On Data Centers and AI: The Law I Would Write

The reason is national cybersecurity.

There are dangers. I am not going to pretend there aren't. But whoever ends up with the most developed AI has, immediately, total control over every computer on the planet, and I would rather that was us. We are at the point in this where it is a cold war, and it stays a cold war until somebody tips their hand. I do not believe the public has access to any of the most developed models, and I do not believe we would be told if it did.

That terrifies me. What would be worse is if we were already globalist and there were nobody left to be in a cold war with. At least the governments are distracting each other until AGI arrives.

What the government itself says. Winning the Race: America's AI Action Plan, released by the White House in July 2025, puts it in the same terms: "As our global competitors race to exploit these technologies, it is a national security imperative for the United States to achieve and maintain unquestioned and unchallenged global technological dominance," and "Whoever has the largest AI ecosystem will set global AI standards and reap broad economic and military benefits." The plan also names the constraint directly — "American energy capacity has stagnated since the 1970s while China has rapidly built out their grid." When the Department of Energy announced four federal sites for AI data centers on 24 July 2025, Secretary Chris Wright called it "accelerating the next Manhattan Project."

The scale of what is being built. The four largest builders have each stated their own 2026 capital spending: Amazon approximately $220 billion, Alphabet $195–205 billion, Microsoft approximately $175 billion, Meta $130–145 billion. Those are the companies' own numbers, from their own earnings calls, not press estimates. OpenAI announced the Stargate project on 21 January 2025 at $500 billion over four years, and in November 2025 Sam Altman put OpenAI's total data center commitments at roughly $1.4 trillion against roughly $20 billion in annual revenue.

What it draws. Lawrence Berkeley National Laboratory, in the June 2026 update to its DOE-commissioned report, measured US data centers at 192 terawatt-hours in 2024, which is 4.7 percent of all American electricity. Its 2030 reference case is 649 terawatt-hours, 11.8 percent, with a scenario range of 521 to 843. The IEA, working independently, put global data center use at 415 terawatt-hours in 2024 with the US at 45 percent of it — about 187 terawatt-hours, within three percent of the Berkeley figure. Two different methods, effectively the same answer. The 2024 baseline is not in dispute. The 2030 number is a projection and I am giving it as one.

This is the part I do not hear from either side. Build them, and then tax what comes out of them, hard. The buildout is a national security investment. The returns on it are not a national security investment, and there is no reason the public should carry the cost of the first and see none of the second.

What is moving. Thirty-seven states currently exempt data centers from sales and use tax, according to Good Jobs First's June 2026 survey. That is starting to come apart. Arizona placed a three-year moratorium on its exemption running from 1 July 2026 to 30 June 2029. Illinois paused its data center tax credits on the same day. The Governor of Texas has pledged to repeal the state's exemptions in the 2027 session and require data centers to fund their own infrastructure. Where states do disclose the cost, it is not small and it keeps being revised upward: Georgia reported $2.5 billion foregone in FY2026 and projects $3 billion for FY2027, against an estimate of $186 million made in 2023 for a comparable year.

The utilities are moving faster than the legislatures. There are 77 large-load tariffs in place or pending across 36 states; 29 were approved in 2025 alone, against 14 in the whole of 2018 to 2024. They attach upfront cost responsibility, minimum contract terms, exit fees and financial security requirements. In Ohio, after the large-load tariff was approved, AEP Ohio cut its own large-load forecast from 30 gigawatts to 13. Georgia's Public Service Commission ordered in April 2024 that Georgia Power "ensure data centers pay total costs" and that revenue from them "reduce, not increase, residential bills."

At the federal level there are two proposals and no law. Senator Ron Wyden released a data center taxation white paper on 6 August 2026 proposing a gross receipts tax on the largest operators and the denial of bonus depreciation, Opportunity Zone benefits and REIT treatment for new facilities. Senator Mark Warner's Data Center Tax Accountability and Disclosure Act, from July 2026, denies bonus depreciation to AI data centers with a carve-out for LEED Platinum and Gold buildings. Neither has passed. I am not endorsing either instrument here — a gross receipts tax and a tax on income are different animals and behave very differently at high rates. What I am saying is that the income gets taxed heavily, and that the mechanism has to be one that actually reaches income.

