Debatetrack
TopicsHyperscale Data Center Moratorium09/01/2026 to 10/31/2026

Resolved: The United States federal government should enact a moratorium on hyperscale data center construction.

The 2026-27 season opens with the infrastructure behind the AI boom. Pro argues the buildout is outrunning the grid, water supplies and local consent, and a national pause is the only way to catch up; Con argues a moratorium would hand the AI race to rivals and knock the legs out from under an economy that AI investment is carrying. This analysis builds on Katheryne Dwyer’s DebateTrack topic lecture.

Topic analysis

This guide draws heavily on the September/October topic lecture by Katheryne Dwyer, director of debate at Civic Debate Academy and a member of the NSDA Public Forum wording committee. Watch the full lecture, linked on this page, for the complete walkthrough. Figures below that come from the lecture are attributed to it. Treat them as starting points and cut the underlying sources yourself, because, as the lecture warns, the evidence on this topic is unusually messy.

The topic is really about heat

A data center does three things: it houses servers, feeds them electricity, and removes the heat they produce. We have had data centers as long as we have had the internet, so the question is what makes this moment different. The answer is AI. The lecture puts it in numbers: a typical non-AI data center rack draws about 8 kilowatts, while an AI rack draws 100 to 130 kilowatts. That is a different scale and a different concentration of energy, and almost every environmental argument on this topic follows from it.

AI uses so much power because a model works by doing arithmetic, constantly, across thousands of specialized chips called graphics processing units (GPUs). Powering that math is what consumes the electricity. And the electricity doesn’t disappear. It becomes heat. The lecture’s rule of thumb is worth memorizing: whenever you see an electricity figure for a rack or a facility, assume roughly the same amount of heat has to be removed. The central clash of this topic is what happens to that heat.

Tight coupling and why density matters

AI models are far too big for one chip, so they are split across thousands, sometimes tens of thousands, of GPUs that have to work together almost constantly to produce one output. The lecture calls this tight coupling, and it is the key difference between AI infrastructure and traditional computing. In a loosely coupled system, like thirty students each doing their own worksheet, one slow student doesn’t slow anyone else. In a tightly coupled system, like thirty students building one Lego tower, everyone waits on the slowest step.

Chips communicate faster the closer they are to each other, so tightly coupled systems are packed as densely as possible. Density is where the environmental problems start. The lecture notes that a standard AI rack weighs about 1.4 metric tons, which old warehouses can’t easily be retrofitted to carry. That is one reason new construction, rather than conversion of existing buildings, is the norm for these facilities, and it is why a moratorium on construction actually bites. Dense racks also concentrate enormous heat in a small space: by the lecture’s comparison, 130 kilowatts from one rack is roughly what 100 houses use.

Cooling: two separate questions

Debaters constantly mix up two different cooling questions, and keeping them apart will win you crossfires. The first is how heat leaves the server. Above about 20 kilowatts per rack, fans and air conditioning can’t move heat fast enough, because air conducts heat poorly. Every AI rack is well above that line, so AI data centers use liquid cooling, piping coolant directly to the chips. Nobody seriously argues against cooling the servers. Without it they fail.

The second question is how heat leaves the building, and that is where the real debate is. Wet cooling towers evaporate water to shed the heat. That is cheap on electricity, which gives companies a strong profit incentive to use it, but the evaporated water leaves the local watershed. Dry cooling keeps the water in a closed loop but uses much more electricity. In practice most facilities are hybrids, running dry most of the year and spraying water on the hottest days.

The lecture makes a point Pro teams should not skip: hyperscale data centers aren’t uniquely wasteful. Smaller data centers often still reject heat through evaporative towers and can be big water users too. Density doesn’t directly cause more water use. What it does is create a much larger heat rejection load, the total heat a building has to get rid of. Because the load is bigger, the choice between wet and dry cooling matters more, since far more water is at stake.

Withdrawal versus consumption

"Water doesn’t disappear" is a common Con response, and it is half right. Evaporated water does come back as rain, but not where it was taken from. That is why the difference between withdrawal and consumption matters. Withdrawal is all the water taken from a watershed. Consumption is withdrawal minus what is returned directly, which is mostly the evaporated portion. Consumption is what the watershed doesn’t get back. In arid regions like Arizona, where groundwater built up over very long periods, heavy consumption can drain a source faster than it can ever recover on a human timescale.

