V. Frontier technology: evidence, costs and windows

What this appendix contains. The detail behind Frontier technology: an evidence card for each of eleven frontier routes (V.1); the arithmetic for protein made from renewable power in Vietnam (V.2); learning rates and cost paths (V.3); carbon prices and carbon cost per tonne of protein (V.4); AI and bio-design (V.5); approvals and capital (V.6); the research pipeline (V.7); and the windows table (V.8).

How to use it. Read a route's card before pricing a deal or programme in it. Test any cost claim against V.2 and V.3. Use V.8 for the condition that opens each window, and X. Drivers and signals for the signposts.

---

#V.1 Frontier map: one card per route

Each card gives the status in 2026, the largest plant, cost evidence, failures, Vietnam's base, the window and its conditions. Full registers: frontier_gas_tech.csv, frontier_bio_tech.csv, feedstock_futures.csv.

#V.1.1 Power-to-protein (hydrogen-oxidising bacteria)

#V.1.2 Methane and steel-gas protein

#V.1.3 Biomass fermentation on sugar and side streams

#V.1.4 Precision fermentation

#V.1.5 AI-assisted bio-design and biofoundries

#V.1.6 Cellulosic (second-generation) sugar

Scale (our calculation, ngf_calc.py): the 2050 S-ALT sugar need of 1.32 Mt of glucose would take 3.0 to 4.3 Mt of straw (6 to 10% of national straw) and USD 2.8 to 4.5 billion of first-generation plants, 15 to 21 of Panipat's size VN-directLow estimate. Feedstock cases: Z. Aquafeed and feedstocks 2050.

#V.1.7 One-carbon (methanol) routes

Our calculation for 2050: supplying the S-ALT sugar need through methanol takes about 1.5 Mt of methanol, 0.29 Mt of green hydrogen (1.5 to 2.9% of the 2050 target) and about 16 TWh; about 49 MWh per t of protein, below direct hydrogen fermentation but with heavy cooling VN-directLow estimate.

#V.1.8 Cultivated meat and seafood

#V.1.9 Molecular farming

#V.1.10 Cell-free protein synthesis

#V.1.11 Algal and oilseed omega-3 (brief)

---

#V.2 Power-to-protein in Vietnam: the arithmetic

Script: ftg_estimates.py. Results in V.2.2 to V.2.6 are our calculations VN-directLow estimate.

#V.2.1 Inputs

Steps: TWh = t of protein x MWh per t / 1,000,000; solar GW = TWh x 1,000,000 / (8,760 x 0.17) / 1,000; land = MW / (75 to 109).

#V.2.2 Energy, hydrogen, CO2, land and grid carbon

The 500 kt rows are in working-papers/wave3/frontier_gas/frontier_gas.md. Against the plans (our calculation from ^2,76,77) VN-directLow estimate:

• 1 Mt of protein a year takes 4.2 to 6.1% of planned 2050 electricity, 15 to 19% of the 2050 solar plan and 3.5 to 6.9% of the 2050 hydrogen target.
• 100 kt a year would take 14 to 69% of the whole 2030 hydrogen target.
• All the protein in 2025 soybean-meal imports (2.62 Mt) would need 168 to 218 TWh, 113 to 146 GW of solar (38 to 50% of the 2050 plan) and 1.8 Mt of hydrogen: physically possible, economically not (V.2.4 and V.2.5).
• Land: about 13 to 22 t of protein per ha of panels, over ten times any staple crop ⁸⁰ generalMedium.

Grid carbon, electricity only: about 46 t CO2 per t of protein at the 2023 factor; 24 to 28 t at the 2030 average implied by the power plan (0.35 to 0.40 t per MWh, an upper bound); 1.1 to 1.4 t in 2050 (0.018 to 0.020), if the plan's 27 Mt cap is met VN-directLow estimate. Dedicated renewables are a precondition until about 2045.

#V.2.3 CO2 point sources

From co2_point_sources_vn.csv (18 rows):

Readings VN-directLow: the three ethanol plants release about 0.21 Mt a year, less than the 0.30 Mt that 100 kt of protein needs but enough for pilots up to about 70 kt; 1 Mt of protein needs about 13% of Hoa Phat's 2025 Scope 1 emissions (22.5 Mt), but that gas already makes power; volumes at fertiliser plants, refineries and Formosa Ha Tinh are unknown.

