The Other Three Floors Lead
A data center isn't an AI thing. That's the first move in the campaign against them, and it's the one that does all the quiet work. Call it an "AI data center" and you get to take every anxiety attached to AI — the job losses, the bubble, the sense that nobody asked for this — and aim it at a slab of concrete. But across the existing stock, AI is about a quarter of the load. The other three-quarters is the ordinary digital economy running on conventional CPU servers, and it is not optional. Think of it as two categories, not one building: the conventional CPU data centers that are most of the installed footprint and run daily life, and the newer, power-dense AI GPU campuses that are most of the growth — different machines, usually in different buildings.
The opposition's own stack is on those floors too. The platforms that activism runs on — the petition sites, the video town halls, the advocacy email tools, the streaming, the search, the generative AI now drafting public testimony — are all hosted in the very buildings under dispute. The campaign to condemn the infrastructure is itself a tenant of it. You can oppose data centers. You just can't do it without one.
The building, floor by floor
The "floors" are shares of the total national compute load, not tenants in one literal building — in the real world the AI floor is mostly its own GPU campuses. A moratorium proposes condemning the whole stack to slow down the newest slice. Select a floor to see what's running there.
- Frontier models, copilots, the tools driving the productivity gains
- A distinct, power-dense GPU category — mostly its own purpose-built campuses, not mixed into the buildings below
- The newest builds increasingly self-generate power and bypass the grid entirely (see behind the meter) — the category least likely to touch your residential rate
- Banking & payments: card networks, ACH/wire clearing, fraud detection, ATM/POS auth
- Communications: email, messaging, video calls, VoIP, the entire mobile-app layer
- Media & knowledge: news systems, streaming, search, every website
- Healthcare: EHRs, medical imaging (PACS), lab results, e-prescribing, scheduling
- Government: tax filing, benefits, DMV/licensing, courts, 911 dispatch, elections
- Logistics & retail: inventory, freight matching, warehouse automation, point-of-sale
- Utilities themselves: grid balancing, SCADA, smart-meter data, outage management
- Petition platforms, town-hall video, advocacy email tools
- The streamed press conference, the reporter's cloud CMS, the search that surfaced the outrage
- The generative AI now drafting the public testimony against data centers
| Floor / sector | What's actually running there | What breaks if you halt the buildout |
|---|---|---|
| Banking & payments | Card networks, ACH and wire clearing, real-time fraud detection, ATM and point-of-sale auth | Cards decline, payroll doesn't settle, the fraud check protecting your account goes dark |
| Healthcare | Electronic health records, imaging (PACS), lab results, e-prescribing, hospital scheduling | Doctors lose your chart, prescriptions stall, imaging can't route |
| Government | Tax filing, benefits, DMV/licensing, courts, 911 dispatch, election systems | The portal you renew your license on; the system that pays out benefits |
| Logistics & retail | Inventory, freight matching, warehouse automation, last-mile routing, point-of-sale | Empty shelves, packages stop tracking, "out of stock" everywhere |
| Communications | Email, messaging, video calls, VoIP, the entire mobile-app layer | Your texts, your work calls, your group chat |
| Utilities themselves | Grid balancing, SCADA, smart-meter data, outage management | The power company can't run the power company |
| Media & knowledge | News content systems, streaming, search, every website | Including the article telling you to fear data centers |
| The opposition's stack | Petition platforms, town-hall video, advocacy email, the AI drafting the testimony | The campaign against data centers — which is hosted in one |
| AI (GPU, the top floor) | Frontier models, copilots, the tools driving the productivity gains — on purpose-built GPU campuses, mostly their own buildings | ~27% of the load — the newest category, getting blamed for the whole stack |
Every modern political campaign is now a digital campaign, which means it is a data-center campaign. The petition reaches you through a hyperscaler. The outrage clip streams from one. The "ban data centers" mailing list is managed in one. This isn't hypocrisy worth a gotcha — it's evidence of the actual claim: this infrastructure is so foundational that even its opposition can't function without it. The thing you're being told to fear is carrying the message that tells you to fear it.
