How much electricity do AI data centers use in 2026?

Global data center electricity consumption is projected to reach 565 terawatt-hours (TWh) in 2026, a 26% year-over-year increase according to Gartner. AI-optimized servers alone account for 31% of that — 175 TWh — with their electricity use surging 84% from 2025. Total global data center power demand is estimated at 132 gigawatts (GW) in 2026.

How much power does xAI's Colossus use?

xAI's Colossus facility in Memphis currently operates at 352.4 MW with over 275,000 H100-equivalent chips, making it the world's most powerful AI supercomputer. xAI announced plans to expand Colossus to a total of 2 GW capacity, housing 555,000 NVIDIA GPUs. The expansion relies on on-site gas-fired power plants to bypass traditional utility grid constraints.

What is the Stargate initiative's power target?

Stargate — the OpenAI/Microsoft/SoftBank/Oracle/MGX joint venture — aims to deploy up to 10 GW of compute capacity across multiple gigawatt-scale AI campuses. Locations announced so far include Wisconsin (Vantage Data Centers' Lighthouse campus at 902 MW IT capacity), Ohio, New Mexico, Abu Dhabi, and Norway. Stargate represents hundreds of billions in investment.

Are AI data centers harming the environment?

The environmental impact is significant and growing. A single large AI training run can consume as much electricity as hundreds of homes use in a year. The shift to inference — using AI models millions of times per day — multiplies this dramatically. Critics point out that Colossus relies on gas-fired plants rather than renewables. The tech industry is investing in nuclear, geothermal, and solar, but capacity isn't scaling fast enough to keep AI growth carbon-neutral.

Quick Answer

AI Data Center Energy Crisis 2026: Colossus, Stargate, and the Race for Power

Q: Can the power grid handle AI data center demand?

This is one of the biggest constraints on AI growth in 2026. Many regions are struggling to approve and build enough power generation and transmission capacity. Data center construction timelines are increasingly driven by power availability rather than capital. Some projects are being delayed or relocated to regions with available capacity. The grid interconnection queue is backlogged by years in several major markets.

Published: July 5, 2026

AI Data Center Energy Crisis 2026: Colossus, Stargate, and the Race for Power

Global data center electricity consumption will hit 565 TWh in 2026 — a 26% year-over-year surge — with AI-optimized servers consuming an estimated 175 TWh (up 84% from 2025), according to Gartner’s June 2026 report.

The numbers are staggering, and they represent one of the most underappreciated stories of the AI boom: we may be approaching the physical limits of power generation and grid capacity faster than we’re approaching AGI.

By the Numbers

Metric	2025	2026	Change
Global data center electricity	~449 TWh	565 TWh	+26%
AI-optimized server electricity	~95 TWh	175 TWh	+84%
Total data center power demand	~105 GW	132 GW	+26%
AI share of total DC electricity	~21%	31%	+10pp

Source: Gartner, June 2026. “Data Center Electricity Demand to Grow 26% in 2026.”

The Major Projects

Colossus (xAI / Elon Musk)

Current capacity: 352.4 MW (Memphis, Tennessee)
Target capacity: 2 GW
Hardware: 275,000+ H100-equivalent GPUs → 555,000 GPUs at full buildout
Power source: On-site gas-fired power plants (not grid-dependent)
Status: Expansion underway; full 2 GW target faces timeline questions

Colossus is notable for bypassing the electric grid entirely with on-site gas generation. This is the Elon Musk approach: if the utility can’t deliver power fast enough, build your own plant. Environmental critics note this means Colossus’s growth is tied to fossil fuels.

Stargate (OpenAI / Microsoft / SoftBank / Oracle / MGX)

Target capacity: Up to 10 GW across multiple campuses
Locations: Wisconsin (902 MW Lighthouse campus), Ohio, New Mexico, Abu Dhabi, Norway
Investment: Hundreds of billions of dollars
Power approach: Grid-connected campuses with renewable energy commitments

Stargate is the most ambitious single infrastructure project in AI history. The 10 GW target would make it comparable to the total electricity generation of multiple nuclear power plants.

Other Major Projects

Meta Hyperion: Multi-campus buildout focused on Meta’s open-source AI (Llama family)
Amazon AWS AI: Distributed data center expansions across Virginia, Ohio, and internationally
Google/Broadcom: Custom TPU infrastructure with efficiency-focused architecture
SpaceX Orbital: Early-stage plans for orbital data centers (Starcloud) — power from solar, but latency and maintenance remain unproven

The Grid Problem

The biggest bottleneck for AI infrastructure in 2026 isn’t chips or capital — it’s power interconnection.

Key issues:

Transformer lead times: Large power transformers have 18-36 month lead times
Interconnection queues: New data centers face multi-year waits for grid interconnection in Virginia, California, and parts of Europe
Renewable intermittency: Solar and wind can’t provide the 24/7 reliable power AI training clusters need without massive battery storage
Nuclear renaissance: Tech companies are investing in small modular reactors (SMRs) and existing nuclear plants, but these timelines stretch to 2030+

The result: Data center construction timelines are increasingly driven by power availability rather than capital availability. Some planned projects are being delayed or moved to regions with spare grid capacity — which often means rural areas or countries with less stringent environmental regulations.

Environmental Impact

The carbon footprint of AI inference is growing faster than training emissions. A single ChatGPT-style query uses roughly 10x the energy of a Google search. Multiply that by billions of queries per day, and inference quickly dominates.

Impact Area	2026 Estimate
AI data center CO₂ emissions	~320 Mt CO₂e (estimated)
Water consumption for cooling	~4-5 billion liters/day (est.)
E-waste from GPU refreshes	Significantly increasing with annual upgrade cycles

The industry response includes:

Efficiency research: Sparse models, quantization, speculative decoding — each cuts energy per token significantly
Location strategy: Building in regions with hydro, nuclear, or geothermal power
On-site generation: Gas plants (Colossus), backup diesel, early SMR planning
Carbon offsets: Widely criticized as insufficient, but used by most AI companies

The Bottom Line

The AI data center energy crisis is real and intensifying. The gap between AI compute demand and available clean power is one of the biggest structural constraints on AI progress in 2026-2028.

For companies building on AI: expect energy costs and availability to become a strategic consideration, not just a utility bill. For regulators: decisions made in the next 12 months about grid interconnection, nuclear licensing, and renewable deployment will determine whether the AI boom is powered by clean energy or fossil fuels.

Published July 5, 2026. Data sources: Gartner June 2026 report, Epoch AI, Digital Journal, Data Center Knowledge, Tom’s Hardware. Power figures are estimates based on publicly reported specifications and may vary with actual operating conditions.