#V.2.4 Cost estimates for 2030, 2040 and 2050

Method: electricity use x Vietnamese supply price (solar cost x 1.4 to 1.54 for balancing), plus the model's non-energy costs x 1.15 to 1.35 for a Vietnamese cost of capital of about 9 to 11%, against 7% in the model ^2,75.

Vietnamese columns VN-directLow estimate; best sites generalHigh projection (Fasihi et al. published model, EUR x 1.2). On grid power at today's USD 75 per MWh, a 2030 plant would cost USD 8,700 to 10,300 per t and emit about 46 t CO2 per t VN-directLow estimate. In the model, cost of capital plus or minus 2 points moves cost 12 to 15%; the high electrolyser-cost case adds only 3 to 6% ² generalHigh.

#V.2.5 Hydrogen-price sensitivity and parity

A plant buying hydrogen pays about USD 690 per t of protein for each USD 1 per kg of hydrogen (0.694 t per t). At USD 1, 2, 3 and 5 per kg the hydrogen line alone is USD 694, 1,388, 2,082 and 3,469 per t of protein; at USD 2 it exceeds the whole price of soybean-meal protein generalHigh.

Breakeven delivered hydrogen price (point-source CO2 at USD 20 to 50 per t; non-hydrogen costs at a Vietnamese cost of capital) VN-directLow estimate:

Parity. The Vietnamese cost reaches the fishmeal-protein band only at its low end and only around 2050, and never approaches soybean-meal protein under published projections VN-directLow estimate. Today's Vietnamese hydrogen costs (USD 2.81 to 3.88 per kg) sit above most breakevens ⁷⁹ VN-directLow.

Frontier protein costs against the prices to beat

USD per kg of protein; ranges by year; fishmeal and soybean-meal protein at 2025 to 2026 prices

#V.2.6 The five conditions

Also needed: CO2 at USD 20 to 50 per t beside power and hydrogen (available now for pilots); and one hydrogen-route plant above 10 kt a year running two years abroad, not before about 2032 ⁵ VN-directLow estimate. Earlier, narrower windows: formulating imported, approved food ingredients from about 2028 if a new-food route exists; a CAP-type co-product in the 2030s; a 100 to 1,000 t research pilot on ethanol-plant CO2 now to 2030 VN-directLow estimate.

---

#V.3 Learning rates and cost paths

#V.3.1 Learning-rate evidence

A 20% learning rate means unit cost falls 20% each time cumulative output doubles. Full table: learning_rates.csv (21 rows).

Complex, site-built and customised plants learn more slowly than mass-produced panels and cells ^93,94 generalHigh. We use 5 to 20% on capital cost and treat solar-style curves for fermented protein as unsupported generalMedium trend. Big historic bioprocess cost falls came from titre and yield (penicillin titres rose 225-fold in 1941 to 1944), not from building identical plants ⁹⁵ generalMedium.

#V.3.2 Fungal feed protein: floor and planning range

Script: ecf_calc.py, part A. Learning applies only to capital charge and maintenance in edition 1.0's cost stack A (10,000 t a year; Economics); glucose, urea, power, labour and land stay at 2026 real prices.

All rows: our calculation VN-directLow estimate. Eight doublings (256 times today's capacity) is an outer bound.

Planning range VN-directLow estimate: about USD 3,000 to 12,500 per t of protein in 2030, 2,400 to 10,800 in 2040 and 2,200 to 9,500 in 2050 (fast-case low end to central-case high end). The low end needs cheap capital, residue carbon and a heat-tolerant strain.

• The floor, not learning, settles the soybean question: about USD 1,870, 2.1 times soybean-meal protein.
• Fishmeal parity needs fast learning or cheap capital. The low end is below the August 2026 fishmeal-protein price from 2030 (central and fast cases) or 2040 (slow); it reaches the 2025 average (USD 2,625) only in the fast case, around 2040.

#V.3.3 Capital-charge sensitivity

Our calculation VN-directLow estimate. Cheaper capital does as much as the central case's learning to 2030, with no technology risk, which is why concessional debt and guarantees are first-order levers; capital grants and guarantees also beat tax holidays in the regional evidence ¹⁰¹ VN-adjacentMedium.

