Resolved: The United States Federal Government Should Enact a Moratorium on Hyperscale Data Center Construction
AI Good Bad Essay | AI Debate Bible |Discussion of what a “Moratorium” is |
All AI Resources -- 1000s of cards
A federal moratorium would not necessarily stop hyperscale data centers from being built. It would relocate most of them. It would also result in the construction of more smaller-sized data centers.
That single fact reorganizes the entire debate, because it means the real question is not whether the compute gets built but where — and who ends up controlling the physical substrate of artificial intelligence.
There are, in other words, two debates hiding inside this one resolution:
The locational / security debate. If the United States pauses, the compute migrates — to permissive states, to the Gulf, to Europe and Southeast Asia, and, within a decade, toward orbit (though the resolution could be interpreted to prevent US companies from building them there). Displacement raises costs and creates insecurity. This essay develops that argument in full, because almost nobody is making it.
The AI link debate. Because hyperscale data centers are the bricks-and-mortar of AI training and inference, a construction moratorium is functionally a brake (unless they are all built elsewhere immediately) on AI scaling (the development of advanced AI by building better models) and inference (the computation that occurs after you type in your prompt). That wires every “AI is good / AI is bad” argument into the topic. This essay maps that link structure but reserves the impact substance for a companion piece.
I. Defining the terms — and why the definitions decide the round
“Moratorium”
A moratorium is a temporary suspension, not inherently a permanent ban — but the line between the two collapses depending on the end-trigger.
The historical analogues make this concrete. After the 2010 Deepwater Horizon blowout, Interior Secretary Ken Salazar imposed a six-month deepwater-drilling moratorium that he described as a pause rather than a stop; it was promptly enjoined, reissued, and litigated for years, the canonical case study in how a “temporary” pause becomes a political battlefield over permanence. The nuclear analogue is even starker: eight states still enforce nuclear-construction moratoria, most of them conditional — California’s bars new construction until a federal waste-disposal solution exists, a condition unmet for decades. A pause keyed to an unmet condition can last a very long time.
That matters because the federal bill actually on the table is conditional.
The Artificial Intelligence Data Center Moratorium Act (S.4214), introduced March 25, 2026 by Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez, lifts its moratorium only once Congress passes laws guaranteeing — per the sponsors’ own framing — that AI is “safe and effective,” that its gains “benefit workers, not just the wealthy owners of Big Tech,” and that data centers do not raise utility prices, harm communities, or damage the environment. The bill text contains no automatic sunset; it ends only when Congress “expressly terminates” it by passing the required safeguards. Since Congress is nowhere near doing so, the honest characterization is that the affirmative defends something between a multi-year pause and a conditional ban. Debaters should own that rather than hide it.
I wrote an entire essay on why determining when the moratorium ends is difficult.
“Hyperscale data center”
There is no single legal definition, but the industry-consensus floor is roughly 5,000+ servers and 10,000+ square feet, plus a power threshold — sources put the entry point anywhere from ~40 MW to 100+ MW of draw.
These are distinct from enterprise data centers (single-company, small), colocation (”colo”) facilities (multi-tenant), and edge data centers (small, distributed, latency-focused). The defining trait is scalability — modular “pods” that add fixed increments of compute, power, and cooling.
Note: This supercharges all of the “ban means only AI for corporations and the military,” as they could build their own enterprise versions.
S.4214 supplies its own, narrower definition, and it is worth memorizing because it is what a topical affirmative would actually defend: per analysis of the bill text, an “AI data center” is a facility (on a single site or commonly owned connected sites) used for AI at scale, or exceeding 20 megawatts and built to deliver at least 20 kilowatts to a single server rack or to use advanced liquid/immersion cooling. That captures essentially every modern AI training facility while exempting ordinary enterprise IT.
The market is dominated by the cloud “hyperscalers.” Per Synergy Research Group’s Q3 2025 data, the big three held 63% of a $107 billion quarterly cloud-infrastructure market — AWS at 29%, Microsoft Azure at 20%, Google Cloud at 13% — trailed by Oracle and the fast-growing “neoclouds” like CoreWeave. The frontier has now blown past “hyperscale” into the gigawatt campus: Meta’s Hyperion in Richland Parish, Louisiana, sized to scale to 5 GW; Meta’s Prometheus in New Albany, Ohio, its first ~1 GW cluster; and the Stargate program (OpenAI/Oracle/SoftBank), a $500 billion, 10 GW target whose Abilene, Texas flagship anchors a $100 billion venture. As IEEE Spectrum notes, the entire U.S. data-center fleet drew roughly 8 GW total in 2014; today single campuses are sized for that and more.
“Construction”
The resolutional verb is ambiguous in a way that changes the stakes. S.4214 reaches both new builds and expansions/upgrades of existing facilities. That distinction is decisive: a narrow “new greenfield only” moratorium is trivially evaded by expanding existing campuses (where most near-term capacity growth happens through modular pods), while a broad “expansions too” moratorium is far more disruptive — and far more exposed to economic-harm and takings arguments. Note the irony that current federal policy points the opposite way: Executive Order 14318 specifically promotes brownfield and Superfund-site reuse.
“United States federal government” — the mechanism problem (this is the big one)
This is the affirmative’s deepest vulnerability, and it is structural. Data-center siting, zoning, and permitting are overwhelmingly state and local. The Congressional Research Service says so directly: data centers “for the most part, have a discrete geographical footprint involving mostly local and state siting authorities,” with only some associated energy facilities falling under federal jurisdiction (CRS R48762). There is no clean federal lever to ban construction nationwide. The plausible — and each contestable — hooks:
Congress under the Commerce Clause. The cloud is interstate commerce; this is the most defensible statutory route and what S.4214 implicitly relies on.
FERC, which shapes wholesale power markets and grid interconnection but does not ban buildings.
DOE / Interior / DOD control of federal land — relevant only to projects sited on federal land, which is exactly what EO 14318 tries to accelerate (the DOE has already named four federal sites: Idaho National Laboratory, Oak Ridge, Paducah, and Savannah River).
