The Usual Suspects

THE NUMBER: 517 of 809. Two-thirds of the AI data centers planned in America are being built in counties the federal government classified as drought-stricken in the last year — 517 of 809, per a Guardian analysis of NOAA data. Read that again. The industry is aiming its thirstiest machines directly at the places that have already run out of water. Lake Mead sits at 24.7% full this month, its lowest in five years. So when the neighbors show up with torches, they are not imagining things. They are just, in most cases, holding the wrong man responsible for the wrong crime.

Hyperscalers are…toast!
A Stanford study found small, local AI models match or beat big cloud models on the vast majority of everyday tasks at 50 to 85% lower energy cost.

Joachim Klement

Round up the usual suspects

On Tuesday, while Florida voted in a primary that turned partly on data centers, a beer company owned by the Kelce brothers released a 90-second ad begging Americans to mail their urine to “the AI data center of your choice.” Jason Kelce, jar in hand, walking through a field toward a server farm, singing about peeing on computers to save humanity. When a mass-market beer brand builds a campaign around your industry’s infrastructure, that is not a fringe position. That is the culture telling you exactly where the anger is pointed.

Here is the tell that should stop you cold: the party that built this thing is now running from it. The National Republican Senatorial Committee circulated a memo this week — first reported by Axios — warning that data centers have become a “sleeper issue” for the entire midterm cycle. In Ohio, former senator Sherrod Brown has made opposition to them the centerpiece of his campaign against Republican incumbent Jon Husted. The GOP’s own memo says the quiet part in plain language: if Husted loses, “the industry will be blamed.” Democrats lead the generic ballot by more than six points. When your own side circulates a note telling its candidates to run away from the thing you have spent two years championing, you are not watching a policy debate. You are watching a coalition price in a loss.

And it is genuinely bipartisan, which is the part that should worry the builders most. Florida’s Byron Donalds, Trump’s pick and the primary winner, ran on a “Protecting Ratepayers Act” even as an Andreessen-backed super PAC funded his campaign. Pennsylvania’s Josh Shapiro, eyeing 2028, signed an executive order Tuesday putting hard standards on data-center development. Texas’s Greg Abbott, who used to brag about landing these projects, is now racing to slap conditions on them while his challenger runs ads calling him a tool of “moneyed interests.” New York’s Kathy Hochul signed the nation’s first year-long moratorium last month. More than half of Americans now say they are more concerned than excited about AI, up from 37% in 2021. This is not a red or blue story. It is a pocketbook story wearing a hard hat.

The wrong man in the lineup

Round up the usual suspect and you have still cuffed the wrong man.

We said it ourselves last night, in The Perfect Storm: the reason your house keeps getting less affordable is not the data center one county over. It is the debt-funded buildout repricing the long end of the bond market — a first-time buyer in Ohio bidding against a hyperscaler for the same thirty-year dollars, and only one of them has an investment bank arranging the financing. That was the real crime scene. The discount rate is up, every long-duration asset in the country got marked down, and the 30-year mortgage now starts with a 7. None of that is the data center’s fault. It is the macro’s.

But that buyer in Ohio is now a voter. And here is the problem with the real culprit: you cannot run a campaign ad against a discount rate. You cannot put the Japanese carry trade in a thirty-second spot. You cannot make a villain out of a $40 trillion refinancing schedule, because it has no face, no parking lot, no cooling pond, no hum. The affordability rage is real and it is nationwide, and it has nowhere to go. So it goes to the one object you can drive past on Route 23 and point at. The data center is not the arsonist. It is the guy standing closest to the fire when the cops arrive.

This is the oldest move in the book. When the crop failed, Salem didn’t blame the weather. It blamed a woman. The anger needs a body, and the diffuse, invisible, systemic cause never has one. Keyser Söze runs the whole operation and the beauty of it is that nobody’s even looking for him — the greatest trick the devil ever pulled was convincing the world he didn’t exist. We spent the summer doing a version of this ourselves, worrying these data centers burned too much water and too much power, before the bigger bill, the cost of money, came due. So we are not throwing stones from a clean house. We are telling you we recognize the move because we made it.

The dirty cop

Now the part nobody on a debate stage will admit.

