Core Insights Review • Updated July 2026 •
Chips were the story in 2024. In 2026, the constraint has moved downstream — to substations, transformers, and transmission lines. Here's what the latest IEA, Goldman Sachs, and EPRI data says about the AI industry's new bottleneck, and how developers are working around it.
In this article
- Power has become the new AI bottleneck
- AI electricity demand is outpacing forecasts
- Goldman Sachs: U.S. demand could double by 2027
- Why grid connectivity is the industry's biggest problem
- Transformer shortages are delaying projects
- Interconnection queues as a strategic risk
- EPRI's warning on grid pressure
- The rise of "Bring Your Own Power"
- Regional winners and losers
- From a chip constraint to a power constraint
Key Takeaways
- AI demand is no longer the biggest challenge for data centers in 2026 — power availability is.
- The IEA projects global data center electricity consumption will rise from 415 TWh in 2024 to 945 TWh by 2030.
- Goldman Sachs Research forecasts U.S. data center power demand will jump from 31 GW in 2025 to 66 GW by 2027.
- Grid interconnection delays of 4–10 years are becoming the primary obstacle to AI infrastructure deployment.
- High-voltage transformers, substations, switchgear, and transmission capacity are now the industry's most critical bottlenecks.
- EPRI estimates data centers could consume 9%–17% of U.S. electricity generation by 2030.
- Developers are increasingly pursuing "Bring Your Own Power" (BYOP), microgrids, fuel cells, and behind-the-meter generation.
- Texas and Georgia are gaining ground as traditional hubs like Northern Virginia face mounting grid pressure.
AI Data Center Power Crisis 2026
Global Data Center Electricity Use (2024)
IEA Forecast by 2030
US Data Center Power Demand by 2027
Potential Grid Connection Wait
Sources: IEA, Goldman Sachs Research, DNV, EPRI (2026)
Power Has Become the New AI Bottleneck
For years, discussions about artificial intelligence focused on GPUs, semiconductors, cloud infrastructure, and computing power. In 2026, that conversation has changed dramatically.
According to Ditlev Engel, CEO of Energy at DNV, writing in the World Economic Forum article "If Electricity and Data Are the New Oil, Is Grid Connectivity the Strategic Bottleneck in the AI Transformation?" (May 18, 2026), AI infrastructure investment is growing faster than electrical grids can support.
"Grid connectivity is increasingly becoming the factor that determines which AI projects move forward and which remain on paper."
While hyperscale AI campuses can often be designed and built within two to three years, obtaining reliable power connections frequently takes four to ten years in many regions.
As a result, electricity infrastructure — not chips, capital, or software — is emerging as the dominant constraint on AI expansion.
➡️ Also read: Forecasting AI Data Centers Electricity Demand: How Much Power Will Artificial Intelligence Consume by 2030?
AI Electricity Demand Is Growing Faster Than Utilities Expected
Virtually every major forecast released in 2025 and 2026 points to one conclusion: AI is creating unprecedented growth in electricity demand.
The International Energy Agency's landmark report, "Energy and AI," projects that global data center electricity consumption will rise from approximately 415 TWh in 2024 to nearly 945 TWh by 2030.
The IEA estimates annual growth of roughly 15% — more than four times faster than overall global electricity demand growth.
Even more important is where that growth originates. According to the agency, AI-accelerated servers — including GPU clusters used for training and inference — are expected to grow approximately 30% annually and account for nearly half of future data center electricity demand growth.
The United States and China alone are expected to contribute roughly 80% of total global demand growth through 2030.
Goldman Sachs: U.S. Data Center Demand Could Double in Two Years
One of the most cited forecasts in 2026 comes from Goldman Sachs Commodities Research.
In the report "US Data Center Power Demand Projected to Double by 2027," analysts Hongcen Wei, Daan Struyven, and Samantha Dart forecast:
- 31 GW in 2025
- 41 GW in 2026
- 66 GW in 2027
This would represent more than a doubling of U.S. data center power demand within just two years, attributed to aggressive AI infrastructure deployment and hyperscaler expansion. Their analysis incorporates construction progress, permitting activity, utility planning data, facility-level development tracking, and satellite imagery.
