Green hydrogen by PEM electrolysis: what it costs to make
A techno-economic analysis of a 50,000 tonne per year green hydrogen plant. The plant does not clear its hurdle rate at the stated sale price.
Read the accuracy range before you read any number. This is an AACE Class 5 screening estimate, and the expected accuracy is -30% to +50%. The inputs come from published literature and public technical targets. No input is a vendor quote for this configuration at a site.
AACE Class 5 answers whether a project is worth studying further. AACE Class 5 does not answer whether to build. Every number on this page carries that range, and the range is wide enough to change the decision. Treat each figure as a midpoint, and never as a quotation.
This is an educational portfolio model. It is screening economics, and it is not a sanctioned FID-grade estimate.
The estimate class is asserted, and it is not demonstrated. AACE assigns a class from the maturity of the scope-definition deliverables. Nobody has run that assessment against this study. The recorded label is AACE Class 5 (screening, order-of-magnitude). The accuracy range above is a project-specific figure. It is not evidence for the class. See change record PC-002, which is open.
The finding: the base case is uneconomic at the stated price
The levelized cost of hydrogen is 11.34 USD/kg. That figure is the break-even price. The stated sale price in this study is 10 USD/kg, and the stated sale price is below the break-even price.
At the stated sale price the project returns NPV -0.40 BUSD and IRR 7.31%, against a 10% hurdle rate. Payback is 12.5 years. The IRR is below the hurdle rate, so the project does not clear it.
This is the result, and it is not a failure to engineer away. No scenario was hunted for that rescues the project and then led with. The plant needs a hydrogen price above the break-even price, or it needs a materially cheaper capital estimate.
The decision this analysis serves
Somebody proposes a 50,000 tonne per year green hydrogen plant. The plant uses PEM electrolysis, and the electrolyzer draws about 344 MW. Two questions decide whether the proposal goes further.
The first question is what the hydrogen costs to make. The answer is the levelized cost of hydrogen. That figure is the price that makes the project break even over its life.
The second question is whether the project clears its hurdle at the price a buyer would pay. This study states a sale price of 10.0 USD/kg and tests the project against it.
A techno-economic analysis that produces one NPV is arithmetic. This analysis also ranks which uncertainty drives the answer, and it attaches a probability to the downside. Those two additions are what make the result usable.
Basis frozen in the project specification on 2026-08-06. Plant: 50,000 t H2/yr, 8,000 operating hours per year, 55 kWh/kg, 0.045 USD/kWh. Finance: 20 year life, 10% discount rate, 21% tax rate.
Capital cost
Total capital investment is 2.50 BUSD, or 7,270 USD/kW. The estimate uses the Lang factorial method with a fluid-processing factor of 4.8, from Peters and Timmerhaus 4th edition.
The method starts from a purchased equipment cost of 443 MUSD. Each equipment cost is escalated to a 2024 basis by CEPCI, and the index used is 800.
The AACE Class 5 range applies to this number directly. At -30% to +50%, the total capital investment runs from about 1.75 BUSD to about 3.75 BUSD. That span is the largest single source of uncertainty in the answer, and the sensitivity section below shows it.
| Item | MUSD |
|---|---|
| Electrolyzer system (stacks, rectifiers, BoP) | 337 |
| Storage | 94 |
| Compression | 12 |
| Total purchased equipment cost | 443 |
| Fixed capital investment (Lang 4.8) | 2,124 |
| Working capital (15% of total) | 375 |
| Total capital investment | 2,499 |
Operating cost
This page reports two annual cost quantities, and it names each one. They differ by the annualised stack-replacement line.
All-in annual cost is 230 MUSD/yr. It includes the annualised stack replacement. The table below sums to this figure, and every share in the table divides by it.
Recurring operating expense is 199 MUSD/yr. It excludes the annualised stack replacement. The DCF model uses this smaller figure, so the levelized cost of hydrogen and the NPV are tied to it.
The annualised stack-replacement line is 30 MUSD/yr. That line is a capital replacement spread over the years between replacements. It is not a conventional operating expense, so the DCF holds it apart from the recurring lines. That is why the two quantities differ.