Even water that is returned isn’t harmless. It carries concentrated dissolved solids, which can affect crops, aquatic life and drinking water quality. None of this is unique to AI or even to tech: any large industrial cooling process has the same issue. The reason it is debatable now is the speed of the buildout. Pro teams who understand this can defend a local, specific water harm. Pro teams who say AI "drains the oceans" will lose to any Con team that explains the water cycle.

Training, inference and the offshoring question

AI has two kinds of computing demand. Training builds a model: one huge, tightly coupled job running across thousands of chips for months. The lecture cites an estimate that training GPT-4 took around 50 gigawatt-hours, about a year of electricity for 40,000 households. Inference is using the model, like asking a chatbot a question. It is loosely coupled, since your answer doesn’t depend on what other chips are doing. The lecture cites about 2.9 watt-hours for a ChatGPT query compared with 0.3 for a Google search.

You might expect training to dominate, but the lecture’s figures show the two are close: about 23.1 gigawatts for training and 20.9 for inference in 2025, with inference projected to pass training and reach around 93.3 gigawatts by 2030. That matters for one of the most important arguments on the topic: offshoring. If the US pauses construction, why wouldn’t companies just build abroad?

The lecture’s answer is that they partly can. Training is latency tolerant, so it can happen far from users. Inference is latency sensitive: how fast your answer comes back depends on how close you are to the data center, which is why companies have historically built near population centers. Whether inference can be offshored effectively is genuinely unsettled, and the lecture describes the research on it as speculative. The practical takeaway is that roughly half of data center energy demand may not be easy to move abroad. Con teams running leakage need to explain which kind of computing leaks. Pro teams can argue that the latency-sensitive half stays home and gets paused.

The grid and who pays

The scale is large. The lecture cites data centers using 4.4 percent of US electricity in 2023, projected to rise to between 6.7 and 12 percent by 2028, and possibly around 426 terawatt-hours, roughly 9 percent of projected demand, by 2030. Those costs reach consumers in two ways, and only one of them is central to this topic.

The first is retail rates, how a utility splits costs among customers. Utilities traditionally assign costs based on each customer’s share of the grid’s single busiest hour of the year, and data centers break that model because they run at full capacity around the clock. But retail rates are regulated by the states, and many states have responded by writing new rate classes for data centers. The lecture’s view is that this works reasonably well and isn’t the core of the federal debate.

The second is wholesale capacity: what grid operators pay generators to stand by so the lights stay on if demand spikes or a plant fails. This is regulated federally by the Federal Energy Regulatory Commission (FERC), and it is where Pro’s grid argument lives. Grid operators forecast demand a few years ahead, and those forecasts set what generators get paid, a cost spread across everyone’s bill. Data centers file requests to connect to the grid, which inflates the forecasts. According to the lecture, those filings aren’t audited or made public, so no one outside the process can check whether the projected demand will ever materialize.

The lecture points to PJM, one of the largest grid operators, as the example. Its capacity auction for 2027-28 cleared at the maximum price FERC allows, around $333 per megawatt-day, and for the first time fell short of its reliability requirement, by roughly 6,000 to 7,000 megawatts. Of roughly $16.5 billion in capacity costs, the lecture cites about $6.5 billion attributed to data centers, and about $6.2 billion of that to data centers that haven’t been built yet. That is the heart of the Pro grid story: households paying more, for a less reliable grid, to cover demand that may never show up.

Reading the resolution: "moratorium"

A moratorium is generally understood as a temporary pause on permits or construction. It isn’t a ban and doesn’t shut down existing facilities. That is an unusual action for a debate resolution, and many debaters are uneasy with it. The lecture argues it actually gives Pro useful flexibility, because real moratoriums already exist at the state and local level with different lengths and conditions. Pro can defend a fixed-length pause or a conditional one.

The most prominent federal model is the bill introduced by Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez. As the lecture describes it, the moratorium isn’t time-limited but lasts until certain conditions are met, and it also bars exporting AI computing infrastructure to countries without equivalent protections, which shuts down the offshoring argument directly. It includes water and energy reporting requirements as well. The lecture is blunt that this is an aspirational bill with no realistic chance of passing. Pro can use it as a model of what a federal moratorium could look like, but should be careful defending it as the likely implementation.

"Hyperscale" and "the United States federal government"

"Hyperscale" has no scholarly consensus definition. Common thresholds in the literature, per the lecture, include at least 5,000 servers, more than 10,000 square feet, 100 or more megawatts of new load, or at least $500 million in cost. State legislation varies too: the lecture notes that bills in states like Oklahoma, Minnesota, Vermont and Wisconsin use a 100-megawatt line, while Virginia’s uses 1 megawatt. Expect definition and topicality debates whether you want them or not. Pick a definition, know why it is reasonable, and be ready to defend which facilities it covers.

