AI supply is expanding unevenly from research to code
A 19-month longitudinal study compares scientific output with visible open-source implementation to measure where AI technology supply is accelerating. Product Hunt is retained separately as a commercial-visibility check.
AI supply is moving up the stack: from models and tools toward agents, automation and trust infrastructure
AI technology supply is broadening, with pronounced sector divergence. The research-and-code supply index changed +40.4% from Q1 2025 to May–July 2026. Agents & automation gained the most representation in the broad AI cohort (+20.1 pp), while models & infrastructure lost the most (-17.1 pp). Product Hunt is treated separately as a commercial-visibility check, not as a third equal market series.
01
Market snapshot
Comparable measurements from independent market surfaces.
205,365
AI-category papers
Complete monthly arXiv counts across seven computer-science categories.
Q1 2025 compared with the latest three complete months; Product Hunt is excluded.
02
Research and open-source supply
The core index shows arXiv and GitHub only. Each series is indexed within its own source; the January–March 2025 average equals 100.
arXiv papersGitHub repositoriesQ1 2025 = 100
03
How sector representation is shifting
Percentage-point change in each overlapping field's representation within equal-weighted arXiv and GitHub cohorts. Sources with fewer than 30 matching items are excluded from that field's comparison. Product Hunt is shown separately below.
A separate, three-month-smoothed visibility proxy for how AI is packaged into ranked launches. It is not included in the supply index or category shares.
Structural findings that a single aggregate number would hide.
+20.1 pp
Test 1 · Category share is shifting up the stack
Agents & automation gained the most research-and-code supply share, moving from 20.6% in Q1 2025 to 40.7% in May–July 2026. Models & infrastructure moved from 51.3% to 34.2% (-17.1 pp). These are equal-weighted shares inside arXiv and GitHub, not shares of raw combined counts.
11m → 12m
Test 2 · Commercial visibility has no stable lead-lag yet
The numerical maxima place GitHub 11 months after arXiv (r=0.6) and Product Hunt 12 months after GitHub (r=0.5), but they rely on only 7/6 paired monthly changes and sit near the edge of the tested 0–12 month range. The current history is too short to distinguish a real translation lag from a small-sample correlation, so Product Hunt remains a commercial-visibility check rather than a causal input.
83.6% overlap
Test 3 · The agent boom includes a measurable category migration
A deterministic random sample of 60 January–June 2025 Developer Tools and 60 May–July 2026 Agents shows 83.6% overlap across coding, quality, workflow, data, computer-use, and deployment functions. The result supports a mixed interpretation: agents add new autonomous supply, while capabilities previously sold as developer tools and assistants are also being repackaged into autonomous systems.
