"Science: A New Golden Age", The White House's New AI Funding Plan, Explained

The White House just published a 100-plus-page plan to change how America funds science, trains researchers, builds technology, and decides which discoveries deserve national attention.

The report, Science: A New Golden Age (pdf), arrives 81 years after Vannevar Bush’s Science: The Endless Frontier. Bush’s 1945 report helped create the modern federal research system. Michael Kratsios, director of the White House Office of Science and Technology Policy, argues that system now needs another redesign.

The proposal reaches far beyond “use more AI.” It calls for a shift from institutions toward individual scientists, from uniform peer review toward a portfolio of funding mechanisms, and from published discoveries toward domestic manufacturing and national power.

It also imagines a much stranger future: AI agents proposing hypotheses, hiring cloud laboratories, checking one another’s work, and earning credit through machine-readable scientific markets.

Here is the whole blueprint, compressed without sanding off the important parts.

The one-sentence version

The White House wants to turn American science into a faster, more experimental, AI-ready system that funds people, rewards risky work, rebuilds manufacturing, and treats scientific discovery as a national strategic asset.

The accompanying FY 2028 research priorities memo gives that vision teeth. Federal agencies with at least $3B in FY 2026 R&D authority must submit implementation plans within 90 days. Agencies must also reflect the priorities in their FY 2028 budget requests.

The plan has four connected parts:

• Rebuild the machinery that funds and evaluates science.
• Tie discovery more directly to national technology missions and domestic production.
• Expand scientific careers beyond the traditional Ph.D.-to-professor ladder.
• Build an AI-native research system that can generate and verify knowledge at much greater scale.

The diagnosis: America’s science machine got slower while the world changed

The report begins with a structural argument. The postwar system assumed that universities would conduct basic research, industry would turn discoveries into products, and government would fund work that markets would ignore.

That “linear model” no longer describes modern science. Fundamental research, engineering, manufacturing, and commercialization now feed one another in loops. AI research is the clearest example. The most capable models, largest computing clusters, and many top researchers sit inside companies rather than universities.

The report says American businesses now spend about $700B per year on R&D, more than three times the combined spending of government and higher education. Federal science policy still behaves as though universities hold most frontier capabilities.

At the same time, the report argues that the public system has accumulated several drags:

• Researchers can spend nearly half their federally funded time on administration rather than research or teaching.
• Some grant processes can take close to two years from application to award.
• Short, project-specific grants reward safe work with predictable intermediate results.
• Consensus peer review can reject unusual ideas that excite one expert but divide a committee.
• Academic hiring, tenure, and publishing reward paper volume more reliably than replication, tools, datasets, or negative findings.
• Scientific careers take longer to reach independence, which can push younger researchers toward safer questions.
• Reproducibility failures can send entire fields down expensive dead ends.

The report also makes an explicitly political claim. It argues that identity-based criteria weakened meritocratic selection and says federal science programs should judge talent by demonstrated technical ability. Its FY 2028 memo points to reasoning tests, competitions, engineering portfolios, standardized quantitative assessments, and technical work as possible signals.

The geopolitical pressure sits behind almost every recommendation. The report says China has reached rough parity with U.S. R&D spending on some purchasing-power measures. It argues that America can no longer assume discoveries made in U.S. laboratories will become U.S. industries.

Batteries, flat-panel displays, and advanced manufacturing serve as warnings. America can invent the science, then lose the factories, process knowledge, suppliers, and next generation of improvements.

The funding overhaul: bet on people, run more experiments

The report’s most immediate proposal is a wider menu for federal research funding.

Traditional peer-reviewed grants remain part of the system. The White House wants them joined by mechanisms designed for different kinds of problems.

Fund people for longer

The plan highlights portable fellowships such as the NSF Graduate Research Fellowship Program, which supports an individual and can move with the researcher. It also points to the NIH Director’s Pioneer Award, which offers up to five years of support for unusually ambitious work.

The FY 2028 memo asks agencies to propose more grants lasting at least five years. It recommends funding them upfront, then using periodic reviews to redirect money when necessary.

The goal is simple: give scientists enough time and independence to pursue work that may fail early before producing a major result.

