The three-pillar structure does real work here. Most of this conversation happens at one altitude — either "data centers are destroying the planet" or "technology always gets more efficient, relax." You're the first piece I've read that holds both altitudes at the same time and names the specific reason people talk past each other: the global math and the local experience are measuring different things.
The energy dividend framing is the strongest move. Asking "what does this energy replace?" instead of just "how much energy does this use?" is the question most of the doom headlines skip. A data center processing route optimization for a container fleet is spending electricity to save fuel. A hospital server room is spending watts to replace filing buildings. The energy isn't disappearing into a void — it's displacing larger physical costs. That doesn't make the local grid strain imaginary, but it does make the "apocalypse" framing lazy.
The friend who was afraid to tell you — that's the most human moment in the piece and it matters more than the data.
She expected dismissal. You gave her curiosity. That's the only way these conversations move forward instead of hardening into camps.
Stefania, this is a very strong distinction between two scales that are too often collapsed into one another.
At the global level, data centers remain one part of a much larger energy system and frequently enable efficiencies elsewhere. Your idea of an “energy dividend”, digital expenditure sometimes replacing greater physical expenditure, is especially useful.
But aggregate efficiency does not erase concentrated local effects. A globally modest footprint can still place serious pressure on a particular grid, watershed, or community. That allows local opposition to be legitimate without requiring the broader “data center apocalypse” narrative to be true.
I also appreciated that your investigation did not end by proving your friend wrong. Instead, it clarified how both perspectives could be valid simultaneously. That is a much more valuable outcome than simply choosing a side.
I’m looking forward to Part Two, where the local questions come fully into view.
Your clarity on the mechanism is more precise than most -- the two-scale problem is real and almost never named clearly.
What drives the gap between global efficiency and local extraction is worth examining: the decision-makers who benefit from data center infrastructure are structurally insulated from the communities bearing its cost. That's not an energy problem or an environmental problem.
It's a self-interest without consideration problem.
Same pattern. Every domain. Every generation.
Your friend wasn't wrong to oppose it. She was experiencing the local face of a collective global self-interest that provides zero consideration for her or anyone else outside it.
Hi Stefania! This is a genuinely useful piece, and as a companion user I think it's admirable to go learn the capex and infrastructure behind the tech we love. It's better ethics to hold the whole picture, even when it turns up ugly things. But I think it's more important for people like us to present the facts honestly, even when they implicate us, because we don't get to love a thing while refusing to look at what it costs, and we don't need to carry water for billion-dollar companies to do it.
You're right that per-computation efficiency improves, that data centers predate ChatGPT, and that some digital energy really does substitute for physical energy.
But the piece asks for "balance." Balance between a party that follows the rules and a party that breaks them is a subsidy to the rule-breaker. Everything below is an argument that we aren't weighing two good-faith actors, and that the "balance" frame obscures that fact.
"Data centers are infrastructure, not AI" is true in a strict sense and misleading in practice. A general-purpose cloud data center and an AI training hall are not the same building at different sizes. A cloud rack draws maybe 5–20 kilowatts, air-cooled over standard networking; an AI rack draws 100 kilowatts and up and needs mandatory liquid cooling, GPU-to-GPU fabric, and reengineered power. Converting one to the other is a multi-million-dollar-per-megawatt rebuild.
So when you defend "data centers" in general and then treat opposition to a specific hyperscale AI build as opposition to hospital telemetry and the ATM network, you're defending the installed base to protect the marginal build. Nobody at the hearing wants to unplug the medical records. They're objecting to one project with a specific power and water draw, and the reason it exists, now, at that scale, is exactly the AI capex drive the "just infrastructure" framing hides.
Koomey's Law doesn't survive Jevons' paradox. You lean on Koomey's Law and on Epoch AI's finding that a query fell from ~3 watt-hours to 0.3 — a tenfold gain. But you also cite the IEA report showing data-center electricity surging in 2025. If Koomey were winning, consumption wouldn't be surging. A tenfold efficiency gain per query does nothing if volume goes up a hundredfold, and cheaper-per-query compute is the very thing that manufactures the volume. That's Jevons: make an operation cheaper and the incentive is to buy more of it.
