W. The protein and feed balance model

What this appendix contains. The documentation of the model behind Protein balance 2050: purpose and boundary, the published projections and targets it uses, each calculation step with its parameters, the four scenarios, results by scenario and year, the sensitivity test, sanity checks, a reality check against other countries, and how to rerun and extend it.

How to use it. W.1 and W.4 say what the scenarios are; W.5 holds every result; W.6 to W.8 help judge how far to trust them; W.9 explains how to rerun the model with your own assumptions. Every value for 2030 to 2050 is our estimate (our calculation, tools/balance_model.py) VN-directThấp ước tính. The 2025 values are calibrated to published data.

---

#W.1 Purpose and boundary

The model answers one supply-side question: how much feed protein, maize and farmland abroad Vietnam would need to 2050 under stated assumptions, and how much a feed-efficiency programme or an alternative-protein build-out would change that. We built it because no published projection for Vietnam goes beyond 2035 (W.2.3).

The model scenarios (S-BASE, S-HIGH, S-EFF, S-ALT) are distinct from the 2035 scenarios of Outlook and the 2050 worlds of Scenarios 2050, which map onto these ranges without new model runs.

#W.2 Published projections and targets

#W.2.1 OECD-FAO Agricultural Outlook 2026-2035

This is the only published model we found with Vietnamese meat, feed and protein-meal projections, and it stops at 2035. The report PDF was blocked to our fetch tool, but the data are open through the OECD SDMX API. All rows: ¹ VN-directTrung bình dự phóng (published model, baseline).

• Comparators, 2035. Pig, poultry and beef food use per person: Viet Nam 61.9 kg, Korea 59.2 kg, China 48.5 kg, Malaysia 46.1 kg ¹ VN-directTrung bình dự phóng. The Thai poultry value (3.4 kg) looks implausible and is not used.
• Caveats. Values were read through a summarising fetch tool. The annual series are smooth, but two commodity labels were swapped in one summary and we reassigned them by magnitude. The poultry base looks like live weight labelled as carcass (W.9.4).
• Recipe. https://sdmx.oecd.org/public/rest/data/OECD.TAD.ATM,DSD_AGR@DF_OUTLOOK_2026_2035,1.1/VNM.A.<commodity>.<measure>..?startPeriod=2024&endPeriod=2035. Commodities: CPC_EX_PK (pigmeat), CPC_EX_PT (poultry), CPC_EX_BV (beef), CPC_EX_AC (aquaculture), CPC_04 (fish), CPC_21920 (protein meal), CPC_21233 (fishmeal), CPC_0112 (maize), CPC_0141 (soybean), CPC_0221 (milk). Measures: QP (production), QC (consumption), IM (imports), EX (exports), FE (feed), CR (crush), FO_PC (food per person), YLD (yield), joined with "+". Replace VNM with KOR, CHN, ARG, BRA or USA for comparators and origin yields, and the dataflow name for a new edition.

#W.2.2 Other published numbers used

Vietnamese household-demand studies find rising meat budget shares and a shift to poultry and beef, but we extracted no projected numbers ^16,17 VN-directTrung bình. All 86 rows are in balance_published_projections.csv.

#W.2.3 What we did not find

We found no FAO, IFPRI, GLOBIOM or peer-reviewed projection of Vietnamese meat, feed or feed protein to 2050 ^1,6,18 VN-directCao. OECD-FAO stops at 2035, USDA's country tables to 2035 could not be read, and official volume targets stop at 2030 ^8,10 VN-directCao. Our 2040 and 2050 values are extensions.

#W.3 Model structure, step by step

Seven steps run for each scenario and year: demand, production, compound feed, protein meals and maize, land abroad, S-ALT substitution, emissions.

Parameter lookup. Each row of balance_assumptions.csv has a parameter, a scenario tag (ALL, S-BASE, S-HIGH, S-EFF, S-ALT, SENS-LOW, SENS-HIGH) and a year (or ALL). The script tries the sensitivity override, then the scenario, then S-BASE, then ALL; between listed years it interpolates linearly. Reported years: 2025, 2030, 2035, 2040, 2050.

#W.3.1 Step 1: demand (macro input)

2025 consumption per person (kg cwe) = 2025 live-weight output x dressing yield / SSR / population.