Where your electric bill actually did go up because of a data center (and in some markets it did), that is not the machine’s crime. It is the councilman’s. Somebody cut a lazy deal. A tax abatement, a ribbon cutting, a photo in a hard hat, a press release with a jobs number in the headline, and buried underneath it, the cost of the new substation and the new transmission lines quietly socialized onto every ratepayer in the district. The developer got the power. You got the bill. And the same official who signed that deal is now leading the torchlight parade against the building he approved. Captain Renault, shocked — shocked — to find that gambling is going on in here, while he pockets his winnings.

The tell that this is fixable, not fated, is that federal regulators are fighting over the exact question right now: should the developer pay for the capacity it demands, or should existing customers eat it? That is the whole ballgame, and it was always the councilman’s job to get it right. Here is the deal a competent official cuts on day one. Make the builder pay soup to nuts — the power it draws, the interconnects it needs, the new generation and transmission it forces onto the grid, all of it, on its own dime. Then tax the thing and route the money to knock down everyone else’s electric bill. And if the developer signs, make it smooth: fast permits, a clear timeline, a real partner on the other side of the table. Do that and the data center stops being a parasite and becomes what every big industrial project used to be — a tax base, a jobs base, an anchor tenant that leaves the town better than it found it.

Pennsylvania just proved this isn’t hypothetical. Shapiro’s order forces data-center developers to pay for all new generation and transmission, hire and train locally, sign community-benefit agreements, and meet real water standards. That is almost verbatim the argument we made ten days ago in Memo to the Governors — stop excoriating these projects and use their demand as leverage to hand your own people cheaper, steadier power. A governor with a national profile signed our thesis into law while we were still writing the follow-up. The tape doesn’t lie. We just read it early.

The one crime the data center actually committed

There is one place, though, where the anger is aimed at exactly the right target, and honesty demands we say so plainly.

The Southwest is running dry, and the data center is a real straw on a breaking camel. Lake Mead is at 24.7% of capacity, the lowest in five years, with Lake Powell right behind it. The Colorado River system that waters power generation, agriculture, and roughly forty million people is in genuine peril, and it was in peril before the AI boom showed up asking for more. Two-thirds of the country’s planned data centers (our 517 of 809) are going up in drought-classified counties anyway. In Arizona, utilities have to roughly double capacity in four to five years just to serve the load; TSMC’s three Phoenix fabs alone will draw a combined 16.4 million gallons a day; and residents are being told to expect power that is both pricier and less reliable during 120-degree peaks. This is the scenario where the mob has it right. When water and power are already scarce and a new entrant walks in demanding more of both, the price of everything gets repriced and rationed — drinking water, farm water, electricity. Push that forward a few years and you don’t get a policy debate, you get fires, an insurance crisis, a real-estate reckoning, and people leaving. The data center didn’t start that fire. But pouring one into the driest county in the state is the kind of thing a jury remembers.

Which brings us to the industry’s actual problem, and it isn’t the water. It’s the wrapper. Nobody marches on granola, even though a single California almond drinks about a gallon and the orchards soak up roughly a tenth of the state’s farm water. Almonds get a pass because they come wrapped in a wholesome story — health, sunshine, a snack you feel good about. AI never bothered to earn that story. Dario Amodei more or less conceded the whole thing on Saturday, posting that the backlash stems from a “crisis of trust,” that “ordinary people don’t trust companies, governments, or the tech industry and always suspect that we are cooking up some new way to screw them over,” and that “AI is just the latest iteration of it.” He’s right. And the fix for a trust problem is not a better ad campaign. It’s the Shapiro deal — the thing that actually makes the town a beneficiary. The granola only works if the snack is genuinely good for you.

The stranded asset

Here’s the irony that should keep the builders up at night more than any moratorium. They may be fighting to the death over a load that never fully arrives.

The most-read piece in AI research this week argued the hyperscalers are, in the author’s word, “toast.” A Stanford study ran small language models (Qwen, Gemma, the open-weight stuff) on ordinary desktops and laptops against the frontier cloud models, and found the small models matched or beat the big ones on the vast majority of everyday tasks, at 50 to 85% lower energy cost, with inference-per-watt roughly seven times better. Tomasz Tunguz, the VC, put it more plainly this week: your laptop can now run a model as capable as nearly anything in the cloud. People are buying Mac Minis specifically to do it — OpenClaw alone reportedly drove a 50% bump in Mac Mini sales this year. If even 70% of the workloads everyone assumed would run in a Pike County megacampus instead run on the machine already sitting on your desk, the demand curve that justifies a $500 billion buildout has a hole in it.

