Used gemini and claude to compile.
I’ve been thinking about physical bottlenecks to AI scaling due to energy infrastructure and whether existing environmental regulations represent an underutilized lever for buying alignment time.
Transit example
In transit systems like MNR's diesel electric locomotives, they use a Cummins diesel engine for propulsion in non-electrified territory which has to comply with CAA Tier 4 EPA regulation for NOx emmisions. To meet the emmission limit the MNR locos have an DEF (diesel exhaust fluid) aftertreatment system that essentially mixes ammonia with exhausted NOx to produce nitrogen and water vapour before it goes into the atmosphere.
The regulatory arbitrage.
The grid is too slow and congested to meet compute demand, so data center builders are generating their own power behind the meter. They use the same family of off-the-shelf industrial engines (Cummins and Caterpillar units) for backup power along with gas turbines for prime power. Depending on how that equipment is classified, it can face much weaker controls than it would on a train:
- Gas turbines for prime power: Mobile vs. stationary. Mobile sources are certified at the factory, and states are generally preempted from setting their own emission standards for them. Stationary sources need state-issued site permits which could require New Source Review/PSD preconstruction permits, with Best Available Control Technology, air quality modeling, public comment, and timelines that often run a year or more.
- Diesel backup generators: Emergency vs. non-emergency. Stationary engines classified as emergency backup generators are held to older, looser emission limits than non-emergency engines. So the engine family that needs DEF aftertreatment on the MNR locomotive (tier 4) can typically be installed at a data center without it (tier 2). The latter is an inference.
EPA’s recent interpretations have leaned toward giving AI facilities flexibility:
- Acid Rain Program guidance for Islanded Power Generating facilities (July 2026): This basically says Clean Air Act Title IV guidance on Sulfur Dioxide and Nitrogen Oxide emission limits don't apply to power generation facilities disconnected to the larger electricity grid (aka not applicable to an "islanded facility" that doesn't sell electricity as a "utility"). Other Clean Air Act obligations, such as construction and operating permits, can still apply.
- EPA Issues Clarification to Help Power Data Centers, Ensure U.S. Is the AI Capital of the World (May 2025): This says EPA clarified that data centers can run their emergency backup generators for up to 50 hours a year in non-emergency demand-response programs and broadened what counts as an emergency.
xAI Colossus Case
For xAI’s Colossus 2 site in Southaven, Mississippi, the NAACP sued, alleging that roughly 27 gas turbines were built and run without the permits and pollution controls the Clean Air Act requires. xAI argues the turbines are temporary mobile units and therefore exempt from stationary-source permitting and Mississippi regulators have accepted that. Under state rules, mobile generators don’t need an air permit if they run for less than a year. In June, the DOJ moved to intervene and dismiss the suit, arguing that the data center is critical to the economy and the military.
Argument
If data centers didn’t have unrestricted access to on-site gas power, and their demand is too large for a quick grid connection, AI development would inevitably slow. Developers could still comply, by adding aftertreatment and emissions controls or by switching to renewables and storage. But either path means longer design-build cycles. To me, this looks like the biggest concrete slowdown lever not being pursued heavily, and it already has a live legal case and existing law behind it.
Is anyone in the AI safety/governance space actively advocating for closing these environmental loopholes to regulate AI energy use from a slowdown perspective?
Again, it seems like one of the most concrete, legally grounded levers available to slow down frontier compute scaling. Why it seems promising:
- Precedent & Existing Frameworks: Unlike bespoke AI safety legislation, which faces massive political friction and definitional hurdles, the Clean Air Act and EPA emissions frameworks already exist. Enforcing strict stationary-source permitting and emissions caps on behind-the-meter data center power doesn't require proving catastrophic AI risk. It just requires applying existing environmental law consistently.
- Forcing Longer Build Cycles: Even if hyperscalers aren't forced to transition to renewables by stricter environmental rules, compliance takes time. Forcing developers to either wait for grid interconnects, build out renewable microgrids, or engineer aftertreatment and emissions control into their on-site generation would noticeably extend design-build cycles. Seems like a win-win for air quality or AI slowdown. Waiting for the grid, building renewables plus storage, or engineering emissions controls all make design-build cycles longer. From my own work, I’d guess that’s months to years, not weeks. The MNR locomotive shows the compliance technology exists and works, but it isn’t free or fast.
- It has allies who aren’t in EA. Environmental justice groups are already bringing these cases.
From an engineering perspective, compute is fundamentally constrained by physical thermodynamics and mineral/energy infrastructure. By the time a chip cluster is turned on, the upstream infrastructure decisions have already been locked in.
Why this might be a bad idea, or a non-starter:
- Displacement: Builders might move to states with looser rules, or abroad, so the net delay could be small. If compute moves to places with weaker oversight, the result could even be worse for safety.
- Mixed motives: Using environmental law mainly as a slowdown tool could damage trust with the communities and lawyers who bring these cases for their own reasons. It could also invite a political backlash that weakens the Clean Air Act itself.
- Small effect: Compute may be limited mainly by chips rather than power, or the delay may amount to months rather than years. In the xAI case, Mississippi has already approved permits for permanent turbines at the site, which suggests enforcement may slow the bridge period more than the build itself.
- Geopolitical Trade-offs: The obvious pushback and the exact reason the DOJ recently moved to dismiss a Clean Air Act suit in Mississippi citing national and energy security is the fear of slowing down relative to China. Trade-off here: do we accept localized environmental deregulation to win a perceived race, or do we enforce systemic boundaries to ensure safe, sustainable scaling?
Following, my questions:
- Has the EA or AI governance community looked seriously at environmental permitting as a compute governance tool? Is it a bad idea? why?
- Has anyone estimated how much of a delay strict enforcement of CAA regulation would cause?
- Is that delay net-positive once you account for compute moving to another state/country?