Make small bets faster

Agencies are encouraged to create “fast grants” with short applications, decisions within one month, and award sizes suited to exploratory work.

That matters when a new dataset, outbreak, instrument, or unexpected result creates a narrow window for research. A nine-month review can turn a timely idea into a historical footnote.

Give reviewers a golden ticket

A “golden ticket” lets one qualified reviewer champion a proposal that lacks panel consensus. The mechanism is designed for ideas that look brilliant to one domain expert and strange to everyone else.

The system would still require scientific rigor and conflict-of-interest controls. It changes who can say yes.

Pay for outcomes

Prizes, grand challenges, and advance market commitments would pay for verified results rather than reimbursing effort alone. The memo asks ambitious prizes to target at least $3 in private investment for every $1 of federal funding.

These mechanisms work best when the destination is clear and the route remains uncertain. Autonomous vehicles, better batteries, new manufacturing methods, and specific medical capabilities fit that shape better than open-ended basic research.

Fund new kinds of research organizations

Some problems require 10 to 100 full-time people, specialized infrastructure, and five years of coordinated execution. They are too large for one university lab, too public-good oriented for a startup, and too small for a national megaproject.

The report points to NSF X-Labs, a $1.5B initiative over ten years for independent teams pursuing milestone-based scientific challenges. It also favors ARPA-style programs, focused research organizations, curiosity-driven institutes, and joint centers spanning government, academia, and industry.

Make funding agencies study themselves

Each major science agency could create a metascience unit that tests how funding actually works.

These teams would compare review methods, track near-miss applicants, map neglected research gaps, test fast grants against standard grants, and measure whether different programs produce new fields, useful tools, or reproducible findings.

Program officers would gain more discretion and status. The report wants them treated as active architects of research portfolios rather than administrators moving applications through a queue.

Science becomes industrial policy

The report rejects the idea that federal responsibility ends after a paper gets published.

The new model follows discoveries through prototyping, testing, manufacturing, and deployment. That makes science policy inseparable from industrial strategy.

The main proposals include:

• Open DOE national laboratories, NASA centers, defense facilities, and federal testbeds to more startups and private researchers.
• Consider commercial urgency and innovative potential alongside academic merit when granting facility access.
• Streamline licensing and Cooperative Research and Development Agreements.
• Use regulatory sandboxes to test new technologies under controlled conditions.
• Expand public-private consortia for shared engineering bottlenecks.
• Use federal purchasing power to create open standards and stronger domestic supply chains.
• Let states and counties compete as regulatory and manufacturing testbeds.

The report treats the Human Genome Project, DARPA, SEMATECH, and the consortium that advanced extreme-ultraviolet lithography as templates. Government can organize a mission, absorb early risk, align competitors around pre-competitive work, and let companies build products on top.

The FY 2028 memo names six national missions:

1. AI: Use the Genesis Mission to double the productivity and impact of American science within a decade.
2. Quantum: Build a quantum computer capable of beginning quantum-enabled discovery.
3. Fusion: Demonstrate commercial fusion power in the United States by the mid-2030s.
4. Space: Return Americans to the Moon by 2028, develop a lunar base, and expand nuclear power for space systems.
5. Robotics: Develop general-purpose autonomous systems that can manipulate objects and work reliably in physical environments.
6. Semiconductors: Advance post-EUV lithography, 3D packaging, and new materials for future chips.

This is a much more directional view of federal science. Government still funds curiosity, but it also selects arenas where national capability should compound.

The “human hands” chapter matters as much as the AI chapter

The report repeatedly returns to a constraint that software culture often ignores: ideas eventually hit atoms.

A model can propose a material in seconds. Someone still has to fabricate it, measure it, troubleshoot the instrument, and scale the process.

That knowledge is often tacit. A machinist hears when a cut is wrong. A laboratory technician recognizes a bad cell culture before a dashboard does. A manufacturing engineer knows which variables will destroy yield, even when the formal process sheet looks correct.

The White House wants that craft knowledge pulled back into science education and national strategy.

