AI Data Centers Put Grid Reliability Risk in Focus for Firms

The boom in artificial intelligence is creating a power challenge that is becoming difficult for utilities, regulators and data center developers to manage. Demand from AI data centers is accelerating, but the transmission infrastructure needed to support that growth often takes years to plan, permit and build.

A new Wood Mackenzie report warns that this timing gap is putting pressure on electricity markets, project planning and grid reliability. In some regions, the transmission upgrades needed to serve large AI campuses may still be five to 10 years away.

That delay is pushing developers to look for faster ways to secure power. Options such as collocated generation, flexible interconnection agreements and bring-your-own-generation models are gaining attention. However, Wood Mackenzie cautions that these approaches are not simple fixes. They introduce technical, regulatory and commercial risks that many projects may be underestimating.

Flexible interconnection can help some projects connect sooner, but it does not offer the same certainty as firm grid service. For facilities built around high uptime, that matters. Conditional access may come with operating limits during grid stress, which can create tension between data center performance requirements and wider system reliability.

The pressure is especially clear in deregulated power markets. In PJM, committed data center load has reportedly reached 78 gigawatts, while accredited generation capacity in the pipeline stands at 36 gigawatts. In Texas, current market prices of around $30 to $40 per megawatt-hour remain below the $78 to $100 Wood Mackenzie says would be needed to support new gas generation investment.

That gap creates a difficult market signal. Higher prices may be needed to bring new supply online, but those increases would affect all electricity customers, not only the data centers driving new load growth.

Collocated Power Helps, But It Does Not Remove Grid Risk

Collocated generation is often presented as a way for data centers to bypass slow grid expansion. The scale of interest is significant, with Wood Mackenzie estimating more than 90 gigawatts of collocated generation in U.S. interconnection pipelines.

Even so, the model is unlikely to work for every developer. Large hyperscalers may have the capital, technical teams and procurement power to manage these projects. Smaller or less experienced operators could face a more difficult path.

The engineering risks are substantial. AI workloads can shift power demand almost instantly, creating stress for gas turbines, reciprocating engines and other onsite generation assets. Batteries may help smooth those fluctuations, but heavy cycling can shorten battery life. Power quality is another issue, as irregular loads from GPUs and cooling systems can create harmonics that damage equipment if not properly managed.

Regulation adds another layer of uncertainty. Grid operators are developing rules for conditional interconnections, but their priority remains system reliability. In some markets, operators may have the right to dispatch collocated generation to support the grid during certain events. That could force a data center to reduce demand to its firm service level, even when onsite generation is available.

ERCOT is also reviewing ride-through requirements for voltage and frequency events. The concern is that data centers could switch to backup systems too quickly during minor disturbances, reducing grid stability. Wood Mackenzie points to a 2024 incident in Virginia, where 60 data centers reportedly dropped off the grid at the same time after a relatively small disturbance.

The wider cost question is still unresolved. Grid operators are planning close to $100 billion in transmission investments linked in part to data center load growth. This includes $11.8 billion in PJM, more than $30 billion in MISO, $33 billion in ERCOT and $8.6 billion in SPP.

Under traditional cost allocation models, much of that spending could be spread across existing ratepayers. That raises a policy question that is likely to become more urgent: who should pay for grid upgrades driven by AI data center expansion?

For developers, investors and utilities, the next phase of data center growth will not be defined only by speed to market. It will also depend on how well projects manage power availability, market exposure, regulatory change and public concern over electricity costs. AI infrastructure may keep expanding, but the power strategy behind it will need to become more disciplined.