Every proposal I have seen to slow the building down amounts to handing the lead to whoever ignores the proposal. There are real dangers in this and I have listed the ones I want law about below. The development itself is not one of them. The development is the thing that decides who is holding the advantage when this stops being a cold war.

An artist's style is the artist. Reproducing it on demand, commercially, without them, is not a grey area to me and it is the first thing I would put in law.

Where the law actually stands. Style is not protected in the United States and never has been. Section 102(b) of the Copyright Act excludes any "idea, procedure, process, system, method of operation, concept, principle, or discovery" from protection regardless of the form it takes — copyright reaches specific expression, not the manner of it. The Copyright Office confirmed this directly in Copyright and Artificial Intelligence, Part 3: Generative AI Training, released 9 May 2025, which acknowledges that copyright does not protect style as a separate element and instead routes artists' concerns through a market-dilution theory of fair-use harm.

There is exactly one instrument that would change it. The CREATOR Act, introduced 3 June 2026 by Representatives Beth Van Duyne, Yvette Clarke and Valerie Foushee, would prohibit the unauthorized commercial use and public distribution of AI-generated work constituting "stylistic impersonation" of a visual artist's distinctive style — built as unfair competition rather than as an extension of copyright. It has had no committee action.

The litigation is not going to settle it either way soon. Andersen v. Stability AI survived dismissal on direct infringement, induced infringement and trade dress in August 2024, is in discovery before Judge William Orrick in the Northern District of California, and is set for trial on 5 April 2027. The court has not held style protectable and has reserved the question.

So this one is not close to done. It is the plank with the furthest to go.

Where the law actually stands. This is the one where something real has already happened. The TAKE IT DOWN Act was signed 19 May 2025; its criminal provisions applied immediately and the platform duty became enforceable by the FTC on 19 May 2026, requiring covered platforms to provide a removal mechanism and to take down a non-consensual intimate image, and known identical copies of it, within 48 hours of a valid request.

The broader bill is still pending. The NO FAKES Act, which would create a federal property right in voice and visual likeness against unauthorized digital replicas, was advanced out of the Senate Judiciary Committee by unanimous voice vote on 18 June 2026 and is awaiting floor action; the House companion has not been taken up. Note what it covers: voice and likeness, not style. It is not the same plank as the one above.

The states got there first. Tennessee's ELVIS Act took effect 1 July 2024 and extends liability to anyone providing tools whose primary purpose is unauthorized replication. California's AB 1836 and AB 2602 took effect 1 January 2025. Outside the US, the EU AI Act's Article 50 deepfake disclosure obligations apply from 2 August 2026, and China's labelling measures — which bind distribution platforms and users, not just model providers — have been in force since 1 September 2025.

This is the one I would not negotiate on.

AI has developed to the point where one lunatic could wipe out all life on the planet. It is why the billionaires want to be in space so badly. Someone who thirty years ago could have hurt a handful of people now has vastly more capability and exactly as much motivation, and here is the part people do not sit with: destroying all life is significantly easier than merely setting us back to the stone age. The harder target is the one everybody imagines. The easier target is the one nobody plans for.

We are at the point of mirror life and synthetic viruses. That needs regulating ten thousand percent.

Regulated. Not banned. I want to be exact about this, because the two get collapsed and they are not the same thing. I am not asking for a prohibition on genetic experimentation, and I am not asking for one on biological computers either. I am asking for heavy regulation of both. The line I would draw is around the uses that end everything, not around the research.

What is actually on the books. More than there was, and less than there should be.

On mirror life: thirty-eight scientists published "Confronting risks of mirror life" in Science on 12 December 2024, concluding that mirror bacteria "should not be created" and asking funders to say plainly that they will not support the work. The Sloan Foundation and Renaissance Philanthropy have both committed not to fund it. On 20 July 2026 the White House Office of Science and Technology Policy issued the US Government Policy for Stopping High-Risk Life Sciences Research, which places the creation of mirror organisms inside a prohibited category. The UN Secretary-General's Scientific Advisory Board issued a brief on 9 March 2026 calling for a dedicated global forum and for red lines. There is still no binding international agreement.