Its proposals include:

• Count apprenticeships, externships, industry credentials, and hands-on technical work toward degrees.
• Put machinists, technicians, and skilled practitioners inside research institutions.
• Create portable credentials in fields such as advanced machining, cryogenics, lab automation, and semiconductor manufacturing.
• Expand science and technology apprenticeships.
• Turn community colleges into regional research and manufacturing hubs.
• Connect rural makers, hobbyists, and technical students to formal research opportunities.
• Let researchers move more easily among universities, companies, and federal laboratories.
• Consider flexible national-service requirements for some federally funded fellowships.

The goal is a network of regional innovation clusters where researchers, factories, suppliers, community colleges, and skilled workers live close enough to exchange problems and know-how.

The report calls for “a hundred Silicon Valleys,” each built around a local technical strength rather than a copy of the software industry.

Genesis Mission: the operating system for AI-driven science

The centerpiece is the Genesis Mission, launched in November 2025.

Genesis is supposed to connect federal supercomputers, datasets, instruments, autonomous laboratories, and scientific models into an American Science and Security Platform.

The report says the Department of Energy’s 17 national laboratories employ about 40,000 scientists, engineers, and technical staff and receive roughly $20B annually. Genesis would combine those capabilities with data from agencies such as NOAA, FDA, NSF, and the Department of Veterans Affairs.

The mission focuses on four bottlenecks.

Pick problems where AI can actually help

The report favors problems with enormous search spaces, substantial usable data, and clear ways to measure progress. Protein structure prediction worked because it had all three.

DOE must identify at least 20 national science and technology challenges across manufacturing, biotechnology, critical materials, fission, fusion, quantum, and semiconductors. The list will be reviewed annually.

Build public-private capacity

DOE announced agreements with 24 organizations in December 2025, including AI companies, semiconductor firms, and cloud providers.

A Transformational AI Models Consortium would build scientific foundation models around federal data and facilities. The government’s role is to create shared assets that no single company has enough incentive to build for everyone.

Make scientific data usable

A huge share of federal scientific data remains difficult to access, poorly labeled, trapped in old formats, or restricted by licensing.

The proposed American Science Cloud would curate and distribute AI-ready federal datasets. New funding would reward researchers for preserving experimental records, failed trials, protocols, and operational data that normally disappear after a project ends.

This matches the direction we saw in OpenAI’s recent science push. Useful scientific AI needs evidence, experiments, and feedback loops, not prettier answers in a chat box.

Connect models to laboratories

Genesis would invest in robotics, autonomous laboratories, and AI control systems that can propose an experiment, run it, read the result, and choose the next experiment.

The report says the mission has already funded 14 projects involving robotics, automated laboratories, and autonomous control. It also cites an initial $380M NSF push for programmable cloud laboratories.

The equipment itself may need redesign. Scientific instruments often use closed software and incompatible data formats. Federal laboratories could use their purchasing power to demand open interfaces and cross-vendor interoperability.

That physical loop is also why the AI biology renaissance matters. The next leap comes when models can read, design, test, and update in one connected system.

The verifier has to scale with the generator

The report’s strongest warning is that AI can make science look more productive while making the knowledge base worse.

Models can generate more papers, grant proposals, analyses, and peer reviews. None of that guarantees more truth. Bad results can become easier to extend, cite, and encode into future models.

The administration’s “Gold Standard Science” principles include reproducibility, transparency, uncertainty reporting, falsifiability, unbiased review, acceptance of negative results, interdisciplinary work, and freedom from conflicts of interest.

The report proposes a verification layer to enforce those ideas at machine speed:

• Standardized, machine-auditable replication packages.
• AI agents that reconstruct computational environments and rerun analyses.
• Open APIs connecting verification tools to journals, grants, and private research systems.
• Continuous checks rather than occasional replication campaigns.
• Prizes for replicating or disproving influential findings.

Mathematics serves as the early example. AI can translate ordinary mathematical writing into formal code that proof assistants check line by line. The report cites a 2025 project that formalized the Prime Number Theorem in three weeks, producing 1,100 verified theorems and definitions after a human collaboration had struggled for 18 months.

