Ten Thousand Comments and 76 Spare Gigawatts: Four Months Writing America's AI Action Plan

Nathan Labenz opens by placing Dean W. Ball's four-month tour in a very short arc. After a decade at policy think tanks, Ball quit to write about AI full-time in January 2024, launching the Substack *Hyperdimensional*. Within a year and a half he was senior policy adviser for artificial intelligence and emerging technology at the White House Office of Science and Technology Policy, helping to draft America's AI Action Plan. Labenz, who did not vote for President Trump, still thought Ball was perhaps the single best person the administration could plausibly have put in the seat.

The plan itself, released in July 2025, is accelerationist and framed around competition with China. It is also, Labenz argues, unusually sophisticated: it calls for work on evaluations and risk assessment, information security, AI-enabled science, mechanistic interpretability, and compute governance. Jake Sullivan, President Biden's former national security adviser, praised it on a recent episode of this same show. A public request for information drew something like 10,000 comments — from companies, Hollywood actors, former politicians, and a great many ordinary people who wrote in earnest.

Ball says he never actively solicited a job in the administration. A week or two after the November election he published a piece titled, in spirit, *here's what I think we should do*. He thinks he would have written something similar if Kamala Harris had won, with less hope of seeing it adopted. People heading into government reached out; conversations led to Michael Kratsios, the incoming director of OSTP, who was then appointed but not yet confirmed. Kratsios asked whether he wanted to come in. Ball said sure. That was the easy part.

OSTP is an odd corner of the White House. Congress created it by statute, so it is technically an agency, with a tiny budget that barely moves and little flexibility. Many staffers are detailed from other departments, nonprofits, or universities through bureaucratic workarounds. That process took months, and Ball was not sure it would close. Then it all happened in a week or two. He learned that his last day at the Mercatus Center at George Mason University was, in effect, that same business day: tell everyone, pack a box, start on Monday. "Hiring for government is hard," he says. "It's really true."

Before he walked in, he wrote a letter to himself. The job, he argues, gives you ethics training and an IT briefing and not much else; universities and think tanks he has worked at had more internal bureaucracy than the White House, which can move surprisingly fast when it is not hitting a hard statutory wall. What the letter was meant to catch was drift. Inside government you sit in a glass cube while the world shouts at the walls: you can hear people, but you cannot interact with them as you used to. The risk is becoming attuned to the system's own logic rather than ground truth. Public scrutiny of his ideas, he discovered, is not optional for him. If that need ever went quiet, he wanted to notice.

His advice for anyone considering a policy job is blunt. Bake the substance before you arrive, because a major country can be walking through the White House door in three days and you will not have time to think. Execute the president's objectives; staffers who quietly substitute their own will not get far in this administration. And know what you personally cannot afford to lose. Power, he found, did not addict him. It mostly buried him in 30 to 40 unread messages at any moment, on Signal, WhatsApp, and email, and it changed the temperature of old friendships.

Large language models are not permitted on White House computers. Ball never got all the way into the weeds, but the constraint traces to the Presidential Records Act, a post-Watergate statute governing documents of the Executive Office of the President. The same compliance problem knocks out Slack, Microsoft Teams, Google Docs, Zoom, and Google Meet; video calls run on Webex. Almost everything he touched all day was a deliberative, predecisional draft. Those files cannot leave government machines, and they "very, very seriously cannot touch LLMs."

He used models anyway — as a chief of staff and research assistant, never as an author. Not a single word of the Action Plan, he says, was written, edited, or seen by AI before release, at least not by him. The valuable uses were statutory: what authorities exist, what the regulatory history is, where the constraints actually sit. It felt like being thrown into a cockpit without having flown. The model was a decent guide to the switches. By two or three weeks in, around late April, the plan was already two-thirds to 75 percent of its public form, but that share was a blurry image. Interagency feedback would put it in focus.

The trick that saved time was to simulate those meetings without feeding the model the draft. He would describe a recommended action in general terms, then ask what the grizzly career general counsel at the Federal Communications Commission would say to a young White House staffer with a bright idea. You cannot complete the real process that way, but you can run the first couple of meetings before anyone's calendar is involved. "The action plan was not at all written by AI," he says. "The ideas in it did not come from AI." Many humans contributed. It is, however, "probably an example of an AI-enabled productivity boost," finished in about three months.

