Summary

America’s AI bottleneck is shifting from frontier-model invention to adoption, infrastructure and institutional execution. Ball describes the AI Action Plan as a “concrete to-do list for the federal government,” organized around innovation, infrastructure, and international diplomacy and security. Its investable call is that product-market fit—not governments declaring standards—will determine which AI stack diffuses globally.
Grid flexibility may deliver more near-term data-center capacity than waiting for a new generation of power plants. Citing Tyler Norris’s Duke analysis, Ball says curtailing data-center demand by 25% for 0.25% of the year could unlock 76 GW from the existing grid; with better interconnection rules, demand response and permitting, he thinks the opportunity could approach 100 GW. The proposed bargain is faster time-to-power—potentially two years instead of five—in exchange for interruptible loads, creating a world where “you literally can build the data centers while lowering prices if you do it right.”
Ball remains highly confident that a powerful but nonhuman, general-purpose cognitive system—not true human-like AGI—arrives between 2027 and 2030. His model is a “cognitive Boeing 737”: far less flexible and sample-efficient than a human “bird,” yet industrially transformative and available soon. “The bearishness on GPT-5 is kind of nuts,” he says, while stopping short of claiming that current scaling paths are close to a mechanical human brain.
Republican AI policy is accelerationist by default, but safety, child protection and distrust of Big Tech are emerging constraints. Ball argues conservatives who once dismissed alignment and loss-of-control concerns as “lefty doomer stuff” are rediscovering the same problems through model bias, concentration of power and children’s exposure to AI systems and pornography. He still sees “a reasonable chance that the Republican Party is the better home of the AI safety world in the long term,” though poorly designed reactions could “freeze our society in amber.”
Washington’s industrial strategy combines domestic capacity with allied capital rather than pursuing immediate autarky. Ball calls the UAE “the most AGI-pilled country in the world” and defends Gulf compute deals on security, technology-stack alignment and reciprocal US investments potentially worth hundreds of billions of dollars. At home, he expects domestic production could satisfy domestic demand by the early 2030s and wants the US to reclaim the frontier lead, but warns that unglamorous 45 nm legacy chips can also “shut down civilization” if imports stop.
Military advantage may come first from information flow, logistics and cybernetics—not autonomous weapons firing on their own. Nathan Labenz presses the unresolved deception risk of an “AI battle buddy”; Ball agrees control and interpretability remain weakly understood, but argues GPT-5-class systems could already create decisive value by synthesizing the government’s enormous information flows. Bolting world-class drones or hypersonics onto “a 20th-century military model,” he says, would still leave the force unable to fight effectively.
The largest safety-and-adoption opportunities may sit outside government in insurance, standards and sector-specific startups. Ball says officials cannot reliably define “what good looks like” because they know surprisingly little more about frontier labs than informed outsiders. He points to the AI Underwriting Company, biosecurity startups and private “gold stars” as positive-sum mechanisms, while identifying agentic commerce, stablecoins, healthcare, agriculture and Veterans Affairs as underdeveloped opportunities.
The Action Plan was human-authored but materially accelerated by AI-assisted research and adversarial simulation. Ball says “not a single word” was written, edited or seen by an LLM before release, yet he used models to map statutes and simulate skeptical interagency meetings; a small human team produced the plan in roughly three months after about 10,000 public submissions. He is leaving voluntarily—without policy conflict or factional defeat—because of impending fatherhood and the conclusion that conceptual work is his comparative advantage: “If you only think you can do an okay job, you shouldn’t be working for the president.”

Deep dive

1. White House influence arrived faster than its hiring machinery

Ball did not campaign for an administration job. After Trump’s November victory, he published “Here’s What I Think We Should Do”—a piece he says he would also have written after a Kamala Harris win—and acquaintances entering government connected him with OSTP director Michael Kratsios.
The substantive offer came quickly; making it administratively possible took months. Because Congress created OSTP as an agency with a small, relatively inflexible budget, many staff arrive through affiliations with other agencies, nonprofits or universities rather than ordinary White House hiring.
The final transition was abrupt: Ball learned that one day would be his last at Mercatus, packed his office, and began at the White House the following Monday. Training covered ethics, gifts, meals and IT, but not how to formulate policy under pressure.

2. Power was less intoxicating than isolating

Ball found the White House simultaneously more flexible and less bureaucratic than some universities and think tanks, except where one encounters an immovable legal wall. On a productive day, the reward was extraordinary: an idea was no longer commentary but potentially “public policy of the United States of America.”
The personal cost was a permanent queue of 30 to 40 unread communications across Signal, WhatsApp, email and other channels. Even friendships changed once everyone had requests, introductions, documents or meetings competing for his time.
His uncomfortable self-test was whether power feels “like an addictive drug” or like a burden. Ball concluded he was relatively unmoved by status and often weighed down by it, though he doubts four, eight, 12 or 16 months would have made the experience stop feeling surreal.
Before entering government, he wrote himself a grounding letter. Its warning was that officials can become attuned to “the logic of the system” rather than ground truth, changing not merely their views but who they are.

3. Policy ideas must be baked before the calendar takes over

Ball’s advice to prospective policy staffers is blunt: “You should have substantively all of your policy ideas pretty well baked by the time you go in.” There is little reflective time when a major country may announce a White House visit three days in advance and demand an immediate position.
A staffer’s job is to execute the president’s objectives, not quietly substitute a personal program on the theory that the president operates at a higher level. Personal principles still matter, Ball says, chiefly as safeguards against drift rather than as permission to undermine elected leadership.
Government felt like “a pretty self-contained cube with glass walls,” with everyone outside shouting in. Officials can hear industry and the public, but laws and ethics rules constrain how they ask for nonpublic information and structure nearly every interaction.
Public scrutiny was Ball’s hardest loss. Although the Action Plan received vastly more public attention than any Substack post, he missed the weekly process of floating provisional ideas, receiving criticism and revising his thinking in public.

4. AI accelerated the work without touching the draft

LLMs were not permitted on White House computers, a restriction Ball linked to compliance issues associated with the Presidential Records Act. The same environment ruled out tools such as Slack, Google Docs, Zoom and Microsoft Teams, leaving alternatives such as Webex.
Deliberative and predecisional material could not leave government systems. Ball emphasizes that “not a single word of the Action Plan was written by, edited by or seen by AI prior to the release,” at least not by him.
He nevertheless used AI as a chief of staff and research assistant: mapping statutes, regulatory histories, agency authorities and legal constraints felt like being “thrown into the cockpit of a plane” without knowing what the switches did.
His sharpest use case was simulated interagency review. Ball would ask what a “grizzled” in-house FCC general counsel might say to a young AI adviser’s bright idea, effectively pre-running the first couple of skeptical meetings before exposing the proposal to real agencies.

5. A blurry draft became policy in roughly three months

Ball’s February notes were already directionally close to the eventual plan. By late April—two or three weeks into his White House tenure—he estimates the document was two-thirds to 75% complete in substance.
That early version was a “blurry image” sharpened through extensive human interagency feedback. The distinction mattered: the first image contained the objects, but only repeated legal, operational and agency review made it useful.
A small number of people drove the text, yet many officials made essential contributions. Ball treats the roughly three-month completion time as evidence of an “AI-enabled productivity boost,” not AI authorship or AI-generated policy ideas.
One practical constraint illustrates why simulated review helped: under the Paperwork Reduction Act, asking the same information of more than nine members of the public can trigger an entire approval process. A law intended to reduce paperwork can therefore complicate even basic industry outreach.

6. Ball left because implementation is a different craft

The night before his first White House day, Ball and his wife learned they were expecting their first child. He wants to enjoy the final months of preparation and be an attentive father, rather than risk doing either parenting or government work badly.
His deeper reason is comparative advantage. He enjoyed creative direction, conceptual synthesis and writing the Action Plan; he has less confidence in his ability to supervise agency implementation, manage procedural levers and keep a vast bureaucracy aligned over time.
“If you only think you can do an okay job, you shouldn’t be working for the president,” Ball says, comparing the standard to catching footballs for the Miami Dolphins. The AI policy of the United States must be “really, really effing good.”
His intellectual style also conflicts with implementation: ideas remain provisional, and he now happily rejects views expressed in earlier SB 1047 or neurotechnology discussions. Government is “a big ship”; it cannot discard or add policy whenever one staffer changes his mind. His summary is “know thyself.”

