Seventy-six percent. That's the share of voters in NBC News' Election Day polling who said they opposed data centers in their communities, according to The Verge. I've spent two years treating compute as the one thing in this industry that only goes up. Tuesday was the first time the people who live next to it got a vote, and they used it.
Read the day's news together and one pattern shows up three times. The AI bill landed in front of three different audiences: voters, consumers and one of the richest companies on earth. All three pushed back.
The ballot box found the data center#
The Verge's tally is blunt. Voters faced data center questions in 45 local ballot measures. As of Wednesday morning, 33 of them came out in favor of restricting development. These are not vague advisory votes:
Wells, Maine backed a 180-day moratorium on large data centers.
Several Ohio cities voted to ban data centers over 25MW.
El Dorado, Kansas banned facilities with peak load over 10MW for 10 years.
Texas is where it gets expensive. Democrat Gina Hinojosa unseated Governor Greg Abbott while campaigning for a statewide moratorium, and James Talarico beat Ken Paxton for the Senate seat. MIT Technology Review lists what now sits in that state's pipeline. Stargate Frontier in Shackelford County is projected at 1,820MW. Nexus Hubbard, built mostly for Anthropic, is planned at 6,400MW, the second-largest US data center under construction after Meta's Hyperion. Both winners made data centers a central campaign issue.
Ohio is the other pressure point. TechCrunch reports that the county where Meta planned a 700,000-square-foot facility just passed a ban, in the same state where OpenAI has eyed a $500 billion build. Goldman Sachs now predicts construction delays in Georgia and Texas.
The money didn't save the industry, either. TechCrunch's rundown of AI PAC spending shows American Mission (Leading the Future's GOP arm) and Building the Future putting $4.8 million behind a South Carolina Senate candidate who lost, plus more than $500,000 each on two House Republicans who also lost. The pro-regulation side had a better night, and it isn't shy about the goal:
“Our goal is to make AI money politically toxic.” — Charlie Blaettler, Guardrails Alliance
Caveats matter here. Both The Verge and TechCrunch note how hard it is to separate data center anger from a broader Democratic wave, and a few pro-build candidates, like Byron Donalds in Florida, still won. Still, even the winners had to campaign on ratepayer protection. My read is that cheap, fast permitting is over in the states that had it. The next fight is over who pays for grid upgrades, and the likely answer is the developer. That flows straight into the inference prices you and I pay.
Uber counted the agent economy and got four#
The consumer side of the bill showed up on an earnings call. Uber told analysts that third-party AI agents booked four rides. During the week of Halloween, about 500 Uber Eats orders came from agents, out of roughly 320 million trips and orders across the platform. TechCrunch and The Verge both carried it, and it echoes DoorDash's few-hundred-orders admission from last week.
Dara Khosrowshahi offered the most useful detail without meaning to. Booking a ride in the app takes about 50 seconds, and re-ordering a favorite meal takes about 90. An agent that has to parse your intent, confirm a pickup point and handle payment is competing with a minute of thumb work, and right now it is losing.
For builders, I'd draw a simple lesson from it. Agents win where the task is long, tedious and multi-step, not where a polished app already exists. If your agent roadmap starts with “book things in apps people already like,” that roadmap has a demand problem, not a model problem.
Even Apple is rationing tokens now#
The third audience was a finance department. Per a Bloomberg report relayed by TechCrunch, Apple cut its daily AI spending cap for engineering teams from $300 to $200 in October. Some allowances are expected to fall to $100, and non-engineering staff could see a limit as low as $25 a day. The “tokenmaxxing” phase, where companies pushed staff to burn as many tokens as possible, is turning into rationing once the invoices arrive.
A per-seat daily cap is a strange instrument, but it tells you what Apple is measuring: spend per person, not output per dollar. If you sell AI tooling into enterprises, expect that question in every renewal.
The cheap side of the ledger#
Against all that, the day's most practical release is a reminder that the price of capability keeps falling. Ricursive Intelligence, the chip-design startup founded by AlphaChip veterans Anna Goldie and Azalia Mirhoseini, released Ripple-65B Preview under Apache 2.0. It's a dense 65B model trained on 12 trillion tokens. Ricursive puts the total training cost at $32 million.
Ricursive's own numbers put it at 80.5 on Terminal-Bench 2.1 and 79.1 on SWE-bench Pro v2.0 Hard. Those are the two benchmarks behind its US open-coding claim. It trails Meta's Muse Spark Open on DeepSWE and several others, and VentureBeat flags that the model card's comparison numbers don't match the launch graphic. Treat it as promising rather than proven. Still, at $0.20 per million input tokens and $0.60 per million output tokens on OpenRouter, it's cheap enough that I'm running it against my own coding evals this week.
Meta, meanwhile, is buying rather than building on the tooling side. It acquired Conductor, the startup that runs teams of Claude and OpenAI agents in parallel workspaces. Conductor's cloud service shuts down November 20 and user cloud data gets deleted, so if you rely on it, export now.
Where this goes next#
Put the four stories side by side. Power is getting politically expensive, consumers aren't adopting agents yet, and enterprise buyers are capping spend. Meanwhile a $32 million model posts frontier-adjacent coding scores. The squeeze rewards efficiency over scale.
My prediction: within six months, at least one of the 1GW-plus Texas or Ohio projects gets formally delayed or resized, and the lab behind it frames the delay as a shift to “inference efficiency.” I'm adjusting my own work the same way. I'm watching cost per finished task more closely than leaderboard rank, and I'm keeping a cheap open model in the loop.