Somebody Has to Host This

I run Claude on a Mac Studio in my home office. The machine hums at about 40 watts, moves no detectable air, and handles AI inference that would have required a rack of servers five years ago. I mention this because somewhere in Maryland, a resident is convinced the data center three miles away is preventing snow from falling in their neighborhood.

We are not having a serious conversation about data centers.

The claim that air above a hyperscale facility runs 70 degrees hotter up to a mile in the sky has been circulating through community opposition meetings and local news coverage for two years. Arizona State University researchers drove temperature sensors around four operating Phoenix-area data centers last summer, the first peer-reviewed study to measure ambient air temperatures directly rather than relying on satellite land surface readings. The findings were specific. Exhaust air runs 14 to 25 degrees above ambient at the point of discharge. The effect is measurable to about a third of a mile at a maximum of 4 degrees above background. At one mile, the signal is gone. The 70-degree figure has no scientific basis anywhere. The snow claim has no scientific basis either, and would require atmospheric physics that do not apply to a building with cooling towers.

The real concerns are serious enough that manufacturing additional ones seems counterproductive. A 100-megawatt facility uses roughly 530,000 gallons of water per day for cooling. US data centers consumed 176 terawatt-hours of electricity in 2023, about 4 percent of national consumption, and that figure is projected to reach between 325 and 580 terawatt-hours by 2028 depending on how aggressively AI workloads scale. The grid load is real. The water draw is real. The land use is real. These deserve honest policy conversations, not viral weather claims layered on top of them.

The Alternative Is Worse

The argument almost never made in the community opposition conversation is the geopolitical one. If the United States does not build domestic AI compute infrastructure, other countries will, and those countries will control the underlying layer of an economy that is becoming dependent on AI the way it became dependent on semiconductors in the 1990s.

Gladstone AI, which consults for the US government on AI risk, published a report in April 2025 arguing that every major US AI data center, including the Stargate facilities, is vulnerable to Chinese asymmetric sabotage and model exfiltration. The concern is operational, not theoretical. The European Union is committing €200 billion to domestic AI compute infrastructure because member states have decided they cannot trust either American or Chinese cloud providers with sovereign data. Every nation capable of doing the math is now making this calculation.

Chips need fabs. Models need compute. Compute needs power, water, land, and physical security. You can oppose all of it on principle, but your data still lives somewhere. The question is whether it lives somewhere you can govern.

The grid load, water draw, land use, and impact on communities that did not sign up to become industrial corridors should be priced honestly, and the people bearing them should be compensated honestly. What they should not be given instead of honest answers is junk science about regional weather patterns.

Follow the Money, Then the Passport

The national security argument for data centers is sound. The people making it have a financial interest worth noting.

Peter Thiel, co-founder of Palantir, a company whose entire business is built on data infrastructure and government data processing, relocated his family to Buenos Aires in late 2025. The timing was precise. A California ballot initiative would have cost him roughly $1.4 billion on his $28 billion fortune. He retains US citizenship and keeps more than 99 percent of his assets in the United States. Argentina is a hedge, not a relocation. He finds good company there in Javier Milei, whose politics align neatly with Thiel's on the subjects of taxation and government regulation.

The gap between what Thiel's US-based investments earn and what he has optimized to contribute back is not small. He is not unusual in this configuration. He is an example of it.

Masayoshi Son, the primary face of the Stargate announcement alongside Sam Altman and Larry Ellison, is a Japanese national. The UAE's MGX fund and Saudi Arabia's sovereign HUMAIN program are significant Stargate investors. Foreign state capital is financing the infrastructure that American communities are being asked to host, on the argument that it is an American national security asset. That argument can be accurate and the incentive structure can still be worth examining at the same time.

Virginia, Texas, and Georgia, the three largest US data center markets, compete for projects by offering property tax exemptions worth hundreds of millions over a project's life. The community accepts the industrial footprint, the grid draw, the water consumption, and a reduced local tax base. The investor accepts the return plus the subsidy, and in some cases structures their residence to minimize further contribution.

The case for building the infrastructure does not depend on the investors being patriotic. It depends on the math being right about what happens if we don't. But the community resistance is not irrational when the costs flow one direction and the returns flow another. The policy question nobody in this debate is asking seriously enough is the simplest one. What does the host community actually get in exchange for what it gives?

The Hostage Problem Nobody Is Talking About

The deeper risk from data center dependency is not the one showing up in town hall meetings. It is the pricing risk.