Goldman Sachs also notes that U.S. data center capacity could exceed 95 GW by the end of 2027, highlighting the extraordinary scale of planned AI infrastructure investment.
Goldman Sachs
US Demand by 2027
IEA
Global Demand by 2030
EPRI
US Electricity Share by 2030
Bloom Energy
US IT Load by 2028
Why Grid Connectivity Has Become the Industry's Biggest Problem
The challenge is not simply generating electricity. The challenge is delivering it.
According to both the World Economic Forum/DNV analysis and the International Energy Agency, transmission infrastructure is struggling to keep pace with AI-driven demand growth. Developers are increasingly encountering:
- Multi-year interconnection queues
- Transmission congestion
- Substation shortages
- Transformer shortages
- Utility approval delays
- Local permitting barriers
The World Economic Forum notes that new AI facilities often require power connections far beyond what existing local infrastructure was designed to support. Many projects are reaching construction readiness before power infrastructure is available — reversing decades of traditional infrastructure planning, where power availability was usually assumed rather than questioned.
What Is Limiting AI Growth in 2026?
Transformer Shortages Are Delaying AI Projects Nationwide
One of the most widely discussed bottlenecks in 2026 is the shortage of high-voltage transformers.
A detailed analysis published by Tech Investments in May 2026 titled "Power Bottlenecks & The AI Data Center" highlighted the issue using data from Sightline Climate. The report noted that:
- Approximately 12 GW of U.S. data center capacity was announced for 2026.
- Only around 5 GW was actually under construction.
- Around 11 GW remained in the announcement stage with limited physical progress.
The article cites transformer lead times that have expanded dramatically. Before 2020, typical lead times were roughly 24–30 months. In 2026, lead times often extend to 3–5 years or longer.
The implication is significant: AI companies may have land, financing, GPUs, and construction permits, yet still be unable to operate because critical electrical equipment has not arrived.
Transformer Crisis Slowing AI Expansion
Pre-2020 Lead Time
2026 Lead Time
Projects Waiting
Interconnection Queues Are Becoming a Strategic Risk
The Center for Strategic and International Studies (CSIS), World Economic Forum, and multiple utility operators now describe electricity access as a strategic issue for AI competitiveness. In major U.S. markets, grid connection timelines can exceed seven years.
Northern Virginia — often called the world's largest data center market — has become a symbol of this challenge, as demand growth outpaces the expansion of transmission and generation infrastructure.
This phenomenon is not limited to America. Europe, Southeast Asia, and several emerging AI markets face similar constraints. The issue is becoming global.
Capacity Growth Is Outpacing Grid Expansion
Research from Bloom Energy's 2026 Data Center Power Report illustrates the scale of the challenge. The company estimates U.S. IT load capacity near 80 GW in 2025, rising to approximately 150 GW by 2028.
The report also predicts one in five AI campuses could reach gigawatt scale by 2030, and one in three by 2035 — unprecedented infrastructure requirements. For perspective, a single gigawatt-scale AI campus can rival the electricity demand of a major metropolitan area.
Utilities must therefore build new transmission lines, new substations, additional generation capacity, backup systems, and grid reinforcement infrastructure — all while maintaining reliability for existing customers.
EPRI Warns About Growing Grid Pressure
The Electric Power Research Institute (EPRI) intensified concerns in its 2026 report "Powering Intelligence." The organization estimates that data centers could account for 9% to 17% of U.S. electricity generation by 2030.
Virginia faces even greater pressure — EPRI projects that data centers could eventually consume between 39% and 57% of the state's electricity supply.
These numbers explain why regulators and utilities increasingly view AI infrastructure as a system-wide planning challenge rather than a niche technology issue.
The Rise of "Bring Your Own Power" (BYOP)
Because utilities cannot always deliver power quickly enough, developers are increasingly pursuing alternative solutions. A major trend identified by HSBC, Bloom Energy, and industry analysts is the emergence of Bring Your Own Power (BYOP).