Naming the whole category after its newest, scariest member lets you import every AI anxiety and aim it at concrete. It's shutting the power plant to slow down crypto. The thing you're afraid of is one class of machine — the GPU campuses — and the ban you're being sold also condemns the conventional CPU data centers that run daily life.
Quick Reference — Claims vs. Reality
The cheat-within-the-cheat: the highest-frequency rebuttals, in one scan.
| What you'll hear | What's actually true |
|---|---|
| "AI data centers are eating the grid." | AI is ~27% of data-center workload. The other ~73% is conventional cloud and legacy enterprise — the digital economy you already use. |
| "A moratorium pauses AI." | It freezes the physical layer of the entire digital economy — the conventional CPU data centers that are most of the stock — to slow down the newer GPU category. |
| "Data centers raise everyone's electric bill." | Only where power is tight and a monopoly smears the cost across the base. Where the load pays its own way, they can lower rates. |
| "Data centers are driving the demand spike." | On ERCOT's forecast, even with zero new data centers, Texas peak demand still climbs ~40% (≈87→121 GW by 2031) on manufacturing, oil & gas, crypto, and electrification. |
| "AI is spiking your bill." | A fraction (AI) of a fraction (data centers) of total demand growth — wearing the costume of the whole. |
| "Data centers are draining and polluting the water." | Cooling water is evaporated or recirculated in closed loops, not contaminated. US data centers' direct use is ~0.01% of national water — a whole year ≈ a few hours of farm irrigation. |
| "This is a runaway luxury nobody asked for." | The buildout added more to US GDP growth over two quarters than all consumer spending combined. |
| "Big Tech is getting a free ride." | Partly true — the tax breaks are indefensible. But that argues for killing the subsidy, not banning the building. |
The Workload Mix (What the Dollars Hide)
The confusion between "data center" and "AI" isn't an accident of language — it's baked into how the money looks.
The function is mostly plumbing; the dollars skew to AI. By 2027 the workload mix is projected at roughly half conventional cloud, a quarter legacy enterprise, and about 27% AI Goldman Sachs proj., Feb 2025. But the capital spending tells a different story, because an AI rack draws far more power than a conventional one: NVIDIA's current rack-scale systems pull around 130 kW, with next-generation designs targeting 250 kW, against 10–15 kW for a conventional rack — and they cost to match. So in dollars, this is unmistakably an AI buildout. In what the installed base collectively does, it's still mostly the machinery of daily life, with a fast-growing new category beside it.
Yes — dedicated AI-only GPU campuses exist, and the marginal new shovel is heavily AI: the great majority of the megawatts now being broken ground on are GPU builds, not conventional cloud. "It's not an AI thing" is a claim about the installed base — what the existing stock collectively runs and who pays for the grid — not a claim that every new building is mixed-use. The stock is mostly conventional; the flow is mostly AI. Both are true at once, and the argument here only needs the first.
Always ask which lens it's using. In dollars, AI dominates — it's where the capex and the power draw concentrate. In function, AI is the minority — most of what these buildings do is the ordinary digital economy. A claim that uses the dollar figure to describe the workload (or vice versa) is doing a magic trick. Watch which cup the ball is under.
The Newest Tenant Is Also the Most Valuable
The top floor everyone's afraid of happens to be the most valuable thing in the economy right now — and most of the value doesn't land where you'd expect.
The value lands on users, not sellers
US consumer surplus from generative AI ran about $172 billion a year by early 2026, up from $112 billion twelve months earlier, with median value per user roughly tripling — and most tools are free or close to it Stanford HAI, Apr 2026 · modeled. Surplus is the part the seller doesn't capture; it lands on the user.
Jobs & wages went up where AI is heaviest
The industries most exposed to AI posted higher productivity, employment, and wages — not lower: 34% vs 24% productivity growth since 2018, 52% vs 36% headcount growth, a 62% AI-skills wage premium PwC 2026 · correlational. If AI were the job-destroyer of the narrative, the damage would show first where the tools are used most. It's the opposite.
"Adds to GDP" is the abstraction. Here's the value in the form people actually feel it — hours back, errors caught, a specialist's help that used to be out of reach:
- Coders ship ~55% faster on the tasks an assistant covers — a controlled GitHub study clocked completion at well under half the time GitHub/Copilot, 2023.
