The Energy Cost of Green Hydrogen

Overview

The thermodynamic floor for a low-temperature stack is 39.39 kWh/kg. Everything sold is measured against it, and almost nothing is measured the same way twice.

Three floors are routinely confused. The heating-value convention alone — lower against higher — moves a headline number by about eighteen per cent, as much as a decade of research. The stack, system and plant boundaries do not enclose the same equipment, and five distinct efficiency points sit between them. ISO 22734 has been withdrawn, its Part 2 is in draft and aimed at grid services, and a 476 mV spread has been measured between laboratories testing the same thing. This report states the boundary, the basis and the load point behind every figure, or states that the source does not.

Part II puts the four technologies on one basis: alkaline and its five different answers to the same question about minimum load; PEM, which shows no persistent efficiency difference from alkaline in the US Department of Energy's own modelling, at four times the current density and with an iridium problem; AEM, with commercial hardware at a published 51.3 kWh/kg and the only degradation curve in the report; and solid oxide, where the steam boundary produces two DOE rows for one machine and an efficiency of 126 per cent that the chapter explains.

Part III is the supply side. Thirteen of the twenty-eight manufacturers surveyed publish no efficiency figure at all. Nobody guarantees an efficiency, and every convention in the guarantees that do exist runs one way. Four insolvencies and impairments in eighteen months are recorded, with what each does to a stack warranty.

Part IV covers what has actually been built: four assets with published output against nameplate, and no measured kWh/kg anywhere in the public record. Part V covers capital cost — which went the wrong way — and the sensitivity grid showing which term dominates under which power price and load factor. Part VI covers the frontier to 2031, including the single claim that would consume 87 per cent of the remaining headroom, and closes with an inventory of what the public record does not contain.

Absences are reported as findings. Where an authority or a manufacturer was checked and found to publish nothing, that is recorded with the source checked and the date. Absence claims are scoped to the sources reviewed as at 31 August 2026; they are not claims about the whole world.

This is an independent publication. It is not commissioned or endorsed by, or affiliated with, any electrolyser manufacturer, any national authority, the International Energy Agency, IRENA, the Joint Research Centre, the US Department of Energy or the Clean Hydrogen Partnership. All organisations and products are named for identification only.

What this edition covers

Twenty chapters in six parts, plus a manufacturer evidence appendix. The measurement problem: scope and provenance, the thermodynamic floor, the heating-value convention, the three boundaries, and the standards vacuum. The technologies: alkaline, PEM, AEM and solid oxide, then the master comparison table and the part-load curve nobody publishes. The suppliers: what manufacturers publish, China and the price-against-current-density trade-off, what a performance guarantee is worth, and the supplier failure record. The evidence from operation: what has been built, and attrition against the FID gap. Economics: capex and what a kWh/kg is actually worth. The frontier: demonstrated against announced to 2031, and what is missing from the public record. Appendix A lists every manufacturer document examined, its date, its retrieval date and what was disclosed.

Who this is for

Developers, investors, lenders, offtakers and engineers who need electrolyser efficiency figures they can compare on one basis, trace to a retrieved source, and reproduce.

What you get

• Alkaline, PEM, AEM and solid oxide restated onto one consistent basis
• Every kWh/kg figure carries its measurement boundary, heating-value basis and load point
• The thermodynamic floor explained, and the three floors that are routinely confused
• 28 manufacturers surveyed — and the thirteen that publish no efficiency figure at all
• What a performance guarantee actually covers, and why every convention runs one way
• A dated record of four supplier insolvencies and impairments in eighteen months
• Four built assets with published output against nameplate
• A sensitivity grid showing which term dominates at which power price and load factor
• Absences reported as findings, with the source checked and the date
• 291 sources, 55 data tables, and a manufacturer evidence appendix
• Independent — not commissioned by any manufacturer, authority or programme

Table of contents

1. Part I — The Measurement Problem: Why two efficiency figures are rarely the same quantity.
2. Chapter 1 — Scope, definitions and method: The provenance scheme, and the three boundaries that do not enclose the same equipment.
3. Chapter 2 — The thermodynamic floor: Three floors routinely confused, and why a low-temperature stack cannot go below 39.39 kWh/kg.
4. Chapter 3 — The convention problem: LHV, HHV, and the eighteen per cent that moves a headline as much as a decade of research.
5. Chapter 4 — The boundary problem: Stack, system, plant — and the five efficiency points that separate them.
6. Chapter 5 — The standards vacuum: ISO 22734 withdrawn, a draft Part 2 aimed at grid services, and a 476 mV spread between laboratories.
7. Part II — The Technologies: Four technologies on one basis.
8. Chapter 6 — Alkaline: Cheap, durable, and five different answers to the same question about minimum load.
9. Chapter 7 — PEM: No persistent efficiency difference from alkaline in DOE's modelling, four times the current density, and an iridium problem.
10. Chapter 8 — AEM: Commercial hardware at a published 51.3 kWh/kg, and the only degradation curve in the report.
11. Chapter 9 — Solid oxide: The steam boundary, two DOE rows for one machine, and 126 per cent efficiency explained.
12. Chapter 10 — The comparison: The master table, the capital-against-energy trade-off, and the part-load curve nobody publishes.
13. Part III — The Suppliers: What the supply side discloses, and what it does not.
14. Chapter 11 — What the manufacturers publish: Thirteen of twenty-eight publish no efficiency figure at all.
15. Chapter 12 — China: The price gap, the current-density trade-off, and what merely circulates.
16. Chapter 13 — Guarantees: Nobody guarantees an efficiency, and every convention runs one way.
17. Chapter 14 — The supplier failure record: Four insolvencies and impairments in eighteen months, and what they do to a stack warranty.
18. Part IV — The Evidence from Operation: What has been built, and what it produced.
19. Chapter 15 — What has been built: Four assets with published output against nameplate, and no measured kWh/kg anywhere.
20. Chapter 16 — Attrition and the FID gap: The dated cancellation record, and a 150 per cent forecast overshoot by one institution in one year.
21. Part V — Economics: What the energy term is worth against the capital term.
22. Chapter 17 — Capex: The cost that went the wrong way, and the boundary ratio that reconciles the sources.
23. Chapter 18 — What a kWh/kg is worth: The sensitivity grid, and which term dominates under which power price and load factor.
24. Part VI — The Frontier: Demonstrated against announced, and what the record does not contain.
25. Chapter 19 — The frontier, 2026–2031: Demonstrated against announced, and the 87 per cent of remaining headroom one claim would consume.
26. Chapter 20 — What is missing: The gaps in the public record, and the instruments that would close them.
27. Appendix A — Manufacturer evidence: Every document examined, its date, its retrieval date, and what each manufacturer disclosed.

Research and sources

No efficiency figure is quoted without its boundary: a figure in kilowatt-hours per kilogram is not comparable until the measurement boundary, the heating-value basis and the load point are known, and every figure states them or states that the source does not. Figures restated onto a common basis are marked as derived and the operation is shown. Every figure carries a provenance class — measured, published, reported, estimated, derived, forecast and none — and these are never tabulated in the same column. Figures that could not be traced to a retrieved primary source are labelled as circulating and never stated as fact. Where authorities disagree both figures are given, and nothing is resolved by averaging. 291 unique sources, every one retrieved and read. Absence claims are scoped to the sources reviewed as at 31 August 2026.