The White House Just Rewrote the 1945 Blueprint for American Science. The New Load-Bearing Wall Is AI.

TL;DR

On July 21 the White House Office of Science and Technology Policy published Science: A New Golden Age, an 80-plus-page redesign of how the United States funds research. It is the first document of its kind since Vannevar Bush's 1945 report Science: The Endless Frontier, the one that produced the NSF and the postwar grant system. Four goals drive it, and the fourth is stated flatly: "we must prepare our research enterprise for the AI revolution." The concrete asks are fully funding the Genesis Mission, opening federal scientific datasets, funding closed-loop autonomous laboratories, and building a verification layer that can keep up with machine-speed generation. Roughly $200 billion a year in federal R&D is the portfolio being repointed.

What actually landed

The report is signed by OSTP Director Michael Kratsios in a letter of transmittal dated July 21, 2026. It comes with an annex: an FY 2028 R&D Priorities Memo telling agencies to fold this guidance into their budget submissions to OMB. That annex is the part with teeth. A report is a report; a budget instruction is a budget instruction.

The diagnosis is blunt. Federal research money has grown while scientific productivity has slowed. The enterprise has become "dependent on a narrow set of legacy institutions." Administrative work now eats nearly half of a researcher's working hours. And the report notes that some grants take nearly two years from submission to award, which it compares, deadpan, to the time it took to design and produce the first Boeing 747.

The four goals: prioritize the individual scientist over legacy institutions, diversify how grants get allocated (golden tickets, fast grants, prize challenges, regranting), set national technology missions and rebuild the manufacturing muscle to land them, and rebuild the whole thing for AI.

Genesis Mission goes from a program to the plan

The Genesis Mission was launched by executive order in November 2025 and directs the Department of Energy to build the American Science and Security Platform, wiring together supercomputers, AI systems, scientific instruments and datasets into one discovery engine, with the stated goal of doubling the productivity and impact of American science within a decade.

Until now it read like one more initiative with a good name. This report makes it the organizing structure. DOE's 17 national laboratories, roughly 40,000 technical staff and about $20 billion in annual funding, become the substrate. DOE has been directed to name at least 20 science and technology challenges of national importance, spanning advanced manufacturing, biotechnology, critical materials, fission and fusion, quantum information science and semiconductors, reviewed annually. In December 2025 DOE announced collaboration agreements with twenty-four organizations, including AI companies, chipmakers and cloud providers.

Two named pieces matter if you build things. The American Science Cloud is meant to have the labs curate and distribute AI-ready federal scientific data to the broader research community, the report's example being the kind of public dataset that already exists because government built it, like NOAA weather data or the Materials Project. The Transformational AI Models Consortium is aimed at domain foundation models trained on DOE data and facilities. The report is explicit that the federal value-add is not "funding AI in the abstract" but pointing it at problems where a breakthrough unlocks whole downstream branches, protein structure prediction being the worked example.

The interesting part: they know the output will be sludge

The strongest section is the one about verification, and it is unusually clear-eyed for a government document.

The argument runs like this. The cost of generating scientific claims has fallen exponentially. The cost of checking them has not. Publishing novelty carries funding and prestige; confirming someone else's result carries neither. Previous large-scale replication efforts died because they needed armies of specialists checking studies by hand, and manual verification does not scale to millions of papers a year. Bolt AI onto that and you do not get more truth, you get more plausible-sounding output faster than anyone can audit it.

Science already has a firehose pointed at a bathtub with one drain. The report's point is that AI is about to widen the hose without touching the drain.

The evidence it cites is close to home: leading AI conferences have seen submission surges of 60% in a single year, with researchers now reviewing nonsensical AI-generated submissions while rebutting low-quality AI-generated reviews of their own work. Anyone who has served on a program committee lately just nodded.

The fix is a spec, and it looks like agent infrastructure

What OSTP wants built is recognizable to anyone who has shipped an eval harness. Open APIs and interoperability standards so verification can plug into journal submission systems, grant reporting platforms and private-sector AI research tools. Standards for replication packages so computational research arrives in machine-auditable form. Prizes for successfully replicating or disproving influential papers.

The mechanism it sketches: specialized agents parse a submitted paper, reconstruct the computational environment, execute the analyses in a sandbox, and compare outputs against the claimed results. That is CI for science. It is also, the report notes, the same infrastructure that would later decide which machine-generated hypotheses are worth spending real lab time on.

The precondition is the May 2025 executive order Restoring Gold Standard Science, whose reproducibility, transparency and data-sharing requirements are exactly the conditions that make automated verification possible in the first place. Compliance paperwork, repurposed as a machine-readable interface. Rare to see that trade go the right way.

Why a builder should care

Three reasons, in descending order of how soon they touch you.

Data. If the American Science Cloud ships anything close to its brief, decades of federal scientific data currently locked behind bureaucracy, restrictive licensing or nobody's budget line becomes AI-ready and public. That is training and evaluation material you cannot buy.

Verification as a product category. Open APIs, machine-auditable replication packages and replication bounties describe a market, not just a policy. Sandboxed execution, environment reconstruction and claim-checking agents are things people are already building badly. A federal standard plus prize money is a demand signal.

Scale context. The report notes that in 2025 alone American companies committed more than $400 billion to AI infrastructure, more than the inflation-adjusted cost of the Apollo Program and the Manhattan Project combined. Against that, $200 billion a year of federal R&D is not the biggest checkbook in the room. It is the one that funds the questions nobody can monetize in four quarters.

The caveats, straight

This is a strategy document plus a budget-priorities memo, not an appropriation. The FY 2028 guidance tells agencies what to weight in submissions to OMB; Congress still writes the checks. "Fully fund and expand the Genesis Mission" is a recommendation, and there is no dollar figure attached to it in the report.

The redistribution is real politics, not neutral plumbing. Prioritizing individual scientists, ARPA-style program managers and new mission-driven organizations over "legacy institutions" means large research universities, which depend on federal funding, come out worse. Expect that fight.

And "AI-native scientific institutions" is currently a heading, not a design. Faster publication, granular credit attribution and new market mechanisms for directing research resources are all sketched at the level of intent. The verification section is the one part with an implementable shape.

Key Takeaways

• OSTP published Science: A New Golden Age on July 21, 2026, the first ground-up redesign of the U.S. research enterprise since Vannevar Bush's 1945 report, covering roughly $200 billion a year in federal R&D.
• Preparing the enterprise for AI is one of four top-level goals, with the Genesis Mission named as the flagship AI-for-science effort and DOE's 17 national labs as its substrate.
• The American Science Cloud would open curated, AI-ready federal scientific datasets; the Transformational AI Models Consortium targets domain-specific scientific foundation models.
• The sharpest section is on verification: generation is cheap, checking is not, and the proposed fix (open APIs, machine-auditable replication packages, sandboxed agent verifiers, replication prizes) is effectively CI for science.
• It cites a real base rate for the problem: one-half to two-thirds of psychology studies and over a third of celebrated experimental economics studies failed replication, with irreproducible preclinical biomedical work misdirecting an estimated $28 billion a year.
• It is guidance, not money. The FY 2028 memo shapes agency budget submissions to OMB; appropriations and the universities-versus-individuals fight are both still ahead.

Sources: The White House, Science: A New Golden Age, full report (PDF), U.S. Department of Energy, Genesis Mission launch, Federal Register, Restoring Gold Standard Science (EO 14303)