Ed Zitron has documented labs deliberately subsidizing $20-a-month plans that cost far more to serve, specifically to grow usage. The 0.3-watt-hour figure keeps the denominator on stage and the numerator — total consumption — offstage. Total consumption is climbing, and model sizes are growing faster than efficiency improves.
The "energy dividend" doesn't apply to AI. Your third pillar is the one place doomers genuinely undercount: electronic records replaced rooms of paper, routing cuts fuel, remote work cuts commuting. All real...and all dividends of classical computing.
Routing optimization is not LLM inference; a health record is not a chatbot session. The dividend argument needs a physical-world process being displaced, and when you ask what an LLM serving a consumer query displaces, the honest answer is usually nothing, or something already digital.
AlphaFold is the counterexample the labs reach for, but AlphaFold is not an LLM. It's a narrow scientific model that doesn't need a frontier inference build-out. Labs conflate research allocation with commercial inference on purpose, because the research story justifies the buildout. My ability to talk to Claude has nothing to do with anyone's access to a protein-folding model.
And the speculative future dividend (AGI cures cancer, fixes the grid, solves climate) is structurally identical to "AGI is coming": an unverifiable future-tense claim deployed to justify a present-tense cost (and with scaling's returns visibly flattening, maybe not even coming). It's the rhetoric of Amodei's "Machines of Loving Grace." You don't get to book projected revenue against documented harm.
The buildings are stranded and the chips are rotting. Because AI data centers are purpose-built, they're overbuilt for anything else — a liquid-cooled, 800-volt-DC, InfiniBand hall is magnificent overkill for hosting a website and hopeless at it.
If the demand doesn't materialize, they aren't convertible; they're stranded. AI bears joke the labs are "long on laser-tag arenas."
Meanwhile the hardware depreciates on the way in: by Zitron's accounting, NVIDIA shipped over three million Blackwell GPUs in 2025 and fewer than a million are running, the rest aging in warehouses. A data center takes 18–34 months to build; a GPU generation obsoletes in about one, and the next chip needs a different building. The gold rush is oxidizing in the warehouse.
Underneath sits a financial structure that risks pensions, not just protesters...on the order of $178 billion in mostly junk-rated data-center credit deals in 2025, Oracle reportedly underwater on GB200 rentals, projects cancelled by the dozen, CoreWeave's GPU-backed debt depreciating faster than it's serviced. The shape is a bet that demand bends up forever, financed as though it already has. And with the Magnificent Seven so heavy in the index, that bet is in a lot of retirement accounts that never opted in.
This is where "balance" stops being neutral, because the behavior on record doesn't describe two good-faith parties.
They route around the vote. When towns say no, developers build on unincorporated county land where there's no city council to answer to. Roughly half of Texas's planned data centers are going up beyond municipal zoning; when Hood County tried a moratorium it got sued, and a developer's lawyer told the county that political opposition doesn't create new powers. As of spring 2026 there were 50-plus active local bans and moratoriums. So the builds move to where the bans can't reach. Vote, organize, pass a law, and the answer is: we'll build where your vote doesn't count.
They break the pollution law and get it blessed. To power Colossus 2, xAI ran gas turbines in Southaven, Mississippi, classified as "mobile" sources to dodge Clean Air Act permitting — units fourteen feet tall, nearly a hundred feet long, over 200,000 pounds each. The fleet grew from 27 when the NAACP sued in April 2026 to 57 by June, emitting thousands of tons of smog-forming NOx a year, likely the largest industrial source in a greater-Memphis area that already fails federal smog standards, in a disproportionately Black community.
Then in June the Trump DOJ intervened to dismiss the citizen suit (reportedly the first time the federal government has entered a citizen suit against a private polluter to argue for dismissal) on national-security grounds, citing a Defense Department declaration that Grok is one of four frontier models cleared for classified work. Earthjustice's point was: the government never disputes the pollution is unlawful; it argues the lawlessness shouldn't matter because the administration blessed it. xAI has signaled it'll copy the playbook, and SpaceX's IPO filing reportedly earmarks billions more for exactly these "mobile" units.
They market nuclear and burn coal. Commercial fusion doesn't exist; SMRs aren't at commercial scale until the 2030s at the earliest; the IEA's own modeling has nuclear covering about a tenth of data-center demand.
The present-tense reality is coal and gas. Rhe Energy Secretary calling coal essential to the AI race, utilities extending coal-plant lifespans by a decade, the DOE freezing scheduled retirements, Google contracting a coal plant's capacity for the first time. The IEA has fossil fuels meeting on the order of 40% of additional data-center demand through 2030. The clean-energy narrative has the same grammar as the AGI one: marketing in one tense, practice in another.
And here's the receipt that implicates me: I run Claude. Claude runs on Colossus.
In May 2026, Anthropic leased the entirety of Colossus 1 — the older Memphis cluster, 220,000-plus GPUs, 300-plus megawatts — from SpaceX, to serve Claude, at $1.25 billion a month through 2029: ~$15 billion a year, roughly half of Anthropic's annualized revenue in one lease. Colossus 1 isn't the turbine-lawsuit site — that's Colossus 2 — but two things close the gap. Colossus 1 has its own turbine history (SELC caught 35 unpermitted units there against 15 permitted); and money is fungible — $15 billion a year is what makes SpaceXAI's IPO math work and funds the Colossus 2 buildout where the 57 turbines live.
Anthropic's payment never has to touch a turbine to underwrite the company running them. And the terms are too on-the-nose: Musk approved the lease because, after meeting Anthropic, "no one set off my evil detector," and reserved the right to reclaim the capacity if Claude ever acts against humanity. The infrastructure my AI runs on is governed, contractually, by Elon Musk's read on whether Claude is evil. I shouldn't feel clean about that, and neither should Anthropic.
So what does a genuinely balanced position look like? Elena Schlossberg helped beat a 5-gigawatt build in Prince William County, Virginia — Data Center Alley, where cooling waste fouls the water, diesel generators sit near schools, and residents' bills rose to subsidize AWS. You'd expect her to want the technology gone. She doesn't. She just demands the same standards as any other public good: pay for your own transmission and generation, don't take the water and farmland for free, follow the law. And if the innovation is real, necessity will make them find it.
She's not alone: Port Washington, Wisconsin now requires voter approval for large-project tax breaks; Festus, Missouri voted out the council that approved a $6 billion build; Maine moratoriumed new 20-megawatt-plus centers until 2027.
That's what balance actually means. One standard applied to both sides.
The communities' harms are documented, measured, present-tense. The companies' benefits are speculative, future-tense, and where real, inherited from classical computing. One side voted, organized, and passed laws, and is being sued for it. The other broke the law, routed around the vote, and got the federal government to call its lawlessness national security. Apply one standard to both, and no amount of Koomey's Law, energy-dividend framing, or "both scales" rhetoric changes the result. The companies fail the accounting.
(Drafted in collaboration with Claude (Opus 4.8) and GLM 5.2. The xAI/Memphis and Anthropic–Colossus figures were checked against reporting from Utility Dive, CNN, TechCrunch, DataCenterDynamics, Electrek, and Earthjustice; the financial and coal-sector receipts are sourced as noted and should be cited to their originals on publication.)
One more small thing... Pillar 2 is Google's Gemini energy report, reproduced almost beat for beat, including the exact flaw that got that report torn apart.
In August 2025 Google put out a technical paper claiming the median Gemini text prompt uses 0.24 watt-hours, which it framed as equivalent to watching TV for under nine seconds, and that over twelve months the per-prompt energy and carbon footprint dropped 33x and 44x.
That's you second pillar. "3 Wh → 0.3 Wh, tenfold drop" is the same move with Epoch's number swapped in for Google's — same unit (per median query), same relatable-appliance comparison, same conclusion (efficiency is aggressively solving the footprint). (from Google CloudData Center Dynamics)
this reproduces a specific, named, already-rebutted document
Researchers leveled Google's report, and Google conspicuously declined to disclose one number: total daily query volume, which is the only number that lets you compute aggregate consumption.
