OpenAI Warns Against Using Natural Gas for AI Data Centers

In October 2024, OpenAI sent a detailed response to the White House about how the country should build out its AI infrastructure. The filing highlighted a concern that companies might turn to natural gas-powered data centers to handle the enormous energy demands of AI systems — and OpenAI argued this would be a mistake for the climate.

The pressure is real: companies running AI systems need to expand their data centers quickly while also meeting commitments to reduce carbon emissions. Since renewable energy like solar and wind can't meet demand everywhere right now, some operators are considering natural gas as a temporary solution. OpenAI's message was clear: that path would undermine climate goals.

How Much Power Does AI Actually Need?

AI systems require a lot of electricity. When you train a large AI model — teaching it to understand language or recognize images — you run the computation across thousands of specialized chips for months on end. Then, once trained, the system needs to serve requests from users around the world with minimal delay.

Think of it this way: a traditional office building might need 5 to 10 kilowatts of power per server rack. An AI training cluster needs 40 kilowatts per rack, and some advanced setups use liquid cooling to handle up to 100 kilowatts. That's a tenfold jump in power density. It strains both the electrical wiring and the cooling systems in most data centers, which weren't designed for this intensity.

Worth flagging: OpenAI explicitly warned that gas-powered data centers built to support this demand would produce carbon emissions equivalent to those of entire nations. That would directly contradict pledges by technology companies to reach net-zero emissions.

The Renewable Energy Alternative

Instead of gas, OpenAI argued for large-scale renewable energy — solar farms and wind installations directly connected to data centers through long-term power purchase agreements with utilities.

The catch is that solar and wind don't produce power constantly. So data centers need batteries and other storage systems to bridge the gaps when the sun isn't shining or the wind isn't blowing. OpenAI mentioned exploring newer storage technologies, like compressed air and gravity-based systems, that can hold energy for longer periods.

The company also noted technical challenges: operating high-power AI systems on electrical grids built for more typical, predictable demand requires careful management of power quality and load balancing. It's solvable, but it demands planning.

Untangling the Regulatory Maze

Building a data center and connecting it to new renewable energy isn't just an engineering problem — it's a regulatory one. Federal agencies, state utility commissions, and local governments all have a say. Right now, the process is slow. Utility-scale solar and wind projects often sit in interconnection queues for years waiting for approval.

Analysis: OpenAI's filing touched on a real bottleneck facing the entire tech industry. Companies want to build AI infrastructure faster than the permitting system can keep up. The company called for clearer federal guidelines and streamlined approval for renewable projects dedicated to AI, while still maintaining environmental reviews.

Why This Matters for the U.S.

OpenAI also raised a geopolitical argument: if the U.S. doesn't have enough domestic data center capacity, companies might build AI systems elsewhere. That could mean losing economic advantage and, because AI is becoming critical infrastructure, losing a strategic edge.

The company also pointed out that some customers — government agencies and large enterprises — need their data and AI systems to stay within U.S. borders for security and compliance reasons. That creates additional demand for domestic infrastructure that the country has to meet.

Building Data Centers the Right Way

OpenAI spelled out what it sees as best practices: purpose-built data centers with liquid cooling, high-speed networking to connect chips together, and modular power systems designed for the job rather than retrofitted old buildings.

The filing also flagged a practical constraint: there aren't enough specialized AI chips to go around. OpenAI recommended policy support for domestic semiconductor manufacturing to ensure a steady supply.

What This Signals

In this author's view, OpenAI's submission marks a meaningful shift in how major AI companies think about infrastructure. By publicly rejecting natural gas solutions and insisting on renewables, the company is staking its competitive position on the bet that sustainable infrastructure will matter — both to customers and regulators.

The detailed technical specifications and policy recommendations OpenAI included could serve as a template for other AI companies facing the same choices. If the industry follows suit, it might accelerate the adoption of renewable-powered data centers across the sector. That would be good for the climate and for the credibility of technology companies' environmental pledges.

We have seen this cycle before. In the 2000s, cloud computing companies initially rushed to expand capacity as fast as possible, then gradually shifted focus to efficiency and sustainability as the technology matured. The AI infrastructure buildout looks to be following a similar arc — starting with raw performance, then increasingly constrained by environmental reality.