Current AI services are priced below cost. Anthropic and OpenAI are buying market share with investor capital. Amazon raised GPU prices 15 percent for certain ML training jobs in January 2026. The companies that cut staff, built workflows on Claude or ChatGPT, and stopped hiring the roles those tools replaced have limited ability to revert when pricing starts to reflect actual infrastructure costs. This is the Oracle and SAP enterprise software lock-in story compressed into three years and applied to cognitive labor instead of workflow software. The workflow is easier to rebuild than the people.

This is an argument for understanding what dependency looks like before you are inside it.

The Moratorium and What It Means

The political resistance is larger than most people following the technology conversation realize. Since mid-2024, $18 billion in US data center projects have been formally blocked and another $46 billion delayed by community opposition. Senators Sanders and AOC introduced a federal moratorium bill in March 2026. There are 188 organized opposition groups operating in 40 states.

The moratorium movement's strongest argument is not about heat plumes or snowfall. It is about process. Communities should have standing in decisions about major industrial infrastructure sited in their neighborhoods. That is a legitimate argument. It is also entirely separate from the claim that data centers are creating regional weather anomalies, and letting the junk science carry the load weakens the legitimate case.

The irony is precise. The communities blocking data center construction are also the communities whose members are using the AI services those data centers host. Someone has to host the models. The question is not whether to build, but where, under what terms, and with what genuine accountability for the real impacts.

What Comes Next

I mentioned the Mac Studio at the start for a reason. Apple's on-device AI runs entirely locally on A-series and M-series silicon. Google's Gemini Nano ships on Android flagships. Meta's Llama 3.2 runs on consumer hardware with 16GB of RAM. The inference workload, which already represents more than half of total AI compute spending, is actively migrating toward edge devices. In April 2026, Span and NVIDIA announced a pilot of residential inference nodes. The hardware is roughly the size of an air conditioner, installed in a side yard, running AI inference at roughly one-fifth the cost of centralized capacity. Homeowners get discounted utilities. NVIDIA gets distributed inference infrastructure. The pilot targets 80,000 nodes by 2027.

The centralized data center, as the exclusive location for AI inference, has a shorter future than the opposition groups realize. The question is what comes between here and there.

Build the data centers. Price the costs honestly. Stop making up the numbers.

Frequently Asked Questions

Is it true that data centers make the air 70 degrees hotter a mile above them?
No. A peer-reviewed Arizona State University study published in 2025 directly measured air temperatures around four operating Phoenix-area data centers. Exhaust air runs 14 to 25 degrees above ambient at the point of discharge. The effect is measurable to about a third of a mile at a maximum of 4 degrees above background. At one mile, no measurable temperature difference exists. The 70-degree figure has no scientific basis.

Do data centers prevent snow from falling?
There is no scientific evidence supporting this claim. Data center thermal plumes are localized to the immediate facility perimeter and do not produce regional weather effects. The claim has no basis in atmospheric physics.

How much electricity do US data centers actually use?
US data centers consumed 176 terawatt-hours in 2023, approximately 4 percent of national electricity consumption. Projections for 2028 range from 325 to 580 terawatt-hours depending on AI workload growth, roughly doubling to tripling from today's baseline.

Why do we need to build more data centers if there are legitimate environmental concerns?
The geopolitical argument is straightforward. AI infrastructure requires physical compute capacity, and that capacity will be built somewhere. If the United States does not build it domestically, competitor nations, particularly China, will control the foundational infrastructure layer of a globally AI-dependent economy. The European Union is spending €200 billion on domestic compute for exactly this reason. The question is not whether to build but under what environmental accountability terms.

Will data centers always be necessary for AI?
Not in their current centralized form. The AI inference workload, currently more than half of all AI compute spending, is actively migrating to edge devices. Apple, Google, and Meta are all shipping on-device AI inference today on consumer hardware. In April 2026, NVIDIA and Span announced a pilot of residential inference nodes targeting 80,000 installations by 2027. The centralized inference model is under real pressure. Training infrastructure will remain centralized for the foreseeable future.

What is the data center moratorium movement?
Since mid-2024, $18 billion in US data center projects have been formally blocked and $46 billion delayed by community opposition. Senators Sanders and AOC introduced a federal moratorium bill in March 2026. There are 188 organized opposition groups in 40 states. The legitimate core of the argument is that communities should have standing in decisions about major industrial infrastructure. The weaker version is the environmental misinformation about heat and weather that often accompanies it.