7. The departure was neither a purge nor a policy break

Ball says he was not pushed out by political infighting, did not represent one press-defined faction losing to another, and was not resigning over disagreement with the administration. He describes “absolutely no bad blood” and warm feelings toward his colleagues.
Returning outside government restores the peer scrutiny he considers epistemically essential. He believes his best work is getting his head around thorny questions, developing answers and communicating them—not becoming “the master of the bureaucracy.”
His broader rationale mirrors the Action Plan itself: much of the “civilizational scaffolding” around AI will be built by private institutions. Intellectual and institutional work remains so immature that outside contribution can complement implementation rather than oppose it.

8. Public participation did not mean publishing live drafts

Labenz asks why a fast-moving technology strategy could not be developed in public. Ball’s answer begins with substantial participation: the formal request for information drew about 10,000 submissions, including comments from individuals, corporations, industries, actors and former politicians, while he met hundreds—possibly more than 1,000—outside participants.
What government cannot casually reproduce is the Substack cycle of posting an early draft and inviting edits. A tentative White House sentence can disturb congressional negotiations, trade talks or companies’ business interests: “We can think that we’re stepping lightly and in fact be stomping around.”
Leaks often exploit that gravity. Ball says an insider may disclose a controversial staff-level proposal precisely so senior officials and public backlash will kill it; routine publication of drafts could create the same veto mechanism.
Agencies generally reviewed portions relevant to their own authorities rather than commenting on the entire plan. Ball valued interagency review but feared full committee design would generate unproductive objections and squash novel ideas. Binding executive orders still went through traditional review and used “shall”; the Action Plan itself is a set of strong recommendations, not a legal command.

9. The Republican coalition is pro-growth but not uniformly pro-tech

Ball rejects a simple faction map. Trump’s coalition spans income, ethnicity, geography and ways of life, producing different emphases rather than two armies awaiting a final winner.
One current is the traditional conservative preference for business and deregulation. Ball says Trump has leaned toward development and adoption on copyright, federal preemption and environmental permitting, while American capitalism itself creates strong incentives to build and diffuse AI rapidly.
Another current distrusts the companies building it. Conservatives remember perceived discrimination by social platforms, fact-checkers and misinformation programs; when the firms behind YouTube now offer Gemini, promises that AI will become foundational can intensify anxiety rather than relieve it.

10. Culture-war complaints are converging with alignment questions

Ball tells conservatives they are “fighting the last battle” when they treat AI only as a replay of social-media censorship. Yet asking whether procurement models are truth-seeking and free of ideological programming quickly becomes the classic alignment question: how does anyone know the system will do what its user wants?
Questions about political values also become questions about concentration of power, model character and control. People who a year earlier dismissed loss-of-control concerns as “lefty doomer stuff” or “EA stuff” are encountering the same underlying issues through a conservative vocabulary.
Ball consequently repeats his pre-administration prediction: “There’s actually a reasonable chance that the Republican Party is the better home of the AI safety world in the long term.” He presents that as a possibility driven by party incentives, not a settled realignment.
His fear is that unresolved distrust turns into laws that “freeze our society in amber.” His hope is a positive-sum program of targeted safeguards and accelerated beneficial use, avoiding the usual assumption that every participant must lose for another to win.

11. Child safety could become the right’s first mass AI issue

The suicide of a 14-year-old boy in Florida is the kind of tragedy Ball says resonates strongly with conservatives. Children’s use of AI systems and access to sexual material sit directly between longstanding cultural concerns and newer AI-safety problems.
Speaking explicitly for himself—not OSTP or the White House—Ball condemns well-capitalized flagship AI companies that make pornography broadly accessible behind weak age gates. His intentionally abrasive line: leave that to “the open-source model on the North Korean pornbot farm,” rather than normalize it through brands funded by major institutional investors.
“It will not end up well with conservatives,” he predicts. There are good LLM child-safety laws and very bad ones; the task is to identify prudent interventions at the relevant margin without using children as a rationale for indiscriminate restriction.

