DeepSeek has been hailed as the greener alternative to Western artificial intelligence (AI) models, but experts say China’s answer to ChatGPT may not herald a carbon-lite era for AI.
To continue reading, subscribe to Eco‑Business.
There's something for everyone. We offer a range of subscription plans.
- Access our stories and receive our Insights Weekly newsletter with the free EB Member plan.
- Unlock unlimited access to our content and archive with EB Circle.
- Publish your content with EB Premium.
Chinese start-up DeepSeek claimed to have matched the performance of established players from the United States but at a fraction of the cost and carbon footprint when it unveiled DeepSeek V3 in January. But while the open-source AI model may have been cheaper to develop – the claimed US$6 million cost to train the new model has been disputed – this does not mean that it is cheaper and greener to use, technology sustainability experts told Eco-Business.
“They [DeepSeek and ChatGPT] are similar in the sense that they are open source large language models. While DeepSeek may have spent less on training their model, it is likely just as expensive and emissions-intensive to run,” said Tatiana Collins, chief executive of Ingenie, a green tech advisory firm.
Though DeepSeek has not disclosed its operational carbon footprint, US tech firms Google and Microsoft have reported massive spikes in emissions due to surging AI use over the last year, casting doubt over these companies’ ability to meet net zero targets.
“
AI firms must actively integrate sustainability into their growth strategies to ensure that efficiency gains lead to actual environmental benefits rather than simply fueling greater AI expansion.
Dr Bernard Leong, founder and chief executive, Dorje AI
DeepSeek may hold promise in limiting the carbon footprint of AI in that its per-query energy consumption is lower than Western rivals, noted Dr Bernard Leong, founder and chief executive of Dorje AI, an enterprise AI startup. However, its perceived superior efficiency may encourage higher usage, which would drive up emissions, he said. This phenomenon is known as Jevons Paradox.
DeepSeek’s strategy of reducing cost per query, with expenses projected to fall five-fold by the end of the year, could rapidly accelerate AI adoption, driving up emissions further, Leong cautioned.
“Although DeepSeek demonstrates that AI can be developed more efficiently, true sustainability depends not just on technological advances but also on systemic shifts in energy sourcing, policy, and responsible deployment,” he said.
“DeepSeek is enhancing compute efficiency, making AI development more cost-effective. However, whether this translates into genuinely greener AI remains uncertain,” he said.
For now, the technology’s breakneck deployment could still surpass the availability of renewables. An analysis published on the World Economic Forum’s blog after the launch of DeepSeek forecasted that AI’s energy demand is still set to grow faster than clean energy expansion, locking in continued fossil fuel dependence.
If AI-related electricity consumption reaches 134 terrawatt hours (TWh) annually by 2027, and only 38 per cent of the demand is met with renewables, that would still leave 83.08 TWh powered by dirty coal and gas, it said.
Regulation gaps
Regulation that keeps AI’s environmental footprint in check by obliging developers to train AI models with efficiency in mind has been lacking, noted Heng Wang, a professor of law at Yong Pung How School of Law, Singapore Management University.
The European Union’s AI Act was the first legal framework on AI, and is the most focused on ethical AI use – but it does not include provisions for the environmental impact of AI.
Neither does the US have clear provisions for AI use. Last month, an executive order from President Donald Trump replaced the Biden administration’s policing of disinformation and bias in AI with policies aimed at removing perceived obstacles to innovation.
Asian countries have mostly taken a business-friendly approach to AI regulation. Japan offers ethical guidelines on AI use, focusing on transparency, societal impact and responsibility. China, which has ambitions to lead global AI development by 2030, has placed regulatory focus on data security and national security.
“Many jurisdictions want to be seen as leaders in AI, and there is concern that regulation will have a chilