’t lost on me. I’m a big believer in “lights out” software engineering, where we write the specs and agents write, test and ship the code. I just didn’t expect the utility company to take “lights out” quite so literally 🙂

So, what would I do on Monday morning as a CTO?

• Sort your AI workloads into “must answer now” and “can wait an hour”. Anything in the second bucket (evals, back-fills, nightly agent runs, bulk summarisation) should go through batch APIs and be happy to be delayed, retried or moved to the small hours.
• Make agent tasks idempotent and resumable. If an agent run dies half way because capacity got shed somewhere in Georgia, it should pick up where it left off, not book the same meeting twice or open the same pull request three times.
• Have a fallback model, from another provider or another region, and test switching to it before you need it. A slightly worse answer from a smaller model beats a timeout.
• Budget AI as a utility with peaks, not as a flat licence fee. Track the cost per task completed, not just the total spend, so you notice when the price of a task starts to drift up.
• Ask your providers how they handle curtailment: which tiers get throttled first, what the SLA says, and whether they’ll warn you in advance. If the answer is vague, that is an answer.

None of this is exotic. It’s the same discipline engineering teams already apply to databases and payment gateways; we just got used to treating the model at the other end of an API call as if it lived in the cloud, in the literal, fluffy sense. It doesn’t. It lives in a very large building with an address, a zoning board and neighbours who vote.

Bottom Line

AI just stopped looking like an infinite, invisible resource and started behaving like what it always was: a utility with peak hours, local politics and a bill. The teams that design their agents for that now will barely notice the next heatwave. The ones that don’t will find out the hard way, probably on the hottest afternoon of the year, with an incident channel full of timeouts and someone asking whether the compute campus has a status page 😉