The Nuclear Regulatory Commission determined this summer that resuming operations at Three Mile Island's Unit 1 reactor — a facility long associated with the 1979 nuclear accident — would not produce a noteworthy environmental impact. Constellation Energy has projected that the facility, to be rebranded as the Crane Clean Energy Center, will resume electricity generation by late 2027, following a 20-year contract valued at $16 billion with Microsoft, under which the reactor's output will be delivered directly to the technology company's data centres rather than fed into the regional grid.

Several weeks afterward, PJM's capacity auction — which sets electricity pricing for roughly 65 million people — reached its price ceiling for the third consecutive year, at $329 per megawatt-day, a figure nearly nine times higher than the 2024 level. According to regulators, data centres account for 63% of the increase, representing $9.3 billion in costs now borne by residential and commercial customers.

Electricity has emerged as the one constraint that the artificial intelligence industry has been unable to bypass. Prior to around 2023, discourse on artificial intelligence competition centred almost exclusively on semiconductor hardware: which companies could secure sufficient Nvidia graphics processors, and which nations Washington's export restrictions would permit to obtain them. The framework retains relevance — the UAE was elevated to the highest approved tier only in July 2026, with semiconductor access remaining limited to a select group of authorised entities.

Training and operating advanced AI models now requires electricity at a scale that goes well beyond hardware considerations alone, with a single next-generation AI facility capable of drawing between 500 megawatts and more than one gigawatt of power continuously — an amount comparable to the consumption of a mid-sized city. From 2025 onward, power grid operators in the United States have faced mounting delays in bringing new high-demand connections online, with waiting periods in data-centre-heavy regions extending between four and seven years, transformer procurement timelines reaching up to five years, and only a portion of announced capacity having moved to construction.

Those investment commitments explain why Microsoft, Google, Amazon, and Meta have each entered into multi-billion-dollar agreements spanning decades to restart dormant reactors or secure capacity from facilities not yet built. Semiconductors are interchangeable commodities that can be acquired through purchase, leasing arrangements, third-party intermediaries, or illicit cross-border transfer. Reliable, continuous power at that scale cannot.

The infrastructure involved includes power transmission lines, transformers with procurement lead times spanning multiple years, and either an established nuclear fleet or a government prepared to fund construction of a new facility. The United States relaxed semiconductor export restrictions for the UAE in July 2026, placing the country within its highest-ranked trade tier.

Funding does not appear to be a limiting factor: HUMAIN, based in Saudi Arabia, has pledged 6.6 gigawatts of AI infrastructure capacity by 2034. As of that same period, approximately 4,000 megawatts of AI infrastructure capacity announced by Saudi Arabia and the United Arab Emirates had yet to record any energisations — the industry term for a facility being connected to the electrical grid. Riyadh currently operates around 467 megawatts of data-centre capacity, a figure that falls well below the city's 6.6-gigawatt development target.

Transformer procurement lead times and utility interconnection agreements — spanning roughly four to seven years — establish the governing timeline for infrastructure buildout, one that semiconductor availability alone cannot shorten. Authorisation from US authorities is needed to procure the chips. The existing measures have proven inadequate.

A parallel limitation holds true domestically as well. According to one industry tracking source, construction at the Stargate project's primary Texas location had recorded little measurable advancement by April 2026, despite the availability of both funding and chip supply throughout the period. Across the country, approximately 12 gigawatts of AI data-centre capacity had been pledged for development in 2026, yet construction commenced on no more than around 5 gigawatts. Grid connection waiting periods across Northern Virginia, Phoenix, and Dallas currently stretch anywhere from four to seven years.

The shift toward nuclear power represents one approach the sector has adopted, structured around regulatory frameworks comparable to those governing export controls. Decisions to reactivate Three Mile Island, place advance orders for TerraPower's Natrium reactors, and align Amazon with X-energy's small modular reactor programme reflect a shared calculation that dedicated generation capacity, even with lengthy lead times, offers a more dependable path than relying on a transmission network unable to keep pace with rising demand.

Deploying that technology internationally follows the same bilateral approval framework applied to semiconductor export tiers: each partner government is required to conclude a "123 Agreement" with Washington on civil nuclear cooperation, and executive orders issued in 2026 established a goal of twenty such agreements by 2028. The majority of small modular reactor concepts have yet to achieve commercial viability, with projected timelines for both TerraPower and Oklo extending to 2032 or later.

Demand for gas turbines already extends several years into the future, while renewable energy sources face limitations in delivering the uninterrupted baseload power that frontier data clusters require, and the pool of dormant nuclear reactors available for recommissioning remains limited. Large technology companies are turning to nuclear power as one pathway to meet their demand for reliable, utility-scale electricity capacity within the coming years.

In the base scenario, assigned a roughly 55% probability, the Crane facility and a few other dormant US reactors return to operation near their projected timelines by 2027–28, while an initial cohort of small modular reactors achieves commercial operation by 2030. Such projections depend on NRC approval timelines remaining on track and no lead-build programme encountering the kind of cost and schedule overruns that affected Vogtle. The United States, alongside hyperscaler-operated infrastructure in a limited number of allied nations, holds a notable advantage in computing capacity.

Announced renewable energy targets across Gulf nations have consistently outpaced actual installed capacity throughout the decade. In a pessimistic scenario, delays to a prominent restart or small modular reactor initiative could coincide with transformer and turbine supply chains remaining the principal bottleneck. New data-centre capacity is projected to cluster within the few regions that have available grid capacity, while consumer opposition to the $9.3 billion cost burden already recorded in PJM may prompt individual states to introduce restrictions on fresh data-centre grid connections.

In nations across the Gulf and other regions where capital is abundant but generating capacity remains limited, the distance between projected power targets and operational output shows little sign of narrowing ahead of 2030. In an optimistic scenario, regulatory streamlining by the Nuclear Regulatory Commission alongside Department of Energy co-location on federal land could bring small modular reactors online earlier than projected, while an accelerated series of new 123 Agreements would extend verified nuclear-and-AI cooperation to a broader set of partner nations.

While chip export controls remain a topic of ongoing debate, analysts point to power grid capacity as the more immediate constraint on who can build next-generation AI systems. Graphics processors can be purchased or obtained through various channels, whereas a gigawatt of reliable nuclear power capacity cannot be acquired so readily. Entities with control over viable nuclear infrastructure — including functioning reactor portfolios, small modular reactor agreements, bilateral energy partnerships, or federally designated sites for co-located development — are positioned to hold the most scarce and sought-after resources in the industry.

A final environmental assessment from the Nuclear Regulatory Commission on the Crane Clean Energy Center is scheduled for release this month, with a ruling on the facility's operating licence anticipated by May 2027.