Electricity is 124 MUSD/yr. Its share depends on which of the two quantities you divide by. Electricity is of the all-in annual cost. Electricity is of the recurring operating expense. Both shares are correct, and each answers a different question.
Electricity dominates the annual operating cost. Electricity does not dominate the uncertainty in the levelized cost of hydrogen. The next section explains why those two statements are both true.
| Line | MUSD/yr | Share of all-in annual cost |
|---|---|---|
| Electricity | 123.8 | 53.8% |
| Insurance and local taxes | 63.7 | 27.7% |
| Stack replacement (annualised) | 30.3 | 13.2% |
| Maintenance | 8.1 | 3.5% |
| Operating labor | 1.4 | 0.6% |
| Plant overhead | 1.3 | 0.6% |
| Water | 1.0 | 0.4% |
| Operating supervision | 0.2 | 0.1% |
| All-in annual cost (includes stack replacement) | 229.8 | |
| Recurring operating expense, used by the DCF (excludes stack replacement) | 199.5 |
The share column divides by the all-in annual cost, and it sums to 100%. The recurring operating expense is shown as a memo line. It carries no share, because it is not the total of the lines above it.
Sensitivity: total capital cost dominates
The dominant uncertainty is total capital cost, at 4.05 USD/kg of impact on the levelized cost of hydrogen. Electrolyzer capex ranks second, at 3.41 USD/kg. Electricity price ranks third, at 1.38 USD/kg.
This result contradicts the project's own recorded prediction. The project README was written before any model code existed. The README argued that electricity would be the unambiguous dominant cost driver. The model says otherwise.
The build report states the outcome plainly. The build report says the result "differs from the README's expectation that electricity would dominate". The build report then says that "the model's answer is reported as found".
The prediction was written down first, and then the model overturned the prediction.
The prediction was left standing in the README. It was not quietly retrofitted to match the result. That sequence is the point. The independent audit recorded the same contradiction as a positive finding rather than as a defect.
The contradiction is also explicable, and the explanation matters so the result does not read as an anomaly. Electricity dominates the annual operating cost, at of it. Uncertainty in the levelized cost of hydrogen depends on the width of each input's range. The AACE Class 5 capital range is a wider swing than the cited electricity price range. Total capital cost therefore moves the answer more.
Both statements are true at once, because the two statements answer different questions. One statement is about the size of a cost, and the other statement is about the width of an uncertainty.
| Parameter | LCOH at low | LCOH at high | Impact, USD/kg | Source |
|---|---|---|---|---|
| Lang factor (total capex) | 8.91 | 15.39 | 4.05 | Peters and Timmerhaus |
| Electrolyzer capex | 9.29 | 14.75 | 3.41 | DOE technical targets |
| Electricity price | 10.51 | 12.71 | 1.38 | DOE technical targets |
| Electrolyzer efficiency | 10.15 | 11.85 | 1.18 | DOE technical targets |
| Discount rate | 10.34 | 12.39 | 1.05 | frozen specification |
| Storage capex | 10.82 | 11.86 | 0.52 | Terlouw et al. 2022 |
| Operating hours | 11.81 | 11.02 | 0.47 | frozen specification |
| Stack lifetime | 11.34 | 10.98 | 0.36 | DOE technical targets |
Read the electrolyzer efficiency row with its field definition. The field is energy consumed per unit of output, in kWh/kg, so a higher value is worse and not better. The row shows LCOH rising from 10.15 to 11.85 USD/kg. That direction is correct. The independent auditor checked this row too, and the auditor confirmed the sign after checking the same field definition.
Monte Carlo risk analysis
The risk analysis runs 20,000 trials with a fixed seed, so the draws reproduce exactly. The spread is the answer here, and the P50 alone is not the answer.
The levelized cost of hydrogen runs P10 9.75, P50 12.13, P90 15.14 USD/kg. The band from P10 to P90 spans 5.39 USD/kg. Report the P50 with the P10 and the P90 beside it, because the P50 on its own hides the width.