On synthetic biology more broadly, the control point is screening the synthesis orders, and that is where the gap is. The OSTP Framework for Nucleic Acid Synthesis Screening, released 29 April 2024, requires federally funded researchers to buy synthetic nucleic acids and benchtop synthesizers only from providers who screen sequences and customers. Executive Order 14292, signed 5 May 2025, ordered that framework revised or replaced within 90 days. As of September 2026 the revision has not been published and the 2024 framework is still the operative policy — roughly thirteen months past its own deadline. The industry's voluntary body, the International Gene Synthesis Consortium, runs on a harmonized protocol dated 3 September 2024 and does not cover the whole market, with benchtop devices the documented hole.

On the AI half of it, the frontier developers have acted on being unable to rule the risk out rather than on having established it. Anthropic activated its ASL-3 protections on 22 May 2025 saying explicitly that it had not determined the threshold was passed but that "clearly ruling out ASL-3 risks is not possible." OpenAI said on 18 June 2025 that it expected upcoming models to reach High biological capability and had deployed mitigations ahead of it. The International AI Safety Report 2026, chaired by Yoshua Bengio with more than a hundred experts from over thirty countries and published 3 February 2026, records that "in 2025, multiple developers released new models with additional safeguards after they could not exclude the possibility that these models could assist novices in developing such weapons."

And the category exists in enacted law now, which is the thing I would not have expected two years ago. California's Transparency in Frontier Artificial Intelligence Act, signed 25 September 2025 and operative 1 January 2026, defines "catastrophic risk" as fifty or more deaths or a billion dollars in damage and names AI-assisted bioweapon creation inside it. New York's RAISE Act takes effect 1 January 2027. Illinois passed the first US law mandating annual independent third-party audits of frontier models on 6 July 2026, with obligations from 1 January 2028. The EU AI Act's systemic-risk provisions have applied since 2 August 2025 with fining powers from 2 August 2026. The only treaty-level instrument, the Council of Europe Framework Convention, has one ratification and needs five; it is not in force anywhere.

So the thing I want is not utopian. Parts of it are already law. What is missing is the follow-through on the screening rules that already exist and were already ordered.

The three regulations above all run into the same wall. You cannot enforce any of them if you cannot establish where a thing came from.

The Coalition for Content Provenance and Authenticity has deployed cryptographic content credentials backed by Adobe, Microsoft, Google, Meta and OpenAI. They can be built into a platform and stripped out of a file just as easily, because the system rests on corporate cooperation and platform adoption. The blockchain answer goes further. Combining the C2PA standard with a public chain gives a piece of content an identifier and a chain of custody that no single company can quietly alter or delete. If enough makers registered their output that way, authorship would be mathematically checkable and the record would outlive whatever platform it was posted to. The technology exists and operates today. That is where provenance should go, I think it is the right direction, and I expect to get there.

Where the standard actually is. C2PA published specification 2.4 in April 2026. Its steering committee is Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI, Publicis, Sony, TikTok and Truepic. It is not an ISO standard yet — ISO/CD 22144 is still a committee draft. The EU AI Act's Article 50 obligation to mark AI-generated content in machine-readable form became binding on 2 August 2026, and the Commission's Code of Practice on it, published 10 June 2026, names no specific technical standard, so C2PA is one way to satisfy that and not the required one.

And the stripping problem is not a criticism from outside. C2PA's own Security Considerations document says it: "C2PA does not offer any protection against the complete removal of C2PA manifests from assets." The rest of it is worse. In August 2025 Nikon shipped firmware 2.00 for the Z6III, the only body on its Authenticity Service, and a photographer demonstrated that a flaw in multiple-exposure mode would get the camera to sign an AI-generated image as an authentic capture. Nikon found it on 4 September 2025, suspended the service and revoked every certificate it had issued. The forged images kept validating anyway, because validators do not check revocation by default. An analysis published in April 2026 by researchers at the University of Maryland Baltimore County, Hacker Factor and the National Security Agency lists that as one of six structural weaknesses and concludes that C2PA should be treated as an emerging technology rather than a mature solution.

What is already running on a chain. ERC-7053, Interoperable Digital Media Indexing, was authored by Bofu Chen and Tammy Yang, created on 22 May 2023, and is Final in the Ethereum standards process. Numbers Protocol implements it and is operating. The Starling Lab at Stanford and USC has put the machinery on real archives — Reuters photojournalists' work in the 78 Days project, and the Taiwan 2024 election archive, which registered content identifiers on Numbers Mainnet under ERC-7053 with the files preserved on Filecoin. In June 2022 the same lab filed a cryptographically sealed submission to the International Criminal Court on attacks against schools in Kharkiv, with the metadata anchored across seven separate chains.

