How the White House Plans to 10x Scientific Productivity | Michael Kratsios | EP #276
Summary
Kratsios divides technology into products “born free” or “born in captivity.” Internet-like technologies need government restraint; commercial drones and AI medical diagnostics need affirmative rule changes before customers can realize their value. For investors, regulation is therefore both a bottleneck and a market-making catalyst.
Washington’s ambition is moving from AI adoption toward national-scale demand creation: humans back on the Moon in 2028, initial lunar-base elements by 2030, a space nuclear reactor by 2028, a scientifically relevant quantum computer by the end of Trump’s term, and AI-driven science through the Genesis Mission. Kratsios said Genesis should probably aim for 10x scientific productivity, not the publicly stated 2x: “We have to aim big. We have to do 10x.”
The administration wants leadership in both closed and open models, but Kratsios concedes that Chinese open-source models are well-performing and currently cheapest. The American AI Exports Program aims to counter that advantage with turnkey packages spanning chips, models, applications, and government-backed financing—a deliberate effort to ensure that “the world should build on America’s AI and tech stack.”
Kratsios strongly defends semiconductor controls, calling the 2019 restrictions on EUV lithography “probably one of the most impactful export controls” in US history. Diamandis argued that restrictions can incubate foreign competitors; Kratsios’s answer was that China already treated semiconductor independence as a strategic priority, while controls throttled its ability to match US frontier systems.
The Golden Age blueprint treats scientific stagnation as an incentive-design failure, not principally a shortage of money: NIH approaches $45 billion, yet researchers reportedly spend about 45% of their time on grant-related administrative work. Proposed repairs include five-year awards, applications reviewed within a month, unilateral “golden tickets” for unconventional proposals, dedicated meta-science units, prizes, and four-year industry-linked PhDs.
The most radical proposal is an AI-native scientific marketplace where funders post bounties, agents identify leads and hire autonomous labs, cryptographically signed results trigger smart-contract payments, and prediction markets guide resources. Humans would still choose the questions and judge the most consequential results, but coordination and experimentation could run continuously “at machine speed.”
Robotics, autonomous laboratories, data centers, and public scientific datasets form the investable physical layer. Kratsios described new limits on non-US humanoid imports, federal support for closed-loop robotic labs, and a requirement that data-center builders “build, bring, or buy” their own power; meanwhile, 70 years of national-lab data should become AI-ready and remain a public good, analogous to NOAA weather data.
Deep dive
1. Regulation must distinguish technologies born free from those born captive
Kratsios’s central taxonomy separates technologies “born free”—best served when government steps back—from those “born in captivity,” whose commercialization depends upon an agency changing rules. The 1990s internet belongs in the first category; commercial drones and AI-powered diagnostics belong in the second.
The warning against premature regulation is the EU AI Act, which Kratsios said was finalized before ChatGPT existed. His conclusion: rules written for an earlier technical paradigm cannot plausibly map cleanly onto today’s large-language-model systems.
For captive technologies, inaction is itself restrictive. Kratsios argued that government must thoughtfully redesign permission structures so products can be “safely deployed to Americans,” because waiting too long can prevent their benefits from materializing at all.
2. Government AI adoption remains fragmented—and surprisingly constrained
Diamandis asked whether the US might appoint a centralized AI minister or deploy AI representations of officials. Kratsios rejected the first idea as a “tall order” because AI affects every agency differently—from drones and medicine to SEC oversight of financial services.
His preferred structure is distributed domain expertise, pushed forward by White House missions such as Genesis. Government will rarely lead the technological frontier, he conceded, but presidential action can force agencies to treat AI-enabled discovery as a whole-of-government priority.
The adoption gap is concrete: large language models were not permitted on White House systems because of Presidential Records Act constraints. Kratsios hoped that would change soon, calling it an illustration of “the pace at which sometimes government tech operates.”
3. AI’s political problem begins with a fear-first narrative
Confronted with claims that roughly three-quarters of Americans fear AI and 71% oppose nearby data centers, Kratsios answered plainly: “AI has a massive PR problem.” He traced part of it to government and industry framing AI almost exclusively through job loss, biological risk, and catastrophe.
His best example was the first UK AI Safety Summit at Bletchley Park, whose organizing premise centered on what could go wrong. If politicians repeatedly describe a technology as dangerous, Kratsios argued, public fear should not surprise anyone.
Healthcare offers the clearest counternarrative because families can directly experience better diagnostics and care. Government can also convene stories about factories, hiring, and domestic supply chains rather than treating AI narrowly as a software product.
Kratsios relayed Jensen Huang’s claim that Nvidia had placed the largest order in Corning’s history, requiring record production. The intended message was tangible: AI investment is creating industrial spillovers across the economy, not merely concentrating revenue inside frontier labs.
4. Better labor data must precede confident claims about AI job losses
Kratsios remains long-term optimistic on employment, recalling that automation fears during the first Trump administration did not produce the forecast collapse and that total employment instead increased. He nevertheless said labor displacement deserves serious monitoring.