EPA under the Clean Air and Clean Water Acts — permitting choke points, not a construction ban.
NEPA review — a delay mechanism, not a prohibition.
The legal landmines cut both ways. On takings, the controlling precedent is helpful to the affirmative but not dispositive: in Tahoe-Sierra Preservation Council v. Tahoe Regional Planning Agency (2002), the Supreme Court held 6–3 that a temporary development moratorium is not a per se taking and must instead be judged under the fact-specific Penn Central balancing test — but the Court also signaled that a moratorium lasting more than a year “should be viewed with special skepticism,” and the dissent argued a years-long pause is functionally a taking. (A historical footnote debaters will enjoy: the lawyer who argued for the moratorium’s constitutionality in that case was John Roberts, now Chief Justice.) The cleanest affirmative mechanism is a Commerce Clause statute conditioning interstate grid interconnection and federal financing on a pause — but even that arguably cannot reach a purely intrastate, behind-the-meter, self-powered facility.
As was pointed out to me by a student, fiat can solve this — the federal government can mandate that there is a moratorium — but this would also link them to the federalism argument — that a moratorium disrupts the balance of power between the states and the federal government.
“Should enact”
Normatively, “should” invokes fiat: the affirmative defends that the federal government ought to do this — defending the policy’s desirability and its consequences, including displacement effects, not merely its likelihood. An affirmative cannot simply assert “data centers harm my community” without defending what happens to the compute that gets displaced.
II. Topic One — Where do they get built instead?
This is the framing most debaters will miss, and it is the spine of Stefan Bauschard’s “The Case for Building Data Centers in Your Community”. The thesis, stated bluntly: nobody is going to stop AI. Turn a project down at the county level and it moves to the next county; turn it down statewide and it moves to the next state; slow it nationally and it gets built in the Gulf, in Asia, and — within a decade — in orbit. A moratorium does not stop AI. It decides which communities get the tax base, the jobs, and the leverage, and it risks producing the most concentrated, most corporate, least democratic version of the AI future, because scarce compute does not flow to the people who need it most — it flows to whoever can pay the most (defense, hyperscalers, big finance), and the remaining compute gets more expensive. The historical analogy Bauschard reaches for: capping power-plant construction in 1925 to wait for cleaner energy would have frozen rural electrification; the answer was to clean it up while building it, not to refuse to build.
The structural claim is that a US moratorium displaces rather than eliminates compute, and displacement raises costs and creates insecurity. Where does displaced compute go?
(a) Elsewhere in the US, around the moratorium
Developers already route around local moratoria, and a federal pause tries to close that exit. But the gaps above — smaller, self-powered, behind-the-meter facilities — mean leakage is likely. The map bifurcates into pro-growth and restricted zones rather than going dark.
The resolution resolves at least most of this concern — they can’t be built anywhere.
(b) Abroad — offshoring
This is the meatiest version of the security argument — the data centers get move abroad.
The Gulf states are the marquee destination, offering electricity at roughly $0.05–0.06/kWh versus a higher US average, sovereign wealth that erases financing constraints, and a permissive regulatory sandbox. The UAE–US AI Campus in Abu Dhabi is a 5 GW project, the largest outside the United States, and Saudi Arabia’s PIF-backed HUMAIN is scaling fast. CSIS frames the strategic stakes precisely: if compute is the new oil, concentrating it in the Gulf inherits the Gulf’s geopolitical exposure. The steelman for offshoring, made in Foreign Policy, is that US domestic infrastructure is already constrained by energy and capital bottlenecks, so allied Gulf compute expands total Western capacity rather than merely relocating it — “compute is not a classified asset; it is a commodity.” Other destinations: Europe (though Dublin and Frankfurt face their own grid limits), Latin America, and Southeast Asia.
Now, the resolution doesn’t add in the United States, so arguably the resolution could be interpreted to mean that the US government could prohibit companies from constructing them elsewhere, but it can’t prohibit foreign companies from constructing them elsewhere and then leasing the data centers to US companies. This same point applies to the space argument that we’ll discuss next.
(c) Space — orbital data centers
Once science fiction, now the subject of serious capital and FCC filings — and a displacement vector a US construction moratorium plainly cannot reach.
Google’s Project Suncatcher, announced November 2025, proposes solar-powered satellite constellations carrying TPUs; the accompanying research argues that a solar panel in the right orbit is up to ~8× more productive than on Earth and that, if launch costs fall below ~$200/kg by the mid-2030s, orbital compute could become cost-comparable to terrestrial, with two prototype satellites slated to launch with Planet by early 2027. The startup Starcloud (formerly Lumen Orbit, valued ~$1.1 billion) already put an Nvidia H100 in orbit aboard Starcloud-1 in November 2025 and ran inference on Google’s Gemma model from space, with a stated ambition to build a 5 GW orbital array. Jeff Bezos predicts gigawatt-scale space data centers within 10–20 years and Blue Origin has been engineering components for over a year, while SpaceX has floated using Starlink for distributed compute and Google has discussed launch support with SpaceX.
Proponents pitch orbit explicitly as an escape from terrestrial water, power, and political constraints. The honest caveat: the economics remain marginal, and space chips still lag terrestrial GPUs by orders of magnitude on a like-for-like basis. And, of course, the resolution could be interpreted to say the US government could try to prevent US companies from constructing them there.
Counterarguments to the displacement frame (steelmanned)
Three rebuttals deserve real weight.
First, maybe offshoring to cleaner, cheaper energy is fine — Gulf and Nordic siting can tap abundant solar and hydro, and allied hosting may be acceptable. A tricky affirmative case by may be to claim that it’s good that the data centers is good.
Second, domestic constraints are real — if US grids genuinely cannot absorb the load without socializing the costs onto ratepayers (see below), essentially pushing more data centers abroad may be as good thing.
Third, displacement is not frictionless — latency, data-sovereignty law, and the national-security premium on domestic compute mean some workloads must stay onshore, so a moratorium does bite the very actors it targets. Whether that is the point or the problem depends on your value framework. Also, a federal moratorium would be a real “kick in the face” to US companies, as no one is expecting a moratorium and the economic impacts of reversing the planned build both for the companies and the their investors. Making plans to relocate abroad would also take a lot of time and resources, threatening any US lead.
