I Fact Checked Every Data Center Lie

Seventy-five percent of voters in a Heatmap Pro / Embold Research poll of 2,045 registered voters (Aug. 8-13, 2026) say they would oppose a data center near their home. On All-In, investor Gavin Baker says the public is getting a distorted story. Farzad’s Sept 8, 2026 long-form says both sides are right and both sides are lying. Some towns really are eating new power costs, water demand, noise, and diesel pollution for buildings that keep almost nobody on payroll. At the same time, one of the most viral water numbers in the debate was off by a factor of about 1,000. X’s Safety team said it found a suspected Chinese bot farm of about 200,000 accounts, and that 200 of those posted claims that AI data centers are driving up household electricity and straining the grid. Farzad wrote Abundance or Collapse, covers this beat for a living, and says he has no interest in AI-company PR or politician fear bait. He wants the claims tested, then a deal structure that stops the tax-break race.

Power bills are real. The mechanism is the story.

Treat a data center as a factory for computation. Chips do the work. Networks move bits. Cooling removes heat. One site can draw like a small city. A giant campus draws more. Lawrence Berkeley National Laboratory estimates data centers could take about 12% of US electricity by 2030. Somebody pays for new plants, lines, and substations. The fight is whether the project pays its share or the cost lands on every other bill.

A recent MIT study of data-center growth from 2010 to 2024 estimated average retail power prices rose about 2.7%, including about 2.1% for homes, with a larger effect at investor-owned utilities and a smaller one at public power and co-ops. Another study found demand can lower rates when a utility sells more power across the same fixed network. Farzad’s reconciliation: spare capacity plus a customer that runs around the clock and pays full freight can spread fixed costs. Billions in upgrades for a customer that under-delivers or leaves do the opposite. Contract, available supply, and rate rules decide which world you get.

Virginia is his example. It hosts the largest US concentration of data centers. A state review found current rates generally recovered cost of service, and also warned that fast future growth could add $14 to $37 a month to a typical home bill by 2040 if the state does not change how new costs are assigned. So yes, data centers can raise bills when execution is bad. They can also lower them when execution is good.

Water: real local pain, viral math that was wrong

Lawrence Berkeley put direct US data-center water use at about 17.4 billion gallons in 2023. Use for the same compute can vary by more than 10,000 times depending on servers, cooling, climate, the local grid, and utilization. Closed-loop cooling recirculates like a car radiator. Evaporative cooling in a dry climate is a different animal. Google’s The Dalles, Oregon site used roughly 355 million gallons in 2021, about 29% of the city’s water. Residents asking about drought days are not inventing a problem.

Then the viral miss. Farzad pins a ~1,000x water error on Karen Hao’s Empire of AI and a proposed Google site in Uruguay. Hao’s published correction, Dec. 17-18, 2025, was about a Chile comparison (Cerrillos, near Santiago): a liters-versus-cubic-meters unit mix-up made the project look like more than 1,000 times local residential use. Uruguay is a separate episode in the book. The bad figure still travels. Closed-loop designs use far less water than the scare posts imply. Golf facilities in the US in 2024 used about 30 times more water than data centers, he says. California almond orchards used about 90 to 100 times more. His punchline: protest bad zoning and desert evaporation, not the category as a whole. Putting a thirsty plant on a strained aquifer is still irresponsible.

Noise, diesel, jobs, and tax theater

Fans, chillers, pumps, transformers, and backup generators make noise. The Virginia review found most sites generated no noise complaints, and also found real complaints at some sites, with a warning that ordinary ordinances can miss low-frequency hum. Backup diesels are usually for testing and outages, but a big campus can host hundreds of them. Regional pollutant share from those generators has been relatively small in Northern Virginia’s review, and can still be a local problem next to homes or schools. Farzad’s fixes are cleaner backups (gas or grid-scale batteries), limits on testing during bad air days, and industrial zoning that keeps the box away from bedrooms.