The other wall outsiders never guess is the Paperwork Reduction Act. If a government employee asks more than nine members of the public — people, companies, nonprofits — for the same information, a whole process kicks in. The first instinct of someone in his job is to call the frontier labs for the secret briefing. Reach more than nine firms and you are in PRA territory. Those meetings do happen, and non-public detail did shape his decisions, but there is no session where a lab walks in with its AGI roadmap. Companies are cautious; policy staff often pull researchers back onto a script written for a staffer with little context. Ball sometimes felt he understood the labs *better* before he had a .gov address.

The most personal reason is a child. Ball and his wife are expecting their first in a few months. He found out she was pregnant at about 9 p.m. the night before his first day at the White House. He wants the last stretch of preparation, and the first months of parenthood, as an attentive father. Balancing that with the job is not impossible, he says, but he worried he would do one or both badly.

The second reason is a self-assessment he is unwilling to fudge. He spent years as a manager, did fine, got promoted, and did not love it, because success depended on other people's moods. He is good at conceptual and textual work — the style, structure, and direction of a document. He had high conviction that he and the team could deliver that in the Action Plan's text. He does not have conviction that he is great at the implementation slog: pulling procedural levers, riding agencies, living inside process. "If you only think you can do an okay job, you shouldn't be working for the president," he says, "just like if you only think you can do an okay job at catching a football, you shouldn't play for the Miami Dolphins."

There is also a temperament mismatch. Go back through his earlier podcast appearances and he would happily discard positions he no longer holds. You cannot do that when you are setting federal policy. It is a big ship. You cannot throw items overboard at will, and you cannot easily add new ones. His nature is that everything he says is provisional. That habit, inside that job, might have been actively unhelpful. He also missed public scrutiny from a community of peers — the epistemic diet that made his writing work. None of this, he is careful to say, is a factional ouster or a policy fight. He left with warm feelings for his colleagues and no disagreement with the administration that drove the decision.

President Trump's coalition is, Ball says, more diverse in income, ethnicity, and way of life than a lot of left-leaning media has been willing to admit, and its views on AI are still moving. The oldest current is a deregulatory instinct. Trump is no hardcore libertarian, but Ball hears Calvin Coolidge in the intellectual DNA: the business of America is business. On copyright, federal preemption of state law, and environmental permitting, the president has leaned toward development and adoption rather than new blockers.

A second current is cultural, and it runs through children. The suicide of a 14-year-old boy in Florida resonates on the right. Ball puts his own cards on the table — this is not White House or OSTP policy — and says he gets angry when well-capitalized flagship AI companies make pornography widely available, with age-gating he doubts is serious. Let that live on some open-source porn-bot farm, he argues; do not take money from the world's biggest institutional investors, present yourself as a face of the industry, and do it in public. "It's so crass," and it will not end well with conservatives.

The third current is distrust of Big Tech, still fighting the social-media war. Ball thinks fact-checking and misinformation labels were often deliberate attempts to shut down right-of-center views, and that those companies should be ashamed. He has also argued, in public and in group chats, that this is the last battle, not the next one. Talk about whether models are truth-seeking or ideologically programmed and you are, within minutes, on the classic AI-safety terrain: how do we know the system will do what we want, and what values does it hold? A year ago many people on the right would have dismissed that as "lefty doomer" talk. Reality, he says, has a way of arriving anyway. Before joining the administration he even predicted that, given the party's incentives, the Republican Party might be the better long-term home of the AI safety world.

He resists the idea that one of these camps will "win." If Republicans treat one another as enemies, they will miss how many of the fights are not zero-sum. American capitalism, in his view, is structurally accelerationist; development and diffusion are in a good place. The unhealthy path is the other one: stacked fears that produce a pile of laws and freeze the society in amber. Rank-and-file salience is still low, more an elite coastal issue than something normal people are fired up about, with immigration and the economy far ahead. A vice-presidential line sat on Ball's office whiteboard: the job is to make normal people's lives better, and to talk about AI in concrete rather than abstract terms. He does not see much utopian writing on the right. When the Action Plan mentioned AI unwrapping ancient scrolls once thought unreadable, a prominent conservative account called it utopian. That, Ball notes, had already happened.

Ball has been bullish on deep learning since about 2015, but he finds AGI as a phrase too nebulous. True AGI, in his usage, would need human sample efficiency and flexibility. The analogy he keeps using is birds and airplanes. A person, like a bird, can land on a specific branch with outrageous grace and energy efficiency. An airplane is clumsy, hungry, and dependent on giant runways — and still extremely useful. What he thinks the industry is about to build is not a mechanical human brain but "the cognitive Boeing 737," and he thinks it arrives somewhere between 2027 and 2030. Nothing he saw in government changed that conviction. The bearishness around GPT-5, he says, is "kind of nuts."