After 2035 the taper is +0.2% a year for pork, +1.0% (to 2040) then +0.5% for poultry, and +1.0% for beef. Other inputs:

• Population: UN WPP medium variant in every scenario and year, including 2025, so growth rates never mix UN and GSO series ⁷ VN-directTrung bình.
• Eggs: 210.6 per person in 2025, rising to 235 (2030) and 275 (2050); egg weight 0.055 kg; no published projection found ¹⁹ VN-directThấp ước tính.
• Milk: 1,300 kt in 2025, OECD-FAO growth to 2035, then +3% and +2% a year to 2050 ¹ VN-directThấp ước tính.
• S-HIGH income multiplier: (GDP ratio of S-HIGH to S-BASE) raised to an elasticity falling from 0.30 to 0.20; 1.06 (2030), 1.11 (2040), 1.15 (2050) ^13,6 VN-directThấp ước tính.

#W.3.2 Step 2: production

Meat (cwe) = demand per person x population x (1 minus the S-ALT food share) x SSR; live weight = cwe / dressing yield. Aquaculture = 2025 NSO output by group x an index. 2025 output in kt, with our fixed export share in brackets: pangasius 1,938.8 (0.90), whiteleg shrimp 994.4 (0.75), other shrimp and crustaceans 387.3 (0.60), other fed fish such as tilapia, snakehead, carps and marine fish 2,096.2 (0.10), and unfed molluscs, seaweed and others 700.1 (0.30); total 6,116.8 kt ²¹ VN-directCao. Exports come to 3.14 Mt (51%), consistent with OECD-FAO fish exports of 4.63 Mt, which also include capture ¹ VN-directThấp.

The S-BASE aquaculture index is 1.076 (2030), 1.200 (2035) and 1.393 (2050): OECD-FAO growth to 2035, then +1% a year VN-directThấp ước tính. S-HIGH follows the 7.0 Mt target for 2030 ¹¹, then +1.5% and +1.0% a year ước tính.

Appendix Z paths. AQ-BASE in Z. Aquafeed and feedstocks 2050 reuses the S-BASE index but lets marine finfish grow faster, reaching 8.86 Mt in 2050 against 8.52 Mt here (aquafeed 9.7 against 9.2 Mt). AQ-LOW (7.42 Mt) matches the low aquaculture sensitivity; AQ-HIGH (11.27 Mt) lies above S-HIGH (8.97 Mt). Because aquaculture growth after 2035 moves 2050 SBM need by only about 1 Mt (W.6), the national result is unchanged VN-directThấp ước tính.

#W.3.3 Step 3: compound feed by species group

Feed = output x FCR x compound-feed share x (1 minus the annual FCR gain) raised to the years since 2025. One scale factor sets 2025 aquafeed to the chosen base. All rows VN-directThấp; paths ước tính.

#W.3.4 Step 4: protein meals and maize

#W.3.5 Step 5: land abroad

Soybean land = SBM x 0.80 (mass allocation; economic about 0.65, gross 1.0) / (bean yield x 0.78). Bean yield 3.16 t per ha in 2025 (OECD-FAO yields for Argentina 2.91, Brazil 3.62 and the United States 3.56, weighted by origin), +0.6% a year. Maize land = imports / 7.28 t per ha (Argentina 7.69, Brazil 6.11, United States 11.68, weighted by 2025 import origin), +0.3% a year ^1,29,28 generalTrung bình dự phóng. Worked 2025 example: 7.2 x 0.80 / (3.16 x 0.78) = 2.34 M ha of soy plus 11.91 / 7.28 = 1.64 M ha of maize, 3.97 M ha in total, about five times Vietnam's maize area of 0.81 M ha (our calculation) VN-directThấp.