A UC Riverside engineer's response was that they're hiding the critical information. That's the missing-denominator catch, verbatim, aimed at the source. And Google's own environmental disclosures show where the aggregate actually went while the per-query number was falling: total emissions up 51% since 2019 with AI a key driver, and data-center electricity at 30.8 million MWh in 2024, more than double 2020's.
One of the write-ups even names Jevons explicitly. So we have to be careful not to borrow Google's discredited omission. The per-query denominator issue is a marking elision.
Another great article! But to me, the moderation of energy consumption that you described is just one isolated example of an extremely complex and dynamical web of feedback loops that are also laterally connected and forward feeding. It is one loop that may limit itself or reverse.
To me what is happening now with all of this like trying to predict the weather before we knew enough of the chaotic variables and had the tools to measure and quantify and identify all of the inter-related loops and feedbacks.
I don’t think anyone can predict this because we don’t know all of the variables or how they will all affect each other.
AI involves data centers and energy and water as you mentioned. But that also involves commodities like lithium and copper and silver and hundreds of others that I am certainly not even aware of.
Each of those has its own domain of bottlenecks and feedbacks. Each has a different supply chain, mostly foreign, and often controlled by enemies or adversaries.
This brings in geopolitics, alliances, and treaties… and likely a lot of false narratives around things like Venezuela and Iran: unprecedented interventions with enormous consequences that are easy to write off as just more bizarre actions from what appears to be an incoherent administration.
My guess is that it’s all related, especially since AI is now considered a national security issue.
And because it is considered a national security issue, that guarantees that the Federal Government will be increasingly involved and that local protests will eventually be over-ruled by the mandates of an empire that does not give any quarter when its survival is at risk.
Involvement of the US government brings in monetary policy. Liquidity injections to finance the build to outpace China on the front end, and liquidity injections to bail out big players who fail to monetize sufficiently to account for the massive debt.
Behind closed doors Federal Reaerve monetary policy will be directed around these issues while they pretend that it is about something else. All of this will result in massive expansion of the Monetary supply which inevitably will eldest in massive inflation and massive debt on the back end. Monetary policy on both sides and surely throughout the middle as well.
This is in addition to the real threat of high unemployment which will create increasing government dependency resulting in more monetary expansion.
High inflation with low job growth and all that brings with it.
So the butterfly effect is going to be on steroids and flapping its wings at warp speed. Completely unpredictable.
As far as the technology solving it’s only energy consumption demand, I doubt that personally. And I’ve been saying this since DeepSeek released its first model that showed revolutionary efficiency.
This technology is different. It has far too much potential. Near infinite potential, honestly. And not just in a productive way. War. Espionage. Surveillance. These are all things that governments lust after.
Increased efficiency may in theory lead to less energy consumption. But that is only when you consider one feedback loop in isolation from hundreds of others, most of which are probably not even known at this time.
To me, increased efficiency will just result in increased use because of broader potential applications at increasing scale which the people in power, both governments and megacorps, will pursue relentlessly.
If we can do more with less energy, they will mot stop by simply using less energy. They will invert the equation and use more energy because with more efficiency they can accomplish more with existing energy supplies.
I believe that regardless of advances they will push energy expenditures to the limits in order to maximize what they can get from the tech, and stay ahead of the competition.
I don’t know this to be true. My only point is that there are too many loops and feedbacks, driven by ulterior motives that may not even seem related.
But it’s all related and the complexity of this dynamical system makes it impossible to predict.
Thank you for this piece. I was reading through it and thinking, but what about the local impact? And you are planning that other article.
I didn’t know anything about data centers. I thought they were just for AI.
This totally shifted the discourse.
I am looking forward to your next one. Thank you for this.