Rather than waiting years for grid connections, companies deploy on-site natural gas generation, fuel cells, microgrids, battery systems, and hybrid power solutions. The goal: operate immediately while grid infrastructure catches up. Some operators now plan to remain partially or entirely independent of the grid indefinitely.
Fuel Cells Are Emerging as a Fast-Track Solution
The Tech Investments analysis highlights growing interest in fuel-cell-based power systems, with Bloom Energy becoming one of the most visible players in this space. According to company disclosures discussed in the report, Oracle's Project Jupiter selected Bloom Energy systems, with the project potentially deploying up to 2.85 GW of power capacity, designed to operate as an islanded microgrid.
The attraction is speed. Traditional utility-scale solutions often require transformer procurement, transmission upgrades, and multi-year approvals — fuel-cell systems can often be deployed significantly faster. Industry leaders increasingly describe "time-to-power" as the most important competitive metric in AI infrastructure development.
Regional Winners and Losers Are Emerging
The AI infrastructure boom is reshaping geography.
According to Goldman Sachs Research, the Mid-Atlantic, Mid-Continent, and Northwest face elevated reliability risks because planned demand growth exceeds expected generation additions. Texas and Georgia are increasingly attractive because additional generation is being built, interconnection opportunities remain stronger, land availability is greater, and regulatory processes are often faster.
Bloom Energy and other analysts also point to broader migration toward power-rich locations. The new competitive advantage is increasingly speed-to-power rather than proximity-to-users.
| Region | Power Availability | Grid Risk | Data Center Growth |
|---|---|---|---|
| Northern Virginia | Low | Extreme | Very High |
| Mid-Atlantic | Moderate | High | High |
| Texas | Strong | Low | Very High |
| Georgia | Strong | Low | High |
| Europe | Moderate | Medium | High |
| Southeast Asia | Moderate | High | Very High |
Gartner: AI Servers Are About to Overtake Traditional Servers
Another important development comes from Gartner's June 2026 forecast. The research firm estimates global data center power demand grows approximately 26% in 2026, AI-optimized servers account for roughly 31% of total consumption this year, and AI servers will surpass conventional servers in electricity consumption during 2027.
This transition matters because AI infrastructure has very different operating characteristics — higher rack densities, larger power requirements, greater cooling needs, and less interruptibility. These characteristics place additional pressure on grid operators.
The Industry Is Moving From a Chip Constraint to a Power Constraint
Perhaps the most important insight from 2026 research is that AI infrastructure constraints have shifted.
How the Industry Is Solving AI Power Bottlenecks
In 2023 and 2024, discussions focused on GPU shortages, semiconductor manufacturing, and AI model development. By 2026, leading organizations — including the International Energy Agency, World Economic Forum, Goldman Sachs Research, EPRI, Gartner, CSIS, Bloom Energy, and Lawrence Berkeley National Laboratory — are increasingly focused on power availability.
The question is no longer whether AI demand exists. The question is whether enough electricity infrastructure can be built quickly enough to support it.
For data center developers, hyperscalers, utilities, investors, and policymakers, the defining challenge of 2026 is not computing capacity — it is energy capacity.
Frequently Asked Questions
Why is power the bottleneck instead of chips in 2026?
GPU supply has scaled significantly, but grid interconnection now takes 4–10 years while a data center campus can be built in 2–3 years — flipping which resource is scarce.
What is "Bring Your Own Power" (BYOP)?
BYOP means a data center operator builds its own generation — natural gas, fuel cells, or microgrids — instead of waiting for a traditional utility grid connection.
Which U.S. regions are best positioned for AI data center growth?
Texas and Georgia currently offer stronger power availability and faster interconnection than Northern Virginia or the Mid-Atlantic, according to Goldman Sachs Research.
Core Insights Review contributors publish research-based analysis and editorial insights on commercial real estate, PropTech, smart infrastructure, sustainable construction, industrial real estate, and emerging technologies shaping the future of the built environment.
Check for more information: Core Insights Review
Try Our Calculators
Follow Core Insights Review