The actor may be the strangest part of the resolution. Zoning and retail electricity rates are state and local powers, so there is no obvious federal mechanism for a construction pause. The lecture lays out the main options. The Commerce Clause is the most direct constitutional route, though a flat ban on construction on private land is aggressive. The Clean Water Act and Clean Air Act could be used to freeze new permits for facilities with water or air discharges. FERC could set conditions on when large loads connect to the grid. Congress could attach spending conditions, withholding funds from states that permit new hyperscale facilities, the way it once tied a share of federal highway funds to the drinking age.

This gives Pro extra jobs: defend a specific implementation, a specific definition of hyperscale, and a specific federal mechanism. But it is also an opportunity. A Pro team that finds a modest, defensible implementation can sidestep many of Con’s standard disadvantages. A narrow FERC interconnection rule, for example, doesn’t obviously trigger a political capital argument the way a sweeping act of Congress would.

The status quo

As of the lecture in July 2026, the trend at the state and local level was strongly toward pauses. The lecture cites 14 states that had considered a moratorium and more than 300 state-level data center bills in early 2026. Maine passed a moratorium that was vetoed in April 2026, and New York became the first state to enact one in June, a one-year pause on hyperscale construction. Locally, the lecture counts 63 proposed moratorium actions, 54 of which passed, and more than $60 billion in AI projects blocked in about a year.

Public opinion is the driver. The lecture cites a March 2026 poll finding 71 percent opposed local AI data centers, 48 percent strongly, higher than opposition to nuclear power, with majorities across party lines. People may like AI in the abstract, but they don’t want the data center in their backyard. Meanwhile the federal government is moving the other way: the lecture notes a July 2025 executive order streamlining permits for facilities of 100 megawatts or more, a later order creating a Justice Department task force to challenge state AI laws, and a voluntary industry pledge in March 2026 to protect ratepayers. That split, between local opposition and federal acceleration, is the uniqueness for most arguments on the topic.

Pro arguments

Grid and ratepayers. Interconnection requests inflate the forecast, the forecast sets the auction price, and everybody pays. A Pro case here argues that a moratorium forces regulators to check the speculation before more costs land on households, improving reliability and slowing price increases. This is the most federal, most evidence-rich Pro advantage, because wholesale capacity is a FERC question.

Water. Hyperscale facilities have large heat rejection loads, the cheapest way to shed heat is evaporation, and evaporated water leaves the local watershed. The impact story runs through drying watersheds, ecosystems and agriculture. The lecture predicts plenty of extinction impacts off this link chain. Win the warrant first, using the withdrawal versus consumption distinction, or the impact won’t matter.

Cyber. Two versions exist. The first is about computing power itself: slowing new capacity slows AI development, which affects offensive and defensive cyber capabilities. That link can run either way depending on your evidence. The second is cleaner for Pro: every new hyperscale facility is a high-value target, and a successful attack on one could disable critical systems or physically damage hardware. Adding more of these targets to American infrastructure is a risk Pro can weigh.

Con arguments

Economy. The lecture cites AI-related investment driving 74 percent of US growth in the first quarter of 2026, with growth roughly flat without it. Con can argue local harms, like construction and operations jobs lost when projects stop, but the bigger link is investor confidence: a federal moratorium could pop the AI investment boom, with recession and spillover impacts.

The bubble turn. Pro will say the buildout is speculative. Con’s smart answer is that the moratorium doesn’t make the forecasts more accurate. It just guarantees the buildout stops. If the projected demand turns out to be real, Pro has popped the boom for no reason, and the economic impacts follow.

Politics. The lecture argues politics runs in only one direction on this topic. One version is federalism backlash: the federal government would need a novel use of its power to pause construction that states and localities usually regulate. The other is the classic political capital disadvantage, which is only as strong as your evidence that this specific policy trades off with another one.

Tech race and hegemony. Per the lecture, the US accounts for more than half of the world’s hyperscale data centers, with Europe and China each holding roughly a third of the rest. Con asks why the US would give up that lead for a pause that doesn’t guarantee a better policy afterward. Pro has answers here too, including provocation arguments: US capacity may itself be driving China’s race to catch up.