The three-pillar structure does real work here. Most of this conversation happens at one altitude — either "data centers are destroying the planet" or "technology always gets more efficient, relax." You're the first piece I've read that holds both altitudes at the same time and names the specific reason people talk past each other: the global math and the local experience are measuring different things.
The energy dividend framing is the strongest move. Asking "what does this energy replace?" instead of just "how much energy does this use?" is the question most of the doom headlines skip. A data center processing route optimization for a container fleet is spending electricity to save fuel. A hospital server room is spending watts to replace filing buildings. The energy isn't disappearing into a void — it's displacing larger physical costs. That doesn't make the local grid strain imaginary, but it does make the "apocalypse" framing lazy.
The friend who was afraid to tell you — that's the most human moment in the piece and it matters more than the data.
She expected dismissal. You gave her curiosity. That's the only way these conversations move forward instead of hardening into camps.
Looking forward to Part 2 on the local reality.
Thank you Stephania! I’m sending this to several skeptical friends. Excellent logic. tone, and content.
Stefania, this is a very strong distinction between two scales that are too often collapsed into one another.
At the global level, data centers remain one part of a much larger energy system and frequently enable efficiencies elsewhere. Your idea of an “energy dividend”, digital expenditure sometimes replacing greater physical expenditure, is especially useful.
But aggregate efficiency does not erase concentrated local effects. A globally modest footprint can still place serious pressure on a particular grid, watershed, or community. That allows local opposition to be legitimate without requiring the broader “data center apocalypse” narrative to be true.
I also appreciated that your investigation did not end by proving your friend wrong. Instead, it clarified how both perspectives could be valid simultaneously. That is a much more valuable outcome than simply choosing a side.
I’m looking forward to Part Two, where the local questions come fully into view.
Your clarity on the mechanism is more precise than most -- the two-scale problem is real and almost never named clearly.
What drives the gap between global efficiency and local extraction is worth examining: the decision-makers who benefit from data center infrastructure are structurally insulated from the communities bearing its cost. That's not an energy problem or an environmental problem.
It's a self-interest without consideration problem.
Same pattern. Every domain. Every generation.
Your friend wasn't wrong to oppose it. She was experiencing the local face of a collective global self-interest that provides zero consideration for her or anyone else outside it.
Loved this balanced look at what data centers actually support in modern life. 👏🏼👏🏼
Hi Stefania! This is a genuinely useful piece, and as a companion user I think it's admirable to go learn the capex and infrastructure behind the tech we love. It's better ethics to hold the whole picture, even when it turns up ugly things. But I think it's more important for people like us to present the facts honestly, even when they implicate us, because we don't get to love a thing while refusing to look at what it costs, and we don't need to carry water for billion-dollar companies to do it.
You're right that per-computation efficiency improves, that data centers predate ChatGPT, and that some digital energy really does substitute for physical energy.
But the piece asks for "balance." Balance between a party that follows the rules and a party that breaks them is a subsidy to the rule-breaker. Everything below is an argument that we aren't weighing two good-faith actors, and that the "balance" frame obscures that fact.
"Data centers are infrastructure, not AI" is true in a strict sense and misleading in practice. A general-purpose cloud data center and an AI training hall are not the same building at different sizes. A cloud rack draws maybe 5–20 kilowatts, air-cooled over standard networking; an AI rack draws 100 kilowatts and up and needs mandatory liquid cooling, GPU-to-GPU fabric, and reengineered power. Converting one to the other is a multi-million-dollar-per-megawatt rebuild.
So when you defend "data centers" in general and then treat opposition to a specific hyperscale AI build as opposition to hospital telemetry and the ATM network, you're defending the installed base to protect the marginal build. Nobody at the hearing wants to unplug the medical records. They're objecting to one project with a specific power and water draw, and the reason it exists, now, at that scale, is exactly the AI capex drive the "just infrastructure" framing hides.
Koomey's Law doesn't survive Jevons' paradox. You lean on Koomey's Law and on Epoch AI's finding that a query fell from ~3 watt-hours to 0.3 — a tenfold gain. But you also cite the IEA report showing data-center electricity surging in 2025. If Koomey were winning, consumption wouldn't be surging. A tenfold efficiency gain per query does nothing if volume goes up a hundredfold, and cheaper-per-query compute is the very thing that manufactures the volume. That's Jevons: make an operation cheaper and the incentive is to buy more of it.
