By Rebecca Stone

Datacenters will burn at least 1,200 terawatt-hours (TWh) of electricity in 2030 — a 3.1× leap from 387 TWh in 2024, according to Gartner forecasts combined with Energy Information Administration (EIA) and Ember data. The 20.8% compound annual growth rate (CAGR) runs roughly seven times faster than the 3% CAGR expected for worldwide electricity consumption.

That mismatch lands directly on the semiconductor industry. Power constraints now sit alongside HBM, DRAM, and flash supply shortages as the gating factor on next-wave GenAI deployment.

How does the consumption stack up?

Gartner's forecast splits datacenter load into three buckets — traditional servers, AI servers, and the rest of the gear (storage, switching, cooling overhead). The firm anchored its model on 2025, projected through 2026 and 2027, then stretched to 2030. The 387 TWh 2024 baseline came from back-calculating Gartner's stated growth rates.

The world's compute footprint looks like this:

• 11,500 to 12,000 very large datacenters worldwide
• ~7 million enterprise, sovereign, and edge dataclosets
• Roughly 75% of datacenter power flows through hyperscalers, cloud builders, neoclouds, HPC centers, and sovereign operators

Small sites keep consolidating into mega-facilities. A handful of campuses now cross the gigawatt threshold.

What share of the grid do datacenters actually take?

Two ratios tell the story:

• Consumption share: 1.3% of world electricity in 2024 → ~3.3% by 2030
• Allocation share: 1.1% of global capacity in 2024 → 2.1% by 2030

Allocated capacity grows from 104 gigawatts (GW) in 2024 to 290 GW in 2030 — an 18.56% CAGR that has utilities scrambling. Worldwide generating capacity rises at 5.9% CAGR, almost twice the rate of consumption but well short of datacenter demand. Total electricity use expands from 30,856 TWh in 2024 to roughly 36,500 TWh by 2030.

Is the panic warranted?

Not necessarily — the same alarm rang during the Dot-Com Boom. Jonathan Koomey, then a staff scientist at Lawrence Berkeley National Laboratory and a Stanford professor, projected in the mid-2000s that datacenters would consume nearly 4% of world power. The actual peak landed under 2%.

The historical share traces a sharp bend:

• 2000: 0.5% of world electricity
• 2010: 1%
• Mid-2010s peak: under 2%
• 2024: 1.3%
• 2030 (forecast): 3.3%

Server virtualization on X86, post-recession capex restraint, and Koomey's public pressure on efficiency bent that line. The same pressure may bend it again — six years into the GenAI transition, silicon vendors are already designing purpose-built architectures.

What could break the curve?

Two forces. First, workload-tuned silicon — accelerators purpose-built for transformer inference and training — promises better perf-per-watt than general-purpose GPUs. Second, fresh nuclear capacity, including small modular reactor commitments, decouples new datacenter campuses from grid constraints.

Neither path is mature. Gartner's 290 GW allocation by 2030 will hinge on which curve bends first.