Construction is the labor boom. A large Virginia project can peak around 1,500 construction workers for 12 to 18 months. Once live, the same facility may support about 50 full-time jobs, roughly half contractors. So when a politician or AI CEO sells “thousands of permanent local jobs,” Farzad wants skepticism. Short-term hiring is real. Steady headcount is thin. A 2026 NBER study found more building and data-processing jobs, firms, wages, reported income, and home values near openings, plus evidence of higher power prices over time. Tax revenue can be huge. Loudoun County says data-center money has funded schools, roads, fire stations, libraries, and parks while helping lower property tax rates. St. Louis approved a project with a binding local deal and projected hundreds of millions in tax revenue over ten years. Georgia’s own DOAA evaluation of its data-center tax break found a negative fiscal return under the assumptions it studied: forgone state sales-tax revenue of about $474 million versus about $41.5 million in credited additional collections. Virginia estimated its exemption saved the industry about $928 million in one fiscal year. Subsidy theater can erase the fiscal win.

The dumbest fix: a national pause

Bernie Sanders has campaigned for pausing new US data centers on AI-danger grounds. Farzad does not call him a bad person. He calls the moratorium one of the worst ideas on offer. Software crosses borders. If America stops building and China keeps going, the US exports construction, energy investment, know-how, tax base, ownership, and influence, then imports the automation shock over the internet anyway. China treats compute as power, consolidates hubs, and has been adding generation at a pace Farzad cites as more solar in a year than the US has total, plus about 543 gigawatts of new capacity turned on in 2025 (Farzad’s about-half-the-US-grid framing; EIA put US utility-scale capacity near 1,280 GW at year-end 2025). A federal pause changes who owns the machines. It does not stop the race.

The buildings are not warehouses for chatbots. They are the factories for agents that can book travel, inspect codebases, write and test software, research, and ship finished work. Tools like Claude Code, Codex, Grok Bot, Hermes, OpenClaw, Replit, and Bolt already show the demand. When that usage scales to hundreds of millions of people and businesses, compute demand is not a vibe. It is a load.

His medical proof point: Moderna and Merck announced Aug. 19, 2026 positive topline Phase 3 results from INTerpath-001 (NCT05933577) for a personalized mRNA melanoma treatment (intismeran autogene / V940 / mRNA-4157, paired with Keytruda). After surgery, the tumor is sequenced, mutations unique to that cancer are identified, a machine-learning system picks up to about 34 neoantigen targets, and a personalized mRNA teaches the patient’s immune system what to attack. The trial enrolled 1,137 patients. The combo beat Keytruda alone on recurrence-free survival (primary) and distant metastasis-free survival (key secondary) at a pre-specified interim analysis. Hazard ratios were not in the topline. Farzad’s point is not that every data center is a cancer lab. It is that calling the category wasteful while personalized oncology is running on this stack is a broken frame.

Cost protection plus an abundance dividend

His deal has two layers. First, cost protection: the company pays for every new power, water, sewer, road, and emergency-service cost the project creates, signs a long-term contract, and eats clear water, noise, and pollution limits. Violate them and pay a heavy fine, ideally a cut of earnings from that site until fixed. Second, an abundance dividend sold in a public bid. One bidder asks for a tax exemption. Another offers to cover baseline electric and water bills for every home in the host town, train hundreds of local apprentices, fund grid upgrades that cut outages, and pour yearly revenue into lower property or sales taxes or a statewide fund for schools and downtown rebuilds. The town is selling scarce land, grid access, water capacity, and trust. The operator is buying the right to put capital and compute there. Today’s system often runs the auction backwards: officials compete on the biggest tax break, deals close behind closed doors, and residents learn when the land behind the house gets cleared.

He is not naive about white-collar automation. Construction jobs will not replace every accounting, legal, support, design, or software seat that agents change. That is another reason he wants public ownership of upside: training, job support, retirement accounts, public funds, cheaper goods, and towns that keep a claim on the dividend. Cars kill about 40,000 Americans a year and we still drive. We add brakes, licenses, crash tests, seat belts, insurance, and roads. Electricity can kill people and we still wire cities. AI, in his framing, needs the same pattern: build the safety system and move faster. A pause hands the stack to other countries and to the few firms that already own it.

This Exclusive is from the long-form at https://www.youtube.com/watch?v=ZQ-rS45A4Gk.

---

Receipt

Event → Farzad proof → book beat → buy link.

• Event: Public Exclusive on the data-center power, water, jobs, and moratorium fight (ZQ-rS45A4Gk).
• Farzad proof: LBNL / MIT / Virginia / Loudoun / Georgia numbers stress-tested against viral water and job claims; abundance-dividend bid structure instead of a national pause.
• Book: AoC — align incentives so locals own upside when compute lands next door. Master Plan Energy — who pays for the electrons that feed the factory.
• Get the books: Abundance or Collapse · Master Plan