Robert Reich and the Moratorium Moment

Robert Reich's argument arrives compressed and effective. In a short video for Inequality Media titled "Are Data Centers Really That Bad?", the former labor secretary lays out a chain that many viewers already feel intuitively: AI data centers are enormous consumers of water and electricity, some are proposed in places already short of water, their power arrangements can pollute the air nearby, they are noisy, they collect large public subsidies while employing relatively few people, and the communities that try to say no can be outspent and outlawyered. The conclusion is a national moratorium while democratic institutions catch up.

The backlash is real. The evidence is uneven. The policy response cannot afford to confuse the two.

That is the premise of this article, the first in a Praevoium series called The Capacity Compact. The goal is not to defend the AI buildout or to endorse Reich wholesale. It is to separate four things that political argument tends to blur: verified concern, fragile statistics, rhetorical framing, and policy judgment — and then to propose a more durable unit of accountability.

The video arrives at a charged moment

Reich is not a technologist, and the video does not pretend to be an engineering assessment. His authority is as a former U.S. labor secretary and a long-running public voice on economic power. His framework treats economic questions as contests over who holds leverage, and in this video the physical facts of data-center construction become evidence in that larger argument about concentrated power.

That framing matters because it explains both the strength and the limits of the piece. The strength is that Reich is describing a genuine mismatch between the scale of what is being built and the capacity of the systems meant to evaluate it. The limit is that a power-and-inequality lens can convert every disputed permit, every high-end number, and every local pause into evidence for a single national remedy. Those are different kinds of claims, and they need to be tested separately.

The argument, reconstructed

Reich's case runs in sequence: large facilities can consume enormous quantities of water and electricity; some are proposed in drought-affected regions; on-site or backup generation can create local pollution; states and localities grant large tax incentives despite limited permanent employment; nearby residents face noise and other effects; public opposition is widespread; and local governments that resist may face litigation or political pressure. From there the argument turns interpretive — these disputes reflect concentrated technology-sector power — and finally prescriptive: a national moratorium is needed.

The first seven links are factual questions that can be checked. The eighth is an interpretation. The ninth is a policy judgment. Reading them as one seamless claim is how a strong argument picks up weak parts.

What Reich gets right

The most durable part of Reich's case is also the least dramatic: governance has not kept pace with scale. Approval systems built for warehouses and office parks are now being asked to evaluate campuses that request hundreds of megawatts, recurring tax exemptions, major water commitments, and large fleets of on-site or backup generation. The public record of the past two years — temporary planning pauses, new large-load tariffs, fresh environmental-review processes, state policy frameworks, and organized local resistance — is itself evidence that ordinary approval systems were often inadequate to the task.

Public opposition is substantial, and it can be stated precisely. Gallup found that 71% of Americans opposed construction of an AI data center in their local area, including 48% who strongly opposed it. That figure supports Reich's headline so long as the wording stays exact: it measures opposition to a facility in one's own area, not opposition to all data centers, to AI itself, or support for a national moratorium. Its real value is what it reveals underneath — a burden-benefit split in which opponents emphasize water, energy, pollution, noise, and quality of life, while supporters emphasize jobs and tax revenue.

Upper-range water demand is real. The Environmental and Energy Study Institute states that some of the largest facilities can approach five million gallons of water per day, comparable to a town of roughly 10,000 to 50,000 people. That is a legitimate high-end figure — and it is not the industry average. Whether it matters at a given site depends on questions the headline cannot answer: whether the number describes withdrawal or consumption; whether the source is potable, reclaimed, surface, or groundwater; what peak summer demand looks like; how the facility behaves under drought restrictions; and whether it uses evaporative, dry, hybrid, or closed-loop cooling.

The drought-location concern holds within its methodology. A Guardian analysis found that 517 of 809 planned U.S. data centers — about 64% — were in places that had experienced drought in the preceding year. That supports Reich's worry, with a caveat: a drought history is not permanent water incapacity. Watershed conditions, aquifer stress, utility capacity, source-water type, reclamation, and drought protocols determine actual risk. The finding is best read as a reason for project-level water review, not a verdict on any single site.

Electricity-demand growth is no longer speculative. Research from Lawrence Berkeley National Laboratory and the Department of Energy documents rapidly rising U.S. data-center electricity use and projects a materially larger share of national demand by 2030. That validates Reich's scale argument. It does not, by itself, prove higher residential bills — rate effects turn on generation availability, transmission, tariff design, minimum-payment obligations, and whether developers bear the cost of the upgrades their loads require.

Subsidy exposure can reach into the billions. Good Jobs First has documented large, and sometimes poorly disclosed, state tax expenditures tied to data-center incentives, and the broad claim that the public is forgoing billions in revenue is supportable at that scale. The sharper phrase — "for little in return" — is not automatically supportable nationally. It has to be tested project by project against taxes actually paid after exemptions, infrastructure costs, permanent employment, construction job-years, local procurement, utility risk, and whether the project would have located there without the incentive at all.

Noise is a genuine governance issue, not merely a complaint. Cooling systems, transformers, substations, generator testing, and on-site turbines can produce continuous tonal or low-frequency sound, and standard commercial noise ordinances often fail to address exactly that: nighttime background levels, tonal character, and the cumulative effect of later phases. Reich's blunt phrasing about noise is rhetorical, but the underlying design-and-enforcement gap is real.

Finally, small governments can face real legal asymmetry. The Saline Township dispute in Michigan supports the claim that a small local government may face expensive litigation after rejecting a large proposal. The neutral formulation is that the developer and landowners challenged the township's denial under exclusionary-zoning theories, and the township later entered a consent agreement allowing the project under negotiated conditions. That is evidence of unequal legal capacity — not proof that the challenge was itself illegitimate.