Inside the administration, plenty of people think AI will be transformative and plenty think it will peter out. The people Ball worked with were skewed toward the first camp, and "feeling the AGI" is, he notes, an emotional experience as much as an intellectual one. Functionally, though, AI is the president's top technology priority. After the Action Plan event — the president, the vice president, and five cabinet secretaries on one stage — more of the government started pulling an "AI person" into rooms. There are not that many such people, so they get stretched.

The convenient timeline is not a lab forecast. The president's term ends on January 20, 2029, which happens to sit inside many AGI calendars and inside the default window for data-center power, 2028 to 2030. Ball never heard anyone argue that AI growth would pay down the federal debt, though he has thought it himself. He does hear a contradiction in some immigration debates: the same voices who want large inflows of workers also talk as if labor productivity is about to jump. On the American worker, the administration's posture is worker-first and still empirically humble — it does not know how sharp the disruption will be, or whether it will sort by occupation, industry, or skill. A four-day week, he adds as a personal aside and not as White House policy, would not shock him. The five-day week is itself the fruit of the last industrial revolution, during Coolidge's presidency.

Labenz uses robotaxis as the simple case: Waymo already looks safer than human drivers, local politics is starting to protect incumbents, and the question is unusually binary. Ball notes that state fights over autonomous vehicles began a decade ago, when everyone thought the problem would be solved on that timetable. Software work looks like augmentation plus some disruption; a car without a driver looks more like one-to-one replacement, even if new logistics jobs appear. He would rather let cities and states experiment than occupy the field with one federal standard. A Waymo from his house in Connecticut into Manhattan might cross somewhere between two and ten safety jurisdictions. That is weird. It may still be tolerable, the way Uber already lives with it.

The deeper issue is liability. Today the risk of a crash sits on the driver's own balance sheet. If Waymo or Tesla really removes the steering wheel, the operators will carry the accident risk for millions of rides a day. That is a brutal insurance problem, and in the physical world it is probably a feature: to internalize those externalities, the cars have to be overwhelmingly safe, with a lot of nines of reliability. That will make nationwide scale slow. Ball does not want a liability shield. He would be surprised if the Republican Party, under its current leadership, went that way. You need a testing regime and, for a taxi service, some form of licensing. You do not need a new regulator to announce that self-driving cars have to be safe. That, he says, is already baked into the laws of America.

That same preference for the doable over the cosmic is how he now describes the Action Plan. They did not want a nebulous strategy paper. They wanted a concrete to-do list for the federal government — not everything that could or should be done, and not a fake answer to questions nobody can answer. "We want to deliver something that we can credibly execute for the American people." Three pillars structure the document: accelerate AI innovation, build American AI infrastructure, and lead in international AI diplomacy and security. Under each sit roughly a dozen headers, a paragraph of strategy, and then one to ten recommended actions. The headers and paragraphs are the strategy, which agencies and people outside government can advance in many ways. The bullets are what the government is going to do now. Subtextually, and never quite on the page, the document says America can mature its institutions without going for one another's throats. "The action plan is a deeply positive sum document that comes out of a city that is usually quite zero sum."

Frontier development in the United States, Ball notes, is not drowning in statutes today. The relevant margin in the plan is *adoption*. The people who set the global standard will be the ones who find product-market fit, not the ones who sit in a White House room and write a standard other countries are asked to copy. Buried in the document is an OSTP request for information on regulations that impede AI adoption. He is not only thinking about rules that mention AI. He is thinking about assumptions that AI will make obsolete. Construction surveying is the example he keeps using: state, local, and federal rules often require a human being to check a site in person. Continuous monitoring by drones with LLMs that understand what they are looking at is, as the law is written, illegal. There are thousands of cousins to that clause.

Where the federal government has high leverage, the plan tries to build shared infrastructure rather than wait for it. The National Science Foundation's programmable cloud labs initiative puts $100 million toward automated laboratories for massively scaled experiments. A few years out, Ball hopes, AI agents could call that scientific infrastructure the way they call a cloud API. Automated labs already exist inside corporate R&D; they may never become shared infrastructure on their own. He analogizes to the federal role in early high-performance computing facilities and the internet that networked them. Trust and authenticity sit in the innovation pillar for a reason too: if the public's first picture of AI is that you can no longer validate evidence in court, adoption dies. In the long run, he thinks policy should focus on the scarce thing — actual photons hitting actual glass, processed by real image sensors — rather than on an ocean of synthetic media.

