The University Is Not the Missing Laboratory

The American university did not fail to become Bell Labs. It was never designed to be Bell Labs.

That distinction matters because universities are now asked to absorb nearly every function shed by other parts of the R&D system. They are expected to produce basic knowledge, train the workforce, maintain expensive facilities, comply with expanding public rules, generate regional startups, license inventions, repair social inequality, and deliver technologies ready for national missions—all while competing for finite project grants and publishing enough visible novelty to sustain institutional prestige.

The result is not simply underfunding. It is a mismatch between the unit being financed and the capability being demanded.

Begin with what academia does exceptionally well

U.S. higher-education institutions reported $117.7 billion in R&D expenditure in FY2024, up 8.1 percent in current dollars from FY2023. The federal government funded 55 percent; the institutions themselves funded 26 percent; nonprofits, businesses, and state and local governments supplied most of the rest. Academic R&D is structurally different from business R&D: 63 percent was basic research, 27 percent applied research, and 10 percent experimental development.

The scale is concentrated. The 30 largest performers accounted for 42 percent of higher-education R&D, and 28 of those 30 had medical schools. Field composition is concentrated too: health sciences and biological and biomedical sciences together accounted for half of total academic R&D. These facts do not diminish the work. They show that “academia” is not one diversified national laboratory. It is a federation of institutions, fields, funders, and principal investigators with very different cost structures and missions. NCSES documents the full distribution. D

Universities preserve three things that no successor R&D system should weaken:

1. permission to publish knowledge openly;
2. intellectual communities not wholly owned by one product or mission customer;
3. education through participation in unresolved problems.

The claim of this article is not that academia is unproductive or corrupt. It is that these virtues do not, by themselves, supply permanent systems engineers, instrument builders, research software maintainers, manufacturing paths, procurement authority, deployment owners, or long-lived mission memory.

The principal investigator as a temporary firm

Much academic research is organized around a principal investigator who assembles a temporary production unit from grants. The grant may support students, postdoctoral researchers, staff, equipment, travel, and part of the investigator’s salary. A successful group wins another grant before the first expires. A very successful investigator operates several overlapping units and spends increasing time raising, coordinating, and reporting their capital.

This resembles entrepreneurship, and often in its best form. It also has three structural consequences.

First, labor continuity and scientific continuity separate. A paper’s question can persist for a decade while the students and postdocs who know the apparatus turn over every few years. The university retains the publication; it may not retain the working team.

Second, shared technical capability becomes residual. A research software engineer, instrument specialist, data steward, or long-term project manager rarely maps cleanly to one hypothesis. If no project can charge the whole role, every project has an incentive to pay only its marginal share and hope another account sustains the person between awards.

Third, proposal competence becomes a survival trait. Writing a persuasive forecast of discovery is not the same activity as making the discovery. Yet the system must select using proposals because the work does not yet exist. This is unavoidable to a point. It becomes destructive when the forecast consumes the work or when deviation from it threatens renewal.

The university can therefore contain brilliant long-horizon researchers while the operating system beneath them remains short-horizon.

Four clocks, one laboratory

An academic group is governed simultaneously by at least four clocks:

• The grant clock: the period of performance, annual reporting, renewal, and the next proposal.
• The publication clock: the interval in which a result must become visible enough to support the next job, promotion, citation, or grant.
• The training clock: the finite degree or postdoctoral appointment, which properly serves a person’s development rather than an institution’s indefinite staffing need.
• The budget clock: the fiscal period in which direct and shared costs must be allocated, recovered, and audited.

None of these clocks is irrational. The problem is their intersection. A technically important activity can be rejected by all four: too uncertain for the next grant, too infrastructural for a paper, too long for a student, and too cross-cutting to charge directly.

Tenure partially solves the problem for some faculty by extending the employment horizon. It does not automatically fund a team, a cleanroom, a longitudinal dataset, or a deployment organization. Nor does it cover the growing share of research labor performed by people who do not hold tenured positions.

Evidence that incentive design changes the direction of research

The strongest claim would be that short project clocks cause worse science. The evidence does not support that statement universally. Fields differ, project grants fund foundational work, and selection into alternative funding programs makes causal comparison difficult.

There is, however, important evidence that contract design changes exploration. Azoulay, Graff Zivin, and Manso compared Howard Hughes Medical Institute investigators—funded with greater freedom, longer horizons, and more tolerance for early failure—with similarly accomplished NIH-funded scientists. Using matching and difference-in-differences methods, they found that HHMI investigators produced high-impact work at a higher rate and moved toward more novel lines of inquiry. The study concerns selected life scientists, not every discipline, and cannot prove that simply lengthening every grant would reproduce the result. It does show that incentives are not neutral containers around creativity.

Wang, Veugelers, and Stephan approached the problem through bibliometrics. Papers making unusual combinations of prior journals had more variable outcomes, greater probability of very high long-run impact, broader cross-field influence, and delayed recognition. They were also associated with lower-than-expected journal impact factors. Short evaluation windows can therefore select against the right tail of novelty precisely because that tail arrives with more failures and longer delay.

Smaldino and McElreath formalized a darker mechanism. In their population model, research groups with methods that generate more publishable findings can culturally outcompete more rigorous groups even without fraud or conscious bad intent. Selection operates on the visible output, not on the invisible truth-seeking process. The National Academies has likewise treated publication pressure and institutional incentives as system-level research-integrity conditions, not only matters of individual character.

Feynman’s warning about cargo-cult science and Shannon’s warning about a bandwagon meet here. A community can preserve the vocabulary, venues, and outward motions of science while reward selection gradually lowers the information those signals carry about truth.

Overhead: the debate is usually wrong before it begins

“Overhead” is often discussed as if a grant contains a pile of money for science and a university removes an arbitrary percentage before the scientist can work. That description is false.

Federal research awards distinguish direct costs from facilities and administrative costs, usually abbreviated F&A. Direct costs can be assigned to a project with high accuracy: project salary, supplies, travel, and some equipment. F&A costs are shared expenses that support more than one activity and cannot be assigned project by project without disproportionate effort. They include building depreciation, operations and maintenance, utilities, libraries, information systems, financial administration, sponsored-research offices, and parts of departmental and general administration.

As the current Congressional Research Service synthesis explains, the negotiated F&A rate is usually applied not to the entire award but to a modified total direct cost base. Major equipment, portions of subawards, and other categories may be excluded. This produces a common arithmetic error. If the applicable base is $100 and the negotiated rate is 60 percent, F&A is $60; it is 37.5 percent of the resulting $160 total, not 60 percent. If excluded direct costs are present, the share of the total award is lower still. D

The rate is also a bundle. Under the federal framework described in the First Circuit’s 2026 opinion, the administrative component for universities has long been capped at 26 percent of the modified direct-cost base, while the facilities component reflects institution-specific costs. A high negotiated rate is not proof that administrators receive that share of a grant. It can reflect a costly scientific plant. Conversely, a negotiated rate does not prove that every shared dollar is well spent or allocated transparently. D

Available federal evidence rejects both simple camps:

• GAO found that indirect costs represented roughly 16 to 24 percent of NSF’s total annual award funding from FY2000 through FY2016.
• GAO reported that facilities reimbursements can support state-of-the-art research infrastructure and also warned NIH to assess whether long-term indirect-cost growth could crowd out the number of grants.
• GAO found that federal requirements themselves create administrative workload and recommended more standardization, delayed pre-award requirements, and greater flexibility.

Shared costs are real. Administrative accretion is also real. “All overhead is waste” and “every negotiated cost is optimal” are competing evasions.