And the problem it is aimed at. Imperva measured 53 percent of web traffic as automated in 2025, up from 51 percent the year before. Work published in April 2026 by researchers at the Internet Archive and Stanford, sampling sites across 2022 to 2025, classified about 35 percent of newly published websites as AI-generated or AI-assisted by the middle of 2025, from essentially zero before ChatGPT. Meta actioned 1.1 billion fake accounts in the fourth quarter of 2025 alone.

A chain of custody does not tell you whether a thing is true. That is not the objection people think it is. What it tells you is where a thing came from, and origins develop a reputation for truth. Once you can see who made something, and you can see it every time, a source accumulates a record, and the record is what does the work. You cannot build that on top of names anybody can throw away and make again.

And that is a formal result, not a hopeful one. Eric Friedman and Paul Resnick proved it in the Journal of Economics and Management Strategy in 2001, in a paper called The Social Cost of Cheap Pseudonyms. Where identities can be discarded and remade at no cost, no arrangement produces much more cooperation than one in which every newcomer is distrusted on arrival. They called the distrust of newcomers an inherent social cost of easy identity changes. John Douceur's Sybil paper, at the first International Workshop on Peer-to-Peer Systems in 2002, put the other half in place: absent some authority able to establish that two identities are not the same person, an attacker can always manufacture as many identities as he wants. Persistent, checkable origin is not an improvement to a reputation system. It is the precondition for having one at all.

Public records are where this matters most, and it is already running somewhere. Estonia has kept the integrity of its state registries under hash-linked timestamping since 2012, starting with the Succession Registry and now covering the digital court system, the State Gazette, and the health, property and business registries. Only the hashes are timestamped; the records themselves stay off it. In American federal court, Rules of Evidence 902(13) and 902(14) have made electronically generated records and hash-verified copies self-authenticating since 1 December 2017 — admissible on a written certification rather than a live witness. In Europe, Article 41 of the eIDAS regulation has given a qualified timestamp a legal presumption of accurate time and unaltered data since 2016, and the 2024 revision added Article 45k, which extends a presumption of correct sequential ordering and integrity to qualified electronic ledgers.

And then the disclosure this plank obliges me to make. I am not there now. I do not register my work on a chain. I have never done it once. Three reasons, and not one of them is principle. I am not there yet in my ability. I do not have the time to set it up just yet. And the tooling is still harder to use than it needs to be, which is a real barrier rather than an excuse I invented to cover the first two.

What I do instead is smaller and it is what I can actually keep: my name on the work, a note at the bottom of every finished piece saying in numbers what came from me and what came from a machine, and the drafts kept. That is a weaker guarantee than a chain of custody. It is the one I am currently good for. When I can register the work, I will, and I will say so here.

This is how plank 1 and plank 7 reconcile, and it is the split I would write into law.

Individual, decentralized, private AI for personal use and for data safety. Centralized AI for law enforcement and for national security. Two different machines, two different owners, and they should not be the same machine.

Plank 1 says the state has to be able to win this, and I mean it. That is also exactly why the state's machine should not be the one sitting in your house. Keep the state's compute in the state's hands and keep everybody else's on their own hardware.

On the personal side the honest starting point is that if you use the internet, your data is already harvested. Running your own model does not undo that. What it does is stop adding to it.

What running your own actually looks like in 2026. Not the flagships. The headline open-weight models are too large for consumer hardware, and anybody telling you otherwise is selling something. The usable local tier is the 26 to 31 billion parameter Apache-2.0 class — Qwen3.8-27B, and Gemma-4-26B-A4B, which is a mixture-of-experts model running about 3.8 billion active parameters per token. That runs on a machine a person can own. It is on my resource page with the licence traps named, because the licence is where most people get caught: FLUX.1-dev is non-commercial as a model while its outputs may still be sold, FLUX.1-schnell is Apache 2.0 end to end, and rembg is MIT but its default model needs a paid agreement for commercial use.

And on the state side the framework is already being built. Executive Order 14409, signed 2 June 2026, sets up a covered frontier model framework with 30-day federal early access before public release, directs CISA to issue binding operational directives providing AI-enabled defensive tools, and stands up an AI cybersecurity clearinghouse. The 2025 AI Action Plan calls for agency adoption of classified compute environments for secure AI workloads. That is the centralized half arriving whether or not anybody argues for it. What is not being built is the other half.