#W.3.6 Steps 6 and 7: S-ALT substitution and emissions

• Food route. A share of meat demand (at 0.15 kg of protein per kg cwe) is replaced; meat production and its feed fall in proportion. Part of the substitute is fermented (mycoprotein-type, on sugar); the rest is plant-based from imported soy.
• Feed route. Shares of SBM protein and fishmeal protein are replaced one for one, on crude protein, by microbial protein, partly from hydrogen gas fermentation and partly from sugar or starch.
• Emissions. Electricity x a grid factor derived from PDP8 targets: 0.659 t CO2 per MWh (2023 official), 0.38 (2030), 0.29 (2035), 0.20 (2040), 0.021 (2050) ^30,13 VN-directTrung bình ước tính. It assumes PDP8 is delivered. Avoided emissions use published averages: 0.85 kg CO2e per kg of SBM protein and 2.72 kg per kg of fishmeal protein ³¹ generalThấp. Cassava farming, steam, urea and land-use change are excluded.

#W.4 Scenario definitions

All values are our assumptions for the years shown VN-directThấp ước tính.

Shared by all: population, SSR, dressing yields, export shares, compound shares, domestic maize output (4.1 Mt), non-feed maize (1.61 Mt) and yields abroad.

• S-ALT shares are what-if levels, not targets. The 40% fishmeal share by 2050 sits within trials that replaced 25 to 60% of fishmeal without growth loss (Feed and aquafeed market, H9) ^26,40 generalTrung bình. The 10% SBM share by 2050 implies large cost falls or policy support: at 2026 costs microbial feed protein costs 4.6 to 5.1 times SBM protein (Economics) ³² VN-directThấp ước tính.
• S-HIGH is a stress case, not a central view.
• A combined S-EFF plus S-ALT case was not modelled, so we give no number for it. The two act through different channels and add up.

#W.5 Results by scenario and year

All tables: our calculation, tools/balance_model.py, from balance_outputs.csv. Values for 2030 to 2050 are VN-directThấp ước tính; 2025 is the calibrated base, identical in all scenarios.

#W.5.1 Compound feed

Compound feed, Mt: total (livestock and poultry / aquafeed)

S-BASE by group, 2025 and 2050, Mt: pigs 12.17 and 16.26; poultry 9.00 and 13.99; other livestock 0.81 and 1.40; pangasius 2.90 and 3.75; whiteleg shrimp 1.29 and 1.67; other shrimp 0.17 and 0.37; other fish 2.13 and 3.45. Aquafeed stays at 21 to 24% of compound feed throughout. S-EFF feed peaks around 2040.

#W.5.2 Soybean meal

SBM import need, Mt (as meal, or as beans crushed in Vietnam)

Soybean-meal need to 2050 in four scenarios

Million tonnes a year, our model; shaded band: the vision range of chapter 19

Beans, intensity and total protein meals

Efficiency against alternative protein. In 2050 S-EFF needs 3.40 Mt less SBM than S-BASE (minus 33%) and S-ALT 1.62 Mt less (minus 16%): efficiency does about twice the work at the shares tested VN-directThấp ước tính.

#W.5.3 Maize, fishmeal and land abroad

• Maize grows more slowly than SBM (+16% against +44% by 2050 in S-BASE) because non-compound maize feeding shrinks as pig farming industrialises and non-feed use is held flat; total S-BASE maize use goes from 16.0 Mt (2025) to 17.9 Mt (2050) VN-directThấp ước tính.
• Fishmeal: S-EFF and S-ALT reach about 151 kt by 2050 through different levers, faster inclusion cuts and 40% microbial replacement VN-directThấp ước tính.
• Land abroad peaks around 2040 in S-BASE because yields abroad keep rising while needs flatten; the S-ALT figure for 2050 excludes 52,000 ha for soy used in plant-based food VN-directThấp ước tính.

#W.5.4 Meat, demand per person and other outputs

Other S-BASE outputs: eggs 235 per person (2030) and 275 (2050); milk 2.75 Mt (2050); aquaculture 6.58 Mt (2030) and 8.52 Mt (2050), 51% exported VN-directThấp ước tính.

#W.5.5 What S-ALT implies

The 2035 column is in balance_outputs.csv (for example 160 kt of microbial feed protein, 2.7% of high-protein feed protein). The share is microbial protein divided by microbial protein plus protein-meal demand x 0.46. Plant equivalents are our arithmetic from the design sizes of Entobel's insect-meal plant in Vietnam (a size benchmark only) and Calysseo's FeedKind plant in China, which halted in 2026 ^41,42,43 VN-directThấp.

Readings. All VN-directThấp ước tính unless tagged otherwise.

