A strange thing happens when a large computing operation stops asking the grid for all of its electricity and instead builds generation beside the machines that consume it, because the electrons may never enter the public power system even though the fuel burned to produce them still adds emissions to the atmosphere. The Environmental Integrity Project’s inventory of 74 proposed or planned gas-fired power projects dedicated to data centers brings that distinction into sharp focus, particularly because the projects would provide electricity directly to computing loads rather than depend entirely on conventional grid interconnection. That arrangement changes the administrative pathway through which electricity reaches a data center, but it does not change the physical chemistry of natural gas combustion, methane leakage, turbine exhaust, or the resulting greenhouse-gas burden.

The useful starting point for lifecycle modeling is therefore not the data center’s utility bill but the physical chain that begins when gas enters the production system and ends when combustion products leave the turbine, because every boundary that molecule crosses can introduce an emissions source that a simple electricity forecast will miss. A credible model would distinguish upstream methane from combustion carbon dioxide, account for transmission and gathering processes, separate routine operating emissions from abnormal releases, and then connect those emissions to the actual runtime of the generation equipment serving the computing load. That approach also requires caution about the difference between permitted emissions, modeled emissions, expected operating emissions, and measured atmospheric emissions, since those categories answer different questions and analysts should never treat them as interchangeable.

Why a private generator still belongs inside a public climate forecast

The deeper issue is not whether a dedicated gas generator technically belongs inside a utility’s decarbonization pathway, but whether a national or corporate climate forecast can afford to leave it outside the physical model when the generator exists specifically because computing demand requires additional electricity. Utility decarbonization claims typically describe the emissions characteristics of electricity delivered through the grid, while dedicated generation can create a parallel supply route whose emissions never enter that same generation mix. A data center can consequently appear increasingly efficient from the perspective of the grid while simultaneously becoming attached to a new fossil combustion source that did not previously exist in the regional generation portfolio. A serious compute-emissions model therefore needs an additional physical layer that sits alongside corporate inventory accounting and asks a simpler question: what fuel is being consumed, where is it consumed, and what emissions result from that consumption? 

A cleaner grid can reduce the emissions intensity of electricity purchased from that grid, while new dedicated gas generation can increase total fossil combustion associated with computing loads outside the grid’s generation portfolio. The conflict emerges when an emissions forecast treats grid carbon intensity as a proxy for the carbon intensity of computing itself, because that proxy assumes that the electricity supply route remains stable as demand grows. NERC’s work on large loads demonstrates that data center demand already creates unusual forecasting requirements because these loads can operate for long periods and can behave differently from conventional commercial demand. The model must instead distinguish between electricity that displaces existing generation, electricity that requires new generation, and electricity that triggers dedicated generation outside the normal grid structure.

Beyond The Smokestack: Modeling The Full Methane to Megawatt Chain

A gas-fired generator does not begin its climate impact when methane reaches the combustion chamber, because the molecule has already traveled through a chain of wells, gathering systems, processing equipment, transmission infrastructure, storage assets, and delivery networks before it becomes electricity. Lifecycle modeling has to preserve that chronology because methane released before combustion can have a materially different atmospheric effect from carbon dioxide released after combustion, particularly across different climate-assessment time horizons. EPA maintains a detailed greenhouse-gas inventory for natural gas and petroleum systems and continually updates its methods as new data become available, which illustrates how upstream estimation remains an evolving measurement problem rather than a fixed coefficient that analysts can apply indefinitely.

The modeling lesson is straightforward: an emissions factor attached only to fuel combustion cannot represent the full climate footprint of gas used for dedicated power generation. A compute forecast that converts projected electricity demand into gas consumption should therefore carry upstream methane intensity as an explicit variable rather than burying it inside a single generic lifecycle factor. Upstream methane deserves particular attention because its distribution across the supply chain is not necessarily smooth, and an average factor can obscure the operational events that create disproportionately large releases. Production sites, gathering systems, processing equipment, transmission infrastructure, storage operations, and other components can experience leaks or abnormal releases that differ substantially from the assumptions used in inventories based on equipment counts and standard emission factors.

