Agrifood AI quantitative panel — Mozilla State of Open Source AI 2026 + Pennells 2025 + EPRS 2023 + Stanford AI Index 2025 + McFadden 2024 + IFPRI GAIA; closing G-356 substantively with what exists, preserving G-366..G-371 where the literature does not carry

Global (panel is cross-economy; agrifood-specific where literature carries; gap-registered where it does not)

Agrifood AI quantitative panel

The substantive empirical anchor for any talk about AI activity in agrifood. Closes G-356 (the substrate-scan’s call for an agrifood-specific quantitative panel) with the literature’s six aggregation currents — and surfaces six specific panel rows (G-366..G-371) where the literature does not yet carry.

Why this unit exists

scans/2026-07-open-source-ai-substrate-v2.md §8.1 surfaced G-356: aggregate quantitative impact of open-source AI in agrifood — the agrifood-specific equivalents of Mozilla’s 3% / 79% / 51% / 4% / 89% figures are not surveyed in published form. This unit addresses G-356 substantively by consolidating what does exist — Pennells 2025 + EPRS 2023 + Stanford AI Index 2025 + McFadden 2024 / GAO 2024 + IFPRI GAIA + Mozilla 2026 — into a single anchor with one panel matrix across four cross-cutting dimensions. It also surfaces six panel rows (G-366..G-371) where the literature does not yet carry, preserving them as named gaps for the next cycle.

The single substantive claim. No single comparable survey aggregates open-source AI in agrifood across sector-position × AI technique × purpose × adoption × deployment × revenue × public-investment. Six distinct aggregation currents exist in the literature; no single panel stitches them into one matrix. This unit stitches them as best the literature allows.

What changed in this scan. The substrate scan (§8.1) named G-356 as the corpus’s substantive missing row — the agrifood-specific quantitative panel that the corpus’s anchor units were citing back to. The fresh-pull scan (scans/2026-07-open-source-ai-agrifood-quantitative-panel.md, July 2026) assembles the panel from six primary sources and surfaces six named gaps within the panel — the cells that the literature does not yet carry. This unit is the panel.

The six aggregation currents

The literature uses six distinct empirical currents to measure activity in AI + agrifood. They are not interchangeable. Each has strengths; each has limits.

Current A — Bibliometric / publication-count (Pennells et al. 2025)

Anchor: Pennells et al. 2025 (Mapping the AI Landscape in Food Science and Engineering, PMC12494660). CSIRO + Leeds + Web of Science Core Collection.

Substantive figures (Pennells 2025, retrieved as primary source):

• 213 peer-reviewed AI-food reviews spanning 2004–2024.
• Bibliometric categorization: food-safety 39 of 213 (18%); process-monitoring 37 of 213 (17%); product-quality 28 of 213 (13%); traceability 23 of 213 (11%); food-security 1 of 213 (0.5%) — substantive gap.
• Country distribution (Pennells 2025 Figure 3): China 35; India 18; Iran 6; USA 5; UK 5; Spain 5 (top six).
• Post-2019 surge — bibliometric activity rises substantially from 2019 onward, consistent with the general-AI capability inflection.

Stance. Bibliometric measures activity in food-AI by peer-reviewed review count. It is a research-temperature reading, not a deployment-scale reading.

Limit. Bibliometric is not deployment. The post-2019 surge is a rising-temperature indicator; it does not measure farms, processors, or retailers actually using AI.

Current B — Adoption-rate / deployment-survey (Stanford AI Index 2025 + Atlanta Fed 2026 + McFadden 2024 / GAO 2024 + EPRS 2023 / Tracxn 2022)

Anchor: Stanford AI Index 2025; Allen / Atlanta Fed 2026; McFadden 2024 / USDA Economic Research Service; GAO-24-105962; EPRS 2023 / Tracxn 2022.

Substantive figures (multi-source, retrieved as primary source):

• Stanford AI Index 2025: AI-firm adoption 78% of organisations; sectoral disaggregation shows agriculture among lowest-AI-adoption sectors (per Federal Reserve / Allen 2026). Public AI investment by country (2024): Canada $2.4B / China $47.5B (semis) / France €109B / India $1.25B / Saudi $100B.
• McFadden 2024 / USDA ERS: 27% of US farms adopted precision-agriculture (broader than AI) as of 2023; adoption substantially higher on large farms, substantially lower on small farms.
• GAO-24-105962: confirms the 27% US precision-ag adoption at the USDA-anchored stratum.
• EPRS 2023 / Tracxn 2022: AI-in-agriculture startups by country — USA 175 / UK 39 / Israel 36 / NL 27 / Brazil 23 / France 19.
• EPRS 2023 §4.1.2: AI-in-agriculture market segmentation — Tier-1 applications ~35%, Tier-2 ~40%, Tier-3 ~25%.
• Allen / Atlanta Fed 2026: AI use by US small businesses — adopted + planned levels by sector; agriculture in lower band.