• Carbohydrate is the binding physical constraint. The 2050 need of 5.37 Mt of fresh cassava roots is 52% of the 2025 harvest (10.24 Mt) and of the cassava area; the sugar alternative, 1.25 Mt, is about 96% of the 2025/26 crop ^36,44. Roots rose about 75% in price in the year to April 2026 and now also feed E10 ethanol ^45,46 VN-directTrung bình.
• Energy volume is not the constraint; energy carbon is. The 11.7 TWh of 2050 is 0.85 to 0.94% of planned commercial electricity ¹³; 185 kt of hydrogen is 0.9 to 1.9% of the official target of 10 to 20 Mt of clean hydrogen a year by 2050 ⁴⁷.
• Carbon break-even. At 41 kWh per kg of protein, gas-route protein beats SBM only below about 21 g CO2 per kWh (0.85 / 41) and fishmeal below about 66 g. The sugar route's electricity (3.5 kWh per kg) beats SBM below about 240 g. The PDP8-derived grid passes 240 g in the late 2030s and reaches about 21 g only around 2050, so before then the gas route needs dedicated renewable power. The break-even depends on the energy factor (W.9.4).
• The food route saves the most imports per tonne. On the 2050 S-BASE feed structure, replacing 1 t of meat protein avoids about 4.5 t of SBM and 9.5 t of maize imports, against about 1.65 t of imported soybeans when half the substitute is plant-based. Microbial feed protein replaces SBM but not maize. Diet change is outside our scope; this is arithmetic only.
• Land. Cassava-based microbial protein needs about 0.80 ha per t of protein (4.0 x 0.926 x 4.4 / 20.5), against about 0.71 ha for imported SBM: it moves land rather than saving it. Hydrogen gas fermentation needs 0.4% of SBM's land per t of protein in one study ³¹ generalTrung bình.

#W.6 Sensitivity: what moves the 2050 result

One assumption, or a paired set, is varied with all else at S-BASE (10.40 Mt of SBM and 13.84 Mt of maize imports in 2050). The last row varies S-ALT. Source: balance_sensitivity.csv; all rows VN-directThấp ước tính, horizon 2050.

What moves 2050 soybean-meal need

Swing around the trend value of 10.4 Mt when one assumption is varied, million tonnes

Readings.

• Formulation and efficiency dominate. Each 1% a year fall in national SBM inclusion is worth about 2.3 Mt of imports in 2050 (10.40 minus 8.09) VN-directThấp ước tính. Yet we found no Vietnamese series of SBM inclusion or FCR by species VN-directCao; measuring them is the first job of a protein statistic (Robust moves).
• Demand assumptions matter less: about 0.6 to 1 Mt each in 2050 VN-directThấp ước tính.
• Maize responds to other levers: FCR (4.0 Mt), domestic output (3.3 Mt) and non-feed use (2.8 Mt) in 2050. Pig industrialisation raises SBM but lowers maize, because the non-compound maize ration shrinks VN-directThấp ước tính.
• The unresolved 2025 aquafeed volume hardly matters (0.02 Mt in 2050), because SBM is calibrated on the 7.2 Mt actually fed. Doubling S-ALT's microbial share cuts 2050 need by another 1 Mt, less than the inclusion trend VN-directThấp ước tính.
• The ranges are illustrative, not confidence intervals; interactions are not tested.

#W.7 Sanity checks

#W.7.1 Checks against anchors and published projections

Saturation. Continuing OECD-FAO's +4.2% a year poultry growth per person to 2050 would give about 38 kg of poultry per person, against 34.3 kg in S-BASE; we taper instead VN-directThấp ước tính. Vietnam becomes an "aged society" by 2035, which argues for slower per-person growth ⁴⁹ VN-directTrung bình dự phóng.

#W.7.2 The 2030 feed target: resolution of OQ-127

On the livestock-only reading, the 2030 target looks out of reach on both our trend and official-growth paths VN-directThấp ước tính. It would be met only if the first-half 2026 growth rate held every year to 2030; the full-year 2026 figure is the signpost.