#V.3.4 Precision-fermentation cost evidence

Our reading: high-value proteins pay now; functional dairy and egg proteins in premium blends in 2030 to 2035; commodity use only if costs near USD 1 to 2 per kg, which no published data yet support generalLow estimate.

---

#V.4 Carbon

#V.4.1 Carbon price paths

Full table: carbon_price_paths.csv (19 rows).

Vietnam's scheme. Decision 263/QD-TTg (9 February 2026) set pilot quotas above 243 Mt CO2e (2025) and 268 Mt (2026), offsets up to 30% and first surrender by 31 December 2027; full operation starts in 2029 ^103,88 VN-directMedium projection (official target). No price is published and agriculture and food are outside, so no feed or food plant faces a domestic carbon price before 2029 VN-directMedium signal. Asian prices in 2025 were about one-eighth of the EU's.

#V.4.2 Carbon cost per tonne of protein

Carbon cost = emission factor x price. It applies only if a price reaches embodied emissions; none does today. Script: ecf_calc.py, parts B and C; full table carbon_cost_per_protein.csv (14 rows).

Fungal footprint inputs per t of product (our assumptions): 2.0 to 2.2 t of glucose at 0.65 to 1.10 t CO2e per t (Thai cassava starch plus hydrolysis); 0.17 to 0.20 t of urea at 1.0 to 2.5 t CO2e per t; 0.9 to 3.1 MWh of power; husk steam counted as biogenic; biogas captured; 45 to 55% protein.

#V.4.3 What a carbon price does and does not change

• For soy, the accounting method is the price. At USD 50 per t, soybean-meal protein gains USD 75 to 78 per t without land-use change (9% of its price) but USD 445 to 480 with it (about half). For Argentine meal, 69% of Vietnam's imports, the two factors differ five-fold; we give both and do not average them generalMedium.
• It barely touches fishmeal: USD 41 to 43 per t of protein at USD 50, 1 to 2% of the price. Scarcity drives the fishmeal gap, not carbon ¹¹⁴ generalMedium.
• It can hurt a cassava-based microbial protein, which emits more than fishmeal (0.8 to 0.9 t per t of protein) and soybean meal without land-use change (1.6). Residue carbon, biogas and clean power come first VN-directLow estimate.
• Protein from power flips from worst to best only on a clean grid, around 2045 to 2050 VN-directLow estimate.
• Nothing prices embodied feed emissions. The EU border carbon mechanism covers urea and hydrogen, both fermentation inputs, but not food or feed ¹⁰⁵ generalHigh. Buyers' scope 3 targets are the likelier channel, 2030 to 2040 generalMedium signal.
• Credits are not a financing route; green-debt labels are. No approved protein-substitution method was found; the Climate Bonds Initiative launched Alternative Proteins criteria in 2025 ^116,117 generalLow.

---

#V.5 AI and bio-design

#V.5.1 Demonstrated and claimed

Full table: ai_biodesign_evidence.csv (21 rows).

Reading. Gains are real but narrow: about 1.6 to 30 times fewer experiments for specific problems. None lowers production cost at scale, and none removes Vietnam's bottlenecks: rules, shared pilot plant, feedstock conversion and scale-up capital generalMedium.

#V.5.2 Compute, policy and the missing biofoundry

Reading. Compute is no longer the constraint; wet-lab automation, data and bio-ML people are VN-directMedium. Design in Vietnam, build and test in a Korean, Singapore or Chinese biofoundry, is the cheap route to 2030 VN-directLow estimate.

---

#V.6 Approvals and capital

#V.6.1 Approvals

Novel-protein approvals counted from edition 1.0's register (Singapore, US, Israel, Australia and New Zealand, China; 2026 to 14 August) ^134,4 generalHigh:

Not counted: 2025 US letters for beta-lactoglobulin (Vivici; Shanghai Changing Biotechnology), lactoferrin (TurtleTree), ovalbumin (Onego Bio) and whey (Verley), with some dates in conflict; a proposed rule to end self-affirmed GRAS; 2024 US molecular-farming clearances; Korea's cell-cultured food standards (June 2026) ^28,29,66,135 generalMedium. Vietnam has no new-food route and no approvals ⁶⁹ VN-directMedium. Full list: approvals_trend.csv (44 rows).

#V.6.2 Capital

Advocacy data; we found no independent projection to 2050 and give none.