Stance. Adoption-rate / deployment-survey measures activity by deployment share among farms, firms, or startups. It is a deployment-scale reading, often stratified.

Limit. Adoption-rate conflates “using at all” with “deploying in production.” Mozilla 2026’s 79% developer-uses / 51%-deployed-in-production gap is the canonical version of this conflation in agrifood-adjacent work.

Current C — Market-size aggregator (Mordor / BCC / FMI / Grand View / GMI)

Anchor: Multiple vendor-and-aggregator market-research reports.

Substantive figures (retrieved, flagged as C-321):

• 2024–25 baseline range: $2.43B–$4.7B (2× spread across vendors).
• Forecast range: $8.5B–$77B by 2030–36 (9× spread).
• Forecast CAGR: ~26% (range across vendors).

Stance. Market-size aggregator measures activity by forecasted total addressable market + growth rate. It is a funding-side reading — the consultants’ revenue model assumes a larger market.

Limit (substantive). Substantively unreliable. 9× spread on forecasts; 2× spread on baselines; no open methodology; vendor-aggregator funding bias. C-321 carries this forward: aggregators are scan-level finding, not unit-level adoption claim. The boardroom-friendly “$77B by 2036” headline is evidence that consultants want you to fund their report; not evidence about deployment.

Current D — Public-investment / capacity-build (Stanford + Mozilla + IndiaAI + EUROPA + EuroStack + Canada AI for All + China “AI Plus”)

Anchor: Stanford AI Index 2025 (national AI investments); Mozilla 2026 (12 new national AI strategies launched last year); IndiaAI Mission; EUROPA; EuroStack white paper; Canada AI for All (June 2026); China Five-Year Plan “AI Plus.”

Substantive figures:

• Mozilla 2026: 12 new national AI strategies launched last year; 47 countries restrict foreign processing for critical workloads.
• Canada AI for All: RAII scaling $200M (Budget 2024) → $500M trajectory.
• Stanford AI Index 2025: Canada $2.4B / China $47.5B semis / France €109B / India $1.25B / Saudi $100B (national AI investments 2024; agrifood carveout not separately available).
• EU EUROPA / EuroStack: strategic digital procurement substrate; agrifood-specific clauses under review.
• China “AI Plus”: national strategy layer covering manufacturing, agriculture, and services; agrifood-specific deployment scale not separately reported.
• IndiaAI Mission: $1.25B aggregate India AI investment; agrifood carveout not separately available.

Stance. Public-investment / capacity-build measures activity by public investment in AI as strategic infrastructure. It is a funding-side reading with institutional-capacity implications.

Limit. Aggregates AI investment in general; agrifood-specific carveout is generally not separately reported. G-370 carries this forward as a named gap.

Current E — Open-source / open-weight AI share of activity (Mozilla 2026)

Anchor: Mozilla State of Open Source AI Report (July 14, 2026; Raffi Krikorian CTO; 950+ developers surveyed by SlashData). See units/mozilla-state-of-open-source-ai-2026.md.

Substantive figures (Mozilla 2026, AI ecosystem-wide, not agrifood-specific):

• 3.3% performance gap between open and proprietary frontier models (March 2026).
• 50× cost reduction in three years.
• 33% of real-world AI usage runs on open models; that 33% captures only 4% of AI revenue.
• 89% adoption in China + East Asia; 70% in Western Europe + Israel; 66% in South America; lower in North America; 39% in South Asia cite security/compliance concerns.
• 79% of developers use open models; only 51% have deployed them in production (vs 63% for closed).
• “The real fight has moved beyond the model” — agentic harness (the software layer between people and models) matters more than the model.
• 93% of users approve AI agent requests by default (consent fatigue).

Stance. Open-source share measures activity by open-model share of deployment, revenue, and developer use. It is a structural reading.

Limit (substantive, and load-bearing for this unit). Mozilla’s figures are AI-ecosystem-wide, not agrifood-specific. The agrifood-specific equivalents (capability gap on agrifood tasks; agrifood deployment rate by open vs closed; agrifood revenue capture by open vs closed; agrifood developer use vs production deployment) are not surveyed in published form as of July 2026. This is G-356 substantively, and G-366..G-370 are the named panel-rows where the literature does not carry.