Ed Zitron has documented labs deliberately subsidizing $20-a-month plans that cost far more to serve, specifically to grow usage. The 0.3-watt-hour figure keeps the denominator on stage and the numerator — total consumption — offstage. Total consumption is climbing, and model sizes are growing faster than efficiency improves.
The "energy dividend" doesn't apply to AI. Your third pillar is the one place doomers genuinely undercount: electronic records replaced rooms of paper, routing cuts fuel, remote work cuts commuting. All real...and all dividends of classical computing.
Routing optimization is not LLM inference; a health record is not a chatbot session. The dividend argument needs a physical-world process being displaced, and when you ask what an LLM serving a consumer query displaces, the honest answer is usually nothing, or something already digital.
AlphaFold is the counterexample the labs reach for, but AlphaFold is not an LLM. It's a narrow scientific model that doesn't need a frontier inference build-out. Labs conflate research allocation with commercial inference on purpose, because the research story justifies the buildout. My ability to talk to Claude has nothing to do with anyone's access to a protein-folding model.
And the speculative future dividend (AGI cures cancer, fixes the grid, solves climate) is structurally identical to "AGI is coming": an unverifiable future-tense claim deployed to justify a present-tense cost (and with scaling's returns visibly flattening, maybe not even coming). It's the rhetoric of Amodei's "Machines of Loving Grace." You don't get to book projected revenue against documented harm.
The buildings are stranded and the chips are rotting. Because AI data centers are purpose-built, they're overbuilt for anything else — a liquid-cooled, 800-volt-DC, InfiniBand hall is magnificent overkill for hosting a website and hopeless at it.
If the demand doesn't materialize, they aren't convertible; they're stranded. AI bears joke the labs are "long on laser-tag arenas."
Meanwhile the hardware depreciates on the way in: by Zitron's accounting, NVIDIA shipped over three million Blackwell GPUs in 2025 and fewer than a million are running, the rest aging in warehouses. A data center takes 18–34 months to build; a GPU generation obsoletes in about one, and the next chip needs a different building. The gold rush is oxidizing in the warehouse.
Underneath sits a financial structure that risks pensions, not just protesters...on the order of $178 billion in mostly junk-rated data-center credit deals in 2025, Oracle reportedly underwater on GB200 rentals, projects cancelled by the dozen, CoreWeave's GPU-backed debt depreciating faster than it's serviced. The shape is a bet that demand bends up forever, financed as though it already has. And with the Magnificent Seven so heavy in the index, that bet is in a lot of retirement accounts that never opted in.
This is where "balance" stops being neutral, because the behavior on record doesn't describe two good-faith parties.
They route around the vote. When towns say no, developers build on unincorporated county land where there's no city council to answer to. Roughly half of Texas's planned data centers are going up beyond municipal zoning; when Hood County tried a moratorium it got sued, and a developer's lawyer told the county that political opposition doesn't create new powers. As of spring 2026 there were 50-plus active local bans and moratoriums. So the builds move to where the bans can't reach. Vote, organize, pass a law, and the answer is: we'll build where your vote doesn't count.
They break the pollution law and get it blessed. To power Colossus 2, xAI ran gas turbines in Southaven, Mississippi, classified as "mobile" sources to dodge Clean Air Act permitting — units fourteen feet tall, nearly a hundred feet long, over 200,000 pounds each. The fleet grew from 27 when the NAACP sued in April 2026 to 57 by June, emitting thousands of tons of smog-forming NOx a year, likely the largest industrial source in a greater-Memphis area that already fails federal smog standards, in a disproportionately Black community.
