Energy Tech and the Importance of Gas/LNG to America's Data Center Buildout

// The AI Compute Boom Is a Power Story

Strip away the model architecture debates, the benchmark competitions, and the enterprise go-to-market strategies, and the AI compute boom ultimately reduces to a single, physical constraint: electricity. A single Nvidia H100 GPU consumes approximately 700 watts. A hyperscaler GPU cluster of 100,000 H100s — now a common configuration for frontier model training — consumes roughly 70 megawatts, equivalent to a small city. Microsoft, Google, Amazon, and Meta have collectively announced over $300 billion in data center capex through 2027. Lawrence Berkeley National Laboratory projects that U.S. data centers will consume 6–12% of total national electricity generation by 2028, up from approximately 4% today. The utility industry, which has spent two decades planning for flat or declining electricity demand, is facing a demand shock it is structurally unprepared to absorb. Transmission interconnection queues already run to seven to ten years in most jurisdictions. The gap between what the AI industry needs — reliable, dispatchable, gigawatt-scale power delivered to specific physical locations within 24 to 36 months — and what the grid can provide is not a policy problem. It is a physics and civil engineering problem, and it has one near-term solution: natural gas.

// Why Renewables Cannot Bridge the Gap

The instinctive response to a power demand problem — build more solar and wind — runs into three structural barriers that are not resolvable within the timeframes the AI industry requires. First: intermittency. Data centers require 24/7/365 "five nines" (99.999%) uptime. Solar generates power roughly 25–30% of the time; wind roughly 35–40%. Battery storage sufficient to bridge multi-day low-generation weather events at gigawatt scale does not exist commercially and will not for years. Second: permitting timelines. Utility-scale renewable projects in the United States now take an average of five to seven years from announcement to operation, driven primarily by environmental review, grid interconnection studies, and transmission buildout. Third: location mismatch. Data centers cluster in specific geographies driven by land cost, fiber connectivity, and tax incentives — Northern Virginia, Phoenix, Dallas, Columbus. These are not always geographies with surplus renewable capacity or available transmission headroom. The Electric Power Research Institute's 2024 load forecast for AI data centers projects an additional 47 GW of demand by 2030 — equivalent to adding the entire generating capacity of California to the grid in six years. Natural gas generation, by contrast, can be permitted, constructed, and commissioned in 18 to 36 months. Combined-cycle gas plants achieve 90%+ capacity factors with dispatchable output. For the specific use case of powering AI data centers at scale, gas is not a compromise; it is the only credible solution available in the required timeframe.

// The Investment Opportunity in Energy-Adjacent Tech

The intersection of the AI compute boom and the power constraint creates investment opportunities across multiple layers of the energy stack. At the macro level, natural gas and LNG infrastructure operators — pipelines, storage facilities, liquefaction terminals — are seeing demand profiles that fundamentally change their long-term economics. In January 2025, the U.S. Department of Energy lifted its pause on new LNG export approvals, unlocking a significant new pipeline of export projects. EIA data shows that U.S. dry natural gas production set a new record in 2024, yet domestic natural gas prices at Henry Hub have remained structurally elevated by historical standards — reflecting the new demand vector from power generation. At the technology layer, the more directly relevant opportunity is in the companies building the operational and optimization software that sits between the energy supply chain and data center operators: power procurement platforms, co-location generation management systems, demand response software, and grid edge management tools. These businesses share the characteristics we find most compelling: they generate proprietary operational data as a byproduct of doing their work, they are deeply embedded in mission-critical infrastructure, and they sit at an AI inflection point where intelligence layered onto their data can create step-change improvements in efficiency and reliability. The AI buildout and the energy transition are not competing narratives — they are converging on the same infrastructure bottleneck, and the companies that help resolve that bottleneck will be among the defining businesses of the decade.