- Support agents resolve ~14% more cases an hour, with the largest gains for the newest staff — AI hands a junior worker a veteran's playbook NBER, Brynjolfsson et al., 2023.
- Doctors get their evenings back: ambient AI scribes cut documentation time and "pajama time" charting, and are linked to measurable drops in physician burnout health-system pilots, 2024–25.
- A free tutor and translator for everyone: writing help, step-by-step math, plain-language explanations of a legal notice or a lab result, and real-time translation — capabilities that used to cost money or a credential, now sitting on a phone for nothing.
The Buildout Is Carrying the Economy
The buildout isn't riding the economy. For the moment, it's carrying it.
AI capex has out-driven all consumer spending. Over the first two quarters of 2025, AI-related capital expenditure added more to US GDP growth than all consumer spending combined Renaissance Macro, 2025. It now runs near 5% of GDP — a leading driver of US economic growth KKR, 2025–26.
Pull the buildout out and you don't land in a world with cheaper power and the same prosperity. You land in the recession it's been papering over. "Pause the data centers" is also "pause the thing currently holding up the economy" — said in a calmer voice.
The Cost Grievance Is Real — and It's a Pricing Problem
None of which makes the cost complaint fake. It isn't.
Residential power is genuinely up. Prices rose about 42% from March 2021 to March 2026 on EIA monthly data PolitiFact, Jun 2026 — and about 27% on EIA's annual-average basis (2019–2024), with more in some regulated markets. Pick either yardstick: the bill went up, and it's a real grievance.
But a pricing problem and a production problem share no solutions. The complaint is about who pays for the grid and how the cost is allocated — a pricing question. The moratorium attacks supply — a production answer. They don't connect. And the pricing fixes are already arriving:
| Where | Mechanism | What it fixes |
|---|---|---|
| FERC (federal) | Show-cause orders (Jun 2026) directing the six grid operators to make large loads bear the cost of upgrades built to serve them tariffs pending | Stops the buildout's grid costs from landing on ordinary ratepayers |
| Oregon & Virginia | Dedicated large-load rate classes — Oregon's POWER Act (2025); a Virginia SCC order (Nov 2025, ≥85% contracted demand) | Separates big-load costs from the residential base |
| Texas (SB6, 2025) | Puts interconnection cost plus curtailment / remote-disconnect duties on loads ≥75 MW | Makes the data center carry the capacity it triggers |
Set the tariff right and the grievance shrinks to a line item. You don't halt the most productive infrastructure wave since electrification to spare someone forty dollars a month that a rate class already fixes.
A Fraction of a Fraction
The grievance leans on a conflation that runs the whole way down.
Data centers are a minority of new electricity demand. AI is a minority of the data centers. On the latest forecast from ERCOT — the operator that runs the Texas power grid — data centers account for roughly 46% of projected electricity demand growth through 2031. Even with no new data centers at all, Texas peak electricity demand would still climb about 40% (from ~87 to ~121 gigawatts) on manufacturing, oil and gas, crypto, and the shift to electric cars and appliances ERCOT forecast, Apr 2025. So the pressure on your power bill mostly isn't data centers, and the data-center share mostly isn't AI.
"AI is spiking your bill" — let's decompose that claim, one layer at a time.
"AI is spiking your electric bill" is a fraction of a fraction wearing the costume of the whole. Strip the costume off and you're left with a line item inside a line item — being blamed for the entire bill.
"They're Draining and Polluting the Water"
The second-most-common attack after electricity — and, like the bill, it's mostly a conflation. Three different things get fused into one scary sentence: water withdrawn vs. water consumed, on-site cooling vs. power-plant cooling, and national totals vs. local stress.
"Pollution" is the wrong word. Cooling water isn't chemically contaminated and dumped. It is evaporated in cooling towers, withdrawn and returned, or recirculated in a sealed closed loop. Evaporated water doesn't vanish — it returns to the hydrologic cycle as rain. Of all the water claims, "polluting our water" is the easiest to retire.
