How Much Electricity Is 11 Billion kWh? Almost More Than We Can Imagine, But It's Only One Measure of AI Demand
Guest Post from Tom Pyle at the Institute for Energy Research.
Over the past year, there have been many stories on the staggering electricity demands from artificial intelligence, or AI. The Wall Street Journal provided a new example of AI’s insatiable electricity consumption. In an article about the challenges OpenAI (the most famous AI company and maker of ChatGPT) is facing in releasing its next-generation AI model (Orion), the WSJ says:
OpenAI has conducted at least two large training runs, each involving months of crunching huge amounts of data, with the goal of making Orion smarter. However, people close to the project say that each time, new problems arose, and the software fell short of the results researchers were hoping for.
They say that, at best, Orion performs better than OpenAI’s current offerings but hasn’t advanced enough to justify the enormous cost of keeping the new model running. Based on public and private estimates of various aspects of the training, a six-month training run can cost around half a billion dollars in computing costs alone.
A half billion dollars in computing costs, the vast majority of which is electricity, is an amazing amount of energy for a single training run—and OpenAI has conducted at least two large training runs for its next-generation AI model. Worse, OpenAI has not finished training the model. In other words, it’s reasonable to assume that OpenAI has spent (or will spend) $1 billion on electricity alone, training its next-generation model.
So how much electricity can you buy for $1 billion? Virginia is the world leader in data centers and, according to the latest data from the Energy Information Administration, industrial electricity rates in Virginia averaged about 9 cents per kilowatt hour (kWh) for the past year. Therefore, $1 billion in electricity costs means the data centers training OpenAI’s next-generation model could have used approximately 11 billion kWhs of electricity.
How much electricity is 11 billion kWh?
1 million homes for a year: The average U.S. home consumes 10,500 kWh a year. That means 11 billion kWh could power one million average U.S. households for a full year.
The U.S. Steel Industry: The entire U.S. steel industry currently consumes about 11 billion kWh a year.
More than the output of a 1 GW nuclear reactor for a year. A 1 GW nuclear reactor can produce about 8 billion kWh per year.
11 billion miles by an EV: A Tesla Model 3 uses about 25 kWh to travel 100 miles. With 11 billion kWh, you could drive that Tesla for a mind-blowing 44 billion miles—roughly three roundtrips to Neptune.
Training vs. Running an AI Model
This is an amazing amount of electricity just to train the model—it isn’t the ongoing electricity costs of running the model. Training a model is the initial phase, where massive datasets are processed using powerful hardware to enable the AI to learn patterns and make predictions. This process takes months and involves vast computational and electricity resources.
Running a model, conversely, refers to deploying the trained AI to perform tasks, such as answering questions or generating text. This isn’t as computationally intensive and, thus, energy-intensive, but as more people use the new AI tool, the electricity demands of running the models will grow. And since many companies and individuals fear being left behind in the march towards the many uses of AI, the “latest and greatest” is likely to attract heavy use and demand for more energy.
Conclusion
This exercise is a rough estimate of some of AI’s cascading electricity needs. While this example represents an enormous amount of electricity, it is not an outlier, as the other major players in the AI industry—Google, Meta (Facebook), Microsoft, Amazon, Anthropic, Mistral, and Alibaba—will likely spend comparable amounts on electricity to train their next-generation AI models as well.
AI has incredible potential, and that potential will only be reached through the use of massive amounts of electricity. The question for policymakers must be what impediments exist to ensure the timely provision of sufficient supplies of adequate and affordable electricity to allow Americans to enjoy the benefits of AI’s promise.
Editor’s Note: AI is in the process of destroying all green energy fantasies based on solar, wind, EVs and the like. They won’t even begin to address future energy demands. Moreover, solar and wind are land hogs and totally unreliable. Battery storage, too, amounts to virtually nothing and is ludicrously expensive and dangerous to boot. There is simply no future for solar or wind in an AI economy. An old Yiddish proverb based on the Talmud says “Mann Tracht, Un Gott Lacht” or, in English, “Man Plans, and God Laughs.” Nothing better explains where green energy fantasies are headed. Only an energy system based on coal, oil, gas and nuclear is realizable.
#AI #Nuclear #Oil #Gas #NaturalGas #EnergyDemand #GreenEnergy #Solar #Wind #BatteryStorage #EVs #ArtificiaIIIntelligence
If we are going to have real improvements in these “AI” Large Language Models, we need competition,which means lots of these models training and running. Even coal and gas generators will have trouble getting enough fuel delivered fast enough to keep up with the massive demand. Only fission has the energy density required to handle this new load with high reliability. The problem with big nuclear plants is those long lead times. We had better start building as fast as we can. Current nuclear technology has an unmatched safety record, but even high tech moguls have taken the “next generation” bait. None of those paper-reactors are ready now, when we should be actively building approved designs. They are also not big enough. We should be building AP1000s and ABWRs by the dozen before it’s too late.
Do we need AI and treats centers?