III. Topic Two — The resolution links to every “AI good / AI bad” argument
(Link map only. Full impact development is reserved for the companion essay.)
Because hyperscale data centers are the physical substrate of AI training and inference, a construction moratorium is functionally a brake on AI scaling. The link structure is uniform:
Moratorium → less / slower compute buildout → slower AI capability growth → triggers whatever AI-good or AI-bad impact you believe in.
That makes the entire AI-impact literature topical as a link into this resolution.
AI-Good impacts (a moratorium is bad because it slows these): medicine and science (drug discovery, diagnostics — Bauschard’s “country of the best doctors”); productivity and GDP growth; education and universal tutoring (the access-equity case); scientific R&D acceleration; US economic competitiveness.
AI-Bad impacts (a moratorium is good because it slows these): existential / catastrophic risk (the 2023 “Pause Giant AI Experiments” letter and the “slow down” statements the bill cites from Musk, Hassabis, and Amodei); labor displacement; autonomous weapons; surveillance (AOC cited ICE–AI partnerships); misinformation and deepfakes; concentration of power.
The important thing to understand about “AI bad” arguments is that that a moratorium would not “on face” avoid them because when the moratorium is lifted the negative impacts of AI would appear. At the same time, a delay may give us time, for example to develop AI safety procedures and align the models to human interests.
The AI-race-with-China overlay sits across both: accelerationists argue the moratorium surrenders the race (so AI-good plus security impacts dominate); pause advocates argue a unilateral US sprint treats speed as the only value. Whichever side better controls the “does slowing compute actually slow capability, and is that good or bad” link will control the round. This essay deliberately does not adjudicate the impacts — that is the companion piece.
The AI development rate/type overlay. This is a “tricky” Pro argument that says that the development of AI with large language models (LLMS) is incredibly inneficient and costly and that we need to have other models (‘world models’) are the most often argued alternative. A moratorium may be good for AI development because it could (a) encourage the development of AI models that are more computational efficient and (b) buy time until those models are developed.
IV. The full argument map
The general arguments in favor of a moratorium assume the moratorium is effectively a ban, at least for a significant period of time.
In this section, I review they key/most direct arguments in favor of a moratorium. The general AI good/bad debate is here.
AFFIRMATIVE / PRO-MORATORIUM
Energy and grid strain. The DOE-funded Lawrence Berkeley National Laboratory 2024 report found data centers consumed about 4.4% of US electricity in 2023 (176 TWh) and projected 6.7–12% by 2028 (325–580 TWh) — load growth that tripled over the prior decade. Globally, the IEA projects data-center electricity roughly doubling to ~945 TWh by 2030, with the US accounting for nearly half of its own electricity-demand growth. Fossil generation is being kept online or built to serve the load: Meta’s Hyperion is served by new Entergy gas plants whose count escalated from three (~2.3 GW) to ten (7+ GW) after a March 2026 agreement, and xAI’s Memphis site trucked in unpermitted gas turbines.
Ratepayer cost-shifting (the strongest affirmative card). In PJM — the grid serving 67 million people across 13 states plus DC — capacity-auction prices went from $28.92/MW-day for 2024/25 to $269.92 for 2025/26 to $329.17 for 2026/27, a roughly tenfold jump that hit the FERC price cap two years running, then cleared at $333.44/MW-day for 2027/28 while falling about 6,600 MW short of the reliability requirement. PJM’s independent market monitor attributed 63% of the 2025/26 increase — roughly $9.3 billion — to data centers, and later found data centers drove 40% of the December auction’s costs and $23.1 billion cumulatively through May 2028. NRDC estimates $100–163 billion in cumulative consumer costs through 2033; Pepco DC residential bills rose about $21/month starting June 2025. States are responding with dedicated hyperscaler rate classes, and the Trump administration brokered a governors’ “Statement of Principles” pushing tech firms to fund their own power. Evenhandedness note: PJM’s own market-services chief cautioned that capacity costs are a relatively small slice of total bills and that the latest auction would mean little change for ratepayers, and a lower load forecast is expected.
Water. Bloomberg found about two-thirds of new US data centers built or in development since 2022 sit in high-water-stress areas, with five states accounting for 72% of those. A Guardian/NOAA analysis put 517 of 809 planned facilities in drought-stricken zones, and a Houston Advanced Research Center study projects Texas data centers using 49 billion gallons in 2025, rising to 399 billion by 2030 — equivalent to drawing Lake Mead down more than 16 feet a year. A single large facility can consume up to 5 million gallons a day, and much of it evaporates rather than returning as treatable wastewater. Concrete fights have erupted from Georgia to Bessemer, Alabama, where a facility was projected to need water equal to two-thirds of the city’s supply.
Community harms and the wave of local moratoria. Per trackers cited by the Columbia/Sabin Center, more than 100 US communities have adopted data-center moratoria. The pace is accelerating: in Q1 2026 alone, backlash delayed or blocked at least 75 projects worth a cumulative $130 billion, on top of more than $64 billion blocked between May 2024 and March 2025. The jobs critique is sharp — data centers produce few permanent on-site jobs against large tax abatements; Good Jobs First documents Georgia losing roughly $2.5 billion a year, Virginia about $1.6 billion, and Texas around $1 billion to data-center tax breaks — and transparency is poor, with 80% of reviewed Virginia jurisdictions bound by NDAs.
Grid reliability. Gigawatt-scale, power-electronic, 24/7 loads stress a grid never designed for them: a July 2024 Virginia disturbance tripped about 60 clustered data centers (~1.5 GW) and nearly forced rolling outages. Interconnection queues are clogged, and NERC has flagged that data-center interconnection delays now complicate demand forecasting itself.
Climate. The IEA estimates data-center CO₂ reaching about 1% of global emissions by 2030 — one of the few sectors with rising emissions — exposing the gap between hyperscaler net-zero pledges and new gas buildouts.