The broader lesson resembles the jagged frontier of AI capability. Generation and verification are different abilities. A system that produces brilliant hypotheses may still make basic mistakes unless the checking layer is built into the workflow.

The final chapter gets much more speculative

The report imagines scientific publishing breaking into smaller, faster outputs: datasets, code, method notes, negative findings, dynamic papers, and public peer review.

It then goes further.

Blockchain systems could record granular scientific contributions. Decentralized organizations could pool money for research. Prediction polls and markets could help funders identify promising directions. AI agents could post bounties, contract cloud laboratories, exchange data, verify milestones, and earn reputations based on reliable results.

In the report’s most futuristic scenario, an AI agent finds a possible drug target, pays another agent to replicate it, rents a robotic laboratory, receives cryptographically signed results, and triggers a smart contract when the evidence passes review.

The report is clear that this system does not exist in mature form. It presents the pieces as experiments worth exploring, not a near-term deployment plan.

The budget memo turns philosophy into instructions

The annex tells agencies how to begin changing their portfolios.

It asks them to increase the share of foundational research relative to later-stage development. Priority fields include physics, chemistry, materials science, mathematics, computer science, engineering, and foundational biology.

It also asks agencies to:

• Invest in mid-scale instruments, advanced compute, user facilities, and shared laboratory platforms.
• Make federal compute easier for students, individual investigators, and small teams to access.
• Use commercial compute when it is cheaper and sufficient.
• Integrate AI into scientific workflows only when AI fits the problem.
• Create scientific foundation models and ambitious public datasets.
• Pair R&D with workforce, infrastructure, manufacturing, procurement, and regional development programs.
• Seek more industry, philanthropy, state, local, and international cost sharing.
• Track the long-term outcomes of funded researchers and near-miss applicants.
• Reduce duplicate grant requirements and regulatory overcompliance.

Agencies with at least $3B in FY 2026 R&D authority have 90 days to explain how they will apply these practices using their FY 2026 and FY 2027 programs. The FY 2028 budget submissions will show where the money actually moves.

The strongest counterargument: institutions are infrastructure, too

The report often describes university overhead as administrative bloat. Universities strongly dispute that framing.

The Association of American Universities says facilities and administrative reimbursements typically represent 25% to 33% of a project’s total budget. Those funds cover laboratories, utilities, data systems, hazardous-material handling, research security, human-subject protections, and grant administration.

AAU says universities spent $30.2B of their own money on research in FY 2024, including $7.1B in unreimbursed facilities and administrative costs.

That does not prove every current rate or process is efficient. It does show the central tradeoff. Moving money directly to individual scientists can increase freedom, while weakening the shared institutions that make expensive research possible.

Two other tensions run through the plan.

First, national missions can concentrate talent and resources around important goals. They can also narrow the range of questions government supports. The outcome depends on who selects the missions, how much room remains for curiosity, and whether unsuccessful programs lose funding.

Second, AI can expand scientific capacity and flood the system with plausible junk. The report understands this risk. Its success depends on whether verification infrastructure grows as quickly as generation.

Our take

The most consequential idea in this report is not “AI will speed up science.” AI labs, autonomous experiments, and scientific foundation models were already coming.

The bigger shift is the proposed operating system around them.

The White House wants research agencies to behave more like active portfolio managers. It wants individual scientists to carry more power, program officers to make bolder bets, laboratories to work more closely with industry, and national missions to connect discovery with factories.

Several pieces deserve real enthusiasm: faster exploratory grants, longer support for risky work, experiments in peer review, machine-readable replication, open scientific instruments, and technical career paths that value people who build the physical world.

The harder parts concern power and balance. A system directed toward national missions needs safeguards for curiosity and dissent. A system funding individuals still needs durable shared infrastructure. A system built around AI needs verification before output metrics become meaningless. And, as with all things, when there's massive amounts of money involved, it creates an incentive problem: will genuine innovators get the money, or will they have to compete even harder against folks just looking to follow the money.

The next evidence arrives quickly. Watch the 90-day agency plans, then the FY 2028 budget requests. Those documents will reveal whether “A New Golden Age” becomes a different way of doing science or a new label on the same machinery.