AI server power density increased 11x between 2020 and 2025 and could multiply another 4x times by 2027.
Efficiency gains are being outweighed by rapidly growing usage and more power-intensive agentic AI tasks.
Power availability is becoming a critical constraint, driving investment across generation, storage, grid equipment, and onsite power.
At its core, AI is a conversion: It turns electricity into intelligence. Every token generated and every model trained is electricity transformed into cognition. As with the technologies that powered earlier eras, from coal during the Industrial Revolution to silicon during the information age, it pays to focus on the input.
A single AI server rack is about the size of a refrigerator but, by 2027, could consume the equivalent power of 65 households. The pace of change is remarkable: Between 2020 and 2025, AI server power density increased 11x and is set to multiply another 4x by 2027.1
This compounding is driven by accelerated computing itself—each successive GPU generation draws vastly more power per rack. In the future, new power architectures, including the conversion of data centers to 800V DC systems, could see individual racks reach one megawatt of power. Growing power density is also pushing data centers to unprecedented scale, with some drawing one to two gigawatts, equivalent to the output of one to two nuclear power plants.
At the hardware and code level, AI is becoming ever more efficient. A gigawatt of power is producing more tokens each month.
Tokens Produced Per Gigawatt of Data Center Capacity
Trillion tokens per GW per month
Yet these gains are being more than offset by two major forces:
First, AI usage is surging, with the number of tokens processed rising rapidly. This is the Jevons paradox in action: Each efficiency gain lowers the cost per token, expanding what AI can be used for and raising total consumption.
Second, new AI tasks are far more power hungry. An agentic AI task involving reasoning can require orders of magnitude more power than a traditional chat-based query.1
GPU Electricity Consumption Across Model Types
Source: International Energy Agency, Apr 2026
As a result, the IEA projects that global data center electricity consumption will more than double to approximately 1,000 TWh by 2030, driven by AI and representing around 3% of global electricity demand. These estimates continue to be revised higher, and some forecasts are even larger. Morgan Stanley estimates that U.S. data center power demand alone could reach 1,100 TWh by 2030.2
The constraint is increasingly physical. Lead times for large power transformers have stretched from 24 to 30 months before 2020 to as much as five years, while high-voltage switchgear can take more than two years. Industry estimates suggest that 30% to 50% of planned U.S. data center openings in 2026 face delays because of power equipment alone.
Electrical equipment represents less than 10% of a data center’s cost, but sits on 100% of its critical path.3
Key Buying and Decision Factors for Data Center Site Selection
% of respondents ranking factor in top 3
Source: Bloom Energy, Jun 2026
Meeting this surge in power demand is reshaping an electrification landscape that has suffered from almost five decades of underinvestment:
Existing technologies are being underwritten in a major way, including the push to revive nuclear power in the U.S.
New technologies, such as small-scale nuclear reactors and even fusion, are being actively explored and funded.
Storage is seeing growth and innovation as it becomes increasingly important for absorbing the large and rapid power swings created by AI data centers.
Grid hardware including transformers, switchgear, high-voltage direct current (HVDC) systems, and cabling, is facing multi-year backlogs.
Onsite, or behind-the-meter, power is booming as data centers turn to turbines, fuel cells, and other solutions to get online faster.
AI runs on electricity, and its growth now depends on a major buildout of power infrastructure. The Tema Electrification ETF (VOLT) offers actively managed exposure across the value chain supporting that expansion, from generation and storage to grid equipment and transmission.