To Fund Applied AI, the NSF Is Cutting the Basic Science That Produced AI. Congress Wrote a Rule to Prevent Exactly This.
Attention, transformers, diffusion — every layer the AI boom rests on came out of the kind of no-product-in-sight basic research the NSF has funded for decades. The NSF just cut that research 20-30% to free up $1.5B for X-Labs, an applied-AI-and-quantum initiative that had $0 originally allocated. It's not new money; it's a zero-sum reallocation inside a flat budget — and it appears to breach a congressional directive written specifically to stop NSF from cannibalizing basic science for its tech arm. Eating the seed corn, possibly against the law.
Every layer the current AI boom rests on — the attention mechanism, the transformer, diffusion models — came out of exactly the kind of unglamorous, no-product-in-sight basic research the National Science Foundation has bankrolled for seventy years. The NSF has now decided to cut that research by 20 to 30 percent to free up $1.5 billion for a new initiative chasing applied AI. It is, close to literally, eating the seed corn to plant a faster-growing crop — and it may be breaking a rule Congress wrote specifically to stop it from doing this.
The official framing is "generational breakthrough science," which is the tell, because the new program arrived with no new money. NSF's overall budget fell about 3% this year; the program, called X-Labs, had exactly zero dollars originally allocated. To conjure $1.5 billion — a roughly ten-year envelope rather than an annual check, in fairness — NSF reached inside an $8.8 billion budget and cut hundreds of existing basic-science grants by far more than that 3%, including something on the order of a 30% reduction to the math and physical sciences directorate and a deep cut to biology. The precise dollar figures are reported inconsistently and I won't pin a number on it; the 30% magnitude is the solid part. "Breakthrough" is doing the quiet work of obscuring "funded by cutting other scientists."
X-Labs itself is modeled on DARPA and ARPA-H: milestone-based funding handed to independent teams of researchers and entrepreneurs working outside the conventional academic lab, aimed at AI-driven sensing and imaging and at quantum systems — both administration priorities. As an idea it's defensible; the ARPA model has produced real things. The problem isn't the program. It's the financing, and what the financing reveals about what NSF is becoming.
Here's the governance hook, and it's not subtle. The FY2026 appropriations bill directs that no NSF directorate may take more than a 5% reduction relative to the FY2024 level. Science reports NSF appears to be defying that directive outright. And the context makes it worse: that 5% language was inserted specifically to stop NSF from growing its applied technology arm at the expense of its basic-science directorates — the exact maneuver now underway. Congress saw this coming and wrote a rule against it, and NSF appears to have done it anyway. The legal question of how much latitude NSF has to reallocate internally is genuinely unresolved, so "appears to defy" is the honest verb. But an agency treating an applied-tech priority as more binding than an explicit appropriations instruction is the kind of precedent that outlives the program that set it.
Now the part that should bother an AI reader specifically, because most coverage files this under science-policy process and misses it. The research getting cut — foundational mathematics, computational theory, the basic-science directorates — is the soil the entire field grew out of. Attention, transformers, diffusion: none of those came from a milestone-based applied program with a deliverable date. They came from people doing curiosity-driven work whose payoff was unknowable in advance, which is the whole point of basic research and the precise thing you cannot fund on a DARPA milestone schedule. NSF is optimizing for near-term applied AI output by defunding the long-horizon work that seeds the next paradigm. The cruel irony is self-contained: you cannot schedule a breakthrough, and the budget line being cut is exactly where the unscheduled ones come from.
Let me steelman the other side, because there is one. Flat, peer-reviewed basic-science grants can be conservative — they fund the fundable, not the wild swing — and DARPA's model works precisely because it gives program managers unusual autonomy to take high-variance bets that committee review would kill. If X-Labs genuinely replicates that culture, it could produce things conventional NSF grants never would. The operative word is "if." Whether X-Labs gets real DARPA-style program-manager autonomy or turns out to be a rebranded small-tech grant with a quantum logo is unverified and unknowable yet, and it's the variable that decides whether this disruption buys anything. The bull case requires both that X-Labs delivers genuine breakthroughs and that the cuts spared the foundational work with downstream AI relevance — protein folding, neuromorphic computing, computational math. Neither is established.
What I'd watch: whether Congress convenes oversight or moves a rescission to force compliance with the 5% rule, and whether the first X-Labs cohort produces anything ARPA-like or just looks like grants in a new envelope. Both land over the next few quarters. My read is that this is a bad trade on the timescale that matters. On a two-year horizon, reallocating from diffuse basic research to a focused applied initiative looks efficient. On the ten-to-twenty-year horizon that actually produced transformers from grant-funded curiosity, it's a bet against the only reliable source of the breakthroughs the applied program is trying to chase. What would change my mind is concrete: X-Labs shipping a genuine, autonomy-driven advance and the cuts demonstrably missing AI-relevant foundational work. Absent that, the NSF has decided the surest way to win the AI race is to stop funding the kind of science that started it.