The 2025 NIH fight exposed the wrong binary

On February 7, 2025, NIH issued guidance replacing separately negotiated university indirect-cost rates with a standard 15 percent rate for existing and new grants going forward. The notice argued that more funding should reach direct scientific work. Universities, medical associations, and states challenged the policy. D

A federal district court permanently enjoined the guidance and vacated it. On January 5, 2026, the U.S. Court of Appeals for the First Circuit affirmed, concluding that the NIH action violated applicable statutory and regulatory constraints. As of this draft, the guidance is not in force. The agency notice and the appellate opinion should be read together; quoting only one side would misstate both the policy and its legal status. D

NIH was not the whole 2025–2026 dispute. A January 2026 Congressional Research Service review records proposed 15 percent policies from NIH, DOE, NSF, and DOD and explains that statutes and FY2026 appropriations constrained their implementation in different ways. It also reports that negotiated university rates commonly range from 30 to 70 percent of the applicable modified direct-cost base—not 30 to 70 percent of the total award. The legal and budget status is agency- and statute-specific; a sentence saying “the federal government capped overhead at 15 percent” would be false. D

The controversy was framed as science versus administration. That is the wrong systems question. A uniform cap does not distinguish an efficient shared facility from duplicative administration, a wet laboratory from a theoretical group, or an institution with accumulated infrastructure from one renting ordinary offices. It can force universities to cross-subsidize federal research from tuition, clinical revenue, philanthropy, or endowment—or to stop doing some research.

The status quo is not therefore vindicated. Negotiated rates are difficult for outsiders and even investigators to interpret. Universities differ in whether recovered F&A returns to the lab, department, school, central administration, or debt service. More grant volume can support more shared infrastructure, but it can also become part of an institutional growth model in which faculty must keep raising external revenue to maintain the system built around raising external revenue.

The useful question is not “What percentage should overhead be?” It is:

Which shared capabilities does the country want to exist, what do they actually cost, who should own them, and which funding stream remains accountable for their performance?

Overhead is the shadow price of projectization

This leads to the article’s most controversial claim.

Much academic overhead is not external to science. It is the accounting shadow cast when a society funds research as temporary projects but expects permanent institutions to make those projects possible.

If the federal government funds a microscope’s experiments one grant at a time, somebody must still pay for the building, calibration, safety system, cybersecurity, procurement, janitorial work, grant accounting, and periods between experiments. Because those costs do not belong cleanly to one hypothesis, they migrate into F&A pools.

Projectization also creates administration endogenously. Sponsors add controls to reduce misuse and improve accountability. Universities hire people and systems to implement those controls. Investigators spend time supplying the required information. The resulting cost justifies larger administrative capacity, while the number and complexity of grants create further coordination work. No malicious administrator is required.

This is a feedback loop, not a conspiracy:

1. uncertain public work is divided into auditable projects;
2. each project creates compliance and coordination requirements;
3. institutions build shared systems to satisfy them;
4. shared systems are recovered through F&A and institutional cross-subsidy;
5. grant volume becomes necessary to sustain the systems;
6. investigators pursue more grants and spend more time in the grant system.

The loop can finance essential infrastructure and still consume too much scientific attention. The defect is not the existence of indirect cost. It is the failure to distinguish three different purchases:

• a research project intended to answer a bounded question;
• institutional capability intended to persist and serve many questions;
• public accountability intended to protect funds, people, data, and safety.

Bundling all three into a percentage on project expenditure makes the debate opaque.

The denominator is wrong: minimize total mission cost, not the rate

The overhead rate is a price-allocation rule, not an efficiency measure. A lower rate can reduce federal reimbursement while increasing the total cost of producing a result if a university closes a shared facility, duplicates compliance inside laboratories, defers maintenance, or forces investigators to assemble infrastructure project by project. A higher rate can also coexist with avoidable layers, weak procurement, or internal cross-subsidy. Neither sign settles efficiency.

The relevant accounting identity is broader:

Total portfolio cost = direct project work + shared scientific capability + required stewardship + transaction and coordination cost + unrecovered institutional contribution.

