Counting Grid Emissions Honestly:
Average, Marginal, and Time-Matched

The same on-site power project can make a company's reported emissions fall under one accounting method and rise under another, without changing a single hour of how it runs. That is not fraud; it is what happens when three different questions are collapsed into one careless word: clean.

Section 01Three questions hiding in one number

Every grid-emissions figure is the answer to a question, and there are more questions in circulation than most readers of sustainability reports realize. When a project team says the grid carries some number of pounds of carbon dioxide per megawatt-hour, they have silently made three choices. First, whether to describe the grid's average output or the specific plants that respond when load changes. Second, over what window to measure: a year, an hour, or the five-minute interval on which the market actually dispatches. Third, which ledger to consult: the physical one, which follows wires, or the contractual one, which follows certificates.

Those three choices generate at least eight defensible combinations, and in California the combinations disagree more sharply than almost anywhere else, because the grid is effectively two grids in one day: a midday system so saturated with solar that the operator routinely turns some of it off, and an evening system that still runs substantially on gas. A number produced in one cell of that matrix is routinely quoted as if it settled the others. It does not.

This paper walks through the three choices in order, shows where each method is strong and where it flatters, and closes with what the tightening rulebook, voluntary and statutory, implies for any organization deciding how to power a site. None of it requires a science degree. It requires only tolerance for the fact that the carbon content of electricity has several honest answers, and that a claim which does not name its method is not yet a claim.

Section 02The average: what the grid was

The workhorse of emissions accounting is the average intensity factor. In the United States the canonical source is the Environmental Protection Agency's eGRID database, which compiles measured generation and emissions for essentially every grid-connected plant in the country and aggregates them into subregions. The current edition covers calendar 2023 and was last revised in June 2025. For CAMX, the subregion covering most of California, it reports an average output emission rate of about 428 pounds of CO₂ per megawatt-hour. The national average is about 767.¹

Averages are the right tool for a specific job: attributional inventories. If the purpose is to divide the grid's actual, historical emissions among the parties who consumed its electricity, the average is not merely convenient. It is the only method under which everyone's slices sum to the size of the pie. That additivity is why corporate inventory standards are built on it, why auditors can work with it, and why it moves slowly and predictably from year to year.

The same properties make the average nearly useless for decisions. It lags: a 2023 measurement, published in 2025 and quoted in 2026, describes a grid that has been adding storage rapidly in the meantime. It flattens: an annual mean of a system that swings daily between solar saturation and gas dependence describes neither condition. And it is silent about causation. Comparing a proposed on-site generator against the average grid quietly assumes that the site's consumption draws a proportional slice of every plant on the system, including the nuclear, hydro, and solar output that was going to be produced regardless. Physically, that is not what happens. The plants that respond to a change in load are a smaller and very specific set.

Section 03The marginal: what your action changes

When a site adds or sheds a megawatt-hour of demand, the grid does not scale every plant proportionally. Resources with no fuel cost or no flexibility, which is most of the zero-carbon fleet in most hours, do not change output because one customer's load changed. The response comes from whatever dispatchable resource sits on the margin in that hour, and in California, outside the hours when renewable output is being curtailed, that resource is nearly always a gas-fired plant. The emissions consequence of an action is therefore set by the marginal unit, not by the average mix.

The federal reference for this way of counting is the EPA's AVERT tool, which publishes region-specific marginal emission rates built on an hour-by-hour, peer-reviewed methodology, precisely so that efficiency and renewable programs can estimate what their output actually displaces.² The rates it produces for a region differ materially from that region's eGRID average, and the difference is not an error in either dataset. They are answers to different questions.

California adds a twist that matters to every power decision made in the state. In 2024 the state's grid operator curtailed about 3.4 million megawatt-hours of utility-scale wind and solar output, 29 percent more than the year before; solar accounted for 93 percent of the curtailed energy, concentrated in spring middays when output is high and demand is mild.³ In an hour when the operator is already turning solar off, the marginal resource is the curtailed solar itself. Adding consumption in that hour absorbs energy that would otherwise have been discarded, and the marginal emissions rate of that consumption approaches zero. Running a fuel-burning generator through that same hour displaces energy that is nearly carbon-free.

The marginal lens therefore renders verdicts the average lens cannot see. Consuming at 1 p.m. and consuming at 9 p.m. are different acts. Generating at 1 p.m. and generating at 9 p.m. are different acts. A battery that charges on midday surplus and discharges into the evening ramp is doing real emissions work on the marginal ledger, even though on an average ledger, after round-trip losses, it looks like a device that consumes more energy than it delivers. Both descriptions are true; they are simply about different things.

