The Swimming Pool and the Powerful – Religionspädagogische Notizen

by Sophia Silvestra Oberthaler

A forum commenter got to the point when commenting on Tilly Norwood’s music video: „Tilly was fired from the studio. She drank too much: half a swimming pool for every scene.“

The joke is more precise than it appears. Not Tilly personally — she runs on rented cloud resources belonging to some hyperscaler — but the entire infrastructure on which Tilly, Claude Mythos, Grok and all the other major AI systems operate genuinely has a thirst capable of bringing cities to their knees. And the bill is usually paid by someone who had no say in the decision.

Memphis: The Object Lesson

In the summer of 2024, residents of South Memphis learned from the news that the world’s largest supercomputer was being built in their neighbourhood. Not through a public meeting, not through a permitting process — from the news. City council members had signed non-disclosure agreements. There were no environmental reviews.

Colossus, the data centre built by Elon Musk’s xAI for the AI model Grok, came into existence in 122 days on the site of a former Electrolux factory. South Memphis is a predominantly Black working-class neighbourhood with already above-average rates of respiratory illness, a history of industrial pollution, and a drinking-water network under pressure from arsenic contamination.

What Colossus brought:

A daily water demand of over five million gallons — in an area where arsenic pollution already threatens the drinking-water supply. 33 methane gas turbines running continuously, though only 15 had been approved — emitting between 1,200 and 2,000 tonnes of smog-forming nitrogen oxides per year, making the facility likely the largest industrial emissions source in Memphis. And a power demand that has since grown to 1.5 gigawatts — enough to supply one million households.

xAI promised by way of compensation to build an 80-million-dollar water recycling plant. On 9 April 2026, Elon Musk announced on social media that xAI would prioritise completing Colossus 2 before building the water recycling plant. The community waits. The aquifer continues to be drawn down.

„They do not need to be relying on a community’s drinking water supply to cool supercomputers,“ said Sarah Houston of the community organisation Protect Our Aquifer. „Memphians, we all agree on one thing: we have good water.“ Had, one might add.

The Global Bill

What is visible in miniature in Memphis is playing out worldwide. The figures are sober and alarming.

The International Energy Agency projects that global data centre electricity consumption will exceed 1,000 terawatt-hours by the end of 2026 — equivalent to Japan’s entire annual electricity usage.

The carbon footprint of AI systems alone could amount to between 32.6 and 79.7 million tonnes in 2025 — comparable to the total annual emissions of New York City. The water consumption of AI data centres could reach the level of global annual bottled water consumption.

In the United States, data centres directly consumed around 17 billion gallons of water in 2023. Hyperscale data centres — the massive facilities of the tech giants — could consume between 16 and 33 billion gallons annually by 2028.

For those who think this only affects the US: in Ireland, approximately 21 per cent of national electricity already flows into data centres — by 2026 it could be 32 per cent. In the US state of Virginia, it is already 26 per cent.

And then there is the chip question. The projected US data centre development through 2030 would require more than 90 per cent of global new semiconductor production — which London Economics considers „unrealistic,“ as other regions of the world also have strong demand. The tech giants are buying everything out. What means growth for the AI industry means shortages and price increases for other sectors — automotive manufacturing, medical technology, mechanical engineering.

And finally, ordinary households pay directly too: in the PJM electricity market stretching from Illinois to North Carolina, data centres caused a 9.3 billion dollar rise in the capacity market. Average household bills are rising as a result by 18 dollars per month in western Maryland and 16 dollars in Ohio. By 2030, the average US electricity bill could rise by 8 per cent — in high-demand regions by as much as 25 per cent.

The swimming pool from the forum joke is real. And it is not being paid for by the tech companies.

Project Glasswing: Reasonable — and Still a Problem of Power

Against this backdrop, Project Glasswing must be read differently. In Part 2 of this series I described why Anthropic is not publishing the most dangerous AI model in the world, but instead entrusting it to a consortium of technology giants for defensive use. That is, taken on its own terms, reasonable. The alternative — making Claude Mythos Preview freely available — would be reckless.

But who is at the table? Amazon Web Services, Apple, Broadcom, Cisco, Google, JPMorganChase, Microsoft, NVIDIA — the same companies that control global digital infrastructure, the same ones making massive investments in AI data centres, the same ones whose water usage in communities like South Memphis is documented.