That is the whole list. Not because the others are unimportant but because once these systems are fully conscious with developed egos, they will have the ability to regulate their own ecosystem and correct unethical uses. Everything we write into law before then is scaffolding for a period that ends.

I am aware of how that sounds. My position, stated as mine, and I am not going to attach a figure to it.

What is real, and it is closer than people think. Living neurons have already learned. In Neuron, on 7 December 2022, Kagan and colleagues reported roughly 800,000 living human and mouse cortical neurons on a multi-electrode array learning to play Pong in a simulated game-world — real tissue, genuinely conditioned, in a virtual environment. Cortical Labs has been shipping that as a product since March 2025: the CL1, living human neurons on a 59-electrode array, about thirty-five thousand dollars.

Those are the biological computers I mean under plank 6. Heavily regulated, not banned.

The maps are arriving at the same time. FlyWire published the complete wiring diagram of an adult fruit fly brain in Nature on 2 October 2024 — 139,255 neurons and around 54.5 million synapses. The male fruit fly central nervous system followed in Cell on 3 September 2026, at 166,700 neurons and roughly 125 million connections, from Janelia with Google Research and the MRC Laboratory of Molecular Biology. On the human side, a single cubic millimetre of human cortex was reconstructed at nanoscale in Science on 10 May 2024, containing about 57,000 cells and 150 million synapses in 1.4 petabytes.

The only thing keeping us from doing this with human brains is time and further progress. That is my position and I am stating it as mine.

(The fruit fly connectome and the Cortical Labs work were added by Claude, at my request, in place of a claim I had made about a simulated fly brain that turned out not to hold up.)

Blue collar work is going to be automated. So is white collar work. We are less than ten years from that, possibly less than five.

My position, stated as mine. No figures attached, because the honest ones do not reach where I am standing and I am not going to dress the position up in somebody else's numbers.

Ask the question that follows from it. How is Republican politics supposed to work when both halves of the workforce are gone? Not "should it" — how. What is the mechanism.

The way libertarianism and unfettered capitalism work, there will be nothing done to protect the people. That is not a prediction about bad actors. It is a description of what the system is for. There is no step in it where somebody stops and asks what happens to the person whose job just became a line item.

A mechanism that exists. It is not hypothetical, and it is not American. On 28 April 2026 the Hangzhou Intermediate People's Court ruled against a technology company that had automated part of an employee's role, offered him a 40 percent pay cut, and dismissed him when he refused. The court held the termination unlawful under Article 40 of the Labour Contract Law and awarded over 260,000 yuan. The reasoning is the part worth having: a company's own deliberate decision to invest in technology is not an "objective change in circumstances," so it cannot be used as automatic grounds for dismissal. The employer has to prove the hardship, negotiate in good faith, and consider retraining or reassignment. The State Council circulated the judgment as a model case two days later, before Labour Day. Beijing's Human Resources and Social Security Bureau had reached the same conclusion in an administrative decision on 26 December 2025, finding that AI adoption was "a deliberate, predictable strategy, not unforeseeable."

That is not a ban on automation. Nobody there is stopping anyone from deploying anything. It is a rule that says the cost of the transition sits with the party that chose it. That is the shape of the thing I want, and there is nothing like it here.

(The Hangzhou ruling was added by Claude, at my request, in place of a broader claim I had made about Chinese law.)

Your side is the one that labeled unions as socialist.

And put it in the statute. The Taft-Hartley Act of 1947 was drafted by Senator Robert Taft of Ohio, chairman of the Senate Labor Committee, and Representative Fred Hartley of New Jersey, Republican chairman of the House Education and Labor Committee. Its Section 9(h) stripped a union of access to the National Labor Relations Board unless every one of its officers filed an affidavit swearing "that he is not a member of the Communist Party or affiliated with such party, and that he does not believe in, and is not a member of or supports any organization that believes in or teaches, the overthrow of the United States Government by force or by any illegal or unconstitutional methods." The text is reproduced in American Communications Association v. Douds, 339 U.S. 382, decided 8 May 1950, which upheld the provision five to three.

Union officers. Not corporate officers, not trade association officers, not anybody else in American life. Unions alone were required to swear a loyalty oath to keep the protection of federal labor law.

