EPRI: US Data Centers Could Eat 17% of National Power by 2030 — The Grid Is Now the AI Bottleneck

The Number That Reframes the AI Buildout

For two years, the AI race has been narrated as a scramble for GPUs. A major utility-industry analysis argues the real ceiling is somewhere else entirely. According to Data Center Knowledge’s coverage of EPRI’s “Powering Intelligence 2026” report, US data centers could consume 9% to 17% of total national electricity by 2030 — more than doubling their current 4-5% share, and representing a 60% increase over EPRI’s own 2024 estimates.

The Electric Power Research Institute modeled three trajectories: a low scenario at 9% of US electricity, a medium at 13%, and a high at 17%. E&E News reported that data centers currently consume “roughly 4.5 percent of U.S. electricity” and that the new estimates are “roughly 60 percent higher than EPRI anticipated two years ago.” In absolute terms, Data Center Knowledge reports the range runs from an estimated 177-192 terawatt-hours consumed in 2024 to roughly 380-790 TWh by 2030. The upper bound would mean data centers alone consuming more electricity than many entire countries.

What makes the figure a turning point is not its size but its implication: adding compute is now gated by the ability to power it. When a single category of building can plausibly claim a sixth of a national grid within five years, electricity stops being a utility bill and becomes the strategic variable that decides where AI infrastructure can physically exist.

Why the Grid, Not the Chip, Is the Constraint

The core problem is a timing mismatch, and EPRI states it plainly. As Data Center Knowledge quotes the report, “AI infrastructure scales in months; generation and transmission take years.” A hyperscaler can order and install racks of accelerators on a timeline measured in quarters. Building the power plants and high-voltage transmission lines to feed them takes the better part of a decade — and no amount of capital compresses the physics and permitting of grid construction into a chip procurement cycle.

That asymmetry is why interconnection — the process of connecting a new large load to the grid — has become the choke point. As a separate Data Center Knowledge analysis of grid constraints puts it, “the single biggest gating item for data center development is not capital or land — it’s power,” and interconnection “could take four to five years or longer,” reaching up to 10 years in some areas for a grid-only site. When generation and transmission move at the speed of infrastructure while demand moves at the speed of software, a backlog is inevitable. The result is that a data center’s limiting factor is increasingly not whether the operator can buy GPUs, but whether the local grid can deliver continuous, high-capacity baseload power to run them.

Geographic Concentration Is Becoming a Grid Crisis

The strain is not evenly spread; it clusters, and the clustering is where the system breaks first. Data Center Knowledge highlights Virginia as the extreme case: data centers there already consume over 25% of state electricity, a figure EPRI’s “Powering Intelligence 2026” executive summary projects could rise to 41-59% by 2030. A single US state may soon route more than half its power to data centers.

The concentration is spreading. The report notes that up to seven additional states — Arizona, Indiana, Iowa, Nebraska, Nevada, Oregon and Wyoming — may exceed a 20% data center electricity share as developers chase locations with available power, land, faster permitting and incentives. This is the mechanism by which “power is the bottleneck” becomes concrete: capacity migrates to wherever the grid can still absorb it, and those regions then approach their own limits. The map of where AI compute can be built is being redrawn around electricity availability, not around traditional advantages like proximity to users or cheap real estate.

The Global Repricing of Location

This reshuffling is not a purely American phenomenon; it is a signal about how the entire global data center map will be drawn. If power availability is the deciding factor for where compute gets built, then the competitive advantage shifts to jurisdictions with abundant, dispatchable generation and spare grid headroom. The US example shows what happens when demand outruns the grid: even the wealthiest hyperscalers hit a wall that money alone cannot move on the required timeline.

For regions that have historically been overlooked in the data center conversation, that repricing is an opening. A place that cannot compete on latency, talent density or capital can still compete on the one input that has become scarce — power that can actually be delivered.

What This Means for Algeria

Algeria’s abundant Saharan solar and natural-gas capacity is a genuine competitive card at exactly the moment power scarcity, rather than land or chips, becomes the deciding factor for where data centers get built.

1. Position surplus generation as a data center attraction, not just an export

The lesson from the US grid strain is that available, deliverable power is now the scarcest input in the compute buildout. Algeria’s gas capacity and Saharan solar potential are assets that could anchor a data center offer to regional and international operators facing grid walls elsewhere — but only if paired with the transmission and interconnection capacity to actually deliver that power to a campus.

2. Build the grid connection before marketing the electricity

EPRI’s core finding is that generation is not the only bottleneck — transmission and interconnection are. Cheap electricity that cannot be delivered to a data center site at the required capacity and reliability is not a competitive advantage. Investment in the substations, high-voltage links and interconnection processes that turn raw generation into deliverable baseload is what would make an Algerian power advantage real.

3. Compete on reliability and baseload, which is what AI actually needs

The report stresses that AI facilities need continuous, high-capacity baseload power. Algeria’s dispatchable gas generation is well-suited to this profile in a way that intermittent-only regions are not. A credible offer emphasizes not just cheap kilowatt-hours but guaranteed, always-on delivery — the specific quality that AI workloads demand and that grid-constrained regions cannot promise.

The Strategic Lesson

EPRI’s report is, on its surface, a warning about American grid strain. Read globally, it is a preview of the single most important variable in the next phase of the AI buildout: which places can actually deliver the electricity that compute requires. The US case demonstrates that even with unlimited capital and first pick of chips, an AI campus cannot exist where the grid cannot feed it — and that grids take years to expand while demand explodes in months.

That reality reshuffles the global competition. The advantage moves toward jurisdictions with surplus, dispatchable generation and the willingness to build the transmission to match. For Algeria, with Saharan solar potential and substantial gas capacity, the opportunity is real but conditional: the country holds a card that has suddenly become valuable, and whether it plays into a genuine data center opportunity depends on building the interconnection and reliability that turn raw power into deliverable compute. The nations that treat electricity delivery as strategic infrastructure — not just a commodity — are the ones that will host the compute of the coming decade.

Frequently Asked Questions

Sources & Further Reading

• EPRI Report: US Data Center Grid Strain Casts Cloud Over AI Race — Data Center Knowledge
• Data Centers’ Share of US Electricity Seen Doubling by 2030 — E&E News (Politico)
• Gridlocked: How Power Constraints Are Shaping the Future of Data Centers — Data Center Knowledge
• Powering Intelligence 2026 — Executive Summary — EPRI