Why “natural gas” cannot be treated as one emissions variable

For compute forecasting, the most useful lifecycle architecture separates the gas system into identifiable stages rather than treating fuel consumption as the only meaningful activity variable. The first stage covers extraction and production, where methane can escape through equipment, wells, and operational events before the gas reaches the gathering network. The next stage follows gathering, processing, transmission, and storage, where compressors, valves, connections, and other infrastructure create additional opportunities for emissions that may not correlate perfectly with the volume ultimately delivered to a generator. The final stages cover combustion and post-combustion emissions, including carbon dioxide produced when the gas burns and other pollutants associated with turbine operation. EPA’s methodology demonstrates that national inventories already divide natural-gas emissions into distinct segments, while the scientific literature shows that measurement campaigns can reveal patterns that equipment-based estimates alone may struggle to capture.

The value of this structure becomes clearer when a data center operator changes its power strategy, because the same computing workload can produce different lifecycle outcomes depending on whether it receives electricity from the grid, uses dedicated gas generation, relies on storage, or combines several resources. A grid-based forecast can estimate emissions from the electricity mix, but a dedicated gas scenario requires the model to introduce fuel demand and then propagate that demand upstream through the natural gas system. A renewable-plus-storage scenario introduces manufacturing, charging losses, storage degradation, and replacement assumptions that belong to a different lifecycle pathway, while a hybrid system creates several interacting pathways that one emissions factor cannot represent faithfully. This does not mean every model must become an enormous simulation with every component represented at maximum resolution.

When Scope 1 Becomes Scope Everyone: The Data Center Accounting Trick

The most important distinction in the accounting debate is that Scope 1, Scope 2, and lifecycle emissions answer different questions, so a technically correct Scope 2 claim can coexist with a physically significant fossil-generation pathway without either statement necessarily contradicting the formal rules under which it was prepared. If a data center obtains electricity from a dedicated generator whose organizational relationship places the generator outside the reporting company’s operational boundary, the direct combustion emissions may not automatically appear as Scope 1 for the data center operator. This distinction becomes increasingly important when companies use market-based Scope 2 accounting to reflect contractual instruments associated with electricity procurement, because the contractual allocation of attributes can differ from the underlying flow of electrons across the power system.

The reporting boundary is not the physical boundary

Market-based accounting exists for a legitimate reason because companies need a standardized method for reflecting contractual choices about electricity supply, yet the method does not claim that certificates physically redirect electrons through a grid or erase emissions from generators operating elsewhere. The GHG Protocol’s Scope 2 framework uses contractual instruments for market-based accounting and distinguishes that method from location-based accounting, so the reported market-based result should not be treated as a direct measurement of every physical electricity flow serving a load. A company may report a lower market-based Scope 2 result when its electricity procurement satisfies the applicable contractual requirements, while separate fossil generation can supply part of the physical load, creating a distinction between the reported procurement attribute and the broader physical emissions pathway. 

The emerging problem is therefore less about discovering a single accounting trick than about understanding the gap between inventory claims and system-level consequences. GHG Protocol’s current revision process reinforces that the treatment of electricity accounting remains an active area of development, with proposed changes addressing issues such as hourly matching and deliverability in market-based reporting. Those proposals matter because they attempt to bring contractual electricity claims closer to the time and place where consumption occurs, even though they remain part of a standards-development process rather than a finalized universal rule. For compute forecasting, the implication is clear: temporal and geographic matching can improve electricity accounting, but neither concept removes the need to identify dedicated fossil generation that physically serves computing loads outside conventional grid supply.

