AI Power Bubble Threatens Consumer Bills

The hum of servers, the whir of cooling fans, and the silent, relentless processing of data — these are the unseen engines of the artificial intelligence revolution.

But beneath the surface of this technological marvel, a less celebrated, and potentially far more costly, phenomenon is taking shape: a hidden electricity bubble that threatens to inflate consumer power bills across the nation.

This isn’t a glitch in the matrix; it’s a systemic issue, driven by a confluence of speculative demand, utility incentives, and a forecasting problem that could leave the public paying for power plants they may never need.

The core of this impending crisis lies in the unprecedented, and often unverified, electricity demands being submitted by data center developers.

From the burgeoning tech hubs of Virginia to the emerging markets in Georgia and Texas, power providers are presenting state regulators with staggering projections for future electricity needs, overwhelmingly attributing this surge to the AI boom. Learn more about US data centers’ energy use amid the artificial intelligence boom.

Their proposed solution? A massive buildout of new power plants and grid infrastructure, the costs of which, under current regulatory frameworks, are inexorably passed on to residential and commercial customers.

Imagine a developer with ambitious plans for a single new data center.

Instead of committing to one location, they approach five different utility companies, each blind to the requests made to the others.

Constellation Energy CEO Joseph Dominguez aptly compares this to a fisherman casting multiple lines into the water, hoping to catch a single fish.

The problem arises when each utility, incentivized to secure new load growth, takes these requests at face value.

They initiate feasibility studies, plan for new capacity, and factor these speculative demands into their long-term infrastructure investments.

The result? Five potential power plants on the drawing board, each project adding to the burden on existing ratepayers, only for the developer to eventually select one site and abandon the other four.

This leaves communities saddled with the costs of idle or underutilized infrastructure, a silent tax on every electron consumed.

The stakes are alarmingly high.

If utilities under-build, the reliability of an already fragile grid could be compromised, leading to blackouts and economic disruption.

But if they over-build, an outcome increasingly likely under current conditions, consumers will be locked into decades of paying for ghost power plants, facilities built on speculative dreams rather than concrete demand.

Mark Dyson, a managing director with the Rocky Mountain Institute, points to a fundamental flaw in the system: utilities ‘investing in the grid and recovering those costs through rates’ possess a ‘structural incentive to forecast demand growth in excess of what it is in order to recoup more returns.’

This creates a dangerous feedback loop, where inflated forecasts beget unnecessary construction, all at the expense of the ratepayer.

The sheer scale of the projected demand is startling.

While McKinsey estimates global data centers will require 219 gigawatts (GW) of electricity by 2030, an analysis by Goldman Sachs reveals that just 26 of the largest investor-owned utilities in the U.S. are forecasting 711 GW for data center projects in the same timeframe.

This figure is almost equivalent to the entire summer peak demand for the continental United States.

Such a colossal discrepancy begs serious questions about the prudence and accuracy of current energy resource planning.

Jeremy Fisher, a principal advisor to the Sierra Club, issues a stark warning: “There’s a real risk that we’re already paying for some of that speculative load, or will be paying for that shortly.”

Further complicating this energy equation is a critical bottleneck in the very technology driving the AI boom: semiconductor chips.

Graphics processing units (GPUs), predominantly from industry giants like Nvidia and AMD, are the computational backbone of AI, often consuming over half of a data center’s energy.

Yet, according to Enverus, an energy market data firm, global GPU production will support only 5 GW of computing power this year, projected to grow to 9.5 GW by 2028.

Even if one doubles that figure to account for cooling and other operational necessities, the total power demand from actual GPU production stands at a mere 19 GW.

This represents a profound chasm between the hardware reality and the hundreds of gigawatts utilities claim they need to provide.

Carson Kearl, an analyst at Enverus, succinctly puts it: “That sort of quantity is extremely detached from what the queues are suggesting.”

For utilities, which have faced decades of stagnant or declining electricity demand due to efficiency gains in lighting and appliances, data centers represent a tantalizing new growth opportunity.

This fuels their willingness to accept ambitious, even speculative, demand forecasts.

The ecosystem of data center development further exacerbates the issue, with a multitude of middlemen – from small operators to large real estate firms – all vying for a slice of the AI pie by submitting power requests in multiple locations.

Astrid Atkinson, cofounder of Camus Energy and a former Google data center engineer, estimates that actual grid connection demand may be five to ten times less than the current requests.

Even some developers, like Ryan Mallory, CEO of Flexential, acknowledge the problem, noting that “some newer developers are out there carpet bombing markets, and that creates a tense situation for everybody.”

Encouragingly, some regulators are beginning to recognize and address this issue.

Ohio’s Public Utilities Commission, for example, approved a landmark settlement requiring data centers to pay for at least 85% of the electricity they reserve, regardless of their actual consumption.

The impact was immediate and dramatic: American Electric Power Ohio, which had previously cited 30 GW of potential demand during the rule-making process, saw requests plummet to 13 GW after the new rules were implemented.

This demonstrates that financial disincentives can effectively filter out the most speculative demands.

Such “large load tariffs” are now under consideration in at least 20 other states, signaling a growing awareness and concerted effort among stakeholders to rein in this speculative frenzy.

Yet, the financial repercussions are already being felt.

Customers served by the PJM Interconnection, which manages the regional electric grid across 13 states including Virginia’s “Data Center Alley,” are facing an additional $7.3 billion in electricity costs, directly attributable to data center concerns.

The total auction for future electricity supplies reached a record $16.1 billion this year, largely driven by the perceived necessity to accommodate AI.

While utilities like Dominion Energy, which serves Virginia’s data center heartland, claim to discount the most speculative requests and rely on broader regional forecasts, the inherent upward bias in utility projections remains a persistent concern.

John Wilson, Vice President at Grid Strategies, acknowledges this bias but also highlights the inadequacy of traditional forecasting methods for these novel, massive, and uncertain loads.

The challenge of accurate forecasting is real, but so is the potential for significant and lasting consumer harm.

The AI revolution holds immense promise, but its foundational energy demands must be met with clear-eyed realism, not speculative exuberance.

Without robust regulatory oversight, innovative forecasting methods, and stringent financial accountability for developers, the dazzling promise of AI could dim under the shadow of inflated electricity bills, leaving ordinary citizens to shoulder the costs of a boom from which they may never truly benefit.

The hidden AI bubble, if left unchecked, threatens to become a very visible drain on household budgets and a lasting burden on the nation’s energy infrastructure.