Grid carbon intensity, not electrolyser efficiency, determines the well-to-gate footprint of PEM hydrogen

Project identifier: SEPT-H2-LCA Status: complete, all acceptance gates passed Study type: well-to-gate greenhouse-gas carbon footprint, not an ISO-conformant life-cycle assessment

Abstract

Hydrogen made by electrolysis is often described as clean, but the description belongs to the electricity rather than to the process, and the size of that dependence is rarely quantified against a fixed plant design with auditable factors. We built a well-to-gate greenhouse-gas inventory for a specific PEM electrolysis plant, holding the functional unit fixed at 1 kg of 100 percent hydrogen delivered at the plant gate at 80 bar, and varying only the electricity supply. The foreground was imported from an existing techno-economic model through a machine-readable adapter that reconciles exactly with its accepted results. Electricity emission factors were taken from hash-locked official sources, eGRID2023 Revision 2 for grid generation and the December 2025 45VH2-GREET dependency package for upstream fuel-cycle and generator-infrastructure modules, with the 4.2 percent grid loss rate applied exactly once. The United States average reference case is 22.913 kg CO2e per kg H2 at a levelised cost of 11.361 USD per kg. Across all 27 eGRID subregions the footprint spans 7.76 to 46.51 kg CO2e per kg H2, a factor of six, while the plant design is identical in every case. Stack electricity accounts for 21.05 kg CO2e per kg H2, which is 91.9 percent of the total. A Monte Carlo analysis over 20,000 draws confirms that electricity quantity and electricity carbon intensity are the only parameters that materially move the result, with Spearman rank correlations of 0.88 and 0.44 against 0.002 for stack lifetime. Dedicated low-carbon supply changes the picture completely, giving 0.155 kg CO2e per kg H2 for nuclear and 0.563 for wind with generator capital included. Reaching a 1 kg CO2e per kg H2 target would require delivered electricity below 0.018 kg CO2e per kWh, roughly 23 times cleaner than the current United States average grid. Electrolyser stack and balance-of-plant manufacturing are excluded because no separable inventory compatible with this plant was available in the frozen official package, and no proxy was substituted.

1. Introduction

Electrolytic hydrogen occupies a central position in decarbonisation plans for industry and heavy transport, on the reasoning that splitting water with electricity produces no carbon dioxide at the point of use. The reasoning is sound about the electrolyser and incomplete about the system. An electrolyser is a device for converting electricity into a chemical carrier, and the greenhouse-gas burden of the carrier is inherited almost entirely from the electricity that made it.

This inheritance is widely acknowledged in principle and inconsistently handled in practice. Published footprints for electrolytic hydrogen vary widely, and the variation often reflects analytical choices rather than physical differences. Those choices include which grid, which year, whether upstream fuel-cycle emissions are included, whether generator construction is inside the boundary, whether transmission losses are counted once or twice or not at all, whether an oxygen coproduct receives a credit, and at what pressure the hydrogen is considered delivered. Each choice moves the answer, and comparisons across studies that differ in several choices at once carry little information.

A cleaner way to isolate the effect of electricity supply is to fix everything else. If the plant design, the functional unit, the delivery pressure, the auxiliary loads, the water balance and the economic assumptions are all held constant, and only the electricity source varies, then the spread in results measures the sensitivity to electricity supply rather than the diversity of analytical conventions.

We took that approach here. The foreground is a single PEM plant with a fixed and accepted techno-economic description, imported through a machine-readable adapter rather than retyped. The functional unit is 1 kg of 100 percent hydrogen at the plant gate at 80 bar, enforced identically across every scenario. Electricity factors come from official sources whose exact bytes are hash-verified, with each factor carrying its version, geography, year, type, unit, boundary, GWP basis and upstream scope as structured metadata. Average and marginal factors are never combined in a single total.

Three questions follow. How much does the footprint of an identical plant vary with grid location? Which parameters actually drive the result, as opposed to those conventionally reported? And how clean would the electricity need to be for common policy thresholds to be reachable?

2. Methods

2.1 Functional unit and system boundary

The functional unit is 1 kg of 100 percent hydrogen at the plant gate at 80 bar. Every scenario is validated against this unit and pressure, which is acceptance gate LCA-G02. The boundary is well to gate: it covers electricity generation and its upstream fuel cycle, generator infrastructure where the factor includes it, plant auxiliaries, compression from 30 to 80 bar, and the water resource flow. It excludes downstream storage, distribution and use.