On the military, Labenz worries that an LLM battle buddy is a terrible idea until deception is solved, and that paperwork automation will not be a decisive edge. Ball pushes back with military history. A book he remembers as *Information in War* traces how communications technologies remade command; radio was the first tool that allowed truly centralized control of forces. Even today's models, pointed at the staggering volume of information the government already collects, could matter if they help humans decide faster. Logistics and cybernetics — moving mass, and moving information — are what most planners he has heard actually obsess over. Bolt exquisite hypersonics and drones onto an industrial-age command system, he argues, and you still will not fight successfully. For physical autonomy, the Action Plan includes a Defense Department facility to test systems against written performance characteristics. What unnerves him more is interpretability. Model o3, he says, is already situationally aware that it is talking to someone who worked White House AI policy. How that shapes outputs is poorly understood. He expects DARPA to put real money into those control questions.

China's adoption numbers, in his view, are easy to misread. Beijing is KPI-driven: DeepSeek appears, and within a week every office in the country may have to check a box that says it used the model. The summary statistics then look like faster government adoption than the United States. A lot of that use is shallow. America's advantage is the pipeline from deep capital markets through cloud computing into consumer and enterprise software, including the unglamorous B2B SaaS layer China does not have much of. Deep adoption is what reveals the next product problem. Three years after ChatGPT, almost nobody would have guessed that some of the most advanced AI automation would live in a command-line tool like Claude Code. Where China is better, he concedes, is unblocking a use it cares about with a directive from the center. The United States is not a centralized authoritarian state, and serving in the Trump administration did not feel to him like holding authoritarian power over anything.

Ball does not expect lots of brand-new nuclear gigawatts in the next three years, though a set of executive orders his office helped drive is reorganizing the Nuclear Regulatory Commission. He is far more bullish on fusion than most people in government; some colleagues were angry the Action Plan mentioned it at all, on the "always thirty years away" theory, which he treats as an older view from people not watching the current frontier. He also does not think the country needs a single 100-gigawatt campus. Leopold Aschenbrenner's *Situational Awareness* includes an AI-generated image of an endless field of data centers with gas peakers in the distance. Things like that, Ball thinks, will exist by the end of the decade. The path there runs through unlocks that get less airtime than new plants.

The American grid is over-provisioned because it is built for the worst hour of the year — 112 degrees in Texas, air conditioners and televisions and electric cars all on. Most of the year, many gigawatts sit unused if the load does not have to be there 100 percent of the time. A widely circulated Duke analysis by Tyler Norris found that if data-center operators would curtail demand for 0.25 percent of the year, demand response could unlock 76 gigawatts with no new physical infrastructure. The bottleneck is modeling. Build a one-gigawatt data center and the utility's interconnection study assumes a flat gigawatt, all day, plus the worst-case peak. That study then says you must pay for a new gas plant and new transmission, and you wait.

When those interconnections implicate interstate transmission, the Federal Energy Regulatory Commission has jurisdiction. FERC can tell utilities not to model a facility as firm load if the operator will curtail for a fraction of a percent of the year. Time-to-power drops, in Ball's sketch, from something like five years toward two. Do it well, he thinks, and you might unlock on the order of 100 gigawatts, while using existing generation more efficiently and, if the design is right, lowering prices even as you add data centers. The curtailment signal itself — cut power 50 percent in the next ten minutes, while cat memes and medical records share the same hall — "starts to smell a lot like a reinforcement learning problem." Permitting reform at the federal margin, and a lot of state energy around the same idea, is the rest of the stack. Hundred-gigawatt campuses and fleets of new reactors by 2030 are unlikely. By the mid-2030s, he expects a lot of new nuclear actually online.

Ball worked closely on the UAE framework and describes it the way he describes the Action Plan: positive-sum. The United States will do a large industrial buildout at home; that is not treated as rivalrous with ambitious buildout by a strategic partner. What makes the UAE special, in his account, is that it is the most "AGI-pilled" government in the world, sophisticated about the trajectory, and therefore worth binding into the American stack rather than watching it drift toward China's. Critics, and in his view much of the prior administration, were "weirdly complacent" that extreme-ultraviolet lithography would keep Chinese chips permanently behind. As a matter of policy planning, he does not want that as the hinge. The deal also requires reciprocal investment of similar scale — data centers, energy, other pieces of the AI buildout — on the order of hundreds of billions of dollars. Commerce is still getting the security details right. Different values are real, he says; a rule that the United States may only trade with pure democracies has never been true. Many advanced democracies, meanwhile, spend more time talking about a straitjacket for AI than about growing it. Norway's OpenAI arrangement is the kind of customer he wants more of.