The immediate problem is weak measurement. The AI Action Plan called for a Department of Labor initiative to gather data from employers that do not typically report it, enabling targeted retraining, reskilling, and labor assistance rather than policy based on anecdotes.
Diamandis proposed a “parachute” incentive requiring companies conducting AI-related layoffs to provide AI upskilling. Kratsios found the concept interesting but questioned what counts as an AI layoff, observing that firms may attribute ordinary restructuring to AI because the story plays better publicly.
Diamandis supplied the market incentive behind that labeling: stocks can rise when companies report more revenue with fewer people. Their exchange preserved an important ambiguity—AI may drive real displacement while also serving as a convenient explanation for cuts planned anyway.
5. National missions are intended to restore ambition and create demand
Kratsios wants government to be “opinionated” about national priorities, reviving the Manhattan Project and Apollo tradition of choosing a North Star that appears impossible, mobilizes institutions, and draws young people into science.
The space timetable is deliberately concrete: man back on the Moon in 2028, first lunar-base elements by 2030, and a nuclear reactor in space by 2028. “That’s crazy,” Kratsios said, before giving the governing rationale: “We’re going to do it because we’re Americans.”
Genesis initially targeted doubling the productivity of the US scientific enterprise over a decade. Another mission seeks a scientifically relevant quantum computer by the end of Trump’s term—not another abstract milestone, but a machine that can “do something,” with pharmaceuticals the application Kratsios finds most compelling.
Fusion remains outside that presidential-term horizon, but Kratsios sees unusually strong private participation. Diamandis counted 37 venture-backed fusion companies, while Kratsios emphasized that private investment in the energy source has never been greater.
6. The US wants an exportable AI stack, not isolated products
Administration policy is categorical: America must lead in both closed and open models. Kratsios acknowledged that the domestic open ecosystem “could be doing better” and that Chinese models are currently well-performing and, for cash-strapped founders, may be the cheapest rational choice.
The American AI Exports Program asked US companies to propose what a complete stack should contain. Commerce and other agencies are converting those submissions into turnkey combinations of chips, models, applications, and tooling, avoiding the burden of making foreign customers assemble seven vendors themselves.
Financing is part of the product. The Export-Import Bank and Development Finance Corporation could make American packages economically viable abroad, turning “a complete package” into an answer to subsidized Chinese technology exports.
Kratsios said the program grew from frustration with Huawei, whose “good enough” telecom stack and PRC subsidies proliferated rapidly. This time, superior Nvidia and AMD chips, American software, open and closed models, and government financing are meant to form one competitive geopolitical offering.
7. Chip controls and humanoid restrictions are industrial policy by denial
Diamandis challenged whether Nvidia restrictions might repeat earlier export-control mistakes by forcing China to build competitors. Kratsios disagreed, calling the restrictions among the administration’s “savviest decisions” because they throttled China’s ability to train models competitive with America’s.
His stronger claim concerned the 2019 EUV lithography controls, which he described as probably among the most consequential in US history. China was already determined to build an independent semiconductor industry, he argued, so withholding enabling technology did not create that ambition.
In robotics, Diamandis contrasted a handful of notable US humanoid companies with more than 150 in China. Kratsios said the administration had just limited imports of non-US humanoids not already shipped, explicitly prioritizing a homegrown industry and secure component supply chain.
He expects that boundary to attract capital, citing a similar action around drones in December followed by dramatic investment in the drone supply chain. His diagnosis is that Chinese robotics receives subsidies and dumping support, while the American sector needs “a bit of a push.”
8. Scientific stagnation reflects process bloat, not insufficient budgets
The Golden Age report begins with declining discovery per dollar. Kratsios highlighted an NIH budget approaching $45 billion alongside rising drug-development costs, arguing that institutions keep repeating the same scientific process and assume additional money will improve its output.
The core prescription is cultural: science policy must become more experimental about doing science. Kratsios found it particularly jarring that a community devoted to experimentation often wants its own funding and publication system preserved as it existed 30 years ago.
Administrative burdens are the clearest loss. A National Academies study reportedly found researchers spend around 45% of their time on grant administration; Kratsios called that “one of the most depressing statistics” because scientists should be working as the country’s “crown jewels.”
Supersonic flight illustrates the regulatory equivalent. Rather than prohibit speeds over Mach 1, the administration is changing the rule toward a noise limit instead of a speed limit: if an operator can fly quietly, “Just fly.”
9. Grants should fit discovery rather than academic bureaucracy
Young researchers remain trapped in a publication-and-tenure ladder designed for slower science. Diamandis cited a startling figure: the median intramural NIH scientist is reportedly 71 years old, while he noted that Nobel-winning work is typically performed around a laureate’s mid-40s.
Most government grants cluster around 18 months, reflecting administrative and academic calendars rather than experimental reality. Five-year grants funded from day one would let scientists pursue long-horizon questions instead of applying for their next award before completing the first.