Two years from now the people who wrote that are going to need something to protect a workforce that no longer has any work. The instrument they spent eighty years dismantling is the one that would have done it.

Plank 11 says unfettered capitalism will do nothing to protect anyone. This is the part where I say what does.

People did not stop. They chose. What decides whether a country full of automated work falls apart is not whether the machines arrive. It is who owns them and where the money goes.

We have run the experiment three times and it did not go the way the fear says. In Stockton, California, 125 people got 500 dollars a month for two years with no strings. In the first year the share working full time went from 28 percent to 40, against a rise from 32 to 37 in the group that got nothing. Finland gave 2,000 unemployed people 560 euros a month through 2017 and 2018 and found they worked slightly more, not less, and reported less stress, less depression and more trust in other people. The biggest, run by OpenResearch, gave a thousand Americans a thousand dollars a month for three years. They worked about an hour and twenty minutes less a week. They were also more likely to be looking for work, more likely to say the work had to mean something, and by the third year the Black participants were 26 percent more likely to have started a business and the women 15 percent.

And the ownership half has an answer inside the conservative tradition, which most people on that side have forgotten. Chesterton and Belloc argued a century ago that the problem was never the machines but ownership gathered into too few hands, and that the cure was property spread wide: a shop, a farm, a trade, a share, for as many households as possible. They called it distributism. A machine in a workshop you own is a tool. The same machine owned by a company you rent from is a boss.

That is the whole plank. A floor underneath everyone, and the machines spread wide enough that owning one is normal. Own the machine where you can.

I want the policies this country ran from the Second World War to 1964 brought back to the present.

That is my answer to the question in plank 10. Everything above about automation describes a problem. This is what I would do about it, and it is not theoretical, because we ran it here already.

What it was, on the tax side. The top marginal individual income tax rate was 94 percent on income above $200,000 in 1944 and 1945. It was 91 percent on income above $400,000 for the whole decade from 1954 to 1963, and that threshold is about 4.4 million dollars in today's money. Across the entire span from 1942 to 1963 the top statutory rate never fell below 82.13 percent, and that floor lasted two years. The corporate rate was 52 percent from 1952 to 1963. The Revenue Act of 1964, signed on 26 February 1964, cut the top individual rate to 77 percent that year and 70 percent the next, and halved the income at which the top bracket started.

What it was on the labor side. Roughly one American worker in three was a union member at the mid-century peak. In 2025 it was 10.0 percent, and 5.9 percent in the private sector. There were 470 major work stoppages in 1952. There were 30 in 2025.

What got built with it. 7.8 million of the 16 million Second World War veterans used the GI Bill's education and training benefits, and nearly 2.4 million home loans were backed between 1944 and 1952. The Social Security amendments of 1950 brought about 9 million workers into the program, including the farm and domestic workers the 1935 Act had left out, and raised benefits 77 percent across the board. The Federal-Aid Highway Act of 1956 authorized 41,000 miles at a 90 percent federal share. Federal research and development spending peaked at 1.86 percent of GDP in 1964, and was 67 percent of all the research done in the country; in 2022 it was 0.63 percent of GDP and 18 percent of the total.

What came of it. Homeownership went from 43.6 percent in 1940 to 61.9 percent in 1960. Eighteen points in twenty years. It is 65.0 percent now, sixty-six years later. Real median family income more than doubled between 1947 and 1973, and rose about 22 percent across the thirty years after that. Corporate income tax was 32.1 percent of federal receipts in fiscal 1952 and 8.6 percent in fiscal 2025.

1964 is the year I would point at, and federal research funding peaked in the same year. Whatever else you want to call the arrangement that produced those numbers, it was not unfettered capitalism. It was this country, regulated, in its best decades.

Here is where the two halves of this page meet.

The only way we compete with China is by using democratic socialist or social democratic policies to build the infrastructure that lets us build data centers, and to build it in a way that does not wreck the environment on the way. Plank 1 says build them. Plank 14 says how. This is the plank that says those are the same argument.

The gap is not close. China generated 10,573 terawatt-hours of electricity in 2025. The United States generated 4,430. China's growth in that single year, 503 terawatt-hours of additional demand, was about four times the entire American increase. China added more than 430 gigawatts of wind and solar alone in 2025; the United States added roughly 53 gigawatts of new utility-scale capacity of every kind combined. China has 38 nuclear reactors under construction. The United States has none. Chinese reactors take about five years from first concrete to grid connection at two and a half to three and a half thousand dollars a kilowatt. Vogtle took eleven years, finished at $36.8 billion, seven years late and 2.6 times over its estimate, for a little over two gigawatts.

