xAI Got Permits for 15 Turbines and Ran 35 — CNBC in Memphis
CNBC's Memphis report is not really about pollution. It is about what happens to an AI buildout when the binding constraint stops being chips and becomes power, permits, and the patience of the people living next to it.
Most coverage of xAI’s Memphis data centers is an argument about air quality, and you already know which side you are on. The more useful way to watch CNBC’s report is as a case study in what actually rate-limits an AI buildout once the GPUs are ordered — and what it costs to discover that the hard way.
The engineering fact underneath the politics
The most interesting thirty seconds of the report have nothing to do with permits. xAI’s original plan was to run Colossus off-grid on leased gas turbines, and it failed for a reason that is not obvious:
“They learned very quickly that that wasn’t going to work because the data center was blowing out a lot of the natural gas turbines because they’re so hard on turbines. That’s when we learned that building stuff off grid didn’t really work.”
The resolution was to go back to the utility:
“So then TVA had figured out a way to upgrade their service and connect the Colossus data center to the grid. So it is now connected to the grid, but it continues to have natural gas units and batteries behind there, just in case TVA needs to turn them down.”
The permitting sequence, in order
CNBC lays out a sequence worth reading as a timeline rather than an accusation:
“First xAI put up a slew of gas burning turbines before it had the required permits at Colossus 1. XAI said at the time that it didn’t need permits because they were temporary. Then it got permits for 15 turbines but was still operating unpermitted turbines.”
What advocacy groups found when they looked:
“Come summer 2025, some of the environmental advocacy groups, they discovered those 35 turbines at that site. We used a thermal camera, and 33 out of the 35 appeared to be operating that day when we flew over.”
An attorney with the Southern Environmental Law Center supplies the scale comparison that makes the number legible:
“I’ve been working in air pollution a long time. I’m used to a power plant that might again, have six of these same turbines, and they’ve got 35 of them with 33 operating.”
By the time of filming the count at Colossus 2 had grown past fifty, and the Department of Justice had filed a motion to intervene.
The bill shows up somewhere else
The part that generalises to any hyperscaler is what this does to the cost of the next project, anywhere. One analyst frames the shift:
“When data centers moved from a thing that got built on the grid, sometimes to this hyperscale era that we’re in now, where overnight we’re seeing data centers permitted that are the electricity load of an entire city dropped into an electricity grid overnight. We immediately saw price spikes from that and grids across the country.”
The numbers attached:
“Electricity prices in Virginia have gone up 200% in the last two years. The people of Pennsylvania have paid $492 million in upgrades to the electricity grid that went to support data centers. Data centers should be paying for that instead.”
And the policy response, which is the actual risk to model:
“New York just became the first state to impose a data center moratorium, and New Jersey has enacted a bill to ensure large data centers pay their fair share for electricity.”
| Constraint | Status in the report |
|---|---|
| Chips | Not mentioned as a limit |
| Grid interconnection | The binding one; off-grid workaround failed on engineering grounds |
| Water | Colossus 1 alone can use 1.28 million gallons/day, not currently recycled |
| Permits and social licence | Now producing moratoria, cost-allocation statutes, and federal litigation |
The contrast the report draws
CNBC notes that Google did not respond to its request for comment, and that Anthropic “is having active conversations with the mayor and the community, which will shape its approach.” Memphis officials also negotiated a community benefit ordinance covering a five-mile radius around the facilities.
Whether that consultative approach is sincere or merely better-advised, it is cheaper. The adversarial path here has produced a DOJ intervention, thermal-camera surveillance by opposing counsel, and named inspiration for federal legislation — Senator Markey’s AI accountability agenda — plus neighbouring towns rewriting zoning specifically to keep the next one out. That is a durable increase in the cost of building, and it lands on everyone in the sector, not just the firm that caused it.
Where to push back
This is advocacy-adjacent journalism and it is edited like it. Residents describe noise and fumes; a lawyer for the plaintiffs supplies the framing; xAI’s position appears mainly as reported denials. The report does not include measured ambient air quality data attributable to the turbines, which is the evidence that would settle the central question, and it does not seriously engage the counterargument that Memphis wanted the investment.
The economics are also presented one-directionally. Virginia electricity up 200% is a real number, but data centers are not the only variable in it, and the report does not attempt to separate them. “Data centers should be paying for that instead” is a policy preference stated as a finding.
None of that touches the two facts worth carrying: the turbines failed for engineering reasons, and the regulatory response is now statutory in multiple states. Both are true regardless of how you weigh the rest.
Worth watching
Twenty-nine minutes. The off-grid failure and the permit sequence run from about 10 to 14 minutes and are the densest part. The grid-economics section around 24 minutes is where the numbers you would put in a model live. Musk’s own answer to all of it — put the data centers in orbit, because “if we go to space, we can go far beyond the electricity generation of Earth” — appears at about 20 minutes, and tells you how binding he considers the terrestrial constraint to be.
Related reading
2026-09-07
300 AI Query Optimizations Went In, 30 Came Out — Datadog at DASH 2026
Most of Datadog's two-hour keynote is a product reel. One slide is not: the vendor selling you AI query optimization discloses that 90% of its model's suggestions failed validation.
2026-09-07
Andrew Ng Won't Sign an AI Contract Longer Than a Year — Interrupt 26
Ng's fireside chat at LangChain's Interrupt has one piece of advice with a number attached, and one example that explains why most enterprise AI projects produce a rounding error instead of growth.
2026-09-07
He Renamed One Function and the AI Did More Work — Alexandrescu at ACCU 2026
Andrei Alexandrescu's ACCU keynote argues abstraction survives AI-generated code for an unfashionable reason: not because humans need it, but because throwing it away is inefficient. He has an experiment to show it.
2026-09-04
'Unmetered Intelligence' Moves the Bill, It Doesn't Remove It — Nadella at Build 2026
Microsoft re-ran its founding slogan for the AI era and pushed inference to the edge. The per-token meter does come off — and reappears as hardware you buy up front.
2026-09-04
Every AI Panic Point Has a Precedent. One Doesn't — Mary Shaw at ICSA 2026
Thirty years after co-writing the book that named software architecture, Mary Shaw walks through the field's AI anxieties and shows most of them are re-runs. Then she names the one that isn't.
Get the best tools, weekly
One email every Friday. No spam, unsubscribe anytime.