Context

Michael Kratsios, director of the White House Office of Science and Technology Policy and the U.S. chief technology officer during the first Trump administration, joins the All-In hosts to defend the administration's science record and unpack a policy report his office released called Science and the New Golden Age. Much of the conversation covers why NSF, NIH, and other federal science agencies have tripled their budgets since the late 1990s without a proportional rise in breakthroughs, and the specific process changes, grant duration, review authority, funding individuals versus ideas, the administration argues will fix that. It also covers politically charged terrain: cuts to DEI-tagged grants and climate research, criticism of Anthony Fauci's COVID-era messaging, and the funding race with China, where Kratsios lays out the administration's reasoning and framing, presented here as his stated position rather than independently verified fact. For a PM, the real substance is a live case study in redesigning a huge, established funding and review process: what happens when an organization's evaluation criteria quietly optimize for the wrong thing, and what it actually takes to re-engineer incentives toward genuine risk-taking.

The Big Idea

Federal science funding has drifted toward optimizing for safe, consensus-backed proposals and a rising budget number, rather than for genuine breakthroughs, and reversing that requires deliberately redesigning who gets review authority, how long funding commitments run, and whether money follows ideas or people.

The administration's "Science and the New Golden Age" report cites NIH's budget more than tripling since 1998 without a proportional rise in breakthrough treatments, an outcome-per-dollar decline researchers call Eroom's Law, as evidence that the existing process, not a lack of money, is the real constraint.

Key Insights

R&D budgets tripled, breakthroughs didn't

NIH's budget grew from $14 billion in 1998 to $47 billion in 2024, more than tripling, but breakthrough treatments did not rise proportionally. Kratsios cites Eroom's Law (Moore's Law spelled backward), the finding that new-drug approvals per billion dollars spent have fallen roughly 80-fold since 1950 and now halve every nine years. His diagnosis is that the default policy response to stagnation has always been to add more money to the same process, rather than asking whether the process itself, who gets funded, for how long, through what review mechanism, is actually producing good outcomes.

Consensus review boards select safe bets

Most NSF grants go through a peer-review panel of three or four reviewers who jointly agree on which proposals are most meritorious. Kratsios argues this structurally favors proposals that are safely "in the strike zone" the whole panel will support, since any one skeptical reviewer can sink an idea, which discourages scientists from proposing genuinely bold, high-risk work in the first place. The administration is piloting a fix, individual reviewers getting limited unilateral authority to fund a proposal outside the panel's consensus, on the theory that a process built for group agreement systematically filters out the kind of high-variance bets that produce real breakthroughs.

Fixed grant length doesn't fit research

• What: NSF grants default to roughly 18 months, an artifact of academic calendars rather than what any given research question actually needs.
• Why it matters: some ideas can be tested in three to six months, others genuinely require three to five years, so a single default duration either forces artificial checkpoints on fast ideas or starves ambitious, longer ones of runway.
• The fix being tested: the administration is piloting "fast-track" short grants that let a scientist test an idea quickly and then apply for a longer grant if it works, alongside explicitly longer multi-year tracks for ideas that need them.

One grant criterion redirected $8B

A Senate committee analysis led by Senator Ted Cruz found that roughly a quarter of NSF grants awarded during the Biden administration, an estimated $2 billion a year across four years, were tied to DEI-related requirements in the application. Kratsios cites this as the report's clearest example of a screening criterion becoming decoupled from the underlying research merit it was meant to be secondary to: once a specific requirement functions as a de facto gate for funding, it can end up shaping a large share of what gets funded regardless of scientific quality. This characterization comes from Kratsios and the cited Senate analysis, not from an independent audit of individual grants' scientific merit.

China's R&D spending grew 19x faster

China's annual R&D spending grew from about $33 billion in 2000 to roughly $670 billion by 2021, a 19-fold increase, versus roughly 3x growth in U.S. spending over the same period. China now publishes an estimated 50% more scientific papers than the U.S. across nearly every field except some life sciences, a reversal from a decade ago when the U.S. published about twice as many papers as China. Kratsios attributes at least part of the U.S. edge in turning research into actual breakthroughs, rather than paper volume, to its decentralized, competitive funding ecosystem, citing China's inability to achieve EUV lithography since 2019 U.S. export controls despite it being a stated top national priority, as evidence that centralized allocation and higher output alone don't guarantee the hardest technical breakthroughs.

STEM PhD pipeline is mostly foreign

Roughly 7 out of 10 STEM PhD recipients in the U.S. are now foreign-born, a reversal from 30 years ago when it was roughly 70% American and 30% foreign in computer science specifically. Kratsios frames this as two compounding problems: too few American students enter STEM at all (he cites the elimination of NSF programs that supported gifted STEM students in K-12 schools, and describes controversy at institutions like UC Berkeley over dropping SAT requirements as unintentionally harming the pipeline), and the U.S. trains large numbers of foreign PhD students, especially from China, without a clear pathway for them to stay and work here after graduation, effectively subsidizing competitors' talent pools. He also notes NIH's own director told him the median age of an NIH intramural researcher is 71.

Mental Models & Frameworks

Fund individuals like a VC portfolio

Instead of only funding specific proposed projects, give exceptional individual scientists open-ended money and let them decide later what to do with it, the way venture capital bets on people rather than fixed plans. Kratsios points to NSF's Graduate Research Fellowship Program, which gives roughly 2,600 of the strongest incoming PhD candidates portable funding they can take to any university, as an existing example, and cites Stewart Butterfield's Slack, which grew out of a failed gaming company he nearly returned investor money on before pivoting to the internal communication tool that became the real business, as the venture-world case for why betting on a person's judgment can outperform betting on their original stated plan. Use this when the value of an idea is highly uncertain but a person's track record or judgment is not: expect most individually funded bets to fail, the same way a venture portfolio expects roughly nine failures for every breakout success, and structure funding so a few outsized wins can justify the losses.

Golden tickets bypass panel consensus

• What: each grant reviewer gets one to three "golden tickets" they can use to unilaterally fund a proposal, without needing the rest of the review panel to agree.
• Why it works: a normal panel process requires every reviewer to sign off, which filters out anything one skeptical reviewer would block; a golden ticket removes that veto for a limited number of picks per reviewer, letting genuinely unconventional ideas get funded even if they wouldn't survive a group vote.
• A secondary effect: giving reviewers real unilateral authority is also expected to make serving on a review panel more attractive to strong scientists, since their individual judgment now visibly matters instead of being averaged away.
• When to use it: in any evaluation process, grant panels, internal idea reviews, resourcing committees, where consensus-based scoring may be systematically filtering out high-variance, high-upside options in favor of ideas nobody objects to.

The market-gap test for public funding

Before committing government capital to a research area, ask whether the private sector or philanthropy is already incentivized to fund it; if so, public money should go elsewhere. Kratsios applies this explicitly to AI: the administration isn't funding AI research generally, since hundreds of billions in private capital already flow into that, but it is funding the application of AI to basic scientific discovery (its "Genesis Mission," aimed at doubling U.S. scientific output) because that specific intersection is a gap the private sector isn't filling on its own. The same test applies to large, decade-long, centrally coordinated bets (a scientifically relevant quantum computer by 2028, fusion by 2035, a moon base by 2030): they're framed as things only a government can coordinate at that scale and time horizon, distinct from the commercial quantum or space activity already happening in the private sector.

Trade-offs & Nuance

Centralized bets vs. market-driven discovery

Kratsios holds two positions that sit in tension: he argues the free-market, decentralized, competitive funding ecosystem is what makes U.S. science outperform China's more centrally planned model, citing China's inability to achieve EUV lithography since the 2019 export controls despite it being a stated top priority, while also championing several large, centrally directed, government-coordinated bets of his own (the Genesis Mission, a 2028 quantum computer, fusion by 2035, a 2030 moon base). His resolution is a portfolio framing: reserve centralized, mission-directed funding for problems only a government can coordinate at scale over a long time horizon (space, energy infrastructure, foundational AI-for-science tools), and leave open-ended discovery science to the decentralized, competitive grant system. The nuance for applying this elsewhere: centralization and decentralization aren't universally better or worse, the right mix depends on whether a problem needs one coordinated multi-year push or benefits from many independent, competing attempts.

Common Mistakes

Treating rising budget as the goal

Kratsios describes the median science lobbyist's approach as fixating solely on whether the R&D budget number goes up, treating that as itself evidence they've "supported science," regardless of whether the spending produces more breakthroughs. He argues the harder, more important question, whether $200 billion a year in federal science and technology funding is actually driving the biggest breakthroughs for the money, or whether it could be deployed differently across organizations, scientists, and time horizons, almost never gets asked, in part because no one inside the system has an incentive to audit whether their own funding is working.

Practical Application

Build a golden-ticket override into review

• Do: give each reviewer or evaluator in your prioritization process, a grant panel, a roadmap review, a resourcing committee, a small number of individual "golden tickets" they can use to unilaterally advance a proposal the rest of the group wouldn't approve by consensus.
• Why it works: consensus scoring structurally filters out high-variance, high-upside options that any single skeptical voter can block; an individual override restores space for genuinely risky bets without abandoning group review for everything else.

Match funding duration to the work

Before defaulting to your organization's standard project or funding cycle, a quarter, a fiscal year, an 18-month grant, ask whether the specific idea actually needs that long, or whether a short fast-track test followed by a longer commitment if it works would use resources better. The administration's fast-track grant pilot, letting a scientist test a six-month idea and then apply for longer funding if it pays off, is a direct model for staging investment by demonstrated progress rather than a fixed calendar.

Audit a screening criterion's real effect

Before letting any single proxy requirement, a certification, a scoring category, a compliance checkbox, become a de facto gate in a review process, check what share of total funding or resourcing it's actually consuming and whether it correlates with the outcome you care about. The episode's example, an estimated quarter of NSF's grant funding tied to a DEI-related application requirement, is presented as a case where a secondary criterion ended up shaping a large share of what got funded; the practical lesson for any organization is to periodically check whether an added screening requirement has quietly become the primary gate.

Ask if you're filling a gap

Before funding, staffing, or building something new, explicitly check whether the market, a competitor, or another team is already incentivized to do it well; if so, redirect resources to the part of the problem nobody else will fund. This is the test Kratsios applies to decide what the federal government funds in AI: not AI research broadly, since hundreds of billions in private capital already covers that, but specifically the application of AI to basic scientific discovery, which he argues the private sector has no incentive to fund on its own.

Questions to Consider

• Does our own review or prioritization process, a hiring committee, a roadmap review, a grant panel, structurally reward proposals the whole group can agree on, the way NSF panels favored ideas "in the strike zone," at the cost of genuinely risky, high-upside bets that one skeptical voice could block?
• Are we forcing every initiative through the same default timeline, a quarterly cycle, a fixed project length, the way NSF defaulted nearly all grants to 18 months regardless of what the work needed, when some efforts would ship value in weeks and others genuinely need years?
• Is there a screening requirement in our own funding or hiring process, a certification, a scoring category, a compliance checkbox, that has quietly become a de facto gate consuming a large share of resources, the way an estimated quarter of NSF's grant funding became tied to a DEI-related application requirement, without us checking whether it correlates with the outcome we actually want?
• Before starting a new initiative, have we actually verified nobody else, a competitor, another internal team, the market, is already solving this well, the way Kratsios argues federal AI funding should only go where hundreds of billions in private capital isn't already flowing?

Bottom Line

Kratsios's core argument is that the constraint on scientific progress isn't primarily funding levels, NIH's budget more than tripled since 1998 without a proportional rise in breakthroughs, but the review, duration, and allocation mechanisms that decide what gets funded, mechanisms that have quietly drifted toward rewarding consensus and rising budget numbers over genuine risk-taking. The fixes he describes, golden-ticket override authority, duration matched to the work, funding people instead of just proposals, and a strict test for where public money should go, add up to a general framework for redesigning any large evaluation process that has stopped selecting for what it was meant to select for.

Resources Mentioned