#W.8 Reality check against international analogues

Model cases against the analogues VN-directThấp ước tính

• S-EFF (7.00 Mt in 2050) is credible. Its minus 1% a year inclusion trend cuts SBM intensity by 22% over 2025 to 2050 (253 to 198 kg per t); China's achieved pace over 25 years would cut it by 20 to 52% (our calculation).
• S-ALT (8.78 Mt) is a stretch. Its microbial share of high-protein feed protein is 1.0% (2030), 2.7% (2035), 4.4% (2040) and 8.4% (2050). Norway's 0.4% makes the 2030 value plausible and the 2050 value unprecedented.
• S-HIGH (11.72 Mt) is an upside risk.

Credible ambitious range. On this evidence, a credible ambitious range for Vietnam's SBM need in 2050 is about 7 to 9 Mt (7.0 to 8.8 Mt, the S-EFF to S-ALT span), against 10.4 Mt on trend VN-directThấp ước tính. Matching ranges are 7.8 to 8.3 Mt (2030), 8.1 to 9.0 Mt (2035) and 7.8 to 9.2 Mt (2040), with microbial feed protein at about 1% of high-protein feed protein by 2030 and 4 to 8% by 2050 ước tính. Holding imports at the 2025 level (7.2 Mt) through 2050 is a stretch case. With no combined S-EFF plus S-ALT run, the range has no value below S-EFF. Vision 2050 builds on this range.

#W.9 Reproduce, extend, limits and disagreements

#W.9.1 How to rerun

• Requirements: Python 3, standard library; nothing is fetched from the web. Input: data/balance_assumptions.csv. Outputs: data/balance_outputs.csv and data/balance_sensitivity.csv, plus a console summary.
• Warning: the script takes no arguments. --help or any argument is ignored, and every run overwrites both output files. To experiment, run a copy: mkdir -p run/tools run/data, copy tools/balance_model.py to run/tools/ and data/balance_assumptions.csv to run/data/, then python3 run/tools/balance_model.py.
• Check: on 24 September 2026 a run in a clean copy reproduced both package files byte for byte.
• Working paper: working-papers/wave3/balance_model/model.py is the same model with stream paths and no record IDs.

#W.9.2 How to extend

#W.9.3 Limitations

• FCRs, inclusion rates and compound shares are assumptions calibrated to national totals; no Vietnamese series by species exists in the sources we reached.
• Self-sufficiency, export shares, domestic maize output and non-feed maize are held constant; population is the same in all scenarios.
• No prices, so no statement on competitiveness (V. Frontier technology).
• No climate impacts on Vietnamese farming or on exporters' yields (Drivers to 2050).
• Emissions cover only electricity for alternative protein; land-use change is excluded.
• Microbial protein replaces SBM and fishmeal one for one on crude protein, with no digestibility or amino-acid adjustment. S-ALT food substitution touches meat only.
• The 2040 and 2050 values extend beyond any published projection.

#W.9.4 Disagreements with published numbers

We give both values and the position taken; these belong in R. Disagreements.

#W.10 Gaps and open questions

These feed Q. Open questions.

#Data files

• balance_assumptions.csv (238 rows, BLA-001 to BLA-238): every input by parameter, scenario tag and year, with basis, sources and confidence, including SENS-LOW and SENS-HIGH rows.
• balance_outputs.csv (1,150 rows, BLO-0001 to BLO-1150): 80 indicators by scenario and year, with unit, foresight type and confidence.
• balance_sensitivity.csv (13 rows, BLS-001 to BLS-013): the tornado table for 2050 SBM and maize import need.
• balance_published_projections.csv (86 rows, BLP-001 to BLP-086): OECD-FAO, USDA, UN, PDP8 and official-target values used for calibration and checks.
• aquaculture_projections.csv (44 rows): the Appendix Z aquaculture paths; AQ-BASE reuses the S-BASE index.
• protein_strategies_benchmark.csv (23 rows): national protein and feed strategies behind W.8.
• Code and notes: tools/balance_model.py; working-papers/wave3/balance_model/balance_model.md; chart data in charts/data/sbm_scenarios_2050.csv and charts/data/sbm_tornado_2050.csv.

Related: Protein balance 2050, Vision 2050, Scenarios 2050, Robust moves, Z. Aquafeed and feedstocks 2050, Feed and aquafeed market, O. Outlook data, U. Futures method.