While headlines focus on the hundreds of billions of dollars being spent on AI infrastructure in 2026, few are asking the critical question: Can we actually power it? According to the AI CapEx Tracker, hyperscalers are building 39 major data centers across 4 countries with 7.1 gigawatts of current power and 35.2 gigawatts of projected power. But power availability — not CapEx budgets — is becoming the real constraint. The IEA confirmed data center electricity demand surged 17% in 2025 alone, with AI-focused facilities growing even faster.
The Scale of the Power Problem
A single large-scale AI training facility now requires between 100 MW and 1,000 MW of dedicated power — equivalent to the electricity needs of 80,000 to 800,000 households. Meta’s Richland Parish AI Campus in Louisiana will draw approximately 2 gigawatts at initial build, with potential to scale to 5 GW. The campus is expected to be operational by 2030, with its dedicated power infrastructure coming online by end of 2028.
The US electrical grid wasn’t designed for this. Grid capacity in major tech hubs — particularly Virginia, California, and Texas — is already strained. In 2026, projects that secured land and financing years ago are stalled waiting for grid connections, transformers, and generation capacity that simply does not yet exist.
The Numbers Tell the Story
- Meta’s Richland Parish: 2.0 GW initial capacity (up to 5 GW potential)
- Meta Hyperion: 0 MW current, 2.262 GW projected
- OpenAI Stargate New Mexico: 0 MW current, 2.210 GW projected
- OpenAI Stargate Shackelford: 0 MW current, 1.960 GW projected
- QTS Cedar Rapids: 0 MW current, 1.587 GW projected
- Combined tracked capacity: 7.1 GW current / 35.2 GW projected across 4 countries
For context, the entire state of New Hampshire consumes approximately 6 GW on average. The AI infrastructure buildout will rival the power consumption of entire states.
Where’s the Power Coming From?
Tech companies are pursuing five main strategies:
1. Nuclear Power Partnerships
Microsoft, Amazon, Google, and Meta have all signed major nuclear deals. The SMR (small modular reactor) pipeline has grown from 25 GW at end of 2024 to 45 GW today — a near-doubling in under 18 months. Nuclear provides baseload power — constant, reliable supply — without the intermittency of renewables. The challenge: most plants are 5-10 years from full operation.
2. Renewable Energy (With Caveats)
Solar and wind are cheaper than ever but share a fundamental problem: intermittency. Data centers need power 24/7/365. Without massive battery storage (still expensive and immature at scale), renewables require backup sources. They work best as part of a portfolio, not a sole solution.
3. Natural Gas — The Interim Fix
As nuclear faces multi-year build timelines, on-site natural gas generation is emerging as a pragmatic bridge. Data center developers constrained by slow grid connections are advancing projects with on-site gas-based power, largely in the United States. Entergy alone plans to build seven new natural gas plants totaling 5,200 MW to support Meta’s Louisiana campus — five times what the entire city of New Orleans uses on an average day. It’s not clean, but it’s available now.
4. Geographic Arbitrage
Building in regions with surplus power. AWS committed €15.7B to Spain. Companies are exploring the Pacific Northwest (hydropower) and the Southeast (nuclear proximity). Trade-off: distance from end users creates latency challenges.
5. Efficiency Improvements
Liquid cooling, custom silicon, and architectural innovations reduce power per compute unit. But these are incremental gains against exponential demand growth. They buy time, they don’t solve the problem.
The Real Constraint
You can spend $200 billion on infrastructure (Amazon’s 2026 projection), but if you can’t secure reliable power, the facility sits idle. The AI CapEx Tracker shows 16 data centers currently under construction. Power constraints will determine how many actually reach planned capacity — and on schedule.
What Happens Next?
2026-2028 is the inflection point:
- Projects stalled on transformers and grid connections create near-term regional shortfalls
- Natural gas bridges the gap while nuclear comes online (2027-2032)
- Regional power prices spike as demand concentrates in tech hubs
- Geopolitical leverage grows for regions with abundant clean energy
- AI workloads may shift to lower-latency-penalty regions with surplus power
The Bottom Line
hundreds of billions of dollars in AI CapEx is meaningless without power. Energy is the overlooked constraint that will determine which companies can actually build at scale. Those who secure reliable, affordable power first win. The AI infrastructure race is fundamentally an energy problem — and that doesn’t get solved by throwing more money at it.