Negotiated F&A reimburses an allocated portion of some middle terms. It does not necessarily equal them, and it does not make the unrecovered contribution disappear. The policy objective should be to reduce total mission cost per unit of independently validated and reusable capability, subject to safety, fairness, and public accountability. That objective is deliberately harder than minimizing a visible percentage.

The decomposition produces three ledgers that current debate often collapses:

Ledger	Examples	Correct test
Scientific commons	Shared instruments, research computing, data stewardship, libraries, calibration, research software, technical staff	Is the capability used, maintained, accessible, and cheaper or better than project-by-project reconstruction?
Public stewardship	Human- and animal-subject protection, safety, financial controls, export and data rules, conflicts, cybersecurity	Is the control proportional to actual risk, standardized where possible, and effective at preventing harm or misuse?
Transaction tax	Repeated proposals, bespoke forms, duplicative reviews, premature detailed budgets, incompatible agency systems	Does the information change a funding, safety, or oversight decision enough to justify researcher and administrator time?

Cutting the first ledger can destroy science while appearing to reduce administration. Defending the third ledger as “the cost of doing research” can protect needless work. The reform target is not a job category. It is any process whose marginal evidentiary or safety value is lower than the technical attention it consumes.

The three ledgers should be published both in dollars and in time. A sponsor that simplifies a form but adds a new assurance may move cost from a central office to faculty and lab staff while claiming administrative savings. Investigator hours are a real research input even when they never appear in an F&A pool.

The missing residual owner

The academic system has several legitimate principals whose objectives diverge at the end of a project:

• the investigator needs a truthful, important contribution and a viable next program;
• the student or postdoc needs learning, credit, and mobility rather than indefinite custodianship;
• the university needs education, scholarly standing, lawful stewardship, and financial sustainability;
• the sponsor needs knowledge or mission value within an authorized award; and
• a future user needs integration, reliability, support, procurement, and maintenance.

The first four can each satisfy their formal obligation while the fifth receives nothing usable. A paper is accepted, a student graduates, the grant closes, the institution passes audit, and the prototype becomes unmaintained. No bad actor is required. The system lacks a residual owner for capability—an organization still responsible after the countable outputs have been delivered.

Technology-transfer offices, licenses, startups, and industry partnerships sometimes create that owner. They work best when a path to appropriation and a capable adopter already exist. They are weaker for shared research software, standards, negative-result archives, public-interest infrastructure, or a technical option whose user will not be identifiable until later.

This is why “commercialize more” and “fund more basic science” are incomplete opposites. One pushes selected outputs toward a private residual owner; the other expands discovery. Neither automatically preserves the broad technical state between them.

What is actually declining?

The expenditure evidence does not show a collapse of academia. Real research activity, publication, and many scientific capabilities remain enormous. A defensible decline claim must identify a variable and a counterfactual.

The proposed deterioration is in relative incentive coverage: as the system demands more compliance, computation, data stewardship, integration, and transition, the marginal rewards remain concentrated on grants, novel publications, and temporary labor. Maintenance, contradiction, shared engineering, and post-award transition grow in importance faster than their durable funding and career status. This is a hypothesis about the widening gap between demanded functions and funded roles—not proof that today’s professors or universities are worse than an earlier generation.

Three observations make the hypothesis plausible but not conclusive. Longer, failure-tolerant investigator funding is associated with more exploratory work in one important life-science comparison. Unusually novel combinations receive delayed and more variable recognition. Selection for publishable results can degrade methods in a formal population model without individual misconduct. None establishes a national time trend in scientific quality. Together they show why counting more grants and papers cannot falsify an incentive-coverage gap.

The most important counterexample is the durable university center that does maintain expert staff, shared instruments, long collaborations, and a path to use. Such centers show that academia can contain the missing layer. The question is what finances it: endowment, philanthropy, medical revenue, state support, a center grant, industry partnership, a national facility, unusual leadership, or repeated cross-subsidy. If the answer is an exceptional local settlement, it is evidence for possibility—not evidence that the ordinary project system supplies the function.