Strategy notes from the lecture

Pro needs to explain what the pause is for. On its own, a construction moratorium doesn’t change how many data centers exist today. It changes the trend. So Pro either argues that disrupting the trend matters in itself, or, more persuasively, that the pause creates time to do something: fix capacity forecasting, set cooling standards that cut consumption relative to withdrawal, or require reporting. This is one of the rare PF topics where defending a specific implementation is both necessary and possible.

Expect bad evidence, and use that. Much of the research comes either from industry-funded journalism or from advocacy groups opposed to AI. Studies mix up withdrawal and consumption, and confuse percent changes with percentage points. The lecture predicts many evidence indictments on this topic. Read your opponents’ cards closely: what exactly is being measured, and by whom?

Con can stay consistent. Every Pro case has to defend a definition, an implementation and a federal mechanism, and each of those is contestable. The lecture suggests a stable Con case paired with a rebuttal that generates a lot of new offense as turns against Pro’s specific implementation.

Warrant everything. Almost every judge has used AI recently and has an opinion about it. "Typing into a chatbot drains the ocean" alienates judges and explains nothing. Explain why dense AI hardware creates a heat problem, where the water actually goes, and why the harm is local rather than global. Teams that do that work will beat teams that rely on slogans, on either side.

Weighing and the solvency fight

Two Con arguments are worth adding to the lecture’s list. The first is threshold leakage: any definition of hyperscale gives builders a line to get under, so projects may be split into smaller facilities that escape the moratorium. The second is efficiency. Large facilities are generally more energy efficient per unit of computing than scattered smaller ones, so pausing the biggest builders could push demand toward less efficient sites and raise total energy use. Both attack whether Pro’s pause reduces the harms it targets, which is often more persuasive to judges than a long AI race link chain.

Most rounds weigh concrete, local, near-term harms (power bills, watersheds, grid reliability) against large, national, more speculative ones (the AI race, an investment-driven slowdown). Con’s best solvency line is that each Pro harm has a targeted fix, such as state rate classes, FERC forecasting reform or cooling standards, that doesn’t require stopping construction. Pro’s answer is timing: the buildout is moving faster than regulators can write rules, and every project approved under today’s rules locks in costs for decades. The team that wins that exchange usually controls the weighing.

Prep downloads

A free PDF brief for this resolution. Block files and sample cases are available to members of the National Debate Club travel team.

PDF brief

Cut cards for this resolution. Coming soon.

Block file

Pre-written answers to the arguments you'll hear every round.

Sample cases

Varsity-level Aff and Neg cases with blocks, written by the travel team.

National Debate Club

Sample cases are a member resource

Read the cases that win rounds.

Full varsity Aff and Neg cases for Hyperscale Data Center Moratorium, with the blocks and frontlines behind them, written by CDA's national circuit coaches.

130+

TOC bids

40+

Championships

3000+

Rounds coached

450+

Students

What members get

  • Semi-private coaching

    Two-to-one with a championship coach, every week of the topic cycle.

  • Full block files & sample cases

    Aff, Neg, frontlines and answers, all written before the tournament starts.

  • National circuit tournaments

    You travel with a team, a coach and prep that is already done.

  • TOC qualification support

    Bid-tournament strategy mapped to your schedule and your budget.

  • Alumni mentorship

    Former NDC debaters now at top programs, in your corner all season.

Taking you to the travel team in 10s

Frequently asked questions

Does the moratorium shut down existing data centers?

No. A moratorium on construction pauses new building. Existing facilities keep running. Pro should say so early, because Con teams sometimes argue impacts that only make sense if existing capacity disappears.

How long does the moratorium last?

The resolution doesn’t say. Pro can defend a fixed period, like New York’s one-year pause, or a conditional one that lasts until specific rules are in place, like the Sanders and Ocasio-Cortez bill. Con should push Pro to commit, because an open-ended pause is easier to attack.

Can companies just build data centers abroad instead?

Partly. Training is latency tolerant and can move far from users. Inference, which is roughly half of data center energy demand, is latency sensitive and wants to be near users, and whether it can be offshored effectively is unsettled. Make the other team say which kind of computing they claim will leak.

What is the difference between water withdrawal and consumption?

Withdrawal is all the water taken from a source. Consumption is the part that isn’t returned, mostly through evaporation. Consumption is what a watershed loses. Many studies and cards on this topic mix the two up, which makes them easy to indict.

How could the federal government actually do this?

There is no obvious single mechanism. Options include the Commerce Clause, freezing permits under the Clean Water Act and Clean Air Act, FERC conditions on grid interconnection, and spending conditions on states. Pick one and be ready to defend it, because Con will press on it.