Clean electricity claims need a physical power check

The physical power check begins with a deceptively simple question: what generator is actually producing the electricity consumed during the hours when the computing load operates? That question shifts the analysis away from annual procurement totals and toward the operational relationship between demand and supply, which becomes especially important for data centers that require continuous and highly reliable electricity. A power purchase agreement or energy attribute certificate can establish a contractual claim under the applicable Scope 2 framework, but the physical system may still require dispatchable generation when renewable output does not coincide with demand or when grid constraints prevent contracted resources from serving the load in the relevant location.

Dedicated gas generation introduces another pathway because it can provide power without requiring the computing load to wait for conventional grid expansion, but it also creates an asset whose emissions trajectory can extend far beyond the annual reporting cycle. A climate commitment that measures progress through annual Scope 2 values can therefore miss the strategic significance of a generation decision that changes the physical energy architecture supporting future compute capacity. The answer is not to replace corporate inventories with a new accounting system, but to place those inventories inside a wider physical model that reveals how procurement, generation, transmission, fuel supply, and computing demand interact.

24/7 Is Not Carbon-Free If Dedicated Generation Runs Continuously

The word ‘backup’ can conceal a major modeling assumption when a generator initially described as resilience infrastructure is also used as a regular source of electricity for a computing load, because a generator that operates routinely has a different emissions profile from one used only during defined reliability events. Traditional backup generation exists to serve exceptional conditions, while a dedicated power plant designed around continuous computing demand can become part of the normal operating architecture of the site. That distinction changes the emissions profile because fuel consumption depends on runtime rather than on the label attached to the generator. A turbine that operates only during outages produces a fundamentally different lifecycle profile from one that runs whenever the computing system needs electricity, even if both assets appear under a broad category such as onsite generation.

The difference between backup power and dedicated baseload

Continuous computing also changes the meaning of utilization because the load itself can remain present even when individual computing tasks fluctuate substantially. AI training, inference, storage, networking, cooling, and supporting systems do not all consume power in identical patterns, but the facility can still maintain a persistent electrical requirement that makes firm generation economically valuable. NERC has specifically noted that data center load factors can be more round-the-clock than other categories of demand, which matters for both grid planning and dedicated generation economics. A fossil generator built beside such a load can consequently operate for reasons that have little resemblance to the peaking role historically associated with gas-fired generation. The climate consequence follows directly from that operating pattern because higher utilization increases fuel consumption and therefore increases combustion emissions, while also extending the period over which upstream gas production and transport support the generator.

The concept of twenty-four-hour clean electricity therefore requires more than a clean annual procurement statement, because continuous demand creates a temporal problem that annual averages can conceal. A data center can contract for renewable energy while still requiring firm power during periods when the contracted resource cannot physically meet the load, and a dedicated gas plant can fill that role if the surrounding system lacks sufficient alternatives. The emissions model should represent that interaction rather than assume that a clean procurement instrument automatically converts every operating hour into a zero-emission hour. GHG Protocol’s proposed hourly matching and deliverability revisions recognize this challenge within the market-based accounting framework, showing that the timing and location of electricity procurement have become central questions in the evolution of corporate electricity reporting.

The utilization curve is the bridge between compute and combustion

The utilization curve should become one of the central variables in any forecast linking AI demand to dedicated gas generation because it translates the abstract growth of computing into the operating behavior of physical equipment. A project with the same nameplate capacity can have radically different emissions outcomes depending on whether it runs intermittently, follows a variable load, operates continuously, or serves as a persistent source of firm electricity. That makes capacity factor more than an engineering statistic; it becomes the bridge between computing intensity and fossil fuel consumption. NERC’s large-load analysis provides a strong reason to treat data center demand differently from conventional commercial demand because these facilities can maintain high loads for long periods and can introduce operating characteristics that require new planning assumptions.