But the split matters, because this is where it gets genuinely interesting rather than just bearish. Intelligence is coming home in two very different halves. The commodity half (writing, summarizing, history, the humanities) goes local fast, for two reasons at once. The raw material is everywhere (the books are all sitting in the training data already), and there is no cheap grader to build a durable lead, so everyone plateaus at “good enough” almost immediately. And good enough is exactly what runs on a Mac Mini. The graded half is a different animal. Math and engineering fall to whoever owns proprietary, self-checking data — where physics or a compiler grades the output for free, every run mints more training signal, and the lead compounds. That is the real reason a shop like the SpaceX/xAI complex has an edge that doesn’t commoditize: it isn’t just the dataset, it’s that the work grades itself. The frontier keeps the problems that check their own homework. Everything else migrates to your desk.

Which leaves the question that actually decides who wins the commodity half: easy, or complex? A trusted, US-based, good-enough model on local iron is worthless to most people until someone wraps it so their aunt can use it without reading a manual. Whoever makes the local model boring to use takes the mass market. Leave it powerful and fiddly and it stays a toy for the people who already know what a token is. Our bet — and it’s a bet, not a fact — is that the first company to ship a genuinely trusted on-device model with the complexity hidden takes an enormous bite, and the natural home for that is Apple paired with a frontier lab doing the heavy lifting. Apple has whiffed on AI before, so don’t write it in ink. But the shovel is getting cheap enough that the little guy can finally dig, and that’s the optimistic read underneath all this noise.

Who wins the drought

One last turn, because it’s the one the pitchfork crowd never sees coming.

If power is the binding constraint — and in the Southwest and half the PJM grid it already is — then the master metric stops being raw capability and becomes computation per watt. And that is Jensen Huang’s game. Nvidia is the most efficient producer of usable compute per unit of power in the merchant market, which means scarcity doesn’t threaten him, it entrenches him. The tighter the grid, the more every operator has to squeeze the most work out of every megawatt, and the more they line up for his chips. The honest asterisk is Google’s TPUs, the one credible perf-per-watt rival, and the open question of whether the SLM shift shrinks the whole pie. But the shape of it is clear: the backlash that looks like a mortal threat to the AI buildout quietly hands the pick-and-shovel winner an even stronger hand. He wins the boom, and he wins the drought.

What it means for you

Three things fall out of this, and they touch you whether or not you own a single AI name.

Price the pitchfork into every data-center and hyperscaler position. Political risk is now a real, datable input, not a vibe. New York has a moratorium, Pennsylvania has hard standards, Texas and Arizona are reversing incentives they used to advertise, and the midterm map is turning the buildout into a ballot question. If your thesis assumed frictionless concrete, cut the growth rate and add a permitting delay. The story can be right and the stock can still stall while the politics catch up.

If a data center is coming to your town, negotiate like Shapiro, not like the guy taking the photo. Demand the whole bill up front — power, interconnects, new capacity — plus a tax that lowers everyone else’s rate, and offer a smooth process in return. A community that does this gets cheaper, more reliable electricity and a real tax base. A community that takes the ribbon cutting eats the surcharge and then votes out the person who signed it.

Move your commodity AI work off the frontier and onto local models now. The writing, the summarizing, the history-and-humanities tasks run well today on a small model on a Mac Mini or its equivalent. Reserve your metered frontier tokens for the graded work — code, math, anything with an answer key — where the edge is real and worth paying for. You cut your spend and you cut your exposure to the very buildout everyone’s fighting about.

The greatest trick the affordability crisis ever pulled was convincing three swing states the culprit was the building with the cooling fans. Round up the usual suspect if it makes you feel better. But the money for a cheaper light bill is in making the builder pay soup to nuts — not in slowing the machine — and the water crisis is real whether or not there’s an election. Get the deal right, and the monster turns back into a neighbor. And like that, he’s gone.

---

The email edition — THE NUMBER, three moves, and the day’s five stories — is in your inbox. If a data center is about to meet your town, your grid, or your P&L, that’s the conversation we run at Outsider Labs.