In late April 2026, the Federal Energy Regulatory Commission took a step that underscored just how dramatically artificial intelligence is reshaping the American power grid: it ordered PJM Interconnection, the nation’s largest grid operator, to rewrite its rules for data centers that want to plug directly into power plants. FERC declared PJM’s existing tariff “unjust and unreasonable” because it never anticipated a world in which a single AI training facility could consume as much electricity as a small city. The order affects a grid serving 65 million people across 13 states and the District of Columbia, and it signals that the collision between Silicon Valley’s appetite for power and the physical limits of the electrical system has become a federal regulatory priority.
The demand surge, by the numbers
The scale of the problem is no longer theoretical. The International Energy Agency reported that global data center electricity consumption surged in 2025, with data centers, AI workloads, and cryptocurrency mining now drawing enough power to materially influence national grids. The IEA’s executive summary on energy and AI found that the largest data center campuses are growing bigger, that AI-focused facilities are claiming a rising share of total electricity use, and that a relatively small number of hyperscale sites account for a disproportionate slice of consumption. Efficiency gains in chips and cooling systems continue, the agency noted, but they are being outrun by the sheer expansion of computational demand.
That mismatch is what makes the grid problem so acute. Interconnection queues in the United States now stretch years, sometimes more than a decade, for new projects seeking to connect to the transmission system. For companies racing to train the next generation of AI models, waiting in line is not a viable strategy. The result is a scramble for alternatives: dedicated power plants wired directly to data center campuses, bypassing the congested public grid entirely.
Nuclear restarts and the Amazon-Talen deal
The most prominent example of that scramble involves Amazon and Talen Energy. Talen announced plans to expand its nuclear energy relationship with Amazon, building on an arrangement to supply power from the facility now known as the Christopher M. Crane Clean Energy Center. That plant is the former Three Mile Island Unit 1 reactor in Pennsylvania, which ceased operations in 2019. The U.S. Nuclear Regulatory Commission maintains a dedicated project page tracking the restart effort’s filings and regulatory milestones.
The deal is significant for what it reveals about Big Tech’s calculus. Rather than wait for new transmission lines or compete for scarce grid capacity, Amazon chose to invest in bringing a dormant nuclear plant back to life. Nuclear generation offers what AI operators prize most: around-the-clock, weather-independent power with minimal carbon emissions. But restarting a reactor that has been offline for years is not simple. It requires extensive engineering reviews, safety assessments, and NRC licensing steps that can stretch over multiple years. The regulatory filings reviewed for this article do not confirm a final restart date or a definitive schedule for major milestones, meaning the timeline for commercial operation remains uncertain.
Amazon is not alone in looking at nuclear. The IEA documented growth in the pipeline of conditional offtake agreements for small modular reactors, with technology companies signing letters of intent for next-generation nuclear designs that have not yet received design certification or construction permits. The word “conditional” matters: the gap between a signed agreement and a functioning reactor delivering electricity could span a decade or more. Interest is real and growing, but the installed capacity is not yet there.
Gas enters the picture
Nuclear is only part of the story. FERC’s order to PJM addresses large loads co-locating with generation broadly, a category that encompasses gas-fired plants alongside nuclear and renewables. Natural gas offers something nuclear currently cannot: speed. A gas peaker plant or combined-cycle facility can be permitted and built in a fraction of the time it takes to restart or construct a nuclear reactor, making gas an attractive bridge for companies that need power within two to three years rather than ten.
Several major technology firms have explored or pursued gas-powered generation for their data center operations. Microsoft has engaged with Constellation Energy on gas-related supply arrangements, and Meta has faced public scrutiny over proposals for gas-fired facilities tied to its AI infrastructure. These moves have drawn criticism from environmental groups and some state regulators who argue that locking in new fossil fuel capacity for decades contradicts the same companies’ net-zero climate pledges.
The tension is real. Gas plants can be built quickly and run reliably, but they produce carbon emissions that complicate corporate sustainability targets. For companies that have publicly committed to carbon-negative or net-zero operations, signing long-term gas contracts creates a credibility problem, even if the immediate engineering logic is sound.
What FERC’s order actually requires
FERC’s directive to PJM is a framework order, not a finished rulebook. The commission found that PJM’s tariff lacked clear guidelines for the co-location arrangements that data center operators are increasingly pursuing, and it ordered PJM to develop a new framework that accommodates these deals while protecting residential and commercial ratepayers from bearing the costs of grid instability.
What the order does not specify, at least in the materials reviewed for this article, is a firm compliance deadline or the precise mechanisms PJM must use to allocate costs between data center operators and existing customers. PJM has not yet publicly detailed its compliance timeline or indicated whether it will seek to narrow portions of the directive. Until those details emerge, the practical effect on electricity prices for households and businesses across PJM’s territory remains an open question.
The stakes are considerable. If co-located data centers draw power that would otherwise flow to the broader grid, system operators could face reliability challenges during peak demand periods. If the costs of reinforcing the grid to handle these new loads are spread across all ratepayers, families and small businesses could end up subsidizing Big Tech’s AI ambitions. FERC’s order acknowledges these risks but leaves the specifics to PJM’s rulemaking process.
The renewable gap no one is ignoring
It is worth noting what this shift toward gas and nuclear does not mean: technology companies have not abandoned renewable energy. Google, Microsoft, Amazon, and Meta collectively hold some of the largest corporate portfolios of wind and solar power purchase agreements in the world. But renewables alone have not been able to keep pace with the concentrated, continuous power demands of AI data centers. Solar generates nothing at night. Wind is intermittent. Battery storage at the scale needed to back up a multi-hundred-megawatt data center campus around the clock remains expensive and, in many regions, unavailable.
That gap is what drives the turn toward gas and nuclear. These are not replacements for renewables but additions to a power strategy that increasingly treats electricity as a strategic resource rather than a commodity to be purchased off the grid.
Where the uncertainty sits
Several critical questions remain unresolved as of May 2026. The nuclear restart timeline for the Crane Clean Energy Center has not been finalized. PJM’s new tariff rules have not been written. The full scope of gas-fired generation contracts tied to AI data centers is not yet visible in public regulatory filings. And the IEA’s documentation of small modular reactor interest, while credible, describes a pipeline of conditional deals that depend on regulatory approvals years away.
The balance between efficiency improvements and demand growth is also unsettled. Chip designers continue to reduce the energy required per computation, and data center operators are deploying more sophisticated cooling and power management systems. But the IEA’s analysis suggests these gains are being overwhelmed by the pace at which AI workloads are expanding. Whether that dynamic holds, or whether a breakthrough in hardware efficiency changes the equation, is something no one can predict with confidence.
What is clear, grounded in binding regulatory orders and international energy data, is that AI’s electricity appetite has already grown large enough to force structural changes in how power is generated, delivered, and governed. The companies building the AI future are no longer content to buy power off the shelf. They are reaching back into the energy system itself, reviving retired nuclear plants, contracting for new gas generation, and pushing regulators to rewrite rules that were designed for a world that no longer exists.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