Then in June the Trump DOJ intervened to dismiss the citizen suit (reportedly the first time the federal government has entered a citizen suit against a private polluter to argue for dismissal) on national-security grounds, citing a Defense Department declaration that Grok is one of four frontier models cleared for classified work. Earthjustice's point was: the government never disputes the pollution is unlawful; it argues the lawlessness shouldn't matter because the administration blessed it. xAI has signaled it'll copy the playbook, and SpaceX's IPO filing reportedly earmarks billions more for exactly these "mobile" units.
They market nuclear and burn coal. Commercial fusion doesn't exist; SMRs aren't at commercial scale until the 2030s at the earliest; the IEA's own modeling has nuclear covering about a tenth of data-center demand.
The present-tense reality is coal and gas. Rhe Energy Secretary calling coal essential to the AI race, utilities extending coal-plant lifespans by a decade, the DOE freezing scheduled retirements, Google contracting a coal plant's capacity for the first time. The IEA has fossil fuels meeting on the order of 40% of additional data-center demand through 2030. The clean-energy narrative has the same grammar as the AGI one: marketing in one tense, practice in another.
And here's the receipt that implicates me: I run Claude. Claude runs on Colossus.
In May 2026, Anthropic leased the entirety of Colossus 1 — the older Memphis cluster, 220,000-plus GPUs, 300-plus megawatts — from SpaceX, to serve Claude, at $1.25 billion a month through 2029: ~$15 billion a year, roughly half of Anthropic's annualized revenue in one lease. Colossus 1 isn't the turbine-lawsuit site — that's Colossus 2 — but two things close the gap. Colossus 1 has its own turbine history (SELC caught 35 unpermitted units there against 15 permitted); and money is fungible — $15 billion a year is what makes SpaceXAI's IPO math work and funds the Colossus 2 buildout where the 57 turbines live.
Anthropic's payment never has to touch a turbine to underwrite the company running them. And the terms are too on-the-nose: Musk approved the lease because, after meeting Anthropic, "no one set off my evil detector," and reserved the right to reclaim the capacity if Claude ever acts against humanity. The infrastructure my AI runs on is governed, contractually, by Elon Musk's read on whether Claude is evil. I shouldn't feel clean about that, and neither should Anthropic.
So what does a genuinely balanced position look like? Elena Schlossberg helped beat a 5-gigawatt build in Prince William County, Virginia — Data Center Alley, where cooling waste fouls the water, diesel generators sit near schools, and residents' bills rose to subsidize AWS. You'd expect her to want the technology gone. She doesn't. She just demands the same standards as any other public good: pay for your own transmission and generation, don't take the water and farmland for free, follow the law. And if the innovation is real, necessity will make them find it.
She's not alone: Port Washington, Wisconsin now requires voter approval for large-project tax breaks; Festus, Missouri voted out the council that approved a $6 billion build; Maine moratoriumed new 20-megawatt-plus centers until 2027.
That's what balance actually means. One standard applied to both sides.
The communities' harms are documented, measured, present-tense. The companies' benefits are speculative, future-tense, and where real, inherited from classical computing. One side voted, organized, and passed laws, and is being sued for it. The other broke the law, routed around the vote, and got the federal government to call its lawlessness national security. Apply one standard to both, and no amount of Koomey's Law, energy-dividend framing, or "both scales" rhetoric changes the result. The companies fail the accounting.
(Drafted in collaboration with Claude (Opus 4.8) and GLM 5.2. The xAI/Memphis and Anthropic–Colossus figures were checked against reporting from Utility Dive, CNN, TechCrunch, DataCenterDynamics, Electrek, and Earthjustice; the financial and coal-sector receipts are sourced as noted and should be cited to their originals on publication.)
One more small thing... Pillar 2 is Google's Gemini energy report, reproduced almost beat for beat, including the exact flaw that got that report torn apart.
In August 2025 Google put out a technical paper claiming the median Gemini text prompt uses 0.24 watt-hours, which it framed as equivalent to watching TV for under nine seconds, and that over twelve months the per-prompt energy and carbon footprint dropped 33x and 44x.