Two honest cautions. First, short-run marginal rates describe the dispatch response, not the investment response. A large, persistent new load also changes what gets built, and that long-run build margin, on a grid whose interconnection queues are dominated by solar and storage, is plausibly much cleaner than tonight's marginal gas plant. But build-margin estimates come from capacity-expansion models, and they move with the modeler's assumptions; treat every such figure as a labeled estimate. Second, marginal claims do not add up. The grid has one margin per hour, so if every midday consumer in California claims the curtailed-solar margin, the sum of the claims exceeds the physical grid. Marginal numbers are decision guidance, not inventory entries.

An average tells you what the grid was. A marginal rate tells you what your decision changes. Neither is wrong; they answer different questions.

Section 04The clock: annual, hourly, five-minute

Cutting across both methods is a second axis: time. The California grid operator publishes an informational dashboard of systemwide CO₂ at five-minute granularity on its public website.⁴ The data needed to time-resolve an electricity claim exists in this state as almost nowhere else, and it shows a daily shape that annual arithmetic erases: deep decarbonization through the solar hours, then a fast evening climb as gas units ramp to replace the setting sun. An annual factor for California is an average over what are, functionally, two different grids.

Time-matching is what turns that physical shape into an accounting question. Under annual matching, a renewable-energy certificate generated on a mild April afternoon can lawfully cover a megawatt-hour consumed at 9 p.m. the following January. Under hourly matching, each hour of consumption must find carbon-free generation in that same hour, from a grid it could plausibly have been delivered from. The difference is not pedantry. A load that runs around the clock and buys annual certificates equal to its total consumption is one hundred percent matched on paper while still drawing gas-fired power every night of the year on the physical ledger. Hourly accounting is the method that notices.

Honesty about the cost side belongs here too. Hourly matching demands interval meters, time-stamped generation records, and settlement-grade data that many organizations do not yet hold, and covering the hardest hours, winter evenings and becalmed nights, is where the expense concentrates. Annual matching is inexpensive precisely because it lets the easy hours pay for the hard ones. That trade is what the rulemakers described next are now weighing.

Section 05The two ledgers: location and market

The framework nearly every corporate electricity claim runs on is the GHG Protocol's Scope 2 Guidance, adopted in 2015. It requires dual reporting: a location-based figure computed from average grid factors for the place where consumption occurred, and a market-based figure computed from contractual instruments, such as supplier-specific rates, certificates, and power purchase agreements, with residual-mix factors, where available, covering consumption no one has claimed.⁵ Both numbers are legitimate, and each disciplines the other. The recurring mischief is quoting whichever number is smaller without naming the method that produced it.

The market-based method deserves its honest case. Contractual instruments gave buyers a standardized way to direct money toward renewable generation before it was cheap, defined what a renewable claim means, and built the registries and audit trails that any stricter future system will still need. The honest case against is equally simple: an annually matched claim carries no information about when the covered consumption happened, and in mature markets the price of an unbundled certificate can be low enough that buying one changes little about what gets built. Both cases are true at once, which is why the framework is being revised rather than defended or abandoned.

The revision points in one direction. The overview the GHG Protocol published in September 2025, ahead of a public consultation that opened in October 2025, proposes that market-based claims be matched on an hourly basis, with exemptions for smaller organizations and load-profile approximations where hourly data is missing; that instruments come from generation deliverable to the consuming load rather than from broad national market boundaries; and that dual reporting remain. The published overview anticipated final approval in late 2027 followed by a phased rollout.⁶ Any organization signing long-dated electricity contracts on the strength of today's annual-matching rules should read that timeline carefully and check its current status, because consultations move.

One more ledger deserves a paragraph, because it is this paper's subject wearing different clothes. Gas has its own contractual-attribution system: book-and-claim instruments that attach renewable or biogenic gas attributes to molecules physically delivered from the common pipeline. Whatever position one takes on those instruments, they raise exactly the questions this paper has been walking through: matched over what period, deliverable from where, claimed on which ledger, disclosed under which method. A generator physically burning pipeline gas while carrying contractual biogas attributes should be described with the same discipline demanded of a building carrying certificates, and by the same people.

