Metered AI

In the last eight years, transformer models and generative AI have changed the world. A survey by Pew Research found that 55% of Americans said they regularly use AI. At this point, AI adoption is more than double that of the Internet in the early 1990s. 

But AI is eating the power grid. Provisioning AI requires the greatest buildout of data center infrastructure ever, far greater than the construction of Internet data centers. With the advent of generative AI, people began by saying that AI has come to take our jobs;  now they’re saying it’s come to take our energy.

The Dell’Oro Group forecasts $1 trillion in data center capex by 2029. Half of that will be focused on AI and accelerated servers. Microsoft has committed to a $100B data center in Wisconsin with its own nuclear reactor; You could build seven aircraft carriers for less. Microsoft also signed an agreement to restart the Three Mile Island nuclear reactor for a data center there. 

Elon Musk tweeted recently that he had “been thinking about the fastest way to bring a terawatt of compute online. That is roughly equivalent to all electrical power produced in America today.” Jensen Huang said many people think they have a technology access problem, but what they really have is an energy access problem. 

And then there’s heat. An NVIDIA DGX H100 AI server rack produces 38,557 BTU per hour. Processors can reach up to 176 degrees Fahrenheit, and there are as many as 30,000 processors in an AI data center. Heat leads to failures, which are punishing given that an AI rack costs $373k. 

New liquid cooling technologies are the only way to keep these data centers cool. Liquid cooling can reduce energy consumption and generate greater performance with the silicon we already have. Better software can be processed more efficiently to skip unnecessary steps, similar to Deep Seek’s LLMs introduced in January. But the best solution for energy efficiency may be in the hands of users. 

Users don’t discriminate when they search, throwing terms like weather forecast or movie times into the biggest LLMs, even though a standard Google search would work just as well. We shouldn’t use big, energy-hogging models to answer simple questions. People don’t need to use generative AI as a calculator; they need to use a calculator as a calculator.

The average number of chat inquiries per user per day on ChatGPT is eight. Each inquiry uses .3 KwH of electricity and 16 ounces of water. Current pricing from local utilities show the cost per inquiry would be $0.00123 for electricity and $0.0058 for water, or about $.01 total per series of chat inquiries. Pay-as-you-go is heresy in tech and unlikely to happen, but usage pricing would share the environmental concerns with users every time they hit the return key.

As AI becomes more integrated into daily life, users’ decisions about when and how to use AI-powered services affect overall demand for energy and water resources. By being selective—choosing smaller, more efficient models, or reducing unnecessary usage—users can help curb the rapid growth in resource consumption. While the industry is working on more efficient hardware, better cooling, and greener infrastructure, these advances alone may not offset the exponential growth in AI demand. Users taking responsibility for their own consumption can complement these efforts and help ensure that the benefits of AI do not come at an unsustainable environmental cost.