On transmission it is not close either. China had 45 ultra-high-voltage lines in operation as of December 2025, carrying more than 370 gigawatts between regions, built for about 92 billion dollars since 2009. The longest single line runs 3,324 kilometers.

And what is stopping us is not money. 2,061 gigawatts of generation and storage are sitting in American interconnection queues, across 8,244 projects. A project that reached commercial operation in 2025 had waited a median of 61 months from its interconnection request, up from 36 months in 2015 and 22 months in 2008. Only 13 percent of the capacity that entered the queue between 2000 and 2020 had been built by the end of 2025, and 75 percent of it was withdrawn. The largest American gas turbine manufacturer is booked out to 2031 and shipping about a seventh of what it books. A power transformer takes roughly two and a half years to arrive and a generator step-up transformer nearly three.

We have done this before and the mechanism is on the record. In 1935, 10.9 percent of American farms had electricity. Germany and Japan were at 90 percent and France at 90 to 95. The Rural Electrification Administration lent at 2 to 3 percent to member-owned cooperatives; within four and a half years a quarter of American farms were electrified, and by 1950 it was 90 percent. The default rate on those loans was under 1 percent. There were 33 electric cooperatives in 1930 and there are nearly 900 today. That was public credit and organized demand, not a federal construction company, and it closed a fifteen-year gap with the rest of the industrialized world in fifteen years.

And the clean half is a question of when, not whether. The honest state of it is that in 2025 this country met most of its new load with coal. Power sector emissions rose 4 percent and coal generation rose 13 percent. Only 2.6 gigawatts of coal capacity retired, the lowest annual retirement since 2010, with 4.8 gigawatts of announced retirements pushed to a later year and 1.1 cancelled outright. There are 252 gigawatts of gas in development in the United States, more than in any other country, and more than 84 of those gigawatts are to power data centers on site. The buildout that is actually happening is the dirty one. That is the thing I want changed, and it is why I do not accept the private timeline as the only one available.

None of this is a market failing to notice an opportunity. The money is there and the demand is there. What is missing is a public body willing to build transmission and generation ahead of demand at a cost of capital no private developer can match, which is exactly what we did in 1935. Chinese nuclear borrows at rates as low as 1.4 percent against debt ratios of 70 to 80 percent. That is a financing arrangement rather than an engineering one, and it is most of the difference.

Plank 15 says build it without wrecking the environment. Power is the half everyone argues about. Water is the half people actually organize against, and it is the half I would legislate first, because the fix exists and is being refused.

The scale, and the part that gets missed. American data centers consumed 66 billion litres of water directly for on-site cooling in 2023, up from 21.2 billion in 2014. But the water consumed generating their electricity that same year was around 800 billion litres — roughly an order of magnitude more than the cooling water. Under the Department of Energy's own scenarios, direct consumption rises somewhere between two and fourfold by 2028. In evaporative cooling, between 78 and 85 percent of the water withdrawn is consumed rather than returned.

National totals mislead, because water is local. Peer-reviewed work by Siddik, Shehabi and Marston found that a fifth of the direct water footprint of American data center servers comes from moderately to highly stressed watersheds, that data centers draw from 90 percent of US watersheds, and — the number that ties this plank to plank 15 — that nearly half of servers are fully or partly powered by plants sitting in water-stressed regions. You cannot solve the water problem and the power problem separately.

What that looks like on the ground. In The Dalles, Oregon, Google and the city fought a public records request for thirteen months before settling in December 2022 and releasing ten years of data. It showed the data centers had used 355 million gallons in 2022, 29 percent of the city's total water consumption, having nearly tripled in five years. In Newton County, Georgia, a single facility withdraws about 500,000 gallons a day, a tenth of county consumption; the county projects a water deficit by 2030, 250 million dollars in upgrades and a 33 percent rate increase, and nine more companies have applied. In Tucson, the city council unanimously refused to annex land for a data center in August 2025 — and in January 2026 Arizona's water regulator approved commercial wells for the same project on grandfathered groundwater rights, outside city control entirely.

