- PJM’s August 2026 ruling mandates AI data centers exceeding 50 MW secure dedicated power by June 2027 or face service curtailment.
- Global data center electricity demand is forecast to reach 945 TWh by 2030, with AI workloads projected to surpass conventional server consumption by 2027.
- US grid demand is tracking toward record highs of 4,268 billion kWh in 2026 and 4,391 billion kWh in 2027, pushing hyperscalers toward on-site generation as grid access becomes a primary business risk.
PJM’s August 2026 ruling gives AI data centers drawing more than 50 MW until June 2027 to secure dedicated power or face curtailment, and it arrived just as US grid demand was already tracking toward back-to-back record highs. The collision between AI’s power appetite and the grid’s physical limits has moved from a planning concern to a commercial barrier, reshaping site selection and operational risk for data center operators across the country.
AI’s Megawatt Hunger
Global data center electricity consumption stood at around 415 TWh in 2024, roughly 1.5% of worldwide electricity use, according to the International Energy Agency. The IEA projects that figure will nearly double to 945 TWh by 2030, growing at roughly 15% per year, a rate far exceeding consumption growth across other sectors. Within that growth, AI workloads are the dominant driver. AI-optimised servers are expected to account for 31% of data center power consumption in 2026, and by 2027, their electricity use is projected to surpass that of conventional servers. Inference, unlike batch processing, demands sustained high-wattage draw rather than periodic peak capacity, a distinction that changes what grid operators must plan for.
The Grid Can’t Keep Up
The primary constraint on AI infrastructure expansion has shifted from computational efficiency to grid-scale power availability. The bottleneck is less about aggregate generation capacity and more about the grid’s ability to transmit and distribute power to concentrated, high-magnitude loads. Interconnection queues for new power projects can stretch beyond three years in some regions. Northern Virginia, Ireland and Singapore, three of the highest-density data center markets, are among the most acute pressure points.
US utilities have announced significant grid upgrade projects tied to data center load growth, and those costs are frequently spread across all ratepayers. Construction costs for power plants are also rising under supply-chain pressure, equipment backlogs, permitting delays and import tariffs. One symptom: shuttered petroleum refineries are being evaluated as candidate sites for power generation, repurposed to close a supply gap that conventional project pipelines cannot fill fast enough. For more on how AI workloads are reshaping power demand projections through 2030 the underlying trajectory is worth reviewing alongside the PJM ruling.
Hyperscalers Build Around the Problem
The response from large operators has been to reduce dependence on grid connection altogether. Construction has begun on large-scale data center projects with on-site generation, with more in pre-construction stages. On the hardware side, AMD has been focused on improving energy efficiency in its rack-scale AI platforms, with future generations targeting gains over current designs. The Wall Street Journal reported in late July that despite a record volume of electricity generation under development in the US, the cost to complete those projects continues to rise, a dynamic that shows no sign of resolving before PJM’s June 2027 deadline.
