Since the boom in artificial intelligence got under way, US tech bosses have demanded a vast expansion of data centres and energy infrastructure to support further progress and widespread uptake of the technology. Now, the shock wave triggered by Chinese company DeepSeek is challenging that view. Some in the industry think DeepSeek’s algorithmic advances could lead to sweeping changes in the way AI models are developed and used, as well as significant energy savings and a lower climate burden. Are they right?
DeepSeek’s R1 model was a shock to US AI companies, and a mystery. How did a team of a few hundred researchers and a reported budget in the several millions of dollars produce a model as capable as OpenAI and Google’s best, with their several thousand-strong teams and billion-dollar budgets? The secret wasn’t down to a single magic ingredient, says at City St George’s, University of London, but a combination of clever engineering tricks that were individually already known about.
One of the most successful methods in AI is what’s called reinforcement learning, where researchers show an AI what success looks like and leave it to figure out the rules using a form of trial and error. This was key to Google DeepMind’s achievements with its chess and protein-folding AI systems, as success in a chess game or predicting a protein’s shape can be easily defined. However, researchers found it more difficult to translate this method to large language models, where success is less concrete.
Advertisement
While OpenAI’s GPT models use a form of reinforcement learning, where they are given feedback about which answers people prefer, this isn’t reinforcement learning in the way that people have traditionally used it. DeepSeek’s R1 was first trained on vast amounts of text from the internet, like the GPT models, but it was then left to work out how to reason by itself using reinforcement learning, without needing human feedback, says Garcez.
To do this, DeepSeek engineers focused on applying reinforcement learning to problems where they could define clear answers, such as in maths and coding, and also had the model produce many answers at once that it could compare side by side. Only then was the model shown human-labelled examples to fine-tune its capabilities in other domains.
After the model had been trained in this way, DeepSeek researchers found a way to transplant its reasoning capabilities into smaller, open-source models that had already been trained, in a step they call distillation. This distillation step is a large reason why so many people are starting to doubt that the US tech companies need as much computing power as they say they do, says Garcez.
“OpenAI and some of its competitors were going a bit crazy on scale. There was this mantra – scale is all you need – and they were scaling up every year,” says Garcez. “What we see with the distillation and the gains that they’re showing is that you don’t really have as much of a reason for scaling up.”
If tech companies need less computing power to train models, this might mean AI doesn’t have to be as damaging to the climate as it currently is, and that plans like a $500 billion AI infrastructure project in the US might not need to go ahead. But US tech companies have been quick to push back against this. Satya Nadella, the CEO of Microsoft, which owns a large stake in OpenAI, that “as AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”
“The CEOs, if you speak to them, of course they have to do the hard sell, and they have to predict that AI will need these data centres,” says at the University of Edinburgh, UK. “But if I can run this stuff on my computer, on one GPU [graphics processing unit], why would I actually pay OpenAI to do anything?”
However, another aspect of DeepSeek’s R1 model might actually increase the energy demands of AI. Like OpenAI’s o1 reasoning models, it uses a method called chain of thought, in which the AI “thinks out loud” and shows its working when asked to respond to a prompt, which researchers have found can improve its performance on some mathematics and coding tasks. If many more people start using AI tools that need to think in this way, it could lead to a greater computational need and cost, as Nadella predicts.
But DeepSeek’s thinking time is divided up between several subsystems that are expert in different fields, such as mathematics or coding, in what is called a mixture of experts model. This will lead to less computational power being needed than using the entire model, says at the University of Sheffield, UK. Also, the vast majority of requests might not need the most computationally intensive “thinking” models, says Aletras, leading to lower overall energy costs.
“If I had to explain every single response to you, we would never finish this meeting,” says Aletras. “Sometimes [chain of thought] is useful… but if I ask a question that’s very straightforward, then I don’t need it there.”
Ultimately, how much of an effect DeepSeek’s innovations will have on the AI industry, and its energy consumption, will depend on whether US tech companies can show that their approach delivers superior results. But with customers able to use DeepSeek’s R1 for less than one-twentieth of the cost of models like OpenAI’s o1, the difference in quality would have to be substantial. “If we don’t have to pay, why would we? And that means that the energy consumption would be effectively lower,” says Lapata.
