The United States is preparing to build more electric generating capacity in a single year than at any point in modern history. According to the U.S. Energy Information Administration’s latest Preliminary Monthly Electric Generator Inventory, developers plan to add 86 gigawatts (GW) of new utility-scale capacity to the grid in 2026. If realized, it would mark the largest annual expansion in decades.
The composition of that growth underscores the scale—and seriousness—of the energy transition:
- 43.4 GW of solar
- 24.3 GW of battery storage
- 11.8 GW of wind
- 6.3 GW of natural gas
Solar accounts for just over half of all planned additions. Battery storage, once marginal, now stands as a core structural component of grid reliability. Wind rebounds after a period of delay. Natural gas, while still present, plays a comparatively modest role in the official expansion cycle.
These figures come from Form EIA-860M, the agency’s monthly update to its Annual Electric Generator Report. The survey tracks existing and proposed generating units at plants with one megawatt or more of nameplate capacity. By integrating monthly filings with the annual EIA-860 dataset and ongoing research on new projects, the agency produces the most authoritative snapshot available of the nation’s current and near-term generating inventory.
The data are rigorous. They are comprehensive within their mandate. They capture the shared grid’s formal expansion. But they do not capture everything — because a new category of infrastructure is emerging beyond the traditional boundaries of grid interconnection reporting.
The Demand Shock
For years, electricity demand in the United States crept upward in modest increments. Efficiency improvements blunted the impact of economic growth. Even as digital services expanded, the power system absorbed new load without requiring a wholesale redesign.
The data centers being built to train and run advanced AI models consume power on a scale that would have seemed implausible a decade ago. A single campus can require several hundred megawatts of electricity—enough to power a small city. Some proposed facilities approach or surpass one gigawatt, rivaling the output of a large power plant.
AI workloads run around the clock. Downtime translates directly into lost value. The economic calculus is unforgiving: the more reliable the power supply, the more valuable the computational output.
But in many parts of the country, plugging into the grid is no longer a straightforward proposition. Interconnection queues stretch years into the future. Transmission projects face permitting battles and rising costs. Utilities struggle to provide firm delivery timelines amid surging demand.
For companies racing to secure advantage in artificial intelligence, waiting for infrastructure upgrades is not a neutral choice. It carries competitive consequences. So a shift is underway. Rather than delay construction, some developers are designing their own power supply into the blueprint—bringing generation onto the campus itself.
The Emergence of a Parallel System
From Texas to Pennsylvania, from Wyoming to West Virginia, a different kind of power buildout is taking shape—one that does not always flow through the traditional channels of the grid.
Developers serving AI data centers are constructing significant volumes of behind-the-meter natural gas generation directly on or adjacent to campus sites. In Ohio and New Mexico, in Tennessee and Utah, gas turbines are being embedded into data center designs as core infrastructure, not temporary backup. Some projects pair gas with on-site solar and battery storage. Others rely primarily on combustion turbines or combined-cycle units built for continuous operation.
These facilities are typically permitted through state environmental agencies or siting boards. They may operate under private-wire arrangements or partially self-supplied configurations. And because many are not seeking full interconnection to wholesale markets, they do not always appear in regional transmission queues.
This is not a lapse in federal data collection. The Energy Information Administration’s mandate is clear: to track utility-scale generation that interconnects with the electric system through formal reporting channels. Behind-the-meter generation serving a specific campus has historically fallen outside that statistical perimeter.
To understand the magnitude of this parallel buildout, AIxEnergy compiled a national dataset drawing from state air permits, siting-board dockets, county commission filings, corporate disclosures and earnings calls, equipment procurement signals and probability-weighted commercial operation timelines. The result is a second ledger—not a correction to EIA data, but an expansion beyond its traditional boundary.
Grid Growth vs. Shadow Growth (2026–2028)
All values in gigawatts (GW).
| Year | Battery Storage | Grid Gas | Expected Shadow Gas | Solar | Wind | Total | Grid Total | Shadow Total | Cumulative Grid Carbon Added | Cumulative Shadow Carbon Added |
|---|---|---|---|---|---|---|---|---|---|---|
| 2026 | 24.3 | 6.3 | 2.4 | 43.4 | 11.8 | 88.2 | 61.5 | 26.7 | 13 | 5 |
| 2027 | 49.8 | 12.3 | 21.7 | 89.3 | 20.1 | 193.2 | 121.7 | 71.5 | 26 | 49 |
| 2028 | 64.6 | 26.9 | 23.0 | 112.7 | 22.6 | 249.9 | 162.3 | 87.6 | 57 | 51 |
| 2026–2028 Cumulative | 139 | 46 | 47 | 245 | 55 | 531 | 345 | 186 | — | — |
Between 2026 and 2028, expected shadow gas additions—47 GW—closely track the 46 GW of gas planned for grid connection. In 2027 alone, probability-weighted shadow gas capacity, at 21.7 GW, exceeds the 12.3 GW slated for the grid.