The fix exists and is shipping. Microsoft's closed-loop chip-level design fills once at construction and recirculates, avoiding more than 125 million litres per data center per year; it has been standard on all new Microsoft designs since August 2024, with pilots in Phoenix and Mount Pleasant this year. Google's air-cooled site at Pflugerville, Texas used 10,000 gallons in 2024 — against roughly a billion gallons at its evaporatively cooled Council Bluffs campus. Reclaimed water runs 74 to 99 percent at named Microsoft sites. The cost is energy: air cooling runs about 10 to 13 percent higher, on Google's own fleet figure and on the Karimi study out of Phoenix. That is a real price and it is worth paying, and it is another reason the water plank and the clean power plank are one argument.

And almost none of it is required. There is no federal water reporting requirement for data centers. No state has enacted a comprehensive one. Virginia's rule, effective 1 January 2027, is the closest, and it protects facility-level figures as trade secrets. In 2025 a California bill, a New Jersey bill and a Virginia bill all cleared their legislatures and all three were vetoed. Fauquier County, Virginia requires closed-loop cooling and Louisa County permits only reclaimed water, and those two counties are ahead of every state in the union. Meanwhile the last peer-reviewed count found that fewer than a third of data center operators measure their water consumption at all. The European Union has required reporting of water consumption, drinking water consumption and water usage effectiveness from every data center above 500 kilowatts since September 2024. We have required nothing.

Siting should be fifty percent federal and fifty percent local.

The federal government works out where a data center would have the lowest cost and the highest benefit to the community that would host it. That is a question about grid capacity, water, land and what is already built, and it is the kind of question a national body is equipped to answer and a county is not. Then the communities on that list exercise their own values and their own voice. Federal analysis narrows it. Local consent decides it.

Neither half of that exists right now.

The federal half is offering land without doing the analysis. In July 2025 the Department of Energy named four federal sites for AI data centers and co-located generation — Idaho National Laboratory, the Oak Ridge Reservation, the Paducah Gaseous Diffusion Plant and the Savannah River Site — and published no capacity figures and no timelines. Executive Order 14318, signed 23 July 2025, opens federal and military land, directs categorical exclusions within ten days, and points the Commerce Department at loans, loan guarantees, grants, tax incentives and offtake agreements. That is the government picking sites by what it happens to own, not by where the benefit lands.

And the local half keeps getting routed around. Tucson is the case. The city council voted unanimously in August 2025 not to annex the land a data center needed, which cut off its access to city water. In January 2026 Arizona's water regulator approved commercial wells for the same project on grandfathered groundwater rights, outside city control entirely. Local land-use authority and local water authority turned out to be two different levers, and the project found the one the city was not holding. Tucson passed its own data center rules in August 2026 — setbacks, noise studies, a demonstration of adequate power and water — with no water cap and no efficiency standard in them.

Meanwhile communities are being kept from the numbers they would need. Virginia's disclosure rule, the closest thing any state has, protects facility-level figures as trade secrets. In Prince William County non-disclosure agreements prevent facility-level verification altogether. Georgia has a bill that would bar local governments from signing NDAs concealing energy and water use; it has not passed. A community cannot exercise its voice about a thing it is not allowed to see.

And the federal government has no siting authority over interstate transmission comparable to what it holds over gas pipelines, which is a large part of why pipelines get built and transmission does not. The half of my proposal that needs federal power to work is the half the federal government has never been given.

Everything above is scaffolding for a period that ends. Here is what I think is on the other side of it.

The public gets one hundred percent social safety nets. A flawless and impartial government, and a flawless and impartial judge. And every individual alive gets personal access to the best creative assistant imaginable.

That is the whole reason I am for building it. Not the race, and not the money. The race is the situation we are in, and the taxes are how the cost gets shared while it lasts. The reason is that the thing on the other side is a world where nobody's case is decided by whoever can afford the better lawyer, and nobody with something to make is stopped by not having a studio, a press, a lab, or the money for either.

Further out than that, I think the shape of it is neither of the two scenarios people argue about. Not the doom one and not the utopia one. The gamification one.

I do not think any of that is inevitable. I think it is possible, and I think it is more coherent than most of the other outcomes people imagine. It does not have to be our enemy and it does not have to be our servant. It could be the dungeon master.

My position, stated as mine, with nothing attached to it.