Market, bubble, and malinvestment. Combined hyperscaler 2026 capex is guided at roughly $660–690 billion. Critics flag circular financing: Nvidia agreed to invest up to $100 billion in OpenAI, which in turn commits to buying millions of Nvidia chips, plus a $6.3 billion Nvidia backstop of CoreWeave. OpenAI has assembled roughly $1.15 trillion in compute commitments across seven vendors against ~$13 billion in 2025 revenue, and HSBC projects a $207 billion funding shortfall by 2030 even as commitments reach $1.4 trillion by 2033 — leading even Sam Altman to concede investors are “overexcited”. A pause, on this theory, guards against stranded assets and a dot-com-style overbuild.
A serious empirical wrinkle helps the bubble argument but cuts against the grid-panic argument: much projected demand may be phantom. A former Google grid engineer estimates speculative interconnection requests run five to ten times the number of real data centers, a former Meta energy director describes the same project bid into multiple utilities at once, and an AWS-commissioned Oxford Economics analysis of Australia’s queue found about six of every seven megawatts were “phantom.” Source-quality flag: the “5–10×” figure is a single expert estimate that Utility Dive itself calls elusive, the Oxford study is Australian and AWS-funded, and LBNL’s authors stress the real constraint is localized siting and timing, not that aggregate demand is fake.
[Note: This last argument is part of the “status quo AI bad” argument I discussed before].
NEGATIVE / ANTI-MORATORIUM
Economic engine. Harvard economist Jason Furman found that information-processing equipment and software — about 4% of GDP — drove the overwhelming majority of US GDP growth in the first half of 2025, with the rest of the economy growing near 0.1% annualized; he posted the underlying figures on X while cautioning it is a share of measured growth, not a counterfactual. Honest counter: Goldman’s Joseph Briggs notes much hardware is imported, shrinking the net GDP boost, and other economists argue consumption was 2025’s real driver, and some analyses find data-center construction’s direct GDP impact was limited. Stargate alone is a $500 billion / 10 GW commitment, and the industry claims hundreds of thousands of direct jobs.
National security and the China race (the strongest negative card). Interior Secretary Doug Burgum called a moratorium the equivalent of waving a “surrender flag” to China — a view echoed across the aisle by Senator John Fetterman, who called the bill “China First” and refused to “help hand the lead in AI to China,” a framing amplified by national-security commentators. On the underlying balance, Foreign Affairs reports US firms control roughly 70% of global AI compute to China’s ~10% (the often-cited “75%” traces to the Biden AI Diffusion Rule’s allied-country threshold, not the US alone). Counter, from The New Republic: China is itself wrestling with AI governance, so a pause is “a claim to a different kind of leadership,” not surrender.
Energy abundance / clean-energy catalyst. Far from only burning gas, hyperscalers are anchoring a nuclear renaissance. Microsoft signed a 20-year PPA with Constellation to restart Three Mile Island Unit 1 — 835 MW, renamed the Crane Clean Energy Center, a $1.6 billion Constellation investment now accelerated to 2027 and confirmed in an SEC filing. Add Google–Kairos Power, Amazon–X-energy, Amazon–Talen Susquehanna, and Meta’s nuclear RFPs — industry tracking counts roughly 9.8+ GW of committed nuclear across more than a dozen deals — plus geothermal and solar-plus-storage. The argument: data centers as anchor tenants fund new firm clean generation and grid upgrades, and “bring-your-own-power” mandates can prevent cost-shifts without a ban. Meta’s Hyperion deal is structured so the company pays its full cost of service, the model the negative would generalize.
Innovation. (Links to the companion essay.) Medical, scientific, educational, and productivity gains all run through compute; a brake on compute is a brake on these benefits.
Federalism / practicality. A federal construction moratorium is arguably the wrong tool wielded by the wrong sovereign. Siting is state and local, and even EO 14318 does not preempt state permitting, zoning, or energy regulation. Targeted instruments — dedicated rate classes, BYO-power mandates, water ordinances, transparency rules, and interconnection reform — address the real harms without a blunt nationwide ban, and states from Maine to Oregon to Virginia are already deploying them.
Free-market distortion. A construction ban picks winners (incumbents with existing capacity benefit from freezing out new entrants), distorts a $600 billion-plus investment cycle, and risks stranding the very capital it claims to protect — while the bubble, if it is one, is arguably best deflated by markets rather than a federal pause.
V. Strategic recommendations
For affirmative debaters
Lead with the mechanism you can defend. Pick a coherent federal lever — a Commerce Clause statute conditioning interstate grid interconnection and federal financing on a pause — and defend it explicitly. The federalism problem is your biggest vulnerability when you rely on fiat; don’t hand-wave it.
Anchor on ratepayers and water. These are the most concrete, near-term, bipartisan harms with the best evidence (PJM’s tenfold capacity jump, the 63% attribution, drought-zone siting). They are domestic harms that displacement cannot fully rebut.
Pre-empt the China card by arguing (a) a conditional pause is not permanent disarmament, (b) much “demand” is phantom, so the race framing rests on inflated numbers, and (c) the moratorium ends precisely when safeguards pass. I don’t think the China argument is super strong if you really work to get underneath it, but it’s not something you can ignore. I think it will be one of the more common Con arguments.
Concede displacement partially and argue some workloads must stay onshore (latency, sovereignty), so the pause still bites where it matters while protecting communities.
Threshold that flips the round to you: show the harms are large, near-term, and not addressable by targeted regulation alone.
For negative debaters
Press topicality / mechanism first — make the affirmative prove a coherent federal construction ban is possible absent fiat that links to federalism.,
Run displacement plus security as a turn: the moratorium doesn’t stop AI, it offshores it to the Gulf and orbit, raising the very ratepayer and security costs the affirmative claims to solve.
Deploy the China and GDP cards as net benefits to that alternative
Threshold that flips the round to you: show the harms are real but addressable by less-restrictive means, making the moratorium gratuitously costly.