Over three years, shadow projects represent 186 GW of private capacity exposure. The EIA data accurately reflects the expansion of the shared grid. The supplemental dataset captures a parallel build cycle unfolding beside it. Taken together, they point to a bifurcated infrastructure moment—two systems growing at once, under different rules.
Invisible Carbon Emissions
What makes the shadow buildout especially consequential is not simply its scale, but its carbon trajectory. Under conservative assumptions — a 60 percent gas capacity factor and an emissions intensity of 0.40 metric tons of CO₂ per megawatt-hour — the numbers are striking.
By 2027, gas plants added to the shared grid since 2025 would account for roughly 26 million metric tons of carbon dioxide per year. Over the same period, probability-weighted shadow gas additions would produce an estimated 49 million metric tons annually.
By 2028, grid-connected gas additions would reach roughly 57 million metric tons per year. Shadow additions would stand near 51 million metric tons annually. In other words, the private, behind-the-meter fleet is adding almost as much new carbon pollution as the officially recorded gas expansion.
These emissions are largely invisible in national decarbonization narratives. They sit outside the planning frameworks used to track grid carbon intensity. They are not always incorporated into state clean energy targets built around utility procurement. They do not appear in the interconnection queues that anchor transmission planning.
Over the past decade, much of the country’s carbon reduction strategy has focused on cleaning the shared grid: retire coal, add renewables, expand storage, and electrify demand. The model assumes that as the grid becomes cleaner, electrification reduces emissions economy-wide. When large new loads instead self-supply with dedicated gas generation, that feedback loop weakens. The grid can decarbonize on paper while a growing share of marginal demand is met by private fossil infrastructure.
AI-driven demand is accelerating into a system designed around shared cost recovery and centralized decarbonization pathways. If a meaningful portion of that demand is met outside those pathways, emissions accounting becomes bifurcated. One ledger shows progress. The other shows persistence.
If shadow gas additions continue to track grid gas additions one-for-one, the carbon impact of the AI era will be determined as much by private permitting decisions as by public clean energy targets. The emissions are not hidden. But without an integrated view, they are easy to underestimate.
System Architecture Shift
For more than a century, the American electric system has operated as a shared enterprise. Utilities build power plants, transmission lines and distribution networks designed to last decades. The costs of that infrastructure are recovered across a broad base of customers—households, businesses and industry alike.
When large data centers partially or fully power themselves, the shared system does not disappear. Transmission lines still require maintenance. Storm hardening and cybersecurity investments still need funding. Legacy generation assets still carry fixed costs that must be recovered. The physical backbone of the grid remains in place.
As some of the largest and fastest-growing loads reduce their reliance on grid-supplied power, the economics of the shared model begin to shift. The system’s obligations persist, but the distribution of who pays—and how much—can evolve.
This is not a judgment on developers’ choices. Interconnection delays are real. Transmission constraints are real. Reliability requirements for AI infrastructure are stringent. Companies racing to deploy high-performance computing cannot afford uncertainty. But the consequence is structural. Power strategy is no longer a downstream utility matter; it is embedded directly into AI campus design and capital planning. The United States is not simply adding generation capacity. It is redefining the relationship between large loads and the grid itself.
Two Systems, One Future
The EIA’s headline figure — 86 gigawatts of planned additions in 2026 — is a genuine milestone. Solar and battery storage are scaling at levels that would have seemed improbable a decade ago. The shared grid is expanding at a pace not seen in a generation.
The supplemental ledger compiled by AIxEnergy does not dispute that progress. It reframes it. Alongside the renewable surge, a parallel build cycle is accelerating — one anchored largely in natural gas and structured outside traditional grid interconnection pathways. It is not recorded in the same way because it is not built in the same way.
Taken together, the two datasets reveal more than rapid growth. They reveal structural divergence. One system is being planned, regulated and decarbonized through public institutions. The other is being permitted, financed and deployed through private infrastructure strategies tied directly to AI deployment timelines. Both are rational responses to real constraints. But they operate under different incentives, different accounting frameworks and, increasingly, different carbon trajectories.
This is not simply a story about generation totals. It is a story about governance. If large-scale self-supply becomes a durable feature of the AI economy, the consequences extend beyond emissions. Cost recovery models may shift. Transmission utilization patterns may change. Ratepayer risk allocation could evolve. Infrastructure planning assumptions — built for a century around shared participation — may need revision.
The United States is entering an era in which electricity is no longer just a utility service. It is a strategic input into computational power, national competitiveness and industrial policy. In that environment, power decisions move from the periphery of corporate strategy to the center.
The country is building a record clean energy expansion. It is also constructing a privately structured generation layer to power artificial intelligence. Those two systems now coexist. How they interact will determine not only carbon outcomes, but the economic architecture of the grid itself. The grid is growing. The shadow is growing with it. The future of American power will be shaped by both.
