Though most customers will be opting for something less fancy than the Superchip, they do buy them in bulk to connect together into a massive AI server and that’s where the hunger for electricity really kicks in. One study published last year looked at the energy consumption required to train a single large-language model used to output text in multiple languages.
BLOOM from startup HuggingFace drew on 176 billion parameters from 1.6 terabytes of data. It took a cluster of 384 Nvidia A100 graphical processors — GPUs — more than 118 days to crunch, according to the study’s authors. The electricity consumption from running so many GPUs for so long likely created 24.7 metric tons (54,000 pounds) of carbon dioxide, they estimated. But the true cost doubles to 50.5 tons when you take into account the network connections and idle time of the entire system.
Even then, training a model is just the start. According to one estimate from Amazon.com Inc., which runs its own AI servers, 90% of the expense from running artificial intelligence comes in the next phase when users query the model to get results — such as asking ChatGPT for chocolate-cake recipes. The energy expenditure from implementing the data, called inferencing, is hard to calculate, but it’s believed to be roughly in the order of 10 times that required in the first training phase — which means 500 tons of CO2. And a single generative AI query may have a carbon footprint four times larger than a Google search, according to one estimate.
Brute-force number crunching is built into Bitcoin’s design, and helps explain why a wave of semiconductors and servers was rolled out around the world in the hope of mining digital gold. An ongoing study at the University of Cambridge estimates that Bitcoin is responsible for 72.5 million tons of carbon dioxide. That figure could be as low as 3 million tons if all Bitcoin mines were run on hydroelectricity.(1)Compared to the wastefulness of cryptocurrency, 500 tons of carbon dioxide from a single round of training and deployment seems like nothing. Yet it’s still equal to driving one million miles in a gasoline-powered car, or 500 flights from New York to Frankfurt.
And it’s still early days. At least a dozen major technology companies are rushing to build and deploy generative AI products, including Amazon.com, Alphabet Inc., Microsoft Corp., OpenAI, Meta Platforms Inc., Baidu Inc., Tencent Holdings Ltd., and Alibaba Group Holding Ltd. Since they’re all in a race to outdo each other, they won’t sit still once a model has been trained; they’ll keep buying power-hungry processors to analyze increasingly large amounts of data. Once that’s done, they’ll compete with each other to serve up the results to consumers in the form of college essays, deepfake videos, and synthetic versions of Pink Floyd music.
Adding insult to injury is the fact that, at present, most AI training is powered by fossil fuels. These server farms have been expanded quickly in existing locations, often thousands of miles away from hydroelectric dams or solar power arrays. Since network latency is an issue when responding to internet requests, they need to be near the end user and not located thousands of miles away.
But Bitcoin has already blazed a trail for the AI industry to follow. Cold climates with plenty of renewable energy became the perfect place to plonk down a power-hungry crypto mine, with the artic air and abundant thermal energy of Iceland making the country an ideal choice. China may also find a new use for the many hydroelectric power stations that attracted mining rigs, but that lost business after Beijing cracked down on the digital currency. Foreign AI providers won’t be able to tap in, naturally, but Chinese technology giants know they have such a resource close by as their power needs rise.
There is another upside to switching out Bitcoin for AI in these server farms. While the cryptocurrency attracted numerous speculators and billions of dollars in investment, it still failed to add much value to the world. Generative artificial intelligence doesn’t have that problem; just ask ChatGPT.
More From Bloomberg Opinion:
• Charting Nvidia’s 30-Year Climb to AI Dominance: Tim Culpan
• Building a Power Station Without Spending a Cent: David Fickling
• Why All Carbon Credits Aren’t Created Equal: Lara Williams
(1) Researchers tend to use the term CO2 equivalent, taking into account the emission of other greenhouse gases which is then converted into an equivalent measure of carbon dioxide.
This column does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.
Tim Culpan is a Bloomberg Opinion columnist covering technology in Asia. Previously, he was a technology reporter for Bloomberg News.
More stories like this are available on bloomberg.com/opinion