#V.6.3 Patents expiring in the mid-2030s

Core families behind heme meat analogues (Impossible Foods, e.g. family 54241231; heme secretion filed September 2014) and animal-free dairy (Perfect Day, family 54105986) were filed in 2013 to 2016, so the earliest members lapse around 2033 to 2036 on a 20-year term (our estimate, not a freedom-to-operate analysis) ¹³⁹ generalLow estimate. Continuations extend some protection. Vietnam's only linked filing is a 2019 US application on yeast biomass ²⁵ VN-directLow; its route to 2040 is process know-how on local feedstocks and strains ³² VN-directLow estimate.

#V.6.4 Distressed equipment

Cheap second-hand tanks cut capital more than any learning curve before 2030, until the cycle turns generalMedium signal. Clear customs and tax rules for used fermenters would let a Vietnamese feed venture use this in 2027 to 2030 (Robust moves, RM-15) VN-directLow estimate.

---

#V.7 Research pipeline

#V.7.1 Global growth by field

OpenAlex articles and reviews; keyword lower bounds; 2026 to 24 September ³² generalMedium trend.

All OpenAlex articles and reviews grew about 0.9% a year (2015 to 2025). Fields with approvals behind them grow fastest; gas fermentation and molecular farming stay thin.

#V.7.2 Vietnam's position

Intensity. Vietnam's alternative-protein research intensity is about a quarter to a third of Thailand's, depending on the database: 4.1 against 13.4 works per 10,000 in OpenAlex, about a quarter in OpenAIRE ^32,141 VN-directMedium. The two databases agree, so we give the range.

Who does the work. Before 2025, mostly aquafeed trials with Australian partners; in 2025 to 2026, mostly domestic food technology (IUH, HUIT, HCMUTE, Phenikaa, VAST), about 70% without foreign co-authors ³² VN-directLow signal. At about 20% a year, output would reach about 100 works a year by 2035, roughly Thailand's 2025 level; a mechanical extrapolation, not a forecast ³² VN-directLow estimate. Institutions and partners: K. Research landscape.

---

#V.8 Windows for Vietnam

From frontier_windows.csv (11 rows). Plays: Plays.

When frontier windows could open for Vietnam

Earliest plausible window by route, our estimate; conditions in the tooltip

Readings VN-directLow estimate:

• The windows open in sequence: feed ingredients and design work to 2030, food ingredients and contract fermentation in 2030 to 2035, new carbon and seafood cells in the 2040s, bulk protein from power last.
• Rules open more windows than technology does. FW-02 to FW-05 all wait on a feed-list entry or a new-food route, both decided in Vietnam in 2026 to 2030.
• The late windows depend on inputs Vietnam can prepare now: residue carbon, clean power at named sites, cheap capital and process skills (Robust moves).

---

#V.9 Gaps and open questions

For Q. Open questions; source conflicts are in R. Disagreements.

---

#Data files

• frontier_gas_tech.csv (10 rows): gas and electro-fermentation routes, costs and Vietnam conditions.
• vn_energy_inputs_2050.csv (50 rows): power, hydrogen and grid inputs to 2050, with our derived needs and costs.
• co2_point_sources_vn.csv (18 rows): CO2 point sources by site.
• frontier_bio_tech.csv (14 rows): bio-design, precision fermentation, cultivated, molecular farming, cell-free and biomass routes.
• ai_biodesign_evidence.csv (21 rows): AI and automation results, demonstrated or claimed.
• approvals_trend.csv (44 rows): approval and policy items (2024 to 2026) and yearly approval counts.
• learning_rates.csv (21 rows): learning rates and our fungal-protein cases.
• carbon_price_paths.csv (19 rows): carbon prices and paths.
• carbon_cost_per_protein.csv (14 rows): emission factors and carbon cost per t of protein.
• feedstock_futures.csv (17 rows): residue, side-stream and one-carbon sources to 2050.
• biblio_trends_frontier.csv (325 rows): OpenAlex counts by topic, country and year, with queries.
• vn_research_profile.csv (53 rows): Vietnam's research counts, institutions and partners.
• frontier_windows.csv (11 rows): the windows table in V.8.

Related: Frontier technology, Economics, Technology fit, Protein balance 2050, Robust moves, Costs and benchmarks, J. Science, K. Research landscape, M. Capital, Z. Aquafeed and feedstocks 2050, X. Drivers and signals, U. Futures method