For the reader / policymaker. The honest synthesis is that the harms are real but the instrument is poorly matched. The evidence for ratepayer cost-shifting and drought-zone water stress is strong and primary-sourced; the evidence that a nationwide construction moratorium is the right fix is weak, given the federalism mismatch, displacement effects, and the targeted tools already being deployed.
VI. Caveats
Live policy, fast-moving facts. Capacity prices, capex guidance, which moratoria are active, and buildout figures change monthly. Several numbers are genuinely contested — Meta’s Hyperion cost has been reported anywhere from an initial $10 billion to over $200 billion, reflecting different scopes — so verify before a round.
Projection uncertainty cuts both ways. Demand forecasts diverge enormously for 2030; the phantom-demand critique undercuts the most alarming grid numbers, but LBNL’s authors caution the real issue is localized timing and siting, not that aggregate growth is illusory. Don’t overstate either direction.
Source-funding flags. Several upbeat figures come from industry (Turner & Townsend, the AWS-commissioned Oxford study); several alarming figures come from advocacy groups (NRDC, UCS, Good Jobs First) or self-published trackers. Weight accordingly.
The two-debates structure is the key insight. Many rounds will collapse into a generic “AI good/bad” clash. The more sophisticated — and more winnable — terrain is the locational/mechanism debate developed here. Full AI-impact adjudication is the companion essay.
VII. The real cruxes, named
Mechanism. Is a federal construction moratorium constitutionally and practically coherent, or is siting irreducibly state and local?
Displacement. Does a US pause stop compute or merely relocate it (and is relocation — domestic, Gulf, orbital — acceptable or dangerous)?
Less-restrictive means. Can targeted regulation (rate classes, BYO-power, disclosure) capture the moratorium’s benefits without its costs?
The link / values question (reserved). Does slowing compute actually slow AI capability — and is slowing AI good or bad? That is the companion essay.
Comprehensive Bibliography
Organized by argument category. Primary sources (statutes, executive orders, court opinions, government and national-lab reports, company filings and announcements) are weighted first within each section where available. Verified working links as of writing.
1. The bill, the executive order, and the federal-policy frame (primary)
S.4214, Artificial Intelligence Data Center Moratorium Act — bill landing page, Congress.gov: https://www.congress.gov/bill/119th-congress/senate-bill/4214
S.4214 full text — GovTrack: https://www.govtrack.us/congress/bills/119/s4214/text
Sanders/Ocasio-Cortez announcement — Senator Sanders (official): https://www.sanders.senate.gov/press-releases/news-sanders-ocasio-cortez-announce-ai-data-center-moratorium-act/
S.4214 plain-English bill analysis (20 MW / 20 kW-per-rack / liquid-cooling definition; export-control provision) — Legisletter: https://legisletter.org/bill/s4214-artificial-intelligence-data-center-moratorium-act
S.4214 summary — Benton Institute: https://www.benton.org/headlines/sen-sanders-rep-ocasio-cortez-announce-ai-data-center-moratorium-act
Executive Order 14318, “Accelerating Federal Permitting of Data Center Infrastructure” — Federal Register (official): https://www.federalregister.gov/documents/2025/07/28/2025-14212/accelerating-federal-permitting-of-data-center-infrastructure
EO 14318 — White House: https://www.whitehouse.gov/presidential-actions/2025/07/accelerating-federal-permitting-of-data-center-infrastructure/
EO 14318 — American Presidency Project: https://www.presidency.ucsb.edu/documents/executive-order-14318-accelerating-federal-permitting-data-center-infrastructure
EO 14318 — govinfo (DCPD-202500788): https://www.govinfo.gov/app/details/DCPD-202500788
“Winning the Race: America’s AI Action Plan” (July 2025) — White House PDF (primary): https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf
Data Center Energy Infrastructure: Federal Permit Requirements (R48762) — Congressional Research Service: https://www.congress.gov/crs-product/R48762
EO 14318 analysis (four named federal sites; agency directives) — Davis Wright Tremaine: https://www.dwt.com/blogs/energy--environmental-law-blog/2025/07/trump-ai-action-plan-data-centers-energy-projects
AI Action Plan analysis — Ropes & Gray: https://www.ropesgray.com/en/insights/alerts/2025/07/winning-the-race-americas-ai-action-plan-key-pillars-policy-actions-and-future-implications
Federal action will drive data center development (states retain authority) — SLR Consulting: https://www.slrconsulting.com/us/insights/federal-action-data-center-development/
2. Moratoria — legal meaning, history, and the takings question
Tahoe-Sierra Preservation Council v. Tahoe Regional Planning Agency, 535 U.S. 302 (2002) — Justia (full opinion): https://supreme.justia.com/cases/federal/us/535/302/
Tahoe-Sierra — syllabus — Cornell Legal Information Institute: https://www.law.cornell.edu/supct/html/00-1167.ZS.html
Tahoe-Sierra — oral argument transcript (John Roberts arguing for TRPA) — Supreme Court: https://www.supremecourt.gov/oral_arguments/argument_transcripts/2001/00-1167.pdf
Tahoe-Sierra — full opinion — FindLaw: https://caselaw.findlaw.com/court/us-supreme-court/535/302.html
Offshore drilling moratorium (”pause button, not the stop button”) — PBS NewsHour: https://www.pbs.org/newshour/show/offshore-oil-drilling-moratoriums-aims-effectiveness-debated
State nuclear-construction moratoria (conditional bans) — Nuclear Energy Institute: https://www.nei.org/advocacy/state-nuclear-moratoriums-101
Local data-center moratoria — municipal considerations — Columbia/Sabin Climate Law Blog: https://blogs.law.columbia.edu/climatechange/2026/05/27/local-moratoria-considerations/