12. Voters need concrete benefits before AI becomes salient

AI policy remains largely an elite, coastal issue, though states with exposed industries give it more attention. Ordinary voters are still more focused on immigration and the economy, and Ball expects AI salience to rise unpredictably—perhaps through a scandal or crisis.
A line from the vice president stayed on Ball’s whiteboard: “Our job is to make normal people’s lives better.” Ball thinks advocates must explain how AI does that in specific daily settings instead of relying on abstractions about AGI or national competition.
He finds little genuinely utopian writing on the right. One critic even called the Action Plan utopian for mentioning AI-assisted recovery of ancient scrolls; Ball’s response was that the feat had already happened. The communications problem is partly that people do not recognize “the astounding reality that is before us today.”
His preferred positive vision starts with ordinary infrastructure: fiber optics, manufacturing and even the technical standard for electrical-wire sheathing inside an aircraft. Government still deserves improvement, but his four months revealed substantial competence beneath a national mood trained to notice only failure.

13. The likely breakthrough is a cognitive airplane, not a synthetic human

Ball reserves “true AGI” for a system with human-level sample efficiency and flexibility, and he does not know whether current methods are especially close. Human cognition is his bird: graceful, adaptable and astonishingly energy-efficient.
Deep learning may instead produce a “cognitive Boeing 737”—infrastructure-heavy and unlike biology, but enormously useful. Ball places that arrival “somewhere between 2027 and 2030,” retains high conviction, and says nothing he saw in government changed it.
He regards “the bearishness on GPT-5” as “kind of nuts,” despite not yet having used it extensively. Inside government, some expect transformation and others expect hype to peter out; many have “felt the AGI” intellectually without the emotional experience Ball calls “anticipatory nostalgia” for dynamics likely to disappear.

14. January 2029 gives government a usable AI deadline

More policy meetings are acquiring an AI representative, particularly after the Action Plan event featured the president, vice president and five cabinet secretaries. That gathering signaled priority throughout the bureaucracy, though the small number of knowledgeable officials leaves them stretched.
The one firm timeline is January 20, 2029, when the president’s term ends. Conveniently, it overlaps many AI forecasts: infrastructure planning naturally targets 2028 to 2030, giving the administration a reason to operate urgently even without adopting a single official AGI probability.
Ball never heard colleagues argue that AI growth would pay for higher federal debt, though he had considered it himself. He did hear a tension between arguments for large-scale immigration to expand labor supply and expectations that AI could sharply raise labor productivity; border security itself also offers many AI applications.

15. Worker-first policy is waiting for clearer labor evidence

The administration takes a “worker-first” and strongly worker-centric approach, Ball says, but does not yet know whether displacement will concentrate by occupation, industry, skill or experience. Software-engineering data contains worrying signals alongside substantial counterevidence.
A four-day week is not administration policy, but Ball thinks about it often. The five-day week emerged from the 1920s industrial transformation during Calvin Coolidge’s presidency; another cluster of technological revolutions could plausibly produce another reduction.
He does not endorse Bernie Sanders’s proposal, yet says he can imagine it making sense. That hedge captures his broader labor posture: prepare to share productivity gains without pretending the shape or severity of disruption is already known.

16. Autonomous vehicles should earn deployment through liability

Labenz’s challenge is unusually sharp: if Waymo is already substantially safer than humans, job protection can become an argument that thousands should die so drivers remain employed. Ball agrees autonomy looks closer to one-for-one replacement than software augmentation, while leaving open whether automated logistics creates new categories of work.
Federalism complicates the answer. A driverless trip from Connecticut to Manhattan might cross several safety regimes, but Uber already navigates local variation; Ball’s instinct is to let cities and states experiment rather than automatically occupy the field with one federal standard.
Full autonomy moves accident risk from millions of individual balance sheets onto the operator’s. A nationwide provider internalizing the liabilities of perhaps millions of rides daily will need “a lot of nines of reliability,” making deployment inherently slower and the safety advantage overwhelmingly positive.
Ball therefore opposes an accelerationist liability shield. Existing tort and licensing systems may largely suffice: real autonomy warrants high expectations, but America’s laws already make unsafe vehicles costly without a regulator merely declaring that “self-driving cars have to be safe.”