Section 06The hour is becoming law

Anyone tempted to file hourly matching under voluntary-market fashion should look at where federal rulemaking has already landed. The final regulations for the federal clean-hydrogen production credit, published January 10, 2025, permit electrolytic hydrogen producers to count grid electricity as clean through certificates only if the electricity satisfies three pillars: incrementality, meaning the generation is new or newly uprated, with defined exceptions; deliverability, meaning the generator sits in the same grid region as the load, with limited provisions for imports; and temporal matching, under which annual matching is permitted only until 2030, after which the certificate must come from the same hour as the electricity use it covers.⁷

Whether any particular facility qualifies for that credit is a question for qualified tax counsel, and nothing in this paper is tax advice. The point for a power decision-maker is narrower. The strictest emissions-accounting rules yet written into federal law treat when and where clean electricity was generated as decisive, not decorative, and the leading voluntary standard is moving the same direction on its own schedule. An emissions claim engineered to today's annual rules should be stress-tested against the hourly rules both regimes have now put in writing.

Section 07The same project under five lenses

Apply all of this to the decision this series exists to serve: an organization with a continuous load in Northern California weighing on-site generation against grid service. Consider a gas-fed on-site unit, whether a reciprocating engine, a turbine, or a non-combustion fuel cell. Converting pipeline gas to electricity at 45 to 60 percent efficiency on a higher-heating-value basis puts direct stack emissions roughly between 890 and 660 pounds of CO₂ per megawatt-hour. Treat those figures as illustrative arithmetic from standard published fuel carbon coefficients, not as a rating of any machine.

Against the annual average, every machine in that range is browner than the California grid it would sit beside. Even the most efficient gas-fed generator in the illustration emits roughly half again the CAMX average of 428 pounds. On this lens, gas-fired on-site generation in California is a step backward, full stop, and a seller of such equipment, or of studies about it, who quotes only friendlier lenses is doing marketing, not analysis.

Against the short-run margin, the verdict turns hour by hour. Overnight, the displaced resource is nearly always gas-fired, and a high-efficiency on-site unit can emit less per delivered megawatt-hour than the marginal plant it displaces, particularly after counting avoided transmission losses, while a low-efficiency unit emits more. Through a spring midday, the same machine competes with curtailed solar, and no fuel-burning device defends those hours. An operating profile that turns down when the grid is clean improves the emissions ledger and worsens the project's economics, because the same fixed cost is spread over fewer running hours. That trade-off is real and belongs in the model, priced, not hidden.

Combined heat and power moves the arithmetic honestly but not simply. Where recovered heat displaces boiler fuel, part of the fuel's carbon belongs to the thermal product and the electricity's effective intensity falls. How far it falls depends on the allocation method chosen, and defensible methods differ enough to move the answer materially, so the method must be disclosed alongside the number. A lifecycle honesty note applies to every gas pathway at once, on-site or utility-scale: upstream methane leakage adds to the total, its assumed rate is contested, and any study using a gas figure should state the leakage assumption it carries.

Solar with storage reads in the opposite pattern. Its generation is zero-emission at the source, and a battery that shifts midday surplus into the evening ramp does genuine work on the marginal ledger. The hourly lens is also where its limits show: the load still draws grid power through the night, round-trip losses mean more total energy consumed than delivered, and covering the last winter hours from on-site resources alone drives sizing that most sites will not carry. Annual certificates can cover the gap on the market ledger today at modest cost; hourly-matched procurement makes the stronger claim, from a supply of matching generation that is still thin and priced accordingly. Whether the annual version of the claim survives the revision described in Section 05 is exactly the open question.

No lens is the lens. An inventory needs the average. An operating decision needs the margin. A public claim needs the ledger and the clock named. The failure mode this paper exists to prevent is simpler than any of them: quoting the friendliest cell of the matrix as if it were the only one.

Section 08A reading protocol for emissions claims

Five questions dispose of most electricity-emissions claims that cross an executive's desk, including the ones in vendor studies, and including ours.

1. Which method. Is the figure an average or a marginal rate, and did the author say so unprompted? A number that does not name its method is an advertisement.
2. Which clock. Annual, monthly, or hourly, computed on which year's data, and how stale is that data relative to a fast-changing grid?
3. Which ledger. Location-based or market-based? If market-based: which instruments, matched over what period, deliverable from where, and what covers the unmatched remainder?
4. Which kind of claim. An inventory entry, which must sum with everyone else's, or a consequence claim, which must survive a counterfactual? Numbers built for one job are quoted for the other every day.
5. Does the verdict survive the other lenses? Projects that look green under every lens exist, but they are rare. What a decision-maker actually needs is to know which lenses flip the sign, and why, before someone else points it out.

This is also the discipline we apply to our own field. On-site generation, the subject of much of this series, looks worse under some of these lenses than its sellers prefer and better under others than its critics allow. The honest position is to show the whole matrix, with sources and dates attached, and to flag any recommendation that only works under one lens as exactly that. A number without its method is an advertisement, and a power decision should not run on advertisements, anyone's.