The explosive growth of artificial intelligence is creating an unprecedented power crunch that traditional electrical grids cannot handle. Global data center power demand is projected to reach 132 gigawatts (GW) by the end of 2026, a 27% increase from 2025, according to recent infrastructure analysis. This surge is forcing a fundamental shift in how data centers operate, moving away from reliance on centralized power grids toward hybrid models that combine on-site power generation, energy storage, and intelligent demand management.

Why Are Data Centers Running Out of Power?

The problem is straightforward but severe. Training and running large language models (LLMs), which are AI systems like ChatGPT that process and generate human-like text, consumes enormous amounts of electricity. An International Monetary Fund study from 2025 estimated that AI-driven annual electricity consumption could reach approximately 1,500 terawatt-hours (TWh) by 2030, equivalent to the current total electricity consumption of India, a nation of 1.4 billion people. Some analysts at consulting firm McKinsey project the figure could exceed 2,000 TWh.

The infrastructure challenge extends far beyond just electricity. Building out the AI ecosystem requires steel and cement for data center construction, fiber optic cables for data transmission, and advanced semiconductor manufacturing. Taiwan Semiconductor Manufacturing Company (TSMC), which produces the world's most advanced chips, alone consumes roughly 10% of Taiwan's entire electrical grid. The island's power supply is dominated by coal and natural gas, making the semiconductor industry highly carbon-intensive.

Multi-year interconnection delays from utilities have created a critical bottleneck. Companies cannot wait years for grid upgrades to connect new data centers, forcing them to seek alternative power solutions immediately.

What Solutions Are Startups Building?

A new class of companies is emerging to address this crisis by delivering power solutions directly at the data center source. These startups are tackling the problem from multiple angles, ranging from on-site power generation to advanced cooling technologies and AI-driven energy management.

• On-Site Linear Generators: Mainspring Energy is commercializing high-efficiency linear generators that run on natural gas and hydrogen, allowing data centers to generate their own power without depending on the grid.
• Flared Gas Conversion: Crusoe Energy co-locates modular data centers with wasted natural gas sources, converting energy that would otherwise be burned off into computing power. The company is expanding its AI data center campus in Abilene, Texas, to 1.2 GW, with the second phase completion scheduled for mid-2026.
• Gigawatt-Scale Infrastructure: American Terawatt acquires and develops large-scale sites with significant power and water access, providing gigawatt-scale energy infrastructure for next-generation data centers.
• Advanced Battery Storage: Energy Plug develops battery energy storage systems (BESS) and AI-driven software to optimize energy use and provide backup power for AI clusters.
• Real-Time Demand Management: Emerald AI, backed by NVIDIA, creates software that helps data centers modulate their energy consumption in real-time to avoid straining the electrical grid.
• Grid Flexibility Platforms: Grid Care enables data centers to participate in demand response programs and act as grid assets, with analysis showing that a 1 GW flexible data center can reduce costs by 5% for all ratepayers on the grid.
• AI-Powered Energy Forecasting: Amperon provides electricity forecasting to help utilities and data centers manage power procurement for volatile AI workloads, with investment from National Grid's venture arm.
• Photonic Interconnects: Celestial AI secured $593.9 million in funding to commercialize its Photonic Fabric, an optical interconnect technology that reduces power consumption and latency between chips, addressing a major energy bottleneck in large AI models.
• Photonic AI Processing: Lightmatter develops AI hardware that uses light for data processing, designed for greater speed and energy efficiency at the chip level.

How Can Enterprises Reduce Data Center Energy Consumption?

Beyond hardware solutions, companies are implementing software-based strategies to make AI infrastructure more efficient. One emerging approach involves AI model compression, which reduces the size of large language models while maintaining performance.

• Model Compression Technology: Multiverse Computing announced a $570 million Series C funding round to scale its CompactifAI platform, which applies tensor network methods from quantum physics to compress large language models by 80 to 95% with minimal accuracy loss. This enables faster, lower-energy AI deployment across cloud, on-premises, and edge devices.
• Hybrid Deployment Strategy: Rather than routing every workload through a hyperscaler data center, enterprises can use compression and routing software to match each workload with the right model and backend, potentially saving hundreds of millions in infrastructure costs while reducing energy consumption.
• Edge Device Processing: Compressed models can run directly on smartphones, factory equipment, and other edge devices without sending data to centralized data centers, eliminating transmission energy costs and latency.

"The AI industry has accepted a false constraint for years, that powerful models require expensive infrastructure. That constraint is gone. We have proven that AI can run at full performance on a smartphone, inside a sovereign data center, on a factory floor with no cloud connection," said Enrique Lizaso, co-founder and CEO of Multiverse Computing.

Enrique Lizaso, Co-founder and CEO of Multiverse Computing

Multiverse's Series C round, co-led by Forgepoint Capital International, BNPP SIVF, and Bullhound Capital, brings total funding to $800 million and positions the company among the fastest-growing AI infrastructure businesses in Europe. The company has achieved 96x year-over-year sales growth in the first quarter of 2026 and is already deploying models across millions of devices including drones, cameras, satellites, vehicles, and telecom infrastructure.

What Is the Broader Environmental Impact?

The energy footprint of AI extends far beyond operational electricity consumption. Researchers analyzing the global impact of AI diffusion distinguish between operational emissions, which come directly from running AI systems, and embodied emissions, which result from manufacturing, transporting, constructing, and installing the physical hardware and infrastructure supporting the AI ecosystem.

Embodied emissions include activities like smelting steel for data center construction, mixing concrete for foundations, cleaning silicon wafers for chip production, and shipping advanced lithography equipment. For example, a single Extreme Ultraviolet (EUV) lithography machine from Dutch semiconductor company ASML requires an average of three cargo planes, 40 freight containers, and 20 trucks to ship disassembled components to Taiwan for reassembly.

"The AI industry has built an extraordinary capability on top of an infrastructure model that wasn't designed to sustain it. The winners in the next phase of this multi-trillion-dollar market will be the ones that make every data center, every GPU, and every enterprise deployment dramatically more efficient," said Damien Henault, Managing Director and Partner at Forgepoint Capital International.

Damien Henault, Managing Director and Partner at Forgepoint Capital International

The data center power market is projected to exceed $50 billion by 2030, reflecting both the scale of the challenge and the investment opportunity for companies solving it. As AI adoption accelerates globally, the race to build efficient, sustainable power infrastructure has become as critical as the race to develop more powerful AI models themselves.