Research map / Large language models / AI and society
Resource Democratization: Is Compute the Binding Constraint on AI Research?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Survey of over 500 U.S. AI researchers (academic and industry) finding that talent and data limit research plans more than compute, with little academia-industry difference in compute use, though heavy compute users want more.
Problem
Whether compute access is the binding constraint on AI research, and how compute resource stratification between academia and industry affects researchers, had not been directly studied by asking researchers themselves.
Approach
Surveys AI researchers directly about project inputs, compute use, research plan changes, and views on a National AI Research Resource, comparing academic and industry subgroups and compute-use levels rather than relying on model-cost estimates or practitioner interviews.
Key results
- 90% of respondents rated specialized knowledge, talent, or skills as very or extremely important for their most significant project vs 52% for large amounts of compute.
- Researchers reject and abandon projects more often due to data or researcher availability than insufficient compute (all p < .001 for compute vs data).
- Academics report spending significantly less on compute than industry, but no significant difference in GPU-hours (p = .200); higher compute users express significantly more concern about future compute access.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Prioritize talent rather than large amounts of compute for project success | AI researchers’ most significant projects · Most significant project success | very or extremely important (%): 90% vs 52% | Mixed |
| Compare insufficient data with insufficient compute as a reason to change project plans | AI researchers · Rejecting, revising, and abandoning projects | frequency of project-plan changes: all p < .001 | Mixed |
| Compare researcher-availability limitations with insufficient compute as a reason to change project plans | AI researchers · Rejecting and abandoning projects | frequency of project-plan changes: both p < .001 | Mixed |
| Shift expectations from compute toward better algorithms as drivers of AI progress | AI researchers’ expectations for the next decade · AI progress over the next decade | strong agreement (%): More compute: 59% -> 40%; Better algorithms: 31% -> 53% | Mixed |
| Compare academic and industry researchers’ GPU-hour use | AI researchers in academia versus industry · Most compute-intensive project in the preceding five years | GPU-hours: p = .200 | No clear effect |
| Compare academic and industry researchers’ compute spending | AI researchers in academia versus industry · Most compute-intensive project in the preceding five years | compute spending: p < .001 | Mixed |
| Compare academic and industry researchers’ concern about future compute access | AI researchers in academia versus industry · Future compute access | concern | No clear effect |
| Use more compute for the most compute-intensive project | AI researchers · Concern about sufficient compute for future research | future compute concern | Helped |
Surprising findings
- Academics report less spending on compute than industry yet similar GPU-hour usage, and compute concern does not differ significantly by sector.
- 43% of respondents report never rejecting a project due to insufficient compute.
Limitations
- Scope limited to U.S. researchers and sample sizes small for some subgroups (e.g., high compute users, startup employees).
- Industry sample drawn from LinkedIn and AI startups list, possibly not representative; response rate modest (410 complete of ~27k+3.9k invited).
Experimental setup
Datasets: survey responses from AI researchers (410 complete, 123 partial) Web of Science author records from 20 AI venues 2016-2021 Revelio Labs LinkedIn data
Metrics: Likert-scale mean ratings percent of respondents chi-squared test of independence Mann-Whitney U test with Bonferroni correction
Builds on: National AI Research Resource proposal
Closest papers by meaning
- AI Progress Should Be Measured by Capability-Per-Resource, Not Scale Alone: A Framework for Gradient-Guided Resource Allocation in LLMs (NeurIPS 2025 · similarity 0.84)
- Position: AI Scaling: From Up to Down and Out (ICML 2025 · similarity 0.84)
- Increased Compute Efficiency and the Diffusion of AI Capabilities (AAAI 2025 · similarity 0.84)
- Consent in Crisis: The Rapid Decline of the AI Data Commons (NeurIPS 2024 · similarity 0.82)
- Rigor in AI: Doing Rigorous AI Work Requires a Broader, Responsible AI-Informed Conception of Rigor (NeurIPS 2025 · similarity 0.82)
- The Unreasonable Effectiveness of Open Science in AI: A Replication Study (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.