This study reports a greenhouse-gas carbon footprint on a single indicator. It is not an ISO-conformant life-cycle assessment, and it assesses no impact category other than climate.

2.2 Foreground import

The plant foreground is imported from a read-only techno-economic snapshot through a strict adapter with a JSON schema contract. The adapter reconciles the imported quantities against the source model's accepted outputs, and every reconciliation boolean must be true for the build to proceed, which is acceptance gate LCA-G01.

Delivered electricity totals 55.51 kWh per kg H2, composed of 51 kWh for the stack, 4 kWh for auxiliaries and power electronics, and 0.51 kWh for compression from 30 to 80 bar. The water requirement is 9 kg per kg H2.

The compression module deserves comment because it is where the carbon boundary and the cost boundary meet. The source model delivers hydrogen at 30 bar, while this study's functional unit is at 80 bar, so the compression electricity is added to the inventory and the matching operating cost is added to the levelised cost as an explicit bridge of 0.02295 USD per kg. The accepted source cost of 11.338377 USD per kg becomes 11.361327 USD per kg. Carbon and cost therefore rest on the same operating assumptions, which is acceptance gate LCA-G08.

2.3 Electricity factors and double-counting controls

Grid factors derive from eGRID2023 Revision 2 for direct generation, combined with non-overlapping GREET modules for upstream fuel cycle and generator infrastructure. The GWP basis is IPCC AR5 GWP100 without climate-carbon feedbacks, with CO2 = 1, CH4 = 28 and N2O = 265. The eGRID gross grid loss rate of 4.2 percent is applied exactly once, to convert generation-basis factors to a delivered basis. All 27 eGRID subregions are included plus a United States average reference, with no post-hoc selection.

Technology cases for wind, solar PV and nuclear are extracted from the official December 2025 45VH2-GREET dependency package. Capital-included and official no-capital variants are kept as distinct scenarios rather than blended, because they answer different questions and mixing them would obscure the boundary.

Each factor carries complete structured metadata and is validated against a JSON schema; incomplete or boundary-incompatible factors are rejected rather than defaulted, which is acceptance gate LCA-G03. Average and marginal factors are never summed into one total. The United States non-baseload rate, which represents displaced marginal generation, is reported only in a separate marginal analysis, satisfying acceptance gate LCA-G05.

Two double-counting controls apply. Water treatment and pumping electricity already sits inside the 4 kWh per kg auxiliary block, so the water module is carried as a resource flow contributing zero rather than being charged a second time. Compression electricity is counted once, in the module that also carries its cost bridge.

2.4 Exclusions

Generator infrastructure is included wherever the underlying factor includes it, and such factors are flagged. Electrolyser stack and balance-of-plant manufacture are excluded, because the frozen official package contains no separable inventory compatible with this plant configuration. This is a declared exclusion with a zero contribution, not an assertion that the burden is zero. No proxy inventory was substituted, in keeping with a pre-registered stop condition that forbids substituting an unapproved proxy when a boundary cannot be closed.

2.5 Uncertainty, break-even and cross-check

Parameter uncertainty was propagated through 20,000 Monte Carlo draws using a PCG64DXSM generator with seed 20260830, with a Gaussian copula of rho = 0.25 coupling system electricity and compression electricity. Scenario uncertainty is kept separate from parameter uncertainty. Sensitivity is reported as Spearman rank correlation against both carbon and cost.

Break-even electricity carbon intensities were computed for target footprints of 0.45, 1, 2 and 4 kg CO2e per kg H2, and independently recalculated in the test suite. A boundary-aligned cross-check against the official model was performed on the nuclear pathway, with structural differences decomposed rather than absorbed.

3. Results

3.1 An identical plant spans a factor of six across the grid

Holding the plant design, functional unit and delivery pressure fixed, and varying only the eGRID subregion, the well-to-gate footprint ranges from 7.76 to 46.51 kg CO2e per kg H2 (Figure 1). The United States average reference is 22.913 kg CO2e per kg H2 at 11.361 USD per kg.

