The big tech companies treat their climate footprint like shoppers with unlimited budgets. Companies like Google, Meta and Microsoft are using more and more energy while using their massive profits to buy clean energy credits to “cancel” it out. This allows for claims that such companies are carbon neutral, even though the companies themselves know this isn’t how emissions from electricity actually work. In reality, there are tradeoffs between electricity consumption and climate impact.
Nothing exemplifies this more than the push for AI. Tech companies seem to be treating the climate impact of AI like they’re shopping in a store that has no price tags with a credit card in someone else’s name. And they’re not showing us the receipts. In our latest video on the climate impacts of AI, we look at what we know - and what we don't know - about the climate impacts of our AI queries.
Through 2021, researchers at companies like Google, Meta, and OpenAI were actually transparent about the energy demands of training their generative models. They published this information along with other research. In 2021, researchers at Google wrote in one study that “We believe it is straightforward for ML [machine learning] practitioners to calculate energy consumption.” They explained it’s little more than doing basic math with simple data points like hours of training, number of processors, and data center efficiency statistics. And then they stopped doing it, along with most of their other industry colleagues.
Companies do not disclose what percentage of a company’s data center capacity is dedicated to AI, how carbon-intensive AI computing is, or what share of their recent astronomical increases in electricity use is due to AI expansion. Nor did they respond to our requests for this information.
Still, we were able to piece together some numbers, albeit an incomplete picture. For example, training GPT-3—a smaller precursor to the models found through ChatGPT, Microsoft’s Copilot, and soon Apple devices—emitted the equivalent carbon dioxide of 254 round trips between New York and Los Angeles by car.
The energy impact of actually using the models is trickier to estimate because companies don’t say how many queries their models get per day or how much electricity it requires to respond to each query. But using the results of a study done by Huggingface in partnership with researchers from Carnegie Mellon University, we can come up with some rough estimates: around half a watt-hour per query. If we estimate roughly 10 million queries per day for a model, that’s around 2.1 tons of carbon dioxide emitted per day by a generative text model, or as much as you’d burn driving a car from New York to Los Angeles, then to Seattle, and then to Denver. And that’s to say nothing of the climate price of manufacturing the special chips required to train and run these models.
To be clear, these are educated guesses based on the best information we have available. Sometimes, journalists struggle with stories where the main takeaway is how little we know. But, the main takeaway here is that we should know more. If AI truly is as revelatory as these companies say it is, and will soon be as ubiquitous and world-altering as they say it will, isn’t it imperative they disclose the climate costs of that technology so we can decide if it’s worth the cost? Shouldn’t we be able to know the price before we buy?