Planning Under Preemption: State Power and Local Authority in the AI Data Center Era — Journal of the American Planning Association: https://www.tandfonline.com/doi/full/10.1080/01944363.2026.2618221
3. The hyperscale frontier — market structure and the gigawatt campus
Cloud Market Share Q3 2025 (AWS 29% / Azure 20% / Google 13%) — Synergy Research Group (primary): https://www.srgresearch.com/articles/cloud-market-share-trends-big-three-together-hold-63-while-oracle-and-the-neoclouds-inch-higher
Cloud market growth Q3 2025 — Synergy Research Group: https://www.srgresearch.com/articles/cloud-market-growth-rate-rises-again-in-q3-biggest-ever-sequential-increase
Big three take two-thirds of $107B cloud market — TechTarget: https://www.techtarget.com/searchcloudcomputing/news/366634757/The-big-three-grab-two-thirds-of-107B-cloud-market-in-Q3
AWS loses share as Azure, Google gain — The Register: https://www.theregister.com/2025/11/20/aws_loses_market_share_azure_google/
AWS vs Azure vs Google Cloud (hyperscale definitions) — Channel Insider: https://www.channelinsider.com/infrastructure/aws-vs-azure-vs-google-cloud/
5 GW data center buildout requires novel engineering (Hyperion 5 GW; Entergy gas; Stargate Abilene 300 MW turbines) — IEEE Spectrum: https://spectrum.ieee.org/5gw-data-center
Hyperion data center — Wikipedia (overview/sourcing): https://en.wikipedia.org/wiki/Hyperion_(data_center)
Meta’s Hyperion & Prometheus — power and ownership — Data Center Frontier: https://www.datacenterfrontier.com/hyperscale/article/55310441/ownership-and-power-challenges-in-metas-hyperion-and-prometheus-data-centers
The power play behind Hyperion (LPSC 4-1 vote; ratepayer transmission costs; Stargate Ohio) — Sherwood News: https://sherwood.news/tech/hyperion/
Meta $200B Hyperion, 10 gas plants, $27B Blue Owl financing — The Next Web: https://thenextweb.com/news/meta-200-billion-hyperion-data-center-louisiana
Meta funds seven new gas plants (7 GW total; “pays full cost of service”) — Tom’s Hardware: https://www.tomshardware.com/tech-industry/artificial-intelligence/meta-will-fund-seven-new-gas-plants-to-power-its-7gw-louisiana-data-center
Meta $27B data center / Louisiana power expansion — Engineering News-Record: https://www.enr.com/articles/62766-27b-meta-data-center-pushes-louisiana-toward-massive-power-expansion
Meta “hundreds of billions”; Prometheus 2026; Hyperion to 2030 — Data Center Dynamics: https://www.datacenterdynamics.com/en/news/meta-to-invest-hundreds-of-billions-of-dollars-into-compute-to-build-superintelligence-with-several-multi-gw-data-center-clusters/
Zuckerberg 5 GW / “footprint of Manhattan” — IT Pro: https://www.itpro.com/infrastructure/data-centres/meta-working-on-a-5gw-data-center-to-supercharge-ai-infrastructure-and-mark-zuckerberg-says-one-cluster-alone-covers-a-significant-part-of-the-footprint-of-manhattan
AI capex 2026: the $690B infrastructure sprint — Futurum Group: https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
4. Energy and the grid
2024 United States Data Center Energy Usage Report — Lawrence Berkeley National Laboratory (primary PDF): https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
LBNL 2024 report — DOI record — https://doi.org/10.71468/P1WC7Q
LBNL report summary (load tripled; doubles/triples by 2028) — Berkeley Lab News Center: https://newscenter.lbl.gov/2025/01/15/berkeley-lab-report-evaluates-increase-in-electricity-demand-from-data-centers/
AI to drive surging electricity demand (~945 TWh by 2030) — International Energy Agency: https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
AI, Data Centers, and the U.S. Electric Grid (Virginia 60-data-center trip) — Harvard Belfer Center: https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid
DOE study: grid can meet long-term demand but localized constraints — Data Center Frontier: https://www.datacenterfrontier.com/energy/article/55019791/doe-study-ai-boom-breeds-localized-energy-constraints-but-grid-can-meet-long-term-demand
NERC: interconnection delays complicate demand forecasting — Utility Dive: https://www.utilitydive.com/news/data-center-interconnection-delays-complicate-demand-forecasting-nerc/820695/
Five charts on data-center energy use and emissions (~1% of global CO₂ by 2030) — Carbon Brief: https://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context/
5. Ratepayer cost-shifting (PJM)
2026/2027 Base Residual Auction report ($329.17/MW-day cap) — PJM (primary PDF): https://www.pjm.com/-/media/DotCom/markets-ops/rpm/rpm-auction-info/2026-2027/2026-2027-bra-report.pdf
PJM 2026/2027 auction press release ($28.92 → $269.92 → $329.17 trajectory) — PJM Inside Lines: https://insidelines.pjm.com/pjm-auction-procures-134311-mw-of-generation-resources-supply-responds-to-price-signal/
PJM 2027/2028 auction ($333.44/MW-day; 5,100 MW of increase from data centers) — PJM Inside Lines: https://insidelines.pjm.com/pjm-auction-procures-134479-mw-of-generation-resources/
PJM capacity prices hit record; 6.6 GW shortfall — Utility Dive: https://www.utilitydive.com/news/pjm-interconnection-capacity-auction-data-center/808264/
Market monitor: data centers 40% of December auction costs; $23.1B cumulative — Utility Dive: https://www.utilitydive.com/news/data-centers-pjm-capacity-auction/808951/
Data-center growth spurs PJM capacity prices by a factor of 10 (63% / $9.3B; Pepco +$21/mo) — IEEFA: https://ieefa.org/resources/projected-data-center-growth-spurs-pjm-capacity-prices-factor-10
PJM 76% wholesale jump; NRDC $100–163B through 2033; ~$70/mo by 2028 — The AI Consulting Network: https://www.theaiconsultingnetwork.com/blog/pjm-data-center-power-76-percent-cre-investors-2026
PJM 2026/2027 auction analysis (reserve margins; Shapiro price cap) — Enel North America: https://www.enelnorthamerica.com/insights/blogs/pjm-2026-2027-capacity-auction-results