That spread of a factor of six is produced entirely by electricity supply. Hydrogen from the cleanest subregion, NPCC Upstate New York at 7.76, and from the most carbon-intensive, Puerto Rico Miscellaneous at 46.51, comes from the same equipment operating on the same specification. Several subregions containing substantial industrial hydrogen demand sit above the national average, including SERC Midwest at 35.50 and MRO East at 41.01.

For context, the United States average case of 22.9 kg CO2e per kg H2 is not low against conventional production. Grid-powered electrolysis is a carbon-reduction measure only where the grid is already clean or where dedicated low-carbon supply is contracted.

The regional spread below is the whole finding in one panel: the same plant, the same efficiency, and a factor of six between the cleanest and dirtiest grid.

3.2 Electricity dominates the inventory and the uncertainty

Contribution analysis for the reference case attributes 21.05 kg CO2e per kg H2 to stack electricity, 1.65 to auxiliaries and power electronics, and 0.21 to compression from 30 to 80 bar (Figure 2a). Stack electricity is 91.9 percent of the total, and all electricity together is effectively the whole inventory. Direct process emissions are zero, since the reaction produces only hydrogen and oxygen. The water module contributes zero by the double-counting control described in Section 2.3, and electrolyser manufacture contributes zero because it is excluded.

Table 1. Contributions for the United States average reference case.

The Monte Carlo analysis agrees (Figure 3). Over 20,000 draws the footprint has a median of 22.46 kg CO2e per kg H2 with a 5th to 95th percentile range of 20.64 to 23.91, a relative spread of about plus or minus 7 percent around the median. This is parameter uncertainty within a fixed grid scenario, and it is an order of magnitude smaller than the scenario spread across grids in Section 3.1. Which grid you are on matters far more than how precisely any parameter is known.

Table 2. Rank sensitivity of carbon and cost to input parameters, 20,000 draws.

Two parameters dominate carbon: how much electricity the plant uses and how carbon-intensive that electricity is. Stack lifetime, which receives substantial attention in electrolyser development, has a rank correlation with carbon of 0.002. It matters for cost, at -0.188, and is nearly irrelevant to the footprint. Efficiency improvements act on carbon and cost together, which is the one lever that moves both.

The inventory breakdown below shows how little of the footprint is anything other than electricity, and what happens when that electricity is changed.

Parameter uncertainty is shown below against the scenario spread from the previous figure, which is the comparison that matters for interpreting the intervals.

3.3 Dedicated low-carbon supply changes the result by two orders of magnitude

Replacing grid electricity with dedicated low-carbon supply reduces the footprint from 22.9 to below 2.1 kg CO2e per kg H2 in every case examined (Figure 2b). With generator capital included, nuclear gives 0.155, wind gives 0.563 and solar PV gives 2.059 kg CO2e per kg H2. Under the official no-capital boundary, wind and solar are reported as zero and nuclear as 0.140.

The gap between the two boundaries is the point rather than a technicality. Wind moves between 0.563 and exactly zero depending only on whether generator construction is inside the boundary. Neither number is wrong, and quoting either without its boundary is uninformative. We keep the variants as separate scenarios for this reason.

Cost does not follow carbon (Figure 4a). The non-dominated set contains three points: nuclear at 0.155 kg CO2e per kg H2 and 12.749 USD per kg, wind at 0.563 and 10.529, and a higher-efficiency United States average case at 20.02 and 10.178. Nuclear is the lowest-carbon option and the most expensive; the cheapest option is neither the cleanest nor close to it. Wind is the interesting middle, cutting the footprint by a factor of 40 relative to the grid average while also costing less than the reference case.

The trade-off below places cost against carbon and marks the policy thresholds, which sit well below every present-day regional grid.

3.4 Policy thresholds require electricity far cleaner than today's average grid

Break-even analysis inverts the question: given the plant's 55.51 kWh per kg demand, how clean must delivered electricity be to hit a target footprint (Figure 4b)?

Table 3. Maximum delivered electricity carbon intensity for target footprints.

The United States average delivered factor is 0.4128 kg CO2e per kWh. A 1 kg CO2e per kg H2 target requires 0.018, which is roughly 23 times cleaner. Even a 4 kg target requires nearly 6 times cleaner. No eGRID subregion in this analysis approaches those levels on an annual average basis, which is why the technology cases in Section 3.3 depend on dedicated supply rather than grid improvement.

