Ohio regulators dropped a bombshell on the hyperscale computing world in July 2025, authorizing a first-of-its-kind tariff that forces large data center operators to underwrite at least 85 percent of the energy capacity they reserve—effectively shifting the financial risk of grid expansion from ordinary ratepayers to the tech giants driving the AI boom. The Public Utilities Commission of Ohio (PUCO) ordered AEP Ohio to impose a 12-year, take-or-pay contract structure on new, large data center customers, complete with four-year ramps, exit fees, and collateral requirements. The move marks the most aggressive attempt yet to solve a problem vexing utilities from Northern Virginia to Texas: who pays for the transmission lines, substations, and generation needed to feed the insatiable power appetite of artificial intelligence?

The answer, increasingly, is the companies building the AI infrastructure themselves. Amazon, Google, Meta, and Microsoft fought the Ohio provision, arguing it would chill economic development and reduce operational flexibility. But consumer advocates and the utility itself warned that without such a mechanism, residential and small-business customers could see bills swell as utilities build capacity for speculative or unevenly used loads. The PUCO’s decision is already being studied by regulators in other states, and it underscores a cardinal principle of cost causation: if your business model requires gigawatts of always-on power, you should pay to make the grid ready for it.

The Scale of the Appetite

The Department of Energy and Lawrence Berkeley National Laboratory pegged data center electricity consumption at about 4.4 percent of the U.S. total in 2023—roughly 176 terawatt-hours. By 2028, under current growth trajectories, that share could nearly triple to between 6.7 and 12 percent, or 325 to 580 TWh. Generative AI and large-language-model training are the primary accelerants. These workloads run near-constant, high-density compute cycles that concentrate power draw in specific geographic clusters—Northern Virginia’s “Data Center Alley,” central Ohio, and swaths of Texas and the Midwest. Traditional utility planning cycles, which look decades ahead, never anticipated tens of gigawatts of new load materializing in five years.

Virginia’s Joint Legislative Audit and Review Commission (JLARC) modeled unconstrained data center growth and found the state would need massive generation and transmission additions just to keep pace. Under some scenarios, residential generation- and transmission-related costs could rise by $14 to $37 per month (real dollars) by 2040. Crucially, that’s a broad range, not the single, often-cited figure of $276 per resident per year by 2030—a number that doesn’t appear in JLARC’s technical documentation. When popular coverage tosses out scary single-year numbers, readers should reach for the primary regional modeling.

Hyperscalers Turn into Energy Titans

Tech companies are not passively waiting for utilities to upgrade the wires. They’ve become major players in electricity markets, signing long-term power purchase agreements (PPAs) for renewables, taking direct stakes in nuclear and gas-fired plants, and building behind-the-meter generation that bypasses traditional rate structures. In some wholesale markets, corporate-owned or contracted generation now accounts for a meaningful share of trading. This vertical integration raises thorny questions about market power and transparency. Bilateral deals are often confidential, and regulators struggle to see whether a hyperscaler’s generation contract distorts pricing, displaces other buyers, or sticks ratepayers with residual costs.

Google’s August 2025 demand-response agreements with Indiana Michigan Power (an AEP subsidiary) and the Tennessee Valley Authority illustrate a newer, more nimble approach. For the first time, a cloud provider will specifically curtail or reschedule non-urgent machine-learning workloads to ease grid stress. Google framed the deals as a way to avoid immediate capital-intensive upgrades and earn compensation for load flexibility. The move builds on a 2024 pilot and signals that AI workloads can be treated as deferrable, not just as relentless baseload. Demand response is no silver bullet—coincident peaks from concentrated data center clusters still require firm capacity—but it buys time and aligns compute with periods of abundant renewable generation.

The Economic Battle Lines

The fight over cost allocation reduces to three basic models. Under socialized cost recovery, utilities spread grid upgrade expenses across all rate classes. That approach speeds data center construction and economic development but risks inflating bills for households and small businesses that don’t use the new infrastructure. The Ohio-style take-or-pay model attaches costs directly to the large users that trigger them, protecting other ratepayers while potentially driving hyperscalers to build in jurisdictions with softer rules. A third, hybrid approach uses staged or conditional service agreements that scale financial commitments as actual load materializes—reducing the risk of stranded assets but demanding sophisticated regulators and reliable load data.

Many local governments actively recruit data centers with tax breaks and infrastructure incentives because they bring construction jobs and, in some cases, high-paying permanent positions. Those benefits are real. But they must be weighed against the long-term utility tab and environmental consequences, including water consumption for cooling and noise from backup generators. The calculus is shifting as energy constraints bite.