The next step is to connect the utilization curve to fuel consumption without assuming that all operating hours carry identical emissions characteristics. Gas consumption varies with generator efficiency, ambient conditions, maintenance requirements, operating load, and technology configuration, while methane emissions upstream vary with the characteristics of the supply system feeding the plant. A technically credible forecast can therefore represent generation as a function of load and operating state rather than multiplying nameplate capacity by an assumed constant runtime and calling the result a lifecycle estimate. This approach also allows the model to test the effects of efficiency improvements without confusing them with reductions in demand, because a more efficient turbine can lower fuel consumption per unit of electricity while still increasing total fuel use if the computing load grows faster than efficiency improves.

The Utilization Curve That Breaks Climate Models

The most consequential variable in a compute-emissions model may not be the amount of computing capacity installed, but the amount of time the power system must remain energized to support that capacity, because a generator attached to a persistent digital workload behaves differently from one designed around occasional reliability events. AI workloads can produce different electricity-demand patterns across training, inference and supporting data-center operations, making the relationship between computing activity and generation response important to represent when a forecast examines dedicated power resources. The International Energy Agency has emphasized that AI-driven data center demand is developing alongside rising power density, changing workload characteristics, and infrastructure bottlenecks, which means electricity demand cannot be treated as a static extension of conventional commercial consumption.

Compute growth needs a combustion forecast beside it

Model training can create sustained demand over long operating windows, inference can produce persistent demand with changing intensity, and newer AI applications can introduce bursts that place additional stress on power electronics, storage, and generation resources. The IEA’s recent work treats changing AI activity and data-center electricity demand as important elements of energy-system analysis, supporting a closer connection between computing demand and the electricity resources required to serve it. Once that connection exists, a forecast can determine whether a dedicated gas generator serves a residual load, follows the computing demand directly, remains available as firm capacity, or operates continuously because the economics of the project depend on persistent utilization.

For climate forecasting, that means the utilization curve cannot remain fixed while compute efficiency changes, because efficiency affects the economics and volume of computing activity that the system can support. A more efficient accelerator can reduce electricity per task while increasing the number of economically viable tasks, which can keep generation assets heavily utilized even as the underlying hardware becomes more efficient. Dedicated gas generation then becomes part of a feedback loop in which computing demand supports new power infrastructure, the availability of that infrastructure supports additional compute deployment, and the resulting utilization determines the fossil fuel consumption that enters the emissions model. The central forecasting task is therefore to model compute growth and generator runtime together, rather than treating one as the cause and the other as an unrelated infrastructure consequence. 

The model should follow the load, not the nameplate

Nameplate generation capacity tells an analyst what a plant could theoretically produce, but it does not explain what the plant will actually do once computing demand, grid conditions, fuel prices, maintenance requirements, storage resources, and contractual arrangements begin interacting. A dedicated gas plant connected to an AI workload can operate as firm generation, residual-load support, or a more flexible resource, and each role produces a different emissions trajectory even when the physical turbine remains unchanged. NERC’s long-term reliability work makes clear that large computing loads require careful treatment in system planning because their size, persistence, and operational characteristics can alter forecasts of demand and resource adequacy.

That approach also prevents a common forecasting error in which analysts assume that a more efficient grid automatically makes every new electricity load cleaner, even when the new load causes additional fossil generation to operate outside the grid’s conventional dispatch structure. The IEA’s analysis shows that natural gas remains part of the electricity supply response to growing data center demand, alongside renewables, nuclear power, storage, and other resources, which means the emissions outcome depends on how the entire supply portfolio evolves rather than on one headline procurement strategy. A compute-linked model should therefore calculate the residual electricity requirement after accounting for physically available clean generation and storage, then determine whether dedicated gas resources fill the remaining requirement.

Stranded Code or Stranded Carbon? The Long Infrastructure Lock-In

A new gas generator can become an unusually durable piece of infrastructure in an industry where the software workload it supports can change far faster than the energy asset itself, creating a mismatch between the lifecycle of computing and the lifecycle of combustion. AI models, accelerator architectures, software systems, and inference workloads can change on much shorter development cycles than power-generation assets, whose economic value depends on electricity demand continuing over the period required to recover and justify the investment. The IEA has highlighted this mismatch between the rapid development cycle of data center technology and the slower planning and construction cycle of energy infrastructure, making the relationship between computing growth and power investment a central issue in the energy-AI transition.