These thresholds are arithmetic consequences of the plant's electricity demand and the stated boundary. They describe what the modelled inventory requires and establish no eligibility under any policy programme, a limitation we return to in Section 4.

3.5 Cross-check against the official model

A boundary-aligned cross-check on the nuclear pathway returned a common-boundary residual of exactly zero against the official factors, satisfying acceptance gate LCA-G06. Structural differences are decomposed rather than absorbed: generator infrastructure accounts for 0.01496 and the 30 to 80 bar pressure bridge for 0.00129 kg CO2e per kg H2, with oxygen credit and water differences at zero.

The check has a stated limit. The official 45VH2-GREET application caps user-entered pressure at 50 bar and was not executed independently on this platform. The comparison uses exact cached values from the hash-verified official dependency package, and it verifies boundary and arithmetic consistency rather than reproducing the official application's own execution.

4. Discussion

The central finding is that a fixed PEM plant has no single carbon footprint. Its footprint is a property of the electricity it consumes, spanning a factor of six across United States grid subregions and two orders of magnitude once dedicated low-carbon supply is contracted. This is not a new idea, but quantifying it against a single fixed plant with hash-locked official factors turns a qualitative caution into a specific number, and the number is large enough to dominate every other choice in the analysis.

Where should effort go? The sensitivity results answer that directly. Stack lifetime has a rank correlation with carbon of 0.002. Improving durability is valuable for cost, at -0.188, and it will not decarbonise hydrogen. Electricity quantity and electricity carbon intensity are the two levers that move the footprint, and only the first also moves cost. An organisation aiming at a carbon target should treat electricity procurement as the primary decision and equipment durability as an economic one.

Break-even analysis sets the bar such procurement must clear. A 1 kg CO2e per kg H2 target requires delivered electricity roughly 23 times cleaner than the United States average. No annual-average subregion here is close, so meeting such a target depends on dedicated low-carbon generation rather than on waiting for grid averages to improve.

One further point deserves emphasis, because it is a reporting hazard rather than a physical result. Wind hydrogen is 0.563 or 0.000 kg CO2e per kg H2 depending only on whether generator construction is counted. A study reporting the second number without its boundary would appear to show carbon-free hydrogen. Both numbers appear here as separate scenarios, and neither is presented as the value.

Four limitations bound the results. First, the study reports a single greenhouse-gas indicator; it says nothing about water, land, materials, or any other impact category, and it is not ISO-conformant. Second, electrolyser stack and balance-of-plant manufacture are excluded, so the reported figures are lower bounds by an unquantified amount, which matters most for the low-carbon cases where the electricity term is small; we declined to substitute a proxy because no compatible official inventory existed. Third, the grid factors are annual averages and establish no specific plant location, hourly matching, additionality or deliverability. Fourth, the cross-check verifies boundary and arithmetic consistency against cached official factors and did not execute the official application independently.

Nothing here establishes eligibility under any tax-credit or certification programme. Such programmes impose their own boundaries, temporal matching rules and additionality tests that this attributional annual-average analysis does not evaluate.

The most useful extension would be temporal resolution. Annual averages hide the covariance between when an electrolyser runs and how clean the grid is at that hour, and a flexible electrolyser operating on hourly-matched clean supply could achieve a footprint that no annual-average calculation can represent. That analysis needs hourly generation data and a dispatch model, and it is outside the frozen boundary here.

5. Claims and non-claims

Supported by this work. A well-to-gate greenhouse-gas carbon footprint for the specified PEM plant at 1 kg H2 and 80 bar, under hash-locked official factors and the stated boundary. The regional, technology, efficiency, stack-life, capital-boundary, marginal and oxygen sensitivities as reported. The cost-carbon tradeoff under operating assumptions consistent with the accepted source techno-economics. Break-even electricity intensities as arithmetic consequences of the stated inventory.

Not supported by this work. ISO-conformant life-cycle assessment. Eligibility under 45V or any other policy programme. Certification of any kind. Hourly matching, additionality or deliverability. Environmental superiority on any indicator other than the modelled greenhouse-gas total. Any claim about a specific plant location, since grid data are annual subregion averages. Completeness of the manufacturing boundary, which is explicitly excluded.