DeepSeek Claims to Be a Sustainable AI Tool. But Is It Really?
The Chinese app consumes significantly less energy than ChatGPT, but training any Large Language Model requires vast amounts of computer power.
DeepSeek's chatbot was first released in January (mundissima / Shutterstock.com)
Have you ever considered the sustainability implications of your artificial intelligence use? If you’re one of the 99% of Americans using at least one AI-enabled product every week, perhaps you should. AI servers are always online and require constant connection to electricity – up to 100,000 watts per hour for each set. Their power is often derived from nonrenewable fossil fuels, like coal and natural gas. Their component parts must be continuously cooled with water – up to one cup per five chatbot responses.
Certainly, AI tools can do a lot of good—in medical research, locating missing persons, and even finding new ways to optimize energy efficiency. However, the average consumer doesn’t employ artificial intelligence for noble purposes. Instead, our top uses of AI models are drafting texts and emails, getting financial advice, and planning travel.
Considering sustainability, I have chosen to actively avoid AI helpers—especially for tasks we all used to do adequately without artificial intelligence. But when I heard a newcomer claimed to have developed a sustainable alternative, I was intrigued. Deep Seek, a Chinese startup, alleges its new V3 generative AI model uses up to 40 times less energy than American counterparts from competitors like Meta and OpenAI. Though Deep Seek’s product is a step in the right direction, calling it sustainable is a step too far.
Can AI ever be sustainable?
V3 is a large language model that, like ChatGPT and similar programs, can generate text responses when presented with a query. The models’ artificial intelligence is derived from feasting on monumental amounts of training data—the “cumulative sum of human knowledge,” in the words of Elon Musk. According to a Deep Seek company report, V3 was exposed to 14.8 trillion unique pieces of information. The models can then algorithmically rework learned patterns into original output.
Indeed, training generative AI models demands staggering upfront energy costs long before the product ever reaches consumers. Favorably for Deep Seek, despite skepticism from some industry leaders, independent research confirms the company’s sustainability claim… to a point. According to findings published in Nature, V3 required 11 times less computational power to train than Meta’s similar large language model Llama 3.1 450B—an auspicious prospect. But a fraction of a very large number is still another very large number.
According to Meta, Llama 3.1 450B was trained over a period of 54 days using 16,000 computer processors. Each processor ran at 700 watt hours (in other words, 700 watts per hour). Therefore, the entire Llama training period represents a power usage of over 14.5 billion watts. For comparison, the average American household consumes about 10 million watt hours of energy each year. Meta’s training period of less than two months is more than what an entire household would use over 1,450 years!
Training generative AI models demands staggering upfront energy costs long before the product ever reaches consumers.
To Deep Seek’s credit, V3 used only 2,048 processors over a similar timeframe of about two months. Despite the company not providing detailed energy use information, we can still calculate an approximation of V3’s power consumption during training. H800 processors, the type V3 was trained on, run using up to 350 watts per hour. Therefore, over a 60-day period, the model’s training may have used about 1.4 billion watt hours of energy. Indeed, this represents power consumption an order of magnitude less than Meta’s Llama model. However, despite V3’s improvements to industry norms, it still used more energy in several weeks than an American household would in 140 years.
What about OpenAI’s ChatGPT, the large language model industry leader? Researchers found that training the GPT-3 program, first released publicly in November 2022, consumed nearly 1.3 billion watt hours of energy (and also released 502 tonnes of carbon dioxide into the atmosphere). That’s better than both Llama and V3—but by now, GPT-3 has been all but replaced by its successor GPT-4, which came out in early 2023. Using 25,000 processors, OpenAI’s latest model required nearly 50 times more energy than its predecessor: a whopping 50 billion watt hours, enough to power an American household for 5,000 years!
300 American households could be powered for a year with the same amount of electricity ChatGPT uses in 24 hours.
To be sure, V3 beats the rest when it comes to training efficiency (even if its energy consumption is still enormous). But what about the impact of consumer AI use, post-release? Within a week of launching in late January, Deep Seek’s artificial intelligence chatbot based on V3 was the most-downloaded free app in the United States. After three weeks, it was installed more than 16 million times globally across platforms. With such a large user base, energy efficiency matters during operation just as much as during training.
This is where Deep Seek’s sustainability claims unravel. Chatting with ChatGPT and Deep Seek’s app both consume the same amount of energy: about three watt hours per response. For reference, a traditional (pre-AI) Google search only requires about one-tenth of that. Though usage statistics beyond total downloads are not yet available for Deep Seek’s program, ChatGPT alone boasts 400 million weekly users who send one billion messages to the chatbot every day. At the going energy rate, 300 American households could be powered for a year with the same amount of electricity ChatGPT uses in 24 hours. As the global AI user base continues to grow and new competitors like Deep Seek join the market, total energy costs will only rise further.
Unavoidable Artificial Intelligence
Even the most efficient generative AI model available uses an unacceptable amount of energy, especially considering most consumers don’t use it to solve cold cases. But as AI tools become ubiquitous helpers for mundane tasks, avoiding them may prove to be near-impossible: new automatic AI integration into Google Search results will consume about 25 trillion watt hours of electricity every year through 9 billion daily searches.
There’s still hope. Researchers at Cornell University found that non-AI, function-specific programs use 30 times less energy to complete tasks than broad-use, generative AI models. As pervasive as artificial intelligence is, more sustainable and energy-efficient tools are out there to help us learn, communicate, and be entertained. We only have to seek them out—just don’t go asking a chatbot for ideas.
Sophia Moore is a graduate of NYU’s College of Arts and Science in Journalism and Romance Languages. She is pursuing an accelerated Masters at NYU in International Relations, and hopes to work for the U.S. Department of State.