6. Water
The AI boom is draining water from the areas that need it most (two-thirds in high-stress; 5 states = 72%) — Bloomberg (primary analysis): https://www.bloomberg.com/graphics/2025-ai-impacts-data-centers-water-data/
Most new US AI data centers built in drought zones (517 of 809 planned) — Tom’s Hardware (on the Guardian/NOAA analysis): https://www.tomshardware.com/tech-industry/most-new-us-ai-data-centers-are-going-up-on-drought-land
Data Drain: land and water impacts (HARC Texas study; Lake Mead) — Lincoln Institute of Land Policy: https://www.lincolninst.edu/publications/land-lines-magazine/articles/land-water-impacts-data-centers/
Data centers and water consumption (5 M gallons/day; 56% fossil) — Environmental and Energy Study Institute: https://www.eesi.org/articles/view/data-centers-and-water-consumption
From energy to air quality: how data centers affect communities (NDAs; $64B blocked; AEP Ohio rate) — World Resources Institute: https://www.wri.org/insights/us-data-center-growth-impacts
Map: data centers in drought-hit areas (Bessemer, AL) — Newsweek: https://www.newsweek.com/map-data-centers-built-drought-hit-areas-11997520
AI water security concerns in high-stress states — Data Center Dynamics: https://www.datacenterdynamics.com/en/news/ai-data-center-growth-deepens-water-security-concerns-in-high-stress-states-report/
AI data center boom and America’s water problem (Guardian analysis; 73B gallons by 2028) — ConstructConnect: https://news.constructconnect.com/ai-data-center-boom-is-running-into-americas-water-problem
Data centers and the water crisis (xAI Memphis; NASA groundwater) — Science and Environmental Health Network: https://www.sehn.org/sehn/2025/8/14/data-centers-and-the-water-crisis
Data center water use (LBNL classification; UC Riverside per-prompt) — MOST Policy Initiative: https://mostpolicyinitiative.org/science-note/data-center-water-use/
7. Local opposition, statewide moratoria, and public opinion
Data center moratorium bills are spreading in 2026 (tax-abatement figures) — Good Jobs First: https://goodjobsfirst.org/data-center-moratorium-bills-are-spreading-in-2026/
Data center opposition goes national ($130B / 75 projects blocked in Q1 2026) — Fortune: https://fortune.com/2026/06/22/data-center-opposition-goes-national-despite-only-8-percent-living-near-one/
Data center moratoriums in 12 states — The AI Consulting Network: https://www.theaiconsultingnetwork.com/blog/data-center-moratorium-bills-states-cre-investors-2026
Gallup: 7 in 10 Americans don’t want AI data centers near them — Washington Times: https://www.washingtontimes.com/news/2026/may/13/gallup-poll-7-10-americans-dont-want-ai-data-centers-near/
More Americans oppose data centers than nuclear plants (Gallup) — Planetizen: https://www.planetizen.com/news/2026/06/137695-more-americans-oppose-data-centers-nuclear-plants-gallup
Americans now overwhelmingly oppose new data centers (49-point swing) — Heatmap News: https://heatmap.news/politics/americans-oppose-data-centers-poll
State data center legislation faces local zoning battles — MultiState: https://www.multistate.us/insider/2026/1/15/state-data-center-legislation-faces-local-zoning-battles
Local regulations gain ground as state bills falter — MultiState: https://www.multistate.us/insider/2026/3/13/local-data-center-regulations-gain-ground-as-state-bills-falter
State data center laws vs. federal AI push: 2026 tracker — MultiState: https://www.multistate.us/insider/2026/4/14/federal-ai-data-center-policy-meets-resistance-from-state-lawmakers
8. Market, bubble, and phantom demand
A fraction of proposed data centers will get built (”5–10×” phantom estimate) — Utility Dive: https://www.utilitydive.com/news/a-fraction-of-proposed-data-centers-will-get-built-utilities-are-wising-up/748214/
Phantom data centers are flooding the load queue — Latitude Media: https://www.latitudemedia.com/news/phantom-data-centers-are-flooding-the-load-queue/
Estimating data centre “phantom demand” (Australia; AWS-commissioned) — Oxford Economics: https://www.oxfordeconomics.com/resource/estimating-data-centre-phantom-demand/
AI circular deals: how Microsoft, OpenAI and Nvidia keep paying each other — Bloomberg: https://www.bloomberg.com/graphics/2026-ai-circular-deals/
OpenAI’s Nvidia, AMD deals boost $1 trillion AI boom with circular deals — Bloomberg: https://www.bloomberg.com/news/features/2025-10-07/openai-s-nvidia-amd-deals-boost-1-trillion-ai-boom-with-circular-deals
OpenAI won’t profit by 2030; $207B shortfall; $1.4T commitments by 2033 (HSBC) — Fortune: https://fortune.com/2025/11/26/is-openai-profitable-forecast-data-center-200-billion-shortfall-hsbc/
OpenAI’s $1.15 trillion infrastructure spend (seven-vendor breakdown) — Tomasz Tunguz: https://tomtunguz.com/openai-hardware-spending-2025-2035/
The circular economy of AI — The Register: https://www.theregister.com/2025/11/04/the_circular_economy_of_ai/
The circular economy of AI: big tech financing itself (Altman “overexcited”) — Calcalist: https://www.calcalistech.com/ctechnews/article/z4lxiqbtw
9. The economic-engine / GDP case (negative ground)
Investment in info-processing equipment drove ~92% of H1 2025 GDP growth — Jason Furman on X (primary):
Without data centers, GDP growth was 0.1% in H1 2025 — Fortune: https://fortune.com/2025/10/07/data-centers-gdp-growth-zero-first-half-2025-jason-furman-harvard-economist/
Furman GDP calculation — Yahoo Finance: https://finance.yahoo.com/news/without-data-centers-gdp-growth-171546326.html
AI wasn’t the biggest engine of US growth in 2025 (import leakage; consumption) — CNBC: https://www.cnbc.com/2026/01/26/ai-wasnt-the-biggest-engine-of-us-gdp-growth-in-2025.html