The infrastructure can outlive the computation that justified it

The carbon consequence of that mismatch depends heavily on what happens after the original computing forecast changes, because an underused gas asset can still create pressure to recover capital through continued operation while a fully utilized asset can become embedded in the local electricity system. The resulting emissions debt does not behave like software depreciation, where an outdated model can simply be replaced and removed from service without leaving a physical energy system behind it. A fossil generator can remain an available operating resource after the computing requirements that helped justify its change, potentially influencing subsequent power procurement, infrastructure planning, fuel arrangements, and reliability decisions. EIP’s inventory is therefore most useful when treated as a forward-looking infrastructure signal rather than as a guarantee that every identified project will operate according to its maximum potential.

A climate commitment has to survive the asset after the model changes

The harder question arrives when a company’s computing strategy changes but the generation asset remains economically and technically capable of operating, because the climate commitment then encounters a physical infrastructure decision that cannot be reversed as quickly as software deployment. Early retirement can reduce future combustion but may create financial losses, while continued operation preserves the economic value of the asset but extends the associated fossil emissions. The same tension can arise when a data center moves workloads, changes hardware, consolidates computing capacity, or adopts more efficient models, because the electricity infrastructure may have been designed around an earlier demand assumption. A credible emissions model should therefore include asset-retirement assumptions and not merely forecast annual emissions from expected operating capacity.

The value of this approach becomes clearer when the model separates emissions already released from future emissions that remain avoidable, because an operating asset does not create an irreversible emissions obligation simply by existing. What can be difficult to reverse is the investment and infrastructure pathway established by a generation project, particularly when the asset later becomes part of the electricity supply arrangement supporting a persistent computing load. A scenario model can therefore test the consequences of different retirement dates, utilization trajectories, clean-power substitutions, and fuel-system improvements without assuming that the plant will operate identically throughout its useful life. The core question is ultimately whether the infrastructure being financed today can remain consistent with the emissions trajectory promised for the years in which that infrastructure is expected to operate. 

From Benchmarks to Burden: Who Really Owns The Molecule?

Assigning emissions to computing becomes difficult when a digital workload creates demand for infrastructure that belongs to another part of the energy system, because several actors can participate in the chain without any single actor controlling every stage. The chip manufacturer enables the computation, the cloud operator provides the service, the data center converts electricity into computation, the generator produces electricity, the gas supplier delivers fuel, and the end user creates the demand that ultimately supports the economic activity. Treating the resulting emissions as belonging entirely to one participant would oversimplify the physical system and could distort incentives for reducing emissions at the stage where mitigation is actually possible. The GHG Protocol’s scope architecture exists partly to address this separation between direct operational control, purchased energy, and broader value-chain activity, although dedicated generation arrangements can create unusual boundary questions that require careful interpretation.

Attribution begins with causality, not blame

Causal attribution can become more useful when the model asks what changed because the computing workload existed, rather than asking who happens to own the generator on paper. Where project documents identify data-center demand as a principal reason for developing new gas generation, the computing load can be treated as a material driver of the associated generation requirement even when the generator belongs to a separate legal entity. That relationship does not mean the entire lifecycle footprint should automatically become the cloud operator’s Scope 1 emissions, because formal reporting boundaries still depend on ownership, control, contracts, and applicable accounting standards. It does mean that a system-level forecast should preserve the connection so that analysts can see how additional compute capacity changes the need for generation and how that generation changes fuel consumption.

The user sits at the end of this chain but should not become the automatic owner of every upstream emission associated with a digital interaction, because attribution must remain proportional to the analytical question being answered. A user-level footprint might reasonably estimate the electricity and associated emissions attributable to a service request, while an infrastructure-level assessment might assign emissions according to generation ownership, operational control, or economic causation. A national forecast requires another approach because it needs to determine how aggregate computing demand changes energy-system investment and fossil fuel consumption across the economy. These are different questions and should produce different attribution outputs rather than one universal carbon number.