Fast-track grants would use a few-page application, a decision within a month, and funding sized to a proof of concept. Kratsios pointed to COVID-era decisions made in hours and argued that crisis-speed funding need not await a crisis.
Under the proposed “golden ticket” system, every reviewer might receive one to three unilateral awards. That gives unconventional ideas a path if “one person believes in you” and could recruit stronger reviewers by granting them genuine decision-making power.
10. Meta-science and agent markets could rebuild scientific incentives
Kratsios insists reforms themselves must be tested. Fast-track grants may sound attractive, but government should compare outcomes, update the mechanism, and abandon what fails; this “science of science” mandate has already produced announced meta-science units at NIH and NSF.
The report’s boldest architecture is a marketplace where funders post bounties, AI agents find promising leads and commission autonomous labs, and cryptographically signed results unlock smart-contract payments. Data, hypotheses, compute, and capital would move through machine-speed microtransactions.
Prediction markets could inform grantmakers, bounty markets could allocate resources toward unsolved problems, and reputation systems could surface reliable agents. The report presented this as possible—not inevitable—while retaining humans for judgment, question selection, and evaluation of major results.
The capital base has changed enough to attempt it. In 1950, Kratsios said 70% of R&D was done by the federal government and 30% by the private sector; that ratio has effectively flipped, while philanthropy now commands unprecedented resources, including an OpenAI Foundation he valued at nearly a quarter-trillion dollars.
11. Prizes, biotech, and longevity expose gaps in the mission portfolio
Diamandis said XPRIZE had deployed $600 million in prizes and driven roughly $30 billion in R&D. Kratsios welcomed public-private pooling around national challenges, particularly where payment can reward a demonstrated result rather than a persuasive proposal.
With federal R&D spending around $200 billion annually and NIH near $45 billion, Kratsios considered a $1 billion government prize difficult to absorb. A prize around $100 million, however, is “certainly doable” when the challenge is important enough.
Diamandis’s $101 million Healthspan XPRIZE has 830 teams seeking to reverse functional age by 20 years across cognition, muscle, and immunity. His economic case: US life expectancy is about 78–79, while healthy life expectancy is 63, leaving approximately 16 years in poor health that could be transformed by extending healthspan.
Kratsios conceded that longevity was missing and “probably should have” been included. Biotech is partly nested within Genesis—about $5 billion of recently announced grants included many biotech projects—and he treated Demis Hassabis’s stated ambition to cure all diseases within a decade as serious.
12. Innovation zones, education, and autonomous labs move experimentation into the physical world
Kratsios’s 2017 drone pilot paired 10 state, local, or tribal governments with UAS operators and granted regulatory room to test deliveries. The administration is applying the model again through an eVTOL pilot, letting willing communities discover what works before nationwide rules harden.
This is federalism as competitive policy: Diamandis’s map ranked Texas first for technological openness, while Kratsios recalled autonomous-vehicle testing leaving California for Arizona after rules changed. Companies and residents “vote with their feet,” turning regulatory arbitrage into pressure for reform.
Data-center operators are similarly expected to earn community acceptance. Under the ratepayer-protection pledge, builders must “build, bring, or buy” their own power and cover associated costs; Diamandis added that golf courses use roughly 30 times and almond farming 50–60 times more water than data centers.
On education, Kratsios put parents first: some families may choose Alpha School’s roughly three to four hours of AI in the morning and social skills afterward, while others prefer no classroom technology. He said most families instead receive a “broken middle” and warned that declining US student interest in STEM threatens national health, security, and growth. Diamandis proposed an AI educational overlay, which Kratsios said was possible if it became economical.
Autonomous labs complete the vision: AI proposes a hypothesis, robots execute it, the model reads the result and designs the next experiment 24/7. Hardware initially limits these systems to narrower domains, but Kratsios wants students and scientists eventually to “go online, put the hypothesis in and hit go.”
13. Genesis is shifting from a 2x promise to a 10x aspiration
Kratsios admitted the 2x Genesis target came from the program’s prior public commitment, and the report retained it partly for consistency. After advances during the subsequent six months, his revised judgment was unambiguous: “We should probably be aiming for 10.”
He cited Opus, Anthropic’s rapid revenue growth, the “Mythos moment,” the “Fable release,” and OpenAI’s coming GPT-6 as evidence of how quickly the premise changes. The exact capabilities remain uncertain, but “the pace of innovation on AI is just insane.”
Genesis’s practical asset is 70 years of national-lab data across physics, chemistry, mathematics, biology, and other fields—much of it neither AI-ready nor incorporated into models. Making those datasets usable could accelerate hypothesis generation across every scientific domain.
Kratsios said breakthroughs built from that federal data should rest on a public-good substrate, analogous to commercial weather applications built atop freely available NOAA data. On more aggressive ideas such as AI personhood, however, he said America is not there yet; the near-term priority is to “let our horses run.”