The probability that NPV is positive at the stated sale price of 10.0 USD/kg is 13.0%. Roughly one draw in eight gives a positive NPV, and roughly seven draws in eight do not.
The band does not capture joint behaviour. The Monte Carlo samples each input independently. Electricity price and availability are correlated in reality, and this model does not correlate them.
Independent sampling can misstate the P10 to P90 band. Correlation moves a band in either direction. The sign of the correlation and the slopes of the two responses decide which way it moves. This study runs no correlated case, so the direction here is not measured. This is trap 6 in the project's frozen methodology. The trap is stated in every report rather than quietly accepted.
| Quantity | Value |
|---|---|
| LCOH P10 | 9.75 USD/kg |
| LCOH P50 | 12.13 USD/kg |
| LCOH P90 | 15.14 USD/kg |
| LCOH mean | 12.33 USD/kg |
| LCOH standard deviation | 2.12 USD/kg |
| P(NPV > 0) at 10.0 USD/kg | 13.0% |
| Trials | 20,000 |
| Convergence: P50 change when trials double | 0.136% |
Separate scenario: a 45V-style production tax credit
This block is not the base case. Everything above this block is the unsubsidised base case. The numbers in this block belong to a separate scenario. The base case does not include any production tax credit.
The separate scenario applies a 45V-style production tax credit of 1.00 USD/kg of hydrogen. Under that scenario the levelized cost of hydrogen falls to 10.63 USD/kg. Under that same scenario NPV is -0.20 BUSD and IRR is 8.75%.
The credit narrows the gap, but the credit does not close the gap. The scenario is still NPV-negative at the stated sale price, and the scenario IRR is still below the hurdle rate.
The base case stays unsubsidised on purpose. The project plan recorded the decision in advance. The plan says a credit inside the base case "would flatter the result and hide the unsubsidized economics". The plan calls the unsubsidised economics "the number a reader actually needs".
| Metric | Base case | Policy scenario |
|---|---|---|
| LCOH, USD/kg | 11.34 | 10.63 |
| NPV at the stated sale price, BUSD | -0.40 | -0.20 |
| IRR | 7.31% | 8.75% |
| Payback, years | 12.5 | 10.9 |
Cash flow over the project life
The plant is built over 2 years and then runs for 20 years. The chart below shows the annual net cash flow and the cumulative cash flow at the stated sale price.
Acceptance gates
The project froze nine numeric acceptance gates before the analysis ran. All nine gates pass. The table reports each measured value beside its threshold, because a verdict without a measured value cannot be checked.
G-1 is the external-validation gate, and it decides whether the rest of the page can be trusted. The DCF engine reproduces two published worked examples: the Turton Chapter 8 example and the Cornell processdesign example.
Both examples are published, and neither example is a self-computed fixture. That distinction matters. An engine validated against its own output tells you nothing, because the engine agrees with itself by construction.
| Gate | Requirement | Measured | Threshold | Status |
|---|---|---|---|---|
| G-1 | DCF engine vs published worked examples | Turton Ch. 8 NPV rel err 3.3e-15, IRR 14.3494%; Cornell processdesign NPV rel err 2.55e-05 and 6.49e-05 | 0.1% | PASS |
| G-2 | Cash flow balance closes every year | residual 0.00e+00 USD | < 1e-6 USD | PASS |
| G-3 | DCF LCOH vs annualized-cost LCOH | rel diff 3.13e-16 | 0.5% | PASS |
| G-4 | Every cost input traces to a citation | zero uncited float literals outside the inputs module | 0 uncited | PASS |
| G-5 | Monte Carlo converged when trials double | P50 moves 0.136% | 0.5% | PASS |
| G-6 | Sensitivity signs physically sensible | all 8 signs correct | correct signs | PASS |
| G-7 | LCOH within the published PEM band | LCOH 11.338 USD/kg inside [4, 18] | in band | PASS |
| G-8 | Test suite passes | 35 passed, exit 0 | exit 0 | PASS |
| G-9 | Full pipeline reproducible | two consecutive runs byte-identical | identical | PASS |
The machine-readable results file carries 11 numeric entries across these nine gates. G-1 folds four measurements into one row and G-7 folds three, because each gate has one verdict. G-4, G-8 and G-9 are process gates, so each one carries its recorded outcome instead of a numeric entry.