Data construction up ~30% with limited GDP impact — Marketplace: https://www.marketplace.org/story/2026/02/27/data-construction-up-nearly-30-in-2025-with-limited-impact-on-gdp
Data center investment drives US economy growth — AI Data Analytics Network: https://www.aidataanalytics.network/data-science-ai/news-trends/data-center-investment-drives-us-economy-growth
10. National security, the China race, and offshoring
Ocasio-Cortez and Sanders push moratorium (Burgum “surrender flag”; Fetterman) — PBS NewsHour: https://www.pbs.org/newshour/politics/ocasio-cortez-and-sanders-push-bill-to-impose-ai-data-center-moratorium
Sanders and AOC launch bill to ban new data-center construction — Fortune: https://fortune.com/2026/03/25/bernie-sanders-and-aoc-launch-bill-to-ban-new-data-center-construction/
A data center moratorium would mean surrendering to China (editorial; industry jobs figures) — Washington Reporter: https://washingtonreporter.news/editorial-a-data-center-moratorium-would-mean-surrendering-to-china/
What Bernie and AOC get wrong about data centers (cross-partisan backlash; counter to “surrender”) — The New Republic: https://newrepublic.com/article/208392/data-centers-aoc-sanders
Fetterman calls the bill “China First” — Capitalism Institute: https://capitalisminstitute.org/fetterman-calls-sanders-aoc-data-center-moratorium-bill-a-gift-to-beijing/
The Myth of the AI Race (US ~70% of global compute to China’s ~10%) — Foreign Affairs: https://www.foreignaffairs.com/united-states/myth-ai-race
Regulatory framework for responsible diffusion (AI Diffusion Rule; 75% allied threshold) — Bureau of Industry and Security: https://www.bis.gov/press-release/biden-harris-administration-announces-regulatory-framework-responsible-diffusion-advanced-artificial
U.S.-China AI competition needs data centers in UAE, Saudi Arabia (offshoring steelman) — Foreign Policy: https://foreignpolicy.com/2025/07/02/data-centers-us-uae-partnership-saudi-arabia-ai/
If compute is the new oil, war in the Gulf raises the stakes — CSIS: https://www.csis.org/analysis/if-compute-new-oil-war-gulf-significantly-raises-stakes
The Middle East’s trillion-dollar bet on AI infrastructure (UAE 5 GW; Gulf power prices) — Introl: https://introl.com/blog/middle-east-uae-saudi-arabia-ai-data-center-boom-2025
Global energy demands within the AI regulatory landscape (Dublin grid limits) — Brookings: https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/
11. Nuclear and clean-energy deals (negative response)
Constellation to launch Crane Clean Energy Center / restart TMI Unit 1 — Constellation (primary press release): https://www.constellationenergy.com/news/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html
Constellation 8-K (835 MW; 20-year PPA; 3,400 jobs) — SEC (primary): https://www.sec.gov/Archives/edgar/data/0001868275/000186827524000058/ceg-202409208kexh991.htm
One year later: Crane ahead of schedule (restart accelerated to 2027) — Constellation: https://www.constellationenergy.com/news/2025/09/one-year-later-crane-clean-energy-center-still-in-the-spotlight-and-ahead-of-schedule.html
Constellation plans TMI restart for Microsoft (600 jobs; Brattle study) — Utility Dive: https://www.utilitydive.com/news/constellation-three-mile-island-nuclear-power-plant-microsoft-data-center-ppa/727652/
TMI returns; Microsoft 20-year 835 MW PPA; $1.6B Constellation investment — Data Center Dynamics: https://www.datacenterdynamics.com/en/news/three-mile-island-nuclear-power-plant-to-return-as-microsoft-signs-20-year-835mw-ai-data-center-ppa/
Constellation to restart TMI, powering Microsoft — World Nuclear News: https://www.world-nuclear-news.org/articles/constellation-to-restart-three-mile-island-unit-powering-microsoft
Nuclear power for AI: Microsoft, Google, Amazon deals — Introl: https://introl.com/blog/nuclear-power-ai-data-centers-microsoft-google-amazon-2025
Every nuclear-powered data center deal (~9.8 GW committed) — SMR Intel: https://smrintel.com/nuclear-data-center-deals/
12. Orbital / space data centers
Project Suncatcher announcement — Google (primary blog): https://blog.google/innovation-and-ai/technology/research/google-project-suncatcher/
Space-based scalable AI infrastructure (8× solar; <$200/kg by mid-2030s; radiation testing) — Google Research (primary): https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/
Project Suncatcher — 81-satellite clusters; two prototypes by 2027 — Data Center Dynamics: https://www.datacenterdynamics.com/en/news/project-suncatcher-google-to-launch-tpus-into-orbit-with-planet-labs-envisions-1km-arrays-of-81-satellite-compute-clusters/
How Starcloud is bringing data centers to space (5 GW array; 10× CO₂ savings claim) — NVIDIA (primary blog): https://blogs.nvidia.com/blog/starcloud/
Starcloud-1 reaches orbit with Nvidia H100 (Bezos gigawatt prediction; Musk, Schmidt) — Data Center Dynamics: https://www.datacenterdynamics.com/en/news/starcloud-1-satellite-reaches-space-with-nvidia-h100-gpu-now-operating-in-orbit/
Nvidia-backed Starcloud trains first AI model in space — CNBC: https://www.cnbc.com/2025/12/10/nvidia-backed-starcloud-trains-first-ai-model-in-space-orbital-data-centers.html
Starcloud launches orbital AI data center with H100 — Data Center Frontier: https://www.datacenterfrontier.com/site-selection/article/55337494/starcloud-launches-orbital-ai-data-center-with-nvidia-h100-gpu
Google and SpaceX discuss launches for orbital data centers (Anthropic/Colossus) — AeroTime: https://www.aerotime.aero/articles/google-spacex-orbital-ai-data-centers
Startups building orbital data centers (Starcloud ~$1.1B; FCC filings) — Quartz: https://qz.com/orbital-data-center-startups-competitive-landscape-061526
Orbital data centers: complete guide (space-vs-terrestrial chip gap; China constellation) — Introl: https://introl.com/blog/orbital-data-centers-space-ai-infrastructure-guide-2025