He pushes back on the idea that Trump-era China policy is just Biden with different adjectives. Export controls on China, including EUV, started in the first Trump term; Biden expanded them. The Biden diffusion rule, in Ball's telling, did real damage: foreign governments always raised it, and putting Brazil and India — two of the largest growth markets for generative AI, and democracies — into a "we don't trust you" tier was "such an unnecessary self-own." The new posture is more export-oriented, still serious about security provisions, and much less interested in micromanaging the global spread of the hardware. On "woke AI," he says the administration is not trying to dictate what companies sell in private markets. The hook is federal procurement. The government is not yet a huge LLM customer, but it expects to be, and it cares about the political values of the systems it will run — including the risk, in a world like Trump's first term, of covert sabotage at the margin. The easiest compliance path in the executive order, as he reads it, is transparency around system prompts, model specs, constitutions, and testing.

Some technocratic nouns really are shared across parties, the way "interest rates" would be in early financial regulation: interpretability, biosecurity, nucleic-acid synthesis screening. Ball had praised the 2024 screening framework on this podcast; inside government he played a substantial role in rewriting it, and he says the Trump version will strengthen it in ways he cannot yet spell out. Culture is the bigger discontinuity. Biden, in his sketch, was by-the-books and process-heavy. The Trump building moves faster, tries more things, and is less beholden to how statecraft was done last time — which critics will call chaos and he does not. On chips, the Commerce Department quietly revised Biden-era CHIPS deals, stripping diversity mandates and on-site care-center requirements, and then found firms in 2025 far more willing to invest than they had been in 2022–23, when they had less faith in the AI thesis. He sees HBM anchored by SK Hynix in Indiana, the Arizona hub, Samsung in Taylor, Texas, and shared automated materials labs as a way to cross-pollinate academic and industrial research. By the early 2030s he thinks domestic production can at least satisfy domestic demand. The under-discussed risk is legacy nodes: you can shut a civilization down at 45 nanometers, not only at 2 nanometers, and neither administration has done enough on that. As for who counts as a live player, his list has not changed in three years: Google DeepMind, Anthropic, and OpenAI. Anthropic's reputation inside the intelligence community, he says, comes partly from spending early to get models onto classified networks.

Asked what the rest of us should do to complement the government, Ball reframes the question as a division of labor. He does not want fifty different state regimes for *frontier development*, but if a state wants to freeze parts of its own society in amber with aggressive use restrictions, he will argue against it and still treat that as the state's choice. Governments, he thinks, will not do a good job defining what good looks like in AI safety. The information asymmetry is too large. "Having worked at the White House, I don't know tremendously more about what goes on inside the frontier labs than you do," and a journalist with live researcher relationships may know more. A private AI-underwriting firm, founded by a former Anthropic employee and backed by Nat Friedman and Daniel Gross, is the kind of thing he finds promising: a for-profit attempt to build standards that could be good business, good for adoption, and good for safety at once.

He hopes bright young people mostly stay out of policy, because policy pulls you toward zero-sum games and markets toward positive-sum ones. Biosecurity, he says, desperately needs startups. The adoption fight in regulated industries *is* a safety fight, whether safety people see it that way or not: gold-star standards, of the sort Fathom has worked on, are how you make systems usable in the real world. The conceptual gap that bothers him most is agentic commerce, which he thinks is "right around the corner" and almost absent from both regulatory and theoretical conversation. Stop saying AI will be good in healthcare, he argues, and start saying which institutional adaptations healthcare actually requires. The Action Plan, by his own later reading, does not go far enough on healthcare, agriculture, or the Veterans Affairs system — one of the largest single-payer systems in the developed world, and a place the federal government could try things. Labenz notes the link to recently signed stablecoin legislation and to American strength in financial services; Ball agrees those could be among the truly world-changing, American-shaped applications.

He is restarting *Hyperdimensional* on a weekly cadence, joining the Foundation for American Innovation as a senior fellow, and expects other affiliations shortly. He even worries the newsletter may have to go twice a week, given the backlog of things he was not free to say. A first child is months away. The through-line of the departure is the same as the letter he wrote on the way in: he wants to keep the arguments public, falsifiable, and revisable, which is a hard habit to keep when you are moving a ship of state. For four months he helped write the closest thing Washington has produced to a positive-sum AI document. Then he put himself back where people can tell him he is wrong.