America’s AI Data Centers Face Ghost Power Demand
💡AI data-center power requests exceed 700 GW—learn how phantom demand could reshape GPU infrastructure planning.
⚡ 30-Second TL;DR
What Changed
Large-load applications in parts of the Midwest, Mid-Atlantic, and South exceed 700 GW, with many potentially duplicated or financially unviable.
Why It Matters
Unreliable load forecasts could cause utilities either to underbuild generation and threaten grid reliability or overbuild infrastructure that customers ultimately fund. For AI companies, access to power and interconnection certainty may become as important as GPU supply.
What To Do Next
Before committing GPU capacity or a new data-center site, require signed customer commitments, ultimate-owner disclosure, and an ERCOT-style load audit in your infrastructure model.
Key Points
- •Large-load applications in parts of the Midwest, Mid-Atlantic, and South exceed 700 GW, with many potentially duplicated or financially unviable.
- •Texas demand requests for data centers and other large users rose from about 48 GW in 2023 to more than 474 GW.
- •Exelon cut its high-probability data-center demand estimate by about 40% to 11 GW after imposing stricter guarantees.
- •Texas now requires disclosure of ultimate data-center owners, taxpayer incentives, water use, and on-site generation plans.
- •PJM capacity auctions have added an estimated $29.4 billion in costs for households and businesses due to existing and expected data-center demand.
🧠 Deep Insight
Background and context from public sources — not the original article. 11 sources cited.
🔑 Enhanced Key Takeaways
- •PJM Interconnection's 2027–2028 capacity auction failed to meet reliability targets by 6,600 megawatts, directly attributing the shortfall to the surge in data center load requests.
- •The lead time for critical electrical infrastructure, specifically large power transformers, has increased to over 160 weeks as of 2026, creating a physical bottleneck for grid expansion.
- •Data centers are projected to consume between 9% and 17% of total U.S. electricity by 2030, a significant jump from the historical baseline of 4% to 5%.
- •Hyperscalers are increasingly pivoting toward 'behind-the-meter' energy solutions, such as on-site small modular reactors and fuel cells, to circumvent grid interconnection delays.
- •Approximately 75 data center projects, representing $130 billion in capital investment, have faced cancellation or indefinite delays as of Q1 2026 due to grid capacity limitations.
🛠️ Technical Deep Dive
- AI data center power density has scaled to 500MW-1GW per campus, with next-generation designs targeting 2-5GW.
- Workload variability is characterized by high-intensity training phases with oscillating power draws versus spike-heavy inference cycles.
- Ghost power is defined by the continuous 24/7 draw of idle equipment combined with reserved but underutilized grid capacity allocations.
🔮 Future ImplicationsAI analysis grounded in cited sources
⏳ Timeline
📎 Sources (11)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: IT之家 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
Weekly AI briefing
One email a week. Unsubscribe anytime.