The bottle-per-email problem

The weakest prominent statistic in the genre is the claim that a single short AI email consumes about a bottle of water. It is worth dwelling on precisely because the underlying concern is legitimate; the statistic is what fails.

The widely repeated comparison traces back to modeling that associated roughly 500 milliliters of water with something like 10 to 50 medium-length responses under specific assumptions. It never established that every 100-word email consumes one bottle. Newer, company-specific estimates can be dramatically lower because they use different models, hardware, utilization rates, cooling, grids, geographies, and — crucially — different system boundaries for what counts as "the water."

The honest formulation is narrow: AI inference has a measurable water footprint, but no universal bottle-per-email conversion exists. Four distinctions carry most of the confusion — modeled versus measured data, direct versus electricity-related water, one service versus all of AI, and a facility average versus a prompt-level estimate. A concern that survives those distinctions is worth acting on. A slogan that collapses them invites a correction that then gets used to dismiss the whole case.

Not every moratorium is the same

"Moratorium" is doing a lot of work in this debate, and it covers policies that are not remotely equivalent.

Seattle's action sits in the first category: the city adopted a temporary, threshold-based planning moratorium on large data-center siting while it studies utility, land-use, water, health, employment, and rate effects. The federal proposals associated with a national pause sit closer to the third. Reich's national framing is closest to the transformational category — which is materially different from a one-year local planning pause, and should be argued on its own terms rather than borrowing credibility from the narrower, more defensible versions.

The cases that need correction

Several examples that circulate alongside the argument are real but mischaracterized, and precision here protects the larger point rather than weakening it.

The Wisconsin referendum is often described as the first strict data-center operating code. It was not. Port Washington voters required direct voter approval for future tax-incremental financing districts above a defined value after a data-center financing controversy — a public-finance consent measure, not a technical ordinance covering water, energy, noise, or emissions.

The Saline Township story is often compressed into "Big Tech forced the town to approve it." The accurate sequence is that the township denied the rezoning, the developer and landowners challenged that denial in court, and the township later settled through a consent agreement permitting the project under negotiated conditions. That is a story about legal asymmetry, not a court order compelling approval.

The xAI Memphis controversy is often cited as a diesel-backup-generator example. It centered substantially on natural-gas turbines — their scale, operating role, emissions, and permitting status — and illustrates how a facility's power arrangements can become a local air-quality dispute. Emergency diesel backup, temporary bridge power, and routine on-site gas generation are different things and should not be merged.

And household electricity bills do not rise automatically. Large loads can raise residential bills when utilities socialize the cost of new generation, transmission, or substations; protective large-load tariffs can shift those costs onto the customer creating them. The decisive question is who pays for new capacity and who bears the risk if projected load never materializes.

The strongest counterargument

A serious case for a broad pause has to reckon with the ways it could backfire. A wide moratorium could advantage incumbent hyperscale operators that already control energized campuses; delay the replacement of older, less efficient infrastructure; constrain universities, research institutions, and smaller competitors that lack existing capacity; push development toward jurisdictions with weaker standards; raise compute scarcity and cost; and encourage behind-the-meter generation that escapes the very oversight the pause was meant to strengthen. None of these is a settled outcome. They are effects to analyze before choosing the instrument — and their existence is one reason a scalpel may beat a hammer.

The Praevoium verdict

Reich is directionally correct that AI infrastructure has outpaced many of the public systems responsible for evaluating and allocating its burdens. His most reliable points concern the scale of electricity growth, upper-range water use, public opposition, significant subsidy exposure, local noise and pollution disputes, and unequal municipal negotiating capacity. His weakest choices are presentational: the bottle-per-email claim, treating high-end values as representative, compressing distinct moratoriums into one category, simplifying the Wisconsin referendum, and imprecise treatment of the Memphis generation dispute.

The pattern is consistent: the diagnosis is strong, several individual numbers are fragile, and the jump from documented local problems to a single national remedy is a judgment that has not yet been weighed against narrower alternatives.

The Capacity Compact

Reich's warning should not be dismissed because one viral statistic is weak, nor accepted wholesale because the underlying backlash is real. The AI buildout requires a better unit of accountability than either approval-without-conditions or prohibition-without-distinctions. It requires a named project, the infrastructure serving it, the public costs assigned to it, and the community asked to host it.

That is the purpose of The Capacity Compact — six tests a large-load project should pass before approval:

1. Power compact. Disclose requested, contracted, energized, and operating load, and assign the risk of grid upgrades and cancellations to the party creating them.
2. Water compact. Disclose source, withdrawal, consumption, peak demand, drought protocol, and cooling design.
3. Ratepayer compact. Use large-load tariffs, minimum payments, exit fees, and stranded-asset protections so existing customers are not made the insurer of speculative load.
4. Fiscal compact. Disclose exemptions, net taxes, cost per permanent job, and clawbacks if commitments are not met.
5. Community compact. Require enforceable noise, emissions, emergency-service, workforce, and community-investment commitments.
6. Consent compact. Require timely public disclosure, independent review, hearings, and enforceable performance reporting.

Reich is right that the old approval system is not adequate to the physical scale of the AI buildout. But the case for stronger governance is more defensible than the case for treating every project, metric, and moratorium as equivalent. The durable answer is not approval without conditions or prohibition without distinctions. It is a public compact that requires project-level disclosure, beneficiary-funded infrastructure, enforceable community protections, and verified performance.

Sources