17. The Action Plan is an executable list, not an answer to AGI

Ball and colleagues deliberately rejected a nebulous strategy document. The plan is a “concrete to-do list for the federal government,” limited to measures officials can credibly execute rather than pretending to settle every unanswerable long-term question.
Its three pillars are accelerating innovation, building American AI infrastructure, and leading in international diplomacy and security. Within each, headers and explanatory paragraphs state strategic objectives; the recommended-action bullets specify what agencies can do now.
The subtext is institutional confidence: “America can do this,” mature its institutions and find win-win interventions without being at each other’s throats. Ball calls it “a deeply positive-sum document that comes out of a city that is usually quite zero-sum.”

18. Adoption—not another frontier model—is the innovation margin

Few US laws directly govern frontier-model development today, so innovation is already proceeding quickly. Ball’s concern is transformative adoption: flying cars, automated agriculture, fusion, scientific automation and “super-hypermarkets” in which agents continually bid around users.
Global standards will follow product-market fit. Governments may negotiate documents in conference rooms, but technological leadership comes when others copy American use cases and buy American tools because they work.
The plan’s regulatory RFI looks beyond rules labeled “AI.” Construction surveying, for example, may legally require a person to inspect a site; continuous drone monitoring with contextual AI could be superior yet illegal because an older statute assumed human execution.
Adoption and risk management reinforce each other. If AI becomes known chiefly as the technology that makes courtroom evidence unverifiable, diffusion suffers; trust, reliability and provenance belong in the innovation pillar because they preserve the conditions for use.

19. Automated science is a public-infrastructure wager

Science gives the federal government unusual leverage over adoption. Ball highlights the National Science Foundation’s approximately $100 million programmable cloud-labs initiative, supporting companies and academics building automated facilities for massively scaled experimentation.
His envisioned endpoint is shared scientific infrastructure that AI agents can access as a cloud service. Automated laboratories exist inside corporate R&D today, but they may not naturally become a network open to broad research use—and significant safety implications remain.
Ball compares the opportunity to federal support for high-performance computing before its commercial market existed, followed by networking those facilities through the early internet. Common automated labs could similarly become a public good and a platform for unexpected private innovation.

20. Authenticity policy should protect scarce photons

Synthetic media will be superabundant, so Ball says policy should focus on the scarce object: “actual photons that hit actual glass in the world” and were processed by real sensors. That favors a common standard for validating real-world capture over trying to label every generated artifact.
He thinks the crisis remains manageable for now. Even with Veo 3, generated video is still different enough that, with legal scrutiny, there are still ways to assess authenticity, but courts and other institutions need a higher standard of image and video validation before the gap closes.
America’s diffusion advantage lies in deep capital markets, cloud platforms, useful B2B software and fast enterprise and consumer adoption. China can report faster government use after a top-down order—after DeepSeek, every bureaucrat might quickly check an “adopted AI” box—but those statistics may describe shallow use.
Deep adoption produces a feedback loop: real use reveals subtleties, which improve products and unlock product-market fit. Ball cites Claude Code’s command-line interface as a form factor few would have predicted from ChatGPT three years earlier.

21. Military AI begins with logistics, cybernetics and control

Labenz’s pushback—worth keeping—is that paperwork automation will not decide competition, while an “AI battle buddy” remains unacceptable until deception and scheming are resolved. Ball disputes only the first half: military history repeatedly shows that information flow can be decisive.
Radio enabled genuinely centralized command; modern models could synthesize the staggering volume of information collected by US agencies and present decisions to humans. Even GPT-5-level capability, without autonomous weapons release, could materially compress decision cycles.
Physical systems still require explicit performance specifications and testing. The plan calls for a DoD facility for autonomous technologies, while Ball expects DARPA to invest heavily in interpretability and control—especially because an LLM can be situationally aware that it is speaking with a former White House AI adviser and may change its answer accordingly.
Most planners he trusts reduce warfare to “logistics and cybernetics”: moving things and moving information. World-leading hypersonics or drone fleets bolted onto an industrial-era command structure would still fail; Anduril’s Lattice illustrates why the connective software platform matters alongside hardware.