The national numbers are a rounding error. US data centers directly consumed about 66 billion liters — roughly 17 billion gallons — of cooling water in 2023 LBNL, 2023. That's on the order of 0.01–0.02% of US water withdrawals. For scale: American agriculture withdraws about 118 billion gallons per day for irrigation — so data centers' entire year of direct water use is a few hours of farm watering.
| Use | Share of US water |
|---|---|
| Thermoelectric power generation | ~41% of withdrawals |
| Irrigation (agriculture) | ~42% of freshwater |
| Public water supply | ~12% |
| Data centers (direct cooling) | ~0.01–0.02% |
Sector shares: USGS national water-use survey (withdrawals). Data-center direct use: LBNL (2023). Power and farming together are ~80% of US water; data centers are a rounding error beside them.
Withdrawal is water taken from a source; most of it is returned. Consumption is the part that evaporates and doesn't come straight back. Thermoelectric power withdraws ~41% of US water but consumes only a sliver of it. A statistic that quotes a big withdrawal number as if it were all "used up" is doing the same magic trick as the electricity bill — counting the gross to imply the net.
Most "AI water" is really power-plant water. A data center's indirect water footprint — the cooling at the power stations feeding it — runs roughly 12× its on-site cooling LBNL, 2023. Which means the water story mostly tracks the grid, not the server hall: shift that power toward nuclear, wind, and solar — which consume little to no operating water — plus efficiency, and the larger share shrinks. The clean-firm-power buildout covered below is also a water story.
Hyperscaler water efficiency (WUE) now runs about 0.12–0.30 liters per kWh, a fraction of older designs. Microsoft's 2024 data-center design uses sealed, chip-level liquid cooling with zero evaporative loss — avoiding more than 125 million liters per facility per year. AWS reports dozens of sites running on 100% reclaimed (non-potable) water. New AI builds trend toward closed-loop and immersion cooling, where the water is a sealed coolant, not something consumed company reports, 2024–25. One honest tradeoff: closed-loop and air cooling can raise energy use, which nudges the indirect footprint up — so it's efficiency, not magic.
The viral line traces to a legitimate peer-reviewed study (Ren et al., "Making AI Less 'Thirsty'"). But it actually estimated that a 500 mL bottle covers roughly 10–50 medium responses — not one — and is highly location- and season-dependent. The famous "700,000 liters to train GPT-3" is the on-site slice of a ~5.4-million-liter total. Cite the concept; don't repeat the bumper-sticker version, which overstates the original several-fold Ren et al., 2023–25.
The honest concession: tiny nationally doesn't mean harmless everywhere. Siting an evaporative-cooled facility in a drought-stressed region, drawing on potable municipal supply instead of reclaimed water, or sending concentrated "blowdown" to a small wastewater plant — those are real, local problems. And the fix is the same shape as the electricity answer: site it where water is available, require reclaimed or closed-loop cooling, and make the load carry its own impact. That's a permitting-and-design problem, not a reason to ban the building.
Data Centers Can Lower Rates
Here's what should retire "data centers raise prices" as a flat claim: they don't, everywhere.
Where load pays its own way, big customers spread fixed costs. In inflation-adjusted terms, North Dakota's residential rates fell between 2019 and 2024 — one of the largest real declines in the country Lawrence Berkeley Nat'l Lab — even as the nominal rate ticked up from 10.3¢ to 11.5¢/kWh, because new large loads spread the grid's fixed costs across more usage. Oncor told Texas legislators that a single Dallas-area project would throw off $160 million in revenue against $143 million in cost — a net gain for everyone else on the system, as long as the load pays for the capacity it triggers Oncor testimony, Apr 2026.
Where power is plentiful and the customer carries its own weight, big loads push rates down by spreading fixed costs. Where power is tight and a regulated monopoly smears the cost across the base, they push up. Same building either way. The variable is the rate-and-energy policy wrapped around it — which means the honest target of your anger is your public utility commission, not the server hall down the road.