The independent audit
An independent auditor reviewed this project. The verdict is PASS, and the auditor recorded no P0 and no P1 findings.
What the auditor re-derived independently
- G-1, from scratch. The auditor re-derived the Turton Chapter 8 worked example in a fresh session. The auditor computed NPV with the textbook's own closed-form annuity formula. The auditor computed IRR by manual bisection, and the auditor used neither scipy nor the project's own solver.
- Three independent computations agree. The textbook's formula, the project's engine, and the auditor's bisection all give the same answer. The auditor matched the IRR to four decimal places.
- G-4, by an independent scan of every module for uncited numeric literals.
- G-6, by independent reasoning over all eight sensitivity signs.
- G-8 and G-9, by re-running them. The auditor re-ran the test suite and ran the full pipeline twice.
The auditor's own stated blind spots
These are reported as the auditor reported them. They are not buried, and they should be weighed by anyone who relies on this work.
- Gates G-2, G-3, G-5 and G-7 were reasoned about by code review. The auditor did not independently re-execute those four gates and did not re-derive them by hand. The audit leans on structural plausibility for those four gates.
- Cited sources were spot-checked for existence, not confirmed against their values. The auditor checked that the citations are real and specific. The auditor did not fetch each source and confirm the number taken from it.
- Two modules were not reviewed line by line beyond the literal scan.
- The full technical report and the figures were not read in that audit round.
Four of the nine gates therefore rest on the build's self-report plus a code read. The auditor's own explanation is that the round went first to the gates most likely to hide a real error.
Limitations
These limitations are load-bearing. They are not a disclaimer, and the first limitation changes the answer.
- AACE Class 5 screening accuracy, -30% to +50%. A total capital investment of 2.50 BUSD means a real range of roughly 1.75 to 3.75 BUSD. The sensitivity section shows this range is the largest single source of uncertainty in the answer.
- Literature and published inputs, not vendor quotes. Cost correlations and reference costs are public and generic. No input is a quotation for this configuration at this site.
- Screening economics, not a sanctioned FID-grade estimate. This is an educational portfolio model, and it must never be presented as an investment-grade appraisal. The boundary is the plant fence, so there is no transport, no distribution, and no end-use economics.
- Single-point technology and scale assumptions. One technology, one plant scale, and one site basis. Efficiency is an input to this model, and efficiency is not an output of it.
- Electricity price is constant, which understates real price volatility. The check on that assumption is the sensitivity range around the price, and the check is not the point value.
- Stack degradation is a periodic replacement, and it is not a continuous efficiency decline.
- The Monte Carlo samples inputs independently, so the P10 to P90 band does not capture joint behaviour between inputs. No correlated case was run, so the direction of that error is not measured.
- A single discount rate is used. There is no financing structure, no debt and equity split, and no WACC derivation.
Method and reproducibility
Citations came before the model. The cited input table was built before any model code existed. One module is the only place where numbers live.
The DCF engine was validated before the plant model existed. The engine is the only externally validatable component, and every downstream number flows through the engine.
One command regenerates every number and every figure from the frozen inputs. The Monte Carlo seed is fixed, so two consecutive runs give byte-identical results.
What this page does not do, and why
This page carries no calculator and no slider. That absence is deliberate.
The estimate is AACE Class 5, and the accuracy range is -30% to +50%. A calculator would imply a precision the method explicitly disclaims. A reader would dial in a value, screenshot the result, and treat the result as a quote.
The tornado chart and the Monte Carlo histogram already communicate the uncertainty, and they communicate it honestly. Both are static figures, so neither invites a reader to manufacture a number the analysis never produced.
Every number on this page is read from the machine-readable results file at build time. 35 project tests pass, plus the page tests. Basis: specification frozen 2026-08-06. Monte Carlo over 20,000 trials, with a fixed seed recorded with the project.