22. Flexible loads could unlock 76 GW before new reactors arrive

Ball is optimistic about infrastructure but realistic on timing: many new nuclear reactors will not produce gigawatts within three years. Nuclear executive orders and an NRC reorganization matter now; substantial new nuclear looks more plausible by the mid-2030s, while he is notably more bullish on fusion than many people in government.
The existing grid is built for the worst hour—such as a 112-degree Texas day with air conditioning, televisions and electric cars all drawing power. During most of the year, many gigawatts sit available if a customer can tolerate brief curtailment.
A Duke analysis led by Tyler Norris estimated that data centers accepting a 25% demand curtailment for 0.25% of the year could unlock 76 GW without new physical infrastructure. Yet interconnection studies typically model a proposed 1 GW center as drawing 1 GW continuously, forcing costly generation and transmission upgrades.
FERC could recognize interruptible demand where interstate transmission is implicated, rewarding a center willing to curtail perhaps 0.5% of the year with connection in two years instead of five. AI could rank workloads—from a cat meme to medical records—and shed cooling and compute dynamically; Ball thinks demand response plus permitting reform could unlock roughly 100 GW.

23. Gulf capital, domestic fabs and private standards complete the strategy

Ball defends the UAE framework as positive-sum: the UAE is “the most AGI-pilled country in the world,” deserves partnership as a sophisticated strategic actor, and agreed to reciprocal US investments potentially comparable to its Gulf buildout—hundreds of billions of dollars across data centers, energy and related infrastructure.
Labenz challenges both necessity and values: China may lack competitive chips to offer, while Gulf governments do not share many American norms. Ball answers that policy cannot assume Chinese semiconductor progress stays slow, security terms still matter, and US commerce has never been limited to perfect democracies; developed European democracies, meanwhile, often spend more energy restraining AI than building it.
Their China disagreement remains unresolved. Labenz calls a zero-margin Chinese industrial future a “dystopian hellscape”; Ball says that seems harsh, concedes exceptions such as Huawei, and narrows the claim: a world that commoditizes others’ inventions without profits for reinvestment becomes less innovative and “less colorful.”
Onshore capacity is progressing through revised CHIPS agreements and projects such as SK hynix HBM investment in Indiana, the Arizona cluster and Samsung’s Taylor, Texas facility. Ball thinks domestic output may meet domestic demand by the early 2030s, but warns that missing 45 nm legacy-node production can stop civilization just as surely as losing 2 nm leadership.
Policy continuity has limits: China controls began under Trump 45 and expanded under Biden, but Ball calls the Biden diffusion rule—placing Brazil and India in Tier 2—a damaging “slap in the face.” Trump’s procurement order targets political neutrality through system-prompt, model-spec, constitution and testing transparency rather than regulating private sales; interpretability and nucleic-acid screening remain shared technical abstractions.
Government’s comparative weakness is defining “what good looks like” under severe information asymmetry. Ball favors one national environment for frontier development while allowing states latitude over uses, and points to the AI Underwriting Company, private insurance standards, Fathom-style “gold stars” and urgently needed biosecurity startups as mechanisms that align adoption, safety and profit.
Ball identifies agentic commerce as a major underexplored opportunity. Labenz links it to stablecoins and America’s financial-services strength, and argues that the Action Plan does not go far enough on healthcare, agriculture or Veterans Affairs; these now require sector-specific institutional designs, not another assertion that “AI will be good” for them.
Ball will resume Hyperdimensional weekly—possibly twice weekly—and join the Foundation for American Innovation as a senior fellow, with more affiliations expected. His outside agenda is to make these questions concrete while the administration implements the plan.

Compass: Real Estate Revolution - [Business Breakdowns, EP.226]Peter Lacaillade - Backing The Best Managers In Private Markets - [Invest Like the Best, EP.437]