AI makes the academic incentive problem nonlinear

Generative AI lowers the cost of proposals, literature summaries, preliminary code, figures, reviews, and manuscripts. That can return time to science. Under unchanged selection rules it can also increase submission volume until scarce reviewers, mentors, instruments, and validation teams become more overloaded than before.

This is a Jevons-like institutional effect, not a claim about energy consumption: when the cost of producing a submission falls, total demand for evaluation may rise enough that the system spends more attention on selection. The equilibrium benefit of AI therefore depends on whether agencies and journals cap or restructure the queue, not only on how many hours one applicant saves.

Academic evaluation should move away from expensive form and toward evidence-bearing state:

• reusable artifacts and complete provenance rather than manuscript polish alone;
• calibrated uncertainty and explicit abstention rather than confident coverage of every criterion;
• contribution records that credit maintenance, replication, data, software, mentorship, and correction;
• triaged review in which short proposals clear a high threshold before full institutional paperwork; and
• portfolio-level limits that prevent cheap generation from turning every plausible idea into an application competing for human review.

AI should also audit the burden it creates. Agencies can log which requested fields affect decisions, detect duplicative assurances, prepopulate stable institutional information, and compare predicted risks with actual findings. A requirement that never changes a decision and rarely catches a failure should face a sunset review.

Why more grants cannot supply the missing systems layer

Suppose every academic project were fully funded at its correct direct and indirect cost. Four functions would still be underprovided.

Long-lived technical labor. Research software engineers, instrumentation experts, data curators, and systems integrators create value across projects. Their careers should not depend on whether one principal investigator has a charge code this quarter.

Negative-result memory. Journals, promotion systems, and grant reports favor successful novelty. The institution needs an internal memory of failed hypotheses, unusable code paths, calibration problems, and boundary conditions. AI trained only on published successes will amplify the same survivorship bias.

Integration authority. A university technology-transfer office can license an invention. It usually cannot compel a product organization, standards body, procurement officer, manufacturer, regulator, and maintainer to follow one technical roadmap.

Mission persistence. Departments and disciplines preserve knowledge communities. They do not necessarily own a public technical mission whose success is evaluated ten years later in the field.

These are not arguments for turning universities into contractors. They are arguments for connecting universities to institutions that are explicitly funded to retain the missing layers.

The strongest counterargument: the university produced the great work

Turing wrote the 1936 computability paper at Cambridge. Universities gave generations of scientists unusual freedom. Tenure, endowed chairs, institutes, sabbaticals, shared facilities, and long collaborations can sustain questions far beyond a grant. Modern academic biology, physics, mathematics, and computer science supply obvious counterexamples to any claim that universities cannot think long. D

The answer is to narrow the proposition. Universities can and do house long thought. The system does not reliably provide the complete path from question to maintained capability, and access to its long-horizon mechanisms is uneven. An endowed theorist, a medical-school laboratory, a soft-money research group, a public-university engineer, and an adjunct instructor do not inhabit the same institution in any economically meaningful sense.

A second counterargument is that university turnover is a feature: students disseminate knowledge by leaving, and competition prevents stagnation. Correct. The problem is not mobility. It is mobility without a funded memory and technical core.

A third is that administrative growth reflects real obligations—human-subject protections, export controls, biosafety, data security, conflicts management, and stewardship of public money. Also correct. The reform target should be duplicated, premature, low-risk, and nonstandard requirements, not accountability itself.

What would falsify this account?

The argument should be weakened if longitudinal, field-normalized evidence shows that ordinary project funding—not exceptional center, endowment, or cross-subsidy arrangements—reliably produces all of the following:

• stable technical careers and team continuity across award boundaries;
• falling investigator time spent on applications and reporting for a comparable risk level;
• transparent shared-cost categories whose growth tracks instrument, safety, data, or mission requirements rather than grant volume itself;
• routine reuse of research software, datasets, facilities, protocols, and negative-result memory;
• high rates of independent reproduction and correction without career penalties;
• accountable transition owners for work that requires integration or maintenance; and
• no systematic disadvantage for long-delay, high-variance, maintenance, replication, or cross-disciplinary contributions.

