Research opportunity · Concentration, Stack & Sovereignty
Does public research compute change what academic AI researchers can do without industry?
Question
Among US academic teams that received National AI Research Resource (NAIRR) pilot allocations, compared with similar teams that did not, does access change reliance on proprietary model APIs, industry co-authorship and the share of compute-intensive work, and how much of the access delivered depends on private-sector in-kind resources?
Causal mechanism not isolatedOne PhD chapter or paperLow costPartly addressed
Why it matters
Public research compute is the main policy answer to academic dependence on a few compute owners. Its scale is now reported (projects supported, estimated capacity), but whether it shifts who academics depend on, or simply routes the same dependence through a public front door, has not been measured. The answer bears on whether "public compute" is a sovereignty lever or a subsidy channel.
The debate lists this among what we still do not know: how much public research compute changes bargaining power
Evidence that motivates it
- NSF reported that NAIRR had supported over 800 research projects by 1 September 2026, while the programme relies materially on private-sector in-kind resources.
NAIRR has supported over 800 projects and relies materially on private-sector in-kind resources; the claim explicitly does not let its dollar scale stand in for bargaining power, which is what this study would measure more directly. - Epoch AI estimated that five U.S. hyperscalers owned about 71% of cumulative global AI compute in Q4 2025, up from about 63% in Q1 2024, measured in H100-equivalents.
Five US hyperscalers owned about 71% of cumulative global AI compute (H100-equivalents) in Q4 2025, so the comparison point for public compute is a concentrated ownership base; the claim does not speak to research access, which is the gap here. - Concentration claims in AI must identify the stack layer, market definition, metric, geography and period; concentration at one layer does not establish concentration or power at another.
Power cannot be inferred from concentration alone; this design asks for a measured downstream outcome (research dependence) rather than a share.
A possible design
Download the public NAIRR pilot allocations list and code each award by resource type and provider (public supercomputer versus private in-kind cloud, model or platform access). Link awardee PIs to OpenAlex and build a matched comparison group of US academic AI researchers with similar field, career stage, institution type and prior output who did not receive allocations. Using a difference-in-differences design over 2022–2027 publications, compare: (a) the share of papers that report using proprietary model APIs versus self-run open-weight models (coded from methods sections with a validated codebook and a human-checked sample), (b) industry co-authorship and industry-affiliated funding acknowledgements, and (c) the share of papers reporting large training or fine-tuning runs. Add 20–30 semi-structured interviews with awardees about what they could and could not do, and on what terms. Report results separately for public-hardware and private in-kind allocations, since those may move dependence in opposite directions.
Data needed
NAIRR allocations CSV (public, from nairrpilot.org); OpenAlex and full-text access for methods sections (arXiv and open-access papers suffice for a first pass); a coding scheme for model use; ethics approval for interviews.
Feasibility
One PhD chapter or paper. Low: open bibliometric data and modest compute for text coding; interview transcription in the low thousands of US dollars. Publication lags mean 2024–2025 awards will show only early effects; selection into NAIRR is not random, so matching quality and pre-trends must be shown. Bargaining power is measured through proxies (dependence on proprietary access, industry ties), and the write-up must say so.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 12 October 2026, checked again on 13 October 2026).
No outcome evaluation of NAIRR awardees against a comparison group was found, and NSF's 2026 progress reporting is descriptive (teams, students, partners) without outcomes or an in-kind share. Trisovic and Sivaloganathan (2026) show, across 104,226 papers, that public funding is associated with more open-weight use and industry cloud credits with less, but this is a cross-sectional association using generic funding acknowledgements, not an evaluation of a public compute programme. Still open: whether NAIRR allocations specifically change awardees' reliance on proprietary APIs, industry co-authorship and compute-intensive work relative to matched non-awardees, and whether public-hardware and private in-kind allocations move dependence in different directions.
How strong this evidence is
- Who searched and checked: Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert; an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 2 of the 8 nearest works below were judged from an abstract, summary or search listing only; each is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web search on the date shown; Google Scholar, SSRN and paywalled journal pages were only partly reachable, and NSF's two-year progress update itself was not opened. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: Scholarly APIs blocked, so no citation-graph forward search; forward search relied on web queries. Read only abstracts or summaries of Trisovic and Sivaloganathan and Hao et al., not full texts, so whether they mention NAIRR in the body is unconfirmed. NSF's linked 2-year progress PDF and the NAIRR pilot project highlights pages were not read. Web search budget for the session ran out before RO-LD11-06 was finished.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Miller & Gelles, "The NAIRR Pilot: Estimating Compute," Center for Security and Emerging Technology (8 May 2024, updated 19 Aug 2024)
Estimates the pilot's first six resources at about 3.77 exaFLOPS (roughly 5,000 H100s) and compares this with industry scale; measures inputs only, not research outcomes or dependence. - Besiroglu, Bergerson, Michael, Heim, Luo & Thompson, "The Compute Divide in Machine Learning: A Threat to Academic Contribution and Scrutiny?," arXiv:2401.02452 (Jan 2024)
Documents the industry–academia compute gap, fewer academic-only teams on compute-intensive topics and academic reliance on industry-built models, and proposes national compute as a remedy; does not evaluate any public compute programme's effect. - Wang, Fox, von Laszewski, Furlani, Gallo, Whitson & DeLeon, "Evaluating the scientific impact of XSEDE," PEARC 2018 (ACM), doi:10.1145/3219104.3219124 Judged from an abstract, summary or search listing only