Compute-linked emissions need a layered ownership model

A practical attribution framework can begin with four distinct layers: physical emissions, operational control, contractual allocation, and demand causation, because each layer answers a different question and prevents one accounting boundary from carrying more meaning than it can support. Physical emissions identify where methane and carbon dioxide enter the atmosphere, operational control identifies who controls the emitting equipment, contractual allocation identifies who purchases or claims the electricity attributes, and demand causation identifies which activity creates the economic requirement for the energy service. A dedicated gas plant serving computing demand can therefore appear in several analytical layers at once without creating a contradiction. The emissions belong physically to the combustion source, the relevant Scope 1 treatment depends on control, the electricity treatment depends on how the power is procured and reported, and the causal linkage to computing belongs inside the broader system model. 

This layered structure also makes scenario analysis more credible because the model can show how emissions change when one participant changes behavior without pretending that the entire system moves in lockstep. A cloud operator can improve workload efficiency, a data center can alter its electricity procurement, a generator can improve thermal efficiency, a gas supplier can reduce methane leakage, a grid operator can add firm low-carbon resources, and a user can change the volume or type of computing service consumed. Each intervention affects a different point in the chain, and a good model should preserve those differences so that decision-makers can see where the largest physical leverage exists. The IEA’s analysis of AI and energy reinforces this systems perspective by examining demand growth, supply responses, efficiency, infrastructure constraints, and emissions together rather than assigning the energy consequences of AI to a single actor.

You Can’t Forecast Compute Without Forecasting Combustion

The energy system cannot be treated simply as an external supply layer for computing demand, because the expansion of data centers can affect generation requirements, transmission needs, fuel demand, storage deployment, and resource-adequacy planning. The IEA’s current analysis shows that data center electricity demand is becoming a significant driver of power-sector investment and that multiple energy sources will participate in meeting the resulting requirement. NERC’s reliability assessments provide the complementary system perspective, demonstrating that large computing loads are changing demand forecasts and creating new planning challenges for electricity systems. EIP’s project inventory adds another layer by identifying dedicated gas generation associated with data center development, creating a concrete reason to incorporate fossil infrastructure into compute forecasting rather than treating it as an external assumption.

The next compute forecast must carry an energy system inside it

The strongest model would therefore connect compute demand to electricity demand, electricity demand to generation behavior, generation behavior to fuel consumption, fuel consumption to upstream methane and combustion emissions, and all of those variables to the evolving lifetime of the infrastructure involved. That architecture would allow analysts to separate announced projects from operating assets, distinguish backup generation from persistent generation, represent grid imports separately from dedicated generation, and test how changes in computing efficiency alter rather than simply eliminate energy demand. It would also allow climate commitments to be evaluated against the physical infrastructure supporting them instead of relying exclusively on annual accounting outcomes. EPA’s continuing updates to its natural-gas emissions inventory demonstrate why the fuel-supply layer requires ongoing methodological attention, while the GHG Protocol’s ongoing Scope 2 revision process demonstrates why contractual electricity claims and physical electricity supply require careful distinction as the standard evolves.

The central issue is ultimately not whether AI is inherently carbon-intensive, because the emissions outcome depends on the electricity system built around the computing demand and on the choices made about generation, fuel supply, efficiency, storage, and grid integration. A workload can become more efficient while total demand rises, a cleaner grid can reduce emissions while dedicated gas generation expands, and a renewable procurement strategy can improve reported electricity attributes while physical generation remains more complicated than the annual accounting suggests. Those outcomes can coexist because the digital system, the electricity system, and the corporate accounting system operate according to different boundaries and timescales. A credible forecast has to preserve those differences while still connecting them through a common physical model.