And the biggest AI campuses increasingly never touch your rate base at all. The newest, most power-dense GPU builds — the exact category the opposition is aimed at — are going behind the meter: self-generating their own power and bypassing the regional grid. More than a quarter of new facilities above 500 MW are projected to self-supply by 2030, up from roughly 1% today; in Texas, nearly half the gas plants now under construction are slated to serve data centers exclusively, without ever connecting to the grid everyone else pays for industry projections, 2025–26. A load that builds its own plant and wires straight into it can't socialize grid-upgrade costs onto residential ratepayers — because it isn't on the residential grid. Draw the CPU/GPU line and this falls out cleanly: the high-density AI category is both the newest and the one least likely to show up on your bill.
Self-supply removes the ratepayer-socialization problem specifically; it does not make a project frictionless. Air permits and gas-pipeline approvals have delayed or blocked several of these builds — including a Stargate site in New Mexico — and a behind-the-meter gas plant trades a grid-cost objection for a local emissions one. So state it precisely: it takes the cost off your bill, not the project out of the fight.
The Patterns Underneath
A local veto over construction. The cost landing on the neighbors. The benefit diffusing across the country. The whole thing dressed up as protecting the character of the place. We've watched that movie for forty years with housing — the same zoning machinery that made it illegal to build homes where people want to live is now being swung at the infrastructure the modern economy runs on. We know how it ends.
Kill the subsidy and the moratorium — in the same motion. One concession, and it cuts against the companies: the tax breaks are indefensible. Texas alone is forgoing at least $1.3 billion this fiscal year — and a projected $3 billion-plus over the 2027–2028 biennium — on its data-center sales-tax exemption alone, separate from any local incentives TX Comptroller, 2025. If a project already pencils out on cheap land and cheap power — which is exactly why these things cluster in Texas — it doesn't need a tax holiday on top. Make them pay market price for power and full freight on taxes, like any other industrial load. That isn't carrying water for Big Tech. It's refusing to socialize either the costs or the gains.
The buildout is financing the nuclear revival fifty years of energy politics couldn't deliver. Every major hyperscaler has now signed nuclear. Customers with 24/7 baseload needs and the balance sheets to underwrite first-of-a-kind reactors are the best thing to happen to American nuclear in a generation.
| Player | Project / site | Scale | Status |
|---|---|---|---|
| Microsoft | Three Mile Island Unit 1 (Crane Clean Energy Center) restart | ~835 MW | 20-year PPA; restart targeted 2027 (accelerated from 2028) |
| Amazon | Talen Susquehanna | ~1.9 GW | Front-of-meter PPA (after FERC rejected the original co-location, 2024–25) |
| Amazon | X-energy Xe-100 SMRs (Energy Northwest) | up to ~960 MW | Development; led ~$500M round |
| Kairos Power SMRs | ~500 MW | Development | |
| Meta | Oklo Aurora SMRs | ~1.2 GW | Development |
| Industry total | Announced hyperscaler + data-center nuclear deals (13) | ~9.8 GW | Mostly announced/contracted near-zero operating |
The strategic floor. The federal push to speed grid connections was framed explicitly as a China race: Energy Secretary Chris Wright urged FERC to help the US "better compete with China for superiority in the fast-growing AI sector," and in June 2026 FERC unanimously directed grid operators to connect large AI loads AP, Jun 2026. Compute is the binding constraint on frontier models, and frontier AI is the strategic technology of the next two decades the way oil was for the last hundred. "Pause the data centers" is "cede the lead," said in a calmer voice. That's a real floor of the building too — the most consequential one — even if it isn't most of the square footage.
The movement that strangled nuclear for decades, then recoiled when gas turbines filled the gap, is now trying to block the load that's financing the clean firm power it says it wants.
Common Mistakes & Anti-Patterns
Each is a trap on the left, the correction on the right.
The Honest Position
The position was never pro-data-center or anti. It's that the thing being built runs the economy we already have, is financing the one we're about to get, and delivers more value to more people than nearly anything else going.
The people selling you the pause aren't protecting your bill. They're using it — and they're reaching you through the very infrastructure they're telling you to fear.
Sources & Methodology
Forecasts are labeled as projections with their vintage. Figures that drift quickly — consumer surplus, GDP share, electricity prices, the ERCOT and tax-exemption forecasts — are dated inline.
- Workload mix (27% AI by 2027): Goldman Sachs Research, "AI to drive 165% increase in data center power demand by 2030," Feb 4, 2025. Projection.
