The 1-Megawatt Arms Race

This summer, two documents landed on opposite sides of the same number. Runware launched its Sonic Inference Pod: a data center packed into a 20-foot shipping container, drawing 1 megawatt of power, installable in a single day. The company pitched it explicitly as a way around "power constraints and construction delays associated with traditional data centers" ^[1]. Weeks earlier, Hamilton County, Tennessee passed a moratorium on new data center development. The trigger: any facility with a peak electric demand of 1 megawatt ^[2]. The same number, chosen independently by an engineering team and a county board. It is not a coincidence. The AI industry, facing a wall of local opposition that has turned data center permitting into a multi-year gamble, is shrinking the unit of deployment below the thresholds that trigger a town's power to say no. And towns, recognizing the strategy, are lowering those thresholds to catch what comes next. The shrinkage ladder starts with the modular pod. Runware's containerized design packs roughly 1,200 GPUs into a box that arrives on a truck. Datalec and PODTECH partnered this year to deliver "pre-staged, deployment-ready AI infrastructure" targeting what they call "commissioning bottlenecks caused by hyperscaler demand" ^[3]. A pod bypasses the construction timeline that gives opposition time to organize. No multi-year build, no foundation pour that triggers a public hearing, no window for a moratorium to pass while the concrete cures. The next rung removes the zoning fight. HPE launched its AI Grid with NVIDIA in March, a distributed edge inference architecture that places compute at telco edge sites rather than centralized campuses ^[4]. Akamai is investing heavily in the same direction, explicitly positioning distributed edge cloud as the alternative to hyperscale data centers ^[5]. A server rack on a rooftop that already has a lease and a power line does not need a zoning variance. It does not even need to be called a data center. Then the retrofit. In Colorado, developers are converting shuttered cannabis cultivation facilities, buildings that already have high-capacity electrical service and climate-control systems, into data centers without triggering new construction permits ^[6]. Vertical Data is doing the same in urban real estate, retrofitting existing buildings for liquid-cooled AI compute. CEO Deven Soni made the economics explicit.

Urban retrofits allow Vertical Edge to deliver capacity faster than new builds, and this market is the kind of location where that timing advantage is most valuable. — Deven Soni

A retrofit inherits the permits the building already has. There is no public hearing for a change of tenant. The next rung skips the grid queue entirely. Google, facing a utility that quoted a 12-year timeline just to complete a mandatory interconnection study, began building data centers directly adjacent to power plants ^[7]. Google's Marsden Hanna identified the bottleneck.

We had one utility who told us 12 years to study the interconnection timeline, which is sort of wild, but that's what we're seeing — Marsden Hanna

Oracle's Stargate project uses behind-the-meter natural gas. xAI deployed temporary gas turbines for its Colossus facility. Burns & McDonnell, the engineering firm, stated the new normal.

To meet this need for speed, on-site generating capacity and behind-the-meter distribution are becoming standard considerations. — Burns & McDonnell

A gas turbine on your own land needs no interconnection study, no transmission line, no queue. Then the jurisdictional escape. The Defense Department is seeking private AI data centers on unused military base land, a 207-acre site at Fort Hood, Texas, and framing them as critical military infrastructure ^[8]. Federal land means federal jurisdiction. A town board has no authority over a base commander's decision. And then orbit. Aetherflux is pursuing "Project Galactic Brain." Google has "Project Suncatcher." Orbital plans to launch its Orbital-1 AI compute satellite in 2027. Eric Schmidt is backing the effort through Relativity Space ^[9][10]. The pitch is energy and cooling, but the structural advantage is jurisdictional: there is no town board in low Earth orbit. None of this means the hyperscale campus is dead. Amazon is simultaneously redesigning its Indiana campus as an "AGI SuperCluster" with more than 6,000 Trainium servers and increasing 2026 capital expenditure to $220 billion ^[11]. The shrinkage strategies are a parallel track: a way to keep building while the big sites fight through permitting. But the towns have read the same playbook. Hamilton County's 1-megawatt threshold is the clearest signal: a number chosen not to catch a 500-megawatt hyperscale campus, which any moratorium would cover, but to catch a shipping-container pod. Little Rock's rejected moratorium had targeted projects over 250,000 square feet or 75 megawatts ^[12]. Hamilton County dropped the floor by two orders of magnitude. The more significant shift is procedural. Towns are no longer waiting for an application to arrive. Pere Marquette Township's Louis Prima stated the position plainly.

I wanted to make sure you knew this wasn’t in response to any type of application for a center. — Louis Prima

Southport's Lawrence Ashley was even more direct.

The moratorium is simply being proactive in the fact the township knows it’s not addressed in our ordinance and we need to do so. — Louis Prima

A proactive moratorium removes the one advantage a pod deployment has over a campus: speed. If the rule is already in place, there is no race to win. The zoning is also getting more granular. Fayetteville, North Carolina is reclassifying data centers as conditional uses requiring energy management plans and water projections. Joplin, Missouri restricts them to heavy industrial districts with noise limits and closed-loop cooling. Pottawatomie County, Kansas is creating overlay districts with height limits and power-line proximity requirements ^[13]. These are filters designed to catch the facility that would otherwise slip through under a generic industrial-use classification: the kind a modular pod or a rooftop rack might claim. New York State raised the stakes further, becoming the first state to implement a one-year statewide moratorium on hyperscale data center permitting ^[14]. A state-level consent wall catches what a town-level one cannot: the developer who crosses the county line to the next jurisdiction. Meanwhile, Energy Secretary Alfie Moon dismissed local opposition outright.

We’ll roll over the opposition to data centers faster than we otherwise should. — Alfie Moon

The federal push to override local consent and the local push to preempt federal bypass are now running simultaneously. The logical endpoint of the shrinkage strategy is no data center at all. Perplexity CEO Aravind Srinivas made the logic explicit.

The biggest threat to a data center is if the intelligence can be packed locally on a chip that’s running on the device and then there’s no need to inference all of it on like one centralized data center. — Aravind Srinivas

Tianrong Internet is already building a decentralized GPU marketplace that lets individuals rent idle gaming PCs for inference, cutting costs 50 to 80 percent against centralized cloud ^[15]. If the intelligence packs onto a chip, the facility disappears. Every point of leverage a town ever had disappears with it. The industry has shown it can go from campus to container, from container to rooftop, from rooftop to orbit. The towns have shown they can go from 75 megawatts to 1, from reactive to proactive, from zoning board to statehouse. Neither side is conceding. Both are aiming at a smaller target. The question is whether the democratic throttle can keep shrinking faster than the deployment unit. And what happens when there is nothing left to permit.