Bibliometric peer comparison of papers using a national compute programme versus same-field and same-issue papers (citation impact); a method template, but it measures citations, not dependence on industry. Abstract read only. - Lee, Puerto, van Aken, Arase, Forde, Derczynski, Rücklé, Gurevych, Schwartz, Strubell & Dodge, "Surveying (Dis)Parities and Concerns of Compute Hungry NLP Research," arXiv:2306.16900 (2023) Judged from an abstract, summary or search listing only
Survey of 312 NLP researchers on compute equity by career stage and academic versus industry affiliation; cross-sectional, no public compute intervention. Abstract read only. - US National Science Foundation, "FY 2026 Annual Evaluation Plan" (fiscal year 2026)
Plans an evaluation of the NAIRR pilot focused on either lessons learned in one priority area or early outcomes of NAIRR-supported research, using interviews, document review and descriptive analysis, subject to available resources; no comparison group and no measure of dependence on industry or proprietary models. - Trisovic & Sivaloganathan, "Open Science, Closed Models: How Funding Shapes AI in Science," arXiv:2609.25347 (Sept 2026)
Closest new work. Codes 104,226 papers (2018 to 2025) for model extension, off-the-shelf use, open-weight versus API-only use, and links these to funding acknowledgements, industry cloud credits and industry co-authorship. Reports (per abstract) that public funding goes with more off-the-shelf and open-weight use, private funding with more model extension, and industry cloud credits with less open-weight use. It is associational, uses generic public funding rather than any specific compute programme, and (from the abstract) does not name NAIRR; it has no matched comparison or before/after design and does not separate public hardware from private in-kind access. Its coding scheme is directly reusable for outcome (a). - Hao, Huang, Zhang, Zhao, Liang, Liang, Zhao, Sun, Kalantari, Zhang & Ghassemi, "The Role of Computing Resources in Publishing Foundation Model Research," arXiv:2510.13621 (Oct 2025)
Reviews 6,517 foundation-model papers (2022 to 2024) and surveys 229 first authors; finds compute correlates with national funding and citations but not strongly with academic versus industry setting. Recommends shared, affordable compute but does not evaluate NAIRR or any public compute programme, and does not measure dependence on proprietary APIs or industry partners. - US National Science Foundation, "NAIRR at 2 years: Advancing American artificial intelligence innovation and leadership" (19 Mar 2026, updated 30 Jun 2026)
Official progress update: more than 600 research teams and 6,000 students supported, 28 private-sector contributors. Reports no comparison group, no publication outcomes and no in-kind contribution share; asserts access enabled otherwise difficult research without data. Confirms the outcome evaluation proposed here has not been published by NSF. A linked 2-year progress PDF was not read.
Literature check
Question stands; related work found. No study evaluates NAIRR pilot awardees against a comparison group, and NSF's own public reporting (NAIRR at 2 years, March 2026, updated June 2026) gives counts of teams and students but no outcome evaluation or in-kind share. However, Trisovic and Sivaloganathan (arXiv:2609.25347, Sept 2026) now link funding acknowledgements in 104,226 papers to whether models are extended, used off the shelf, open-weight or API-only, and find public funding associated with more open-weight use and industry cloud credits with less. That covers outcome (a) as a cross-sectional association, so the verdict moves from OPEN to PARTLY_ADDRESSED, but the NAIRR-specific matched design and the public-versus-in-kind split remain open.
Searched: web (extended); arxiv.org; ssrn.com; osf.io; nber.org; dl.acm.org; sciencedirect.com; academic.oup.com; journals.sagepub.com; tandfonline.com; nsf.gov; nature.com; science.org. Checked by Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert.
The screen before this check (verdict changed)
The search found no study that does the core design (first-pass screen of 12 October 2026).
No outcome evaluation of NAIRR awardees against a comparison group was found. Existing work estimates NAIRR's input capacity (CSET), documents the academic–industry compute gap in general (Besiroglu et al.; Lee et al.), or evaluates older public compute programmes by citation impact (Wang et al. on XSEDE). None measures whether public compute changes academic dependence on proprietary models or industry partners, or separates public hardware from private in-kind access. NSF's FY 2026 evaluation plan proposes a NAIRR pilot evaluation of lessons learned or early outcomes using interviews, document review and descriptive analysis, subject to resources; as described, it has no comparison group and does not measure dependence on industry.
Search queries used
National AI Research Resource NAIRR pilot evaluation researcher outcomes compute accesspublic research compute academic dependence on industry cloud bargaining power studycompute divide academia industry survey of academic researchers GPU access 2025NAIRR 2026 evaluation in-kind contributions private sector dependency academic researchers studycausal effect of supercomputer allocation on research productivity regression discontinuity compute grants AI researchXSEDE OR ACCESS allocation impact on publications citations researchers study matched comparisonNAIRR pilot evaluation outcomes awardees study impact academic researchers 2026
Literature check, 13 October 2026:
NAIRR pilot awardees impact publications proprietary models industry collaboration analysisacademic researchers compute access industry dependence foundation models (arxiv.org, osf.io, ssrn.com)effect of cloud computing credits grants on academic research output difference-in-differences"compute divide" academia industry 2026 study de-democratization AI research industry coauthorshipnational research cloud public compute provision academic AI researchers evaluation (dl.acm.org, sciencedirect.com, nber.org, academic.oup.com, journals.sagepub.com, tandfonline.com, nature.com, science.org)"Surveying (Dis)Parities and Concerns of Compute Hungry NLP Research" follow-up 2025 survey academic GPU access open-weight API relianceNAIRR pilot report 2026 outcomes "in-kind" partners share of resources allocations evaluation"The Compute Divide in Machine Learning" Besiroglu cited by 2026 academic compute provision policy evidence
Provenance
RO-LD11-01 v2. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-11 debate on 2026-10-12. Checked by an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.
Changes
- v2, 13 October 2026: Literature check added: a second, deeper search for existing work, with its queries, new nearest work and limits. The verdict changed from "No study found doing the core design" to "Related work exists" because related work turned up; the question stands. See v1 as published.