- Rack power density (~130 kW AI today, 250 kW roadmap; 10–15 kW conventional): NVIDIA technical blog (GB200/GB300 NVL72); Schneider Electric, "The 1 MW AI rack is coming," Oct 2025; Uptime Institute Global Data Center Survey 2024.
- Consumer surplus ($172B / $112B; value-per-user tripled): Stanford HAI, 2026 AI Index Report, Ch. 4 (Economy), Apr 2026. Modeled willingness-to-pay estimate.
- AI-exposed industries (productivity / employment / wages): PwC, 2026 Global AI Jobs Barometer. Correlational.
- Tangible productivity (developers ~55% faster; support agents +14%/hr; clinical scribes & burnout): GitHub, "Quantifying GitHub Copilot's impact on developer productivity," 2023; Brynjolfsson, Li & Raymond, "Generative AI at Work," NBER w31161, 2023; health-system ambient-documentation pilots (e.g., Kaiser Permanente, The Permanente Medical Group), 2024–25. Task-level / pilot estimates.
- AI capex > consumer spending (H1 2025): Renaissance Macro Research (Neil Dutta), 2025, via DataCenterDynamics, Aug 6, 2025.
- ~5% of GDP: KKR (Henry McVey), "Beyond the Bubble: Why AI Infrastructure Will Compound," 2025–26.
- Residential prices (+42% Mar 2021–Mar 2026 monthly / +27% annual 2019–24): PolitiFact fact-check, Jun 12, 2026; U.S. EIA Electric Power Monthly.
- FERC large-load cost recovery: FERC show-cause orders, Docket RM26-4, Jun 18, 2026. Tariffs pending.
- Rate classes: Oregon POWER Act (2025); Virginia SCC order, Nov 2025.
- Texas SB6 (≥75 MW): Texas SB 6, 89th Legislature, signed Jun 20, 2025.
- ERCOT load growth (46% data centers; ~40% / ≈121 GW ex-data-centers by 2031): ERCOT 2025 Long-Term Load Forecast, Apr 2025; data-center share derived via electricityplans.com. Projection.
- North Dakota real-dollar rate decline: Lawrence Berkeley National Laboratory; nominal prices per U.S. EIA. Inflation-adjusted.
- Oncor ($160M vs $143M): Brian Lloyd testimony, Texas House State Affairs Committee, Apr 2026. Project forecast.
- Behind-the-meter / self-supply (>25% of >500 MW facilities self-supplying by 2030, up from ~1%; ~half of Texas gas plants under construction serving data centers off-grid; New Mexico Stargate permitting delays): industry analyses and trade reporting on data-center power procurement, 2025–26. Projections; figures vary by source.
- Texas data-center sales-tax exemption ($1.3B/yr; $3B+/biennium): Texas Comptroller, Tax Exemptions & Tax Incidence, 2025.
- Hyperscaler nuclear (~9.8 GW; TMI; Susquehanna; X-energy / Kairos / Oklo): SMR Intel deal tracker (Jun 2026); Constellation Energy (Sep 2024); Utility Dive (Jun 2025). Mostly announced/contracted, near-zero operating.
- China-competitiveness framing: Energy Secretary Chris Wright / FERC action, Associated Press, Jun 18, 2026.
- Data-center water (direct ~66B L in 2023; indirect ≈12× direct; WUE): Lawrence Berkeley National Laboratory, "2024 United States Data Center Energy Usage Report," Dec 2024. Modeled.
- US water-use shares (irrigation / thermoelectric / public supply; withdrawal vs. consumption): USGS, "Estimated Use of Water in the United States" (latest national survey). Withdrawals.
- Closed-loop / zero-evaporation cooling & reclaimed water (WUE 0.12–0.30 L/kWh): Microsoft "zero-water" data-center design announcement, Dec 2024; Amazon/AWS water-use disclosures, 2025. Company reports; some figures forward-looking.
- "500 mL per 10–50 responses" / GPT-3 training water: Li, Yang, Islam & Ren, "Making AI Less 'Thirsty'" (arXiv:2304.03271), 2023–25. Estimated; widely misquoted.