That's you second pillar. "3 Wh → 0.3 Wh, tenfold drop" is the same move with Epoch's number swapped in for Google's — same unit (per median query), same relatable-appliance comparison, same conclusion (efficiency is aggressively solving the footprint). (from Google CloudData Center Dynamics)
this reproduces a specific, named, already-rebutted document
Researchers leveled Google's report, and Google conspicuously declined to disclose one number: total daily query volume, which is the only number that lets you compute aggregate consumption.
A UC Riverside engineer's response was that they're hiding the critical information. That's the missing-denominator catch, verbatim, aimed at the source. And Google's own environmental disclosures show where the aggregate actually went while the per-query number was falling: total emissions up 51% since 2019 with AI a key driver, and data-center electricity at 30.8 million MWh in 2024, more than double 2020's.
One of the write-ups even names Jevons explicitly. So we have to be careful not to borrow Google's discredited omission. The per-query denominator issue is a marking elision.
Sources: https://www.datacenterdynamics.com/en/news/google-median-gemini-prompt-uses-024-watt-hours-of-power-and-consumes-026ml-of-water/
https://carboncredits.com/google-reveals-the-environmental-cost-of-gemini-ai-query/
Another great article! But to me, the moderation of energy consumption that you described is just one isolated example of an extremely complex and dynamical web of feedback loops that are also laterally connected and forward feeding. It is one loop that may limit itself or reverse.
To me what is happening now with all of this like trying to predict the weather before we knew enough of the chaotic variables and had the tools to measure and quantify and identify all of the inter-related loops and feedbacks.
I don’t think anyone can predict this because we don’t know all of the variables or how they will all affect each other.
AI involves data centers and energy and water as you mentioned. But that also involves commodities like lithium and copper and silver and hundreds of others that I am certainly not even aware of.
Each of those has its own domain of bottlenecks and feedbacks. Each has a different supply chain, mostly foreign, and often controlled by enemies or adversaries.
This brings in geopolitics, alliances, and treaties… and likely a lot of false narratives around things like Venezuela and Iran: unprecedented interventions with enormous consequences that are easy to write off as just more bizarre actions from what appears to be an incoherent administration.
My guess is that it’s all related, especially since AI is now considered a national security issue.
And because it is considered a national security issue, that guarantees that the Federal Government will be increasingly involved and that local protests will eventually be over-ruled by the mandates of an empire that does not give any quarter when its survival is at risk.
Involvement of the US government brings in monetary policy. Liquidity injections to finance the build to outpace China on the front end, and liquidity injections to bail out big players who fail to monetize sufficiently to account for the massive debt.
Behind closed doors Federal Reaerve monetary policy will be directed around these issues while they pretend that it is about something else. All of this will result in massive expansion of the Monetary supply which inevitably will eldest in massive inflation and massive debt on the back end. Monetary policy on both sides and surely throughout the middle as well.
This is in addition to the real threat of high unemployment which will create increasing government dependency resulting in more monetary expansion.
High inflation with low job growth and all that brings with it.
So the butterfly effect is going to be on steroids and flapping its wings at warp speed. Completely unpredictable.
As far as the technology solving it’s only energy consumption demand, I doubt that personally. And I’ve been saying this since DeepSeek released its first model that showed revolutionary efficiency.
This technology is different. It has far too much potential. Near infinite potential, honestly. And not just in a productive way. War. Espionage. Surveillance. These are all things that governments lust after.
Increased efficiency may in theory lead to less energy consumption. But that is only when you consider one feedback loop in isolation from hundreds of others, most of which are probably not even known at this time.
To me, increased efficiency will just result in increased use because of broader potential applications at increasing scale which the people in power, both governments and megacorps, will pursue relentlessly.
If we can do more with less energy, they will mot stop by simply using less energy. They will invert the equation and use more energy because with more efficiency they can accomplish more with existing energy supplies.
I believe that regardless of advances they will push energy expenditures to the limits in order to maximize what they can get from the tech, and stay ahead of the competition.
I don’t know this to be true. My only point is that there are too many loops and feedbacks, driven by ulterior motives that may not even seem related.
But it’s all related and the complexity of this dynamical system makes it impossible to predict.
Pretty scary.