Review: AI Data Centre Growth: Challenges and Opportunities for the US Power System
by Tamunoimim Kalada-GreenView the associated event for this review
Data centres are often discussed as symbols of digital progress. In Pramod Khargonekar’s seminar, however, they appeared more demanding: a test of whether ageing electricity systems can absorb new industrial load without weakening reliability, raising household bills, or slowing decarbonisation.
Khargonekar, Distinguished Professor of Electrical Engineering and Computer Science at the University of California, Irvine, focused on the United States, where artificial intelligence (AI) data centre (DC) growth has become a major power system concern. His central message was that if planned and priced well, DCs could become a catalyst for smarter grid investment. But if handled poorly, they could intensify scarcity and raise costs for consumers.
The starting point is the unusual position of the US power system. For much of the past two decades, electricity consumption has been relatively flat at about 4,430 TWh annually. Yet this period of limited growth is now giving way to a major investment cycle. The largest technology companies are committing hundreds of billions of dollars to AI infrastructure, with a significant share going into electrical and power systems. The difficulty is that no one knows with confidence how large future AI electricity demand will be, a point highlighted by the widespread forecasts for 2030 DC electricity use.
This uncertainty is already visible in US power markets. In PJM, the regional transmission organisation covering parts of the eastern United States, capacity prices have risen sharply, and a recent auction fell 6.6 GW short of its target. DC load growth has been identified as a major driver of these conditions. This led PJM to ask whether reliability should remain a common good, paid for collectively, or whether different users should receive and pay for different levels of reliability.
Texas illustrates the scale of the challenge even more dramatically. ERCOT is tracking around 410 GW of large loads seeking interconnection by 2030, with DCs accounting for most of that pipeline. In response, Texas Senate Bill 6 has pushed ERCOT towards new interconnection standards for large loads, including controllable load pathways and bring-your-own-generation approaches. These reforms point to a wider shift: DCs may no longer be treated as ordinary customers that simply request power from the grid. They may increasingly be expected to bring flexibility, generation, or financial commitments with them.
A central problem is timing. DCs can be built in two to three years, while transmission lines, generation projects, substations and other grid infrastructure can take much longer. Supply chains for gas turbines, transformers and other critical equipment are also constrained. This creates a mismatch between the speed of AI investment and the slower pace of grid expansion.
Flexibility is one proposed answer. Rather than building the grid solely around peak DC demand, operators could make better use of existing capacity by allowing large new loads to reduce demand during constrained periods. Recent analysis suggests that close to 100 GW of new flexible load could be served with very limited annual curtailment. Importantly, this would not usually mean shutting facilities down, as in most curtailment hours, at least half of the load could remain online. But Khargonekar cautioned against treating this system-level estimate as automatically available everywhere. Local grid constraints still matter. Separate distribution-level modelling shows that flexibility could increase the extra load a local grid can safely absorb by ~90% without network constraints, but by ~61% once those constraints are included.
The seminar also showed that DC flexibility is technically plausible. GPU power-capping can reduce electricity demand while preserving much of the underlying computing performance, especially for workloads that can tolerate slower processing. A demonstration in Phoenix showed an AI cluster reducing power during a demand-response event and then returning to normal operation, while overall job performance stayed close to its target range. The key point is that flexibility depends less on switching facilities off than on managing workloads intelligently.
Yet technical flexibility does not automatically become commercial flexibility. For high-value AI workloads, curtailing computation may be economically unattractive, even when electricity prices are high. This means flexibility will likely depend on contracts, tariffs, and faster interconnection, not only voluntary demand response.
This is where tariff design becomes crucial. If utilities build grid capacity for DCs whose future demand is uncertain, stranded costs could fall on ordinary ratepayers. Khargonekar’s proposed solution is a ratcheted minimum billed demand. In simple terms, if a DC reserves capacity, it should pay for a substantial share of that capacity even if it later uses less than expected. Higher minimum billed demand makes DCs financially responsible for the infrastructure built on their behalf.
The Q&A broadened the discussion. Questions explored the role of power electronics in managing congestion and smoothing rapid DC load changes, as well as the shifting geography of DCs as inference becomes more distributed. On decarbonisation, Khargonekar offered a balanced view: near-term DC demand could encourage fossil generation in the US but improved solar and battery economics may also create opportunities for cleaner expansion.
The seminar’s strongest insight was that AI data centres are testing how power systems allocate scarcity. If grid capacity is a public good, who should pay to preserve reliability? If reliability becomes differentiated, which loads should receive firm service, which should accept interruption, and how should those choices be priced? Done right, DC investment could help modernise the grid. Done poorly, it could leave the public paying for private uncertainty.