The overhead diagnosis should also fail locally. If a university’s shared costs are transparent, benchmarked, demonstrably used, and cheaper than reasonable alternatives—and if administrative time falls after sponsor requirements are standardized—then “bloat” is not an adequate account of that institution. Conversely, if a flat cap preserves capability and reduces total mission cost without shifting burden into hidden subsidies or technical labor, the case for negotiated allocation weakens.

A different funding architecture

The remedy is not one larger grant. It is an explicit separation of functions.

1. Fund projects, capabilities, and accountability separately

Project awards should buy bounded research. Multi-institution capability awards should maintain shared instruments, research software, data, technical staff, and negative-result archives. Sponsors should budget the compliance regime they require and standardize it across agencies where the underlying risk is the same.

2. Add long, reviewable people-and-team funding

Expand funding that gives selected investigators and stable teams seven-to-ten-year horizons, permission to change direction, and evaluation against a portfolio rather than annual linear progress. Early failure should be tolerable; concealment and uncorrected methodological weakness should not be.

3. Replace opaque F&A arguments with a capability ledger

For major research institutions, publish comparable categories showing:

• facilities operation and depreciation;
• research-computing and data infrastructure;
• safety, security, and regulatory compliance;
• sponsored-project administration;
• general and departmental administration;
• institution-funded research and the source of that subsidy;
• the internal distribution of recovered F&A.

This should not expose sensitive salaries or force false field comparisons. It should allow a sponsor and researcher to see what permanent capability an indirect dollar maintains.

4. Reduce proposal waste before reducing scientific support

Use short preproposals, reusable institutional information, common federal forms, later submission of detailed budgets, longer renewals, and experiments with lotteries among proposals that clear a high merit threshold. The objective is not to remove judgment. It is to stop demanding full transaction costs from applicants who have not passed an initial screen.

5. Create durable technical careers

Fund research engineers, instrument scientists, data stewards, reproducibility specialists, and transition leads as first-class professions with promotion paths. Their value should be evaluated across a portfolio of projects and over maintenance horizons, not only by first-author papers.

6. Reward contradiction, maintenance, and transition

Create dedicated budgets and prestigious roles for replication, red teaming, benchmark repair, software maintenance, dataset stewardship, standards work, and negative-result synthesis. These are the error-correcting codes of the research system.

7. Keep universities in a larger institutional network

Universities should remain sources of open inquiry and talent. FFRDCs are formally designed to supply sponsor continuity, access, independence, systems engineering, and trusted analysis. Mission laboratories can maintain facilities and long-lived teams. Companies and startups can supply product integration, manufacturing, and users. The policy task is to fund the interfaces and retained state, not pretend one node can perform the whole graph.

Five different verdicts

• Scientific success: Universities remain the country’s largest institutional home for basic research and produce knowledge that no firm can fully appropriate.
• Technical success: They build major instruments, methods, software, and prototypes, with large variation by field and institution.
• Transition success: Licensing and spinouts work for some technologies, but the system lacks a universal owner for integration, procurement, deployment, and maintenance.
• Institutional success: Departments and tenure preserve some memory; project labor, proposal churn, and fragile technical roles lose other critical state.
• Public-value success: Academic research creates immense spillovers and trains the national workforce, while its full costs and downstream benefits remain poorly aligned with any one payer.

Academia is not the weak link in an otherwise complete R&D chain. It is an overloaded link being asked to impersonate the chain.

The constructive conclusion is not to give universities less responsibility for discovery. It is to stop using grant volume and an opaque overhead percentage as substitutes for an institutional design. Fund the project. Fund the capability. Fund the accountability. Then connect the university to an organization that is actually responsible for what happens after the paper.