Risks That Keep Regulators Up at Night

Several systemic risks demand attention. First, stranded-infrastructure risk: a utility builds a $500 million transmission line for an anticipated data center campus that never materializes—or materializes at a fraction of the subscribed load—and the sunk cost lands on everyone else’s bills. Second, market opacity: exclusive clean-energy contracts, behind-the-meter plants, and privately owned generation assets muddy wholesale price signals. The public often cannot see who is selling to whom or at what price. Third, geographic concentration: clustering creates localized reliability headaches—voltage fluctuations, thermal overloads—that national aggregates hide.

One claim bolting through media coverage—that tech-affiliated generation sold “more than $2.7 billion” in wholesale electricity over a decade—cannot be substantiated in FERC filings or public regulatory summaries based on available records. The figure needs documented sourcing before being treated as established fact. Responsible reporting demands cross-checking with original market-monitoring data.

Engineering Fixes Inside the Fence

While fighting policy battles, hyperscalers are squeezing more compute out of each watt. Power usage effectiveness (PUE) continues to improve thanks to immersion cooling, AI-driven thermal controls, and better airflow design. Shifting non-time-critical training and preprocessing jobs to off-peak hours—when electricity is cheaper and cleaner—reduces grid strain and integrates renewables more smoothly. Site selection is also pivoting: cooler climates, coastal seawater cooling, and regions with abundant wind or solar capacity are gaining favor. Still, efficiency gains alone cannot offset the staggering growth in raw compute demand; that’s why grid planning and regulatory frameworks remain at center stage.

Policy Levers That Could Balance Innovation and Fairness

A fair transition requires a coordinated set of policy moves. First, make cost causation explicit: utilities should document which grid investments are driven by which customers, and tariffs should target those responsible for material planning changes. Second, demand transparency: long-term generation contracts, behind-the-meter ownership stakes, and special tariffs must be disclosed to regulators and market monitors, even if some commercial terms remain confidential. Third, stage financial approvals: conditional service agreements tied to demonstrated load growth, backed by performance bonds or collateral, can reduce stranded-investment risk. Fourth, value demand response properly: flexible AI workloads should be compensated as grid resources, with clear market rules. Fifth, accelerate transmission and firm clean capacity: permitting reform and federal-state collaboration are essential to build the long-lead infrastructure that data center clusters demand. Finally, shield small consumers: opt-in protections and targeted cost caps can insulate residential ratepayers from speculative industrial-scale builds.

None of these levers is sufficient alone; a robust regime mixes conditional commercial terms, regulatory transparency, and interstate coordination to prevent jurisdictional arbitrage.

What IT Leaders Must Do Now

For enterprise architects and cloud planners, the energy grid is no longer a background concern—it’s a frontline factor in site selection and workload design. RFPs for new data center space or cloud capacity should scrutinize local utility tariff structures, interconnection timelines, and the regulatory posture of the host jurisdiction. Expect new surcharges or pass-throughs linked to suppliers’ energy economics to show up in cloud bills for AI-heavy workloads. Architects should design AI deployments to exploit workload elasticity, scheduling training during low-cost windows and using multi-region failover to avoid dependence on a single stressed grid. Operational flexibility is becoming a competitive differentiator.

The Road Ahead

Tech capital is pouring into clean energy and dispatchable generation, and that investment is genuinely additive—it finances renewables and sometimes firm capacity that might not otherwise get built. Google’s demand-response deals with I&M and TVA show that even AI workloads can be made grid-friendly. But opaque contracting and the potential for private generation to warp wholesale markets create governance gaps current rules don’t close. If regulators socialize the costs of speculative or opaquely documented demand, households and small businesses will pay for capacity that hyperscalers triggered and never fully used.

Demand projections remain model-dependent. Different studies put data centers’ share of U.S. electricity anywhere from a modest single digit to well into the double digits, with timing varying by years. Anyone citing a single absolute figure should be met with skepticism until the underlying scenario assumptions are unpacked.

AI is remaking the cloud and the power grid simultaneously. Hyperscalers have the capital and motive to optimize around energy, but the public interest—especially the protection of residential ratepayers—hangs on transparent markets, clear regulatory guardrails, and cost-allocation rules that reflect who creates the need for long-lived infrastructure. Ohio’s new tariff is an early, influential template. Virginia’s JLARC study quantifies the local price tag. Google’s demand-response experiment offers short-term relief. The pieces are on the table. It’s now up to policymakers, utilities, and tech companies to assemble them into a framework that keeps the lights—and the models—running without sticking the bill to those least able to pay.