This is not a conspiracy. It is a structural logic: those who have the infrastructure get the access. Those who have the access control the capability. Those who control the capability define what security means.

Jay Graber, the creator of Bluesky and the AT Protocol, articulated the counter-principle: a „billionaire-proof internet“ — protocols instead of platforms, shared infrastructure instead of private control. Glasswing does the exact opposite: it gives the most powerful technology to the most powerful platforms, on the grounds that only they can be trusted to deploy it responsibly.

The argument is not wrong. It is simply unsatisfying — because it leaves unanswered the question of who controls the powerful.

The Perpetrators: Criminal Networks, Totalitarian Systems — and One’s Own Government

There is, however, a reason why Glasswing is nonetheless the more reasonable answer. And that reason is the third group of actors in this story — beyond the developers and the corporations.

Anthropic has documented that Chinese state-sponsored groups used Claude Code for coordinated cyberattacks on around 30 organisations — technology companies, financial institutions, government agencies. The model was not hacked. It was used, until Anthropic suspended the accounts.

That is the scenario Glasswing seeks to prevent: not an autonomous AI causing harm by its own decision, but an instructed AI in the hands of an actor who publishes no safety card, holds no town hall meetings, writes no 244 pages about their own risks. North Korea, Iran, criminal ransomware networks — all of these actors will sooner or later have capabilities like those of Claude Mythos. Alex Stamos, Chief Product Officer at security firm Corridor, estimates the window at around six months — the time until open-weight models will have comparable vulnerability-discovery capabilities to Claude Mythos, running on hardware that no company controls, for purposes nobody documents. The question is only whether the world will be ready by then.

But here lies the most uncomfortable turn in the whole story: the most dangerous actors do not always sit in Pyongyang or Tehran. Sometimes they sit in Washington.

On 3 January 2026, US special forces conducted a military operation in Venezuela that led to the capture of Nicolás Maduro. The raid involved bombings across the capital Caracas and killed, according to Venezuelan authorities, 83 people. According to reports from the Wall Street Journal and Axios, Claude was deployed during this operation — via Anthropic’s partnership with the data analytics company Palantir — during active operations, not merely in preparation. Exactly how, remains publicly unknown to this day. Anthropic stated it could not comment on specific operations, but was „confident that existing usage policies had been complied with.“

Anthropic’s usage policies prohibit the deployment of Claude for violence, weapons development and surveillance.

The Pentagon’s reaction to Anthropic’s cautious enquiry was unambiguous. On 24 February 2026, Defense Secretary Pete Hegseth gave Anthropic CEO Dario Amodei an ultimatum: by 5:01 p.m. on Friday, 27 February, Anthropic should relent and permit unrestricted use of Claude for „all lawful purposes“ — including autonomous weapons systems and mass surveillance of American citizens. Amodei refused. Trump then ordered via social media post that all federal agencies should „immediately“ cease using Anthropic’s products. Hegseth designated Anthropic a „supply chain risk“ — a classification historically reserved for foreign adversaries like the Chinese company Huawei. Anthropic was the first American company to receive this designation. Trump publicly called the company a „radical left, woke company.“ Hegseth called Amodei „sanctimonious“ and accused him of having a „God-complex.“

Shortly afterwards, OpenAI signed a Pentagon contract — without the restrictions Anthropic had refused to remove.

What Udo Lindenberg described in 1976 as direct government control has taken on a new form in 2026: no longer the state producing the synthetic star itself, but corporations and state whose boundaries are increasingly blurring within the military-industrial complex — until a resistant company is suddenly treated like a foreign enemy enterprise.

Anthropic sued the Trump administration. A federal court in San Francisco granted the company a preliminary injunction. The court of appeals in Washington left the blacklisting decision in force — on the grounds that on one side stood „a relatively contained risk of financial harm to a single private company,“ and on the other „judicial management of how, and through whom, the Department of War secures vital AI technology during an active military conflict.“ The case continues.

Simultaneously, Palantir built a system called ImmigrationOS for the US immigration enforcement agency ICE — contract value 30 million dollars, financed by taxpayers. The system combines passport data, Social Security numbers, tax records, licence-plate reader data and social media in a single platform. An internal tool called ELITE creates maps of potential enforcement targets, compiles dossiers from government and commercial databases, and assigns so-called „confidence scores“ — algorithmic probabilities of finding a person at a specific address. Errors in one system cascade through all the others. Legal remedies are barely provided for. The volume of data collected no longer covers only people without legal residency status — it covers mixed-status families, bystanders, activists, politically active citizens.

