AI energy demands to stall 30% of data center expansions
Surging AI power demands are projected to stall more than 30% of data center expansions by 2026 due to electrical grid constraints. Infrastructure operators are adapting by moving toward on-site power generation, liquid cooling, and distributed edge computing architectures to mitigate energy bottlenecks.
Key Takeaways
- Data centers are projected to account for 10% of total U.S. electricity demand by 2028.
- Proposed hyperscale facilities now require up to 5GW, equivalent to the power needs of 3.5 million households.
- Advanced liquid cooling and AI-driven thermal automation can reduce cooling energy consumption by 40%.
- On-site generation is shifting from a redundancy measure to a primary source via natural gas microgrids and modular reactors.
- Edge computing is being utilized to offload AI inferencing, spreading demand across multiple local power grids.
Why It Matters
The shift from treating power as a utility to a strategic constraint forces a vertical integration of energy and infrastructure. Platforms must now incorporate facility engineering and cooling efficiency directly into their technology stacks to prevent hardware bottlenecks from stalling product roadmaps. In the fragmented streaming ecosystem, operators who secure "behind the meter" power or optimize software for fewer computational cycles will maintain a competitive advantage in latency and cost. Watch for the 2027-2028 delivery window of dedicated nuclear-to-data-center projects as a signal for the next phase of localized hyperscale capacity.
Additional Context
The urgency in securing dedicated energy sources is evidenced by multi-billion dollar commitments from cloud giants. Per Reuters and Energy Digital in June 2026, the Federal Energy Regulatory Commission approved a waiver for Constellation Energy to fast-track the restart of the 835MW Unit 1 reactor at Three Mile Island. Now rebranded as the Crane Clean Energy Center, the facility is on track to supply carbon-free power exclusively to Microsoft data centers by late 2027. This follows Amazon’s June 2025 announcement of a $20 billion investment in Pennsylvania data centers, which includes a power purchase agreement with Talen Energy to pull up to 1,920MW directly from the Susquehanna nuclear plant through 2042. Energy demand is being driven by a generational shift in hardware density. According to analysis from TrendForce and Navitas in late 2024, NVIDIA’s Blackwell GPU racks can require up to 120kW per rack, compared to just 4-6kW for traditional setups. With fewer than 5% of existing global data centers capable of supporting densities above 50kW, the industry is witnessing a massive transition toward liquid cooling. Global Market Insights reported in June 2026 that the AI data center cooling market is projected to grow identifying direct liquid cooling as the dominant technology for managing high thermal design power in modern server clusters. Regulators are also responding to the strain on the public sector. Per IEA reporting in April 2026, data center electricity demand climbed 17% in 2025, significantly outpacing the 3% growth in global demand. This has led to localized bottlenecks in major hubs like Northern Virginia and Malaysia’s Johor state, where connection wait times reached seven years in 2025. Consequently, Gartner forecasts that the global data center electricity consumption will reach 565TWh by 2026, with 31% of that power dedicated solely to AI-optimized servers.
Read full article at computerweekly.com