Seventy-one percent. That’s the increase in planned non-renewable energy capacity additions for 2025-2026, per analysis from the Department of Energy and the American Action Forum. Not a projection. A count of what’s already been planned and permitted. The grid is being built right now, and it runs on gas.
That one figure carries a consequence that the briefing format can’t fully unpack. So this piece will.
The Numbers: What DOE and AAF Actually Reported
Natural gas’s share of planned US grid additions climbed to approximately 18.1% for the 2025-2026 period, up from 11.1% in 2024, per AAF data. Renewable capacity growth, by contrast, ran at approximately 2% over the same period, per AAF analysis. The DOE’s modeling puts AI data center electricity demand on a trajectory toward up to 12% of total US electricity consumption by 2028, not a measurement, a projection, but one anchored in the capacity addition data being built today.
The cost differential driving this is not subtle. Per AAF’s analysis, natural gas grid connection costs approximately $24 per kilowatt. Solar runs approximately $253 per kilowatt, a roughly 10x gap. AAF is a market-oriented policy advocacy organization, and its data selection reflects that orientation. But the directional finding aligns with documented interconnection queue challenges: the line to connect utility-scale solar to the grid now stretches years in most major markets.
One note on AAF’s sourcing: these figures should not be treated as neutral engineering data. They come from an organization with a stated policy preference for market-based energy solutions and skepticism of renewable mandates. That said, the interconnection queue problem they describe is widely documented by parties with no such orientation, including grid operators and the DOE itself. The direction of the cost gap is reliable. The specific dollar figures carry the advocacy-source caveat.
Why Renewables Stalled
The AI workload problem is not a power-quantity problem. It’s a power-quality problem. AI training and inference clusters demand 24-hour, 365-day baseload power at high reliability. They can’t tolerate intermittency. Wind and solar, without storage, are intermittent by definition.
Interconnection queues compound the problem. The pipeline for utility-scale solar and wind connections to the US grid has grown to multi-year backlogs in most regions. A data center operator who needs power in 18 months cannot wait for a solar project that requires 36 months to interconnect. Natural gas plants, by contrast, often connect faster and predictably. That’s not an environmental preference; it’s a procurement constraint.
Combined, these factors explain the 2% renewable growth against a 71% non-renewable surge. The grid is responding to what AI operators can actually buy on the timelines they need, not what ESG frameworks say they should buy.
The ESG Commitment Gap
Several major hyperscalers have made explicit public commitments to 100% renewable energy operations or net-zero carbon footprints by specific dates. The precise language of those commitments matters: some are structured as power purchase agreements for renewable capacity matching, not as real-time renewable sourcing. Others rely on renewable energy certificates that represent clean generation somewhere on the grid, not necessarily electricity flowing to a specific facility.
Operators with public renewable commitments face new pressure from this data. That framing, from the Filter’s qualified language instruction, is the right one. The DOE/AAF data does not demonstrate that any specific operator is in breach of its stated commitments. It does demonstrate that the grid being built to power AI does not match the profile those commitments assumed.
The mechanism is structural. If natural gas accounts for 18.1% of new grid capacity while renewables add at 2%, the overall generation mix serving data centers, including those with renewable commitments, shifts toward gas unless operators are sourcing power through dedicated renewable facilities operating as physical islands rather than grid-connected assets. Very few can do that.
The Operational Paths Forward
Operators facing this gap have four realistic options, none of which resolves the tension cleanly.
*Nuclear.* Long lead times make new nuclear capacity irrelevant to 2026 or 2027 demand. Small modular reactors are years from commercial deployment at scale. Existing nuclear plants near data center clusters are valuable but finite, and already being contracted by hyperscalers who moved early.
*Fuel cells.* Oracle’s Project Jupiter is the most visible current example: a decision to run on 2.45 gigawatts of Bloom Energy fuel cells rather than gas turbines. That April brief documented the fuel-cell-versus-gas choice as an operational decision with ESG implications, fuel cells can run on natural gas, but with higher efficiency and lower direct emissions than turbines, and they’re compatible with hydrogen blends as that market develops. This is the most credible near-term alternative to gas turbines for operators with aggressive carbon timelines.
*Carbon offsets and RECs.* Renewable energy certificates and carbon offsets allow operators to claim renewable equivalence without physically sourcing renewable power. These instruments are defensible under many current ESG frameworks but face increasing scrutiny from regulators, investors, and environmental groups as the physical gap between claimed and actual performance widens. Voluntary carbon markets have also faced credibility challenges on additionality.
*Demand-side efficiency.* Reducing the energy intensity of AI workloads through inference optimization, model compression, and hardware efficiency upgrades is the one path that doesn’t depend on grid policy or carbon markets. It’s also the one least visible in infrastructure announcements. The energy-per-token metric will matter more as AI demand scales.
Investor and Regulatory Implications
Energy investors should read the 71% figure as confirmation, not signal, that the gas infrastructure buildout for AI is underway and backed by committed capacity. The investment thesis that KKR reportedly acted on this week, and that NextEra’s April federal approval reflects, is consistent with this data. The question for energy investors is duration: how long does gas remain the preferred AI power source before nuclear, fuel cells, or transmission upgrades change the economics?
For AI infrastructure operators, the near-term regulatory watch is the House Subcommittee on AI power demand and permitting reform. The April 29 hearing addressed grid cost and ratepayer protection, the gap between what AI energy demand costs and who bears that cost. If permitting reform accelerates gas approvals, the capacity addition trend steepens further. If ratepayer protection provisions impose cost-sharing requirements on data center operators, the economics of gas-powered AI infrastructure change.
ESG-focused investors and sustainability officers face a harder question. Carbon accounting frameworks for AI operations are not yet standardized. A data center that procures renewable certificates against gas-grid power has a very different emissions profile than one with dedicated renewable generation, but both can report “100% renewable” under common frameworks today. As reporting standards tighten, the distinction will matter more.
TJS Synthesis
The 71% natural gas surge is not evidence that hyperscalers are abandoning their sustainability commitments. It’s evidence that the grid those commitments depend on isn’t being built to fulfill them. That’s a different problem, one that RECs and carbon markets can paper over in annual reports but can’t resolve in the physical infrastructure.
Operators who want to close this gap have a short list of credible options: fuel cells (Oracle’s path), dedicated renewable-plus-storage microgrids, or long-dated nuclear power purchase agreements with plants that already exist. Everything else involves accepting that the claimed renewable profile and the actual grid profile diverge, and managing the reporting risk that comes with that gap widening publicly.
The DOE data gives that gap a number. It will be harder to ignore in the next sustainability report.