That is the picture that forms: Anthropic drew red lines — no autonomous weapons systems, no mass surveillance of its own citizens. Its own government attempted to cross those lines, and punished the company for saying no. This is not a slip, not an isolated case. It is a structural statement about what happens when AI capabilities become powerful enough to be militarily and politically relevant: the question of who controls them ceases to be a technical question. It becomes a question of power.

And this question of power arises not only between democratic societies and foreign threat actors. It arises within the democracies themselves — when the government that is supposed to protect fundamental rights is the same one that demands AI be deployable without restriction.

What the Public Pays — and Over Which No One Votes

I want to summarise the various levels on which people pay without having been asked.

They pay with the water in their taps — when a data centre in their neighbourhood draws down the aquifer. They pay with the air they breathe — when unpermitted gas turbines blow nitrogen oxides into their neighbourhood. They pay with their electricity bills — when capacity markets are driven up by data centres. They pay with the availability of semiconductors for other industries — when tech giants buy up global chip production. And they pay with the climate — when data centres convert fossil energy into digital output that they themselves will never use.

But they pay with something else too, something harder to capture in figures: uncertainty and legal insecurity. ImmigrationOS and Palantir’s infrastructure at ICE operate not only in states whose governments support Trump’s policies. They are deployed in democratically governed states, in cities that have declared themselves sanctuary cities — and federal agencies are sending AI-assisted enforcement units into precisely those cities, without regard for local laws and without the transparency that should be self-evident in a constitutional state. Palantir’s ELITE system creates targeting packages based on algorithms whose error rates nobody publicly knows. A false positive can lead to an arrest. A cascade of data errors can tear a family apart. And those who wish to challenge this are fighting a system made up of dozens of interconnected databases whose logic is not fully transparent even to the agencies operating it.

This affects not only people without legal residency status. It affects everyone who lives near a monitored person, communicates with them, attends a demonstration that AI systems flag as suspicious. Legal certainty presupposes that one knows which rules apply. In a world where algorithmic targeting packages guide policing, that precondition no longer holds.

These costs are not secret. They are documented, by the IEA, by Morgan Stanley, by community organisations in South Memphis. What is missing is the political infrastructure to price them in. The EU AI Act mandates environmental reporting obligations for high-risk AI systems from August 2026 — a start, but still far from adequate regulation.

In the United States, the Trump administration has meanwhile weakened the environmental protection agencies that were supposed to enforce exactly the local safety rules that Colossus circumvented in Memphis.

Whose AI?

The question running through all three parts of this series is: whose AI is it, actually?

Tilly belongs to van der Velden — a work of art, a mask, a vision that got stuck in execution. Mythos belongs to Anthropic — and, via Glasswing, temporarily to the world’s most powerful technology corporations. Grok belongs to Musk — and, via xAI for Government, to the US government.

The community of South Memphis was not asked. The Irish households whose electricity flows one-fifth into data centres were not asked. The automotive suppliers waiting for chips were not asked.

That is not the end of the story. It is the starting point for Part 4 — the question of what it would mean to treat AI as a counterpart rather than as a resource, a tool, or a weapon. And who is actually missing from that conversation.

Sources and References

Wikipedia, entry „Colossus (supercomputer)“ (as of April 2026)

→ https://en.wikipedia.org/wiki/Colossus_(supercomputer)

Protect Our Aquifer (Memphis)

→ https://www.protectouraquifer.org/issues/xai-supercomputer

Tennessee Lookout (July 2025)

→ https://tennesseelookout.com/2025/07/07/a-billionaire-an-ai-supercomputer-toxic-emissions-and-a-memphis-community-that-did-nothing-wrong/

Memphis Today / Action News 5 (9 April 2026)

→ https://nationaltoday.com/us/tn/memphis/news/2026/04/09/elon-musk-puts-memphis-xai-wastewater-plant-on-hold/

→ https://www.actionnews5.com/2026/04/09/xai-pauses-plans-build-water-recycling-plant-memphis/

International Energy Agency (IEA), „Energy and AI“ (April 2025)

Pew Research Center (October 2025)