Current F — Multilateral / public-good / open-research (CGIAR + IFPRI GAIA + AU Continental AI Strategy + FAO)

Anchor: units/cgiar-agrillm-ai-global-south.md (CGIAR + AgriLLM + UAE AI71 + Gates); units/agra-alliance-green-revolution-africa.md (AGRA 15M smallholders / 562 seed varieties / 42 policy reforms); IFPRI GAIA project; AU Continental AI Strategy (July 2024); FAO digital-agriculture surveys; CGIAR Open and FAIR Data Assets Policy.

Substantive figures:

• CGIAR: operational since 1971; AgriLLM launched June 2025; chatbot prototype target COP30; deployment in Bihar / Kenya / Mexico at S1–S2 pilot scale.
• AU Continental AI Strategy: adopted July 2024 in Accra; 13 named national AI strategies across African Union member states.
• AGRA: 15 million smallholder farmers trained; AGRA-PASS 562 new seed varieties commercialised; 42 public policy reforms advocated across 13 African countries.
• IFPRI GAIA: responsible-GAI methodological framing (Jones-Garcia 2026 C-H-A-T framework); 11-country blog-experiment (Keenan 2026); cross-Africa deployment via FAO partnership.
• Mozilla Common Voice African-languages: Common Voice 23.0 — 357 hours Spontaneous Speech across 51 languages; substantive African deployment via Maseno University + Africa Next Voices + Africa’s Talking Kiswahili Hackathon.

Stance. Multilateral / public-good / open-research measures activity by multilateral-deployment scale + recipient count. It is a deployment-of-record reading.

Limit. Multilateral deployment is concentrated (CGIAR + AGRA + FAO + WIPO); aggregate scale of smallholder-reach across all multilateral-adjacent programmes is not published.

The panel matrix (sector-position × quantitative dimension)

The substantive output. Each cell gives source → value → flag for vendor-reporting / aggregation / direct measurement / gap-registered.

Sector-position × dimension (where it exists in literature)

AI technique × agrifood deployment evidence

Open-source / open-weight × agrifood-specific quantitative

Country × open-source AI adoption (mapped from Mozilla + Stanford)

Substantive cross-reading. Mozilla 2026’s developer-survey-based open-source adoption rate is region-stratified (Asia 89% vs West ~70%) but does not itself survey agrifood specifically. The Mozilla “Asia at 89%; West behind” finding maps to the Stanford AI Index 2025 optimism finding “China 83% optimism vs Canada 40% vs US 39%.” Neither the adoption rate nor the optimism rate is agrifood-specific.

The field guide’s Africa + China + LAC + Canada + EU cycles have deployment-side numbers (Mila DISA Rwanda, PolArctic Sanikiluaq, DJI 222M tons water saved, Bayer Climate FieldView 250M acres) but these are vendor-reported and not panel-comparable.

The substantive status (§14 of the fresh-pull scan, restated)

Six specific panel-rows where the literature does not yet carry (G-366..G-371)

These are the named cells where the panel framework wants a figure but the literature does not yet provide one. Each carries forward as a corpus gap; each is a candidate research target for the next cycle.

• G-366 — agrifood-AI capability gap × open weights vs proprietary. Mozilla 2026 has 3.3% capability gap for AI ecosystem; the agrifood-specific equivalent (capability gap on agrifood tasks: weed-id, plant disease classification, yield prediction, food-safety imaging) is not surveyed.
• G-367 — agrifood-AI adoption rate by sector-position × farm size. McFadden 2024 gives US precision-ag (27%) stratified by farm size; no equivalent for AI specifically or for processing / distribution / retail / consumption.
• G-368 — agrifood-AI revenue capture by vendor tier (open vs proprietary). Mozilla 2026 has 4% revenue capture for open AI; the agrifood-specific vendor-tier breakdown (OpenAI / Anthropic / Cohere / Mistral agrifood revenue vs agrifood vendor revenue from open-source like Microsoft FarmVibes.AI + OMB) is not surveyed.
• G-369 — Published agrifood-AI deployment-scale metrics by farm size × region. McFadden 2024 gives US precision-ag; the cross-country panel (US / China / Brazil / EU / Africa / India / Canada farm-level adoption rates by farm-size quintile) does not exist as a single published survey.
• G-370 — Public-investment in open-source AI in agrifood specifically. Stanford AI Index 2025 gives Canada $2.4B / China $47.5B semis / France €109B / India $1.25B / Saudi $100B as aggregate national AI investments; the agrifood carveout is not available at the same granularity for any of these.
• G-371 — IFPRI Jones-Garcia 2026 framework for responsible GAI in agricultural extension — quantitative assessment. Jones-Garcia 2026 paper presents the C-H-A-T framework and argues for participatory design; a quantitative panel (number of agricultural extension programmes using GAI by region × by 2027) is not yet specified.

Substantive synthesis (closing G-356 substantively with what exists; preserving G-366..G-371 as gaps)

G-356 is closed for cells where literature carries, and preserved for cells where it does not. This unit does not paper over the gaps; it names them in the panel rows above (G-366..G-371) and in the §14 status table. The substantive state of the agrifood AI quantitative panel as of July 2026 is: bibliometric + adoption-rate + market-size-aggregator + public-investment + open-source-share + multilateral-deployment evidence exists across six distinct aggregation currents; no single panel stitches them into one matrix.

The substantive state of the panel is partially populated, structurally honest, and named where it is not populated. That is the field guide’s discipline: aggregating what the literature carries, surfacing what it does not, refusing to fill the gaps with vendor-aggregator figures or fabricated metrics.

Three substantive findings

1. The six aggregation currents are not interchangeable. Mixing Stanford’s 78% AI-firm adoption with Pennells’s 213-review count with McFadden’s 27% US precision-ag with Mozilla’s 89% East Asia adoption would produce a non-comparable composite — not a panel. Each current measures a different thing at a different level. The panel keeps them separate.

2. The market-size aggregator (Current C) is not load-bearing evidence. It is funding-side evidence; it tracks consultants’ revenue model, not deployment. The 9× forecast spread is a structural signal that the market-size aggregators cannot agree on basics. Worth surfacing this to any audience that arrives expecting “$77B by 2036” headlines as evidence.

3. The Mozilla 33% / 4% asymmetry (Current E) is the load-bearing insight for the open-source + smallholder + multilateral talk (archetype 07). Open-source AI has reached a structural turning point in deployment share — but the value capture has not followed. The Global-South deployment-of-record (CGIAR + AGRA + Mozilla Common Voice African-languages + Ushahidi / ODK / Digital Green + OSSI) is the corpus’s most-substantive evidence that the deployment share is real and operational.

Why this matters for talks

For archetype 01 (vendor sweep): the panel is the empirical anchor for “what’s actually deployed” claims. Aggregator figures are not what the panel carries; Pennells + McFadden + Stanford are.

For archetype 02 (Canadian data-sovereignty) and archetype 03 (Canadian adoption-diagnosis): McFadden’s 27% US precision-ag and the Federal Reserve / Allen 2026 sectoral disaggregation (agriculture among lowest-AI-adoption sectors) are the comparative anchors for Canada’s 1.8% adoption figure; both Canada and the US precision-ag sector are below the federal-AI-firm 78% baseline that audience pre-conceptions may carry. Archetype 03 should surface McFadden explicitly.

For archetype 04 (cooperative alternative) and archetype 06 (regional cluster comparison): the public-investment / capacity-build axis (Current D) is the substrate for the cooperative-state-vendor-cluster comparison. JoinData / WAGRI / AgriStack are the data-substrate leg; the public-investment leg is the layered-in comparison. Archetype 06’s cluster-pattern observation depends on the panel for empirical backing.

For archetype 05 (critical-lens Indigenous): the IDSov-and-open-source overlay (PolArctic + Salmon Vision + SIKU + Mozilla Common Voice African-languages deployment) is the deployment-of-record read for the Indigenous-led agrifood AI work. The aggregate scale (cells G-356-deferred under IDSov) is itself a panel row that the next cycle should target.

For archetype 07 (open-source + smallholder + multilateral): this panel is the empirical spine of the Mozilla turning-point talk. Current E (open-source / open-weight share) + Current F (multilateral / public-good / open-research) together produce the cross-cutting structural case the talk delivers.

Why this matters for the methodology

This unit is the methodology-layer empirical anchor the substrate scan’s G-356 called for. It is not a survey of agrifood AI activity (that would be a vendor-catalog or market-aggregator framing); it is the cross-cutting empirical panel that the corpus’s anchor units cite back to. Every other unit that names a “33%” or “89%” or “27%” figure now cites back to this unit for the panel context.

The substantive state of the panel — partially populated, structurally honest, named where it is not — is the load-bearing methodological discipline. Mozilla 2026’s 33% / 4% / 89% figures are AI-ecosystem-wide, not agrifood-specific. McFadden 2024’s 27% is US precision-ag, not AI specifically. Stanford’s 78% is firm-level, not agrifood-specific. Pennells’s 213 is reviews, not deployment. EPRS’s 175 US startups is company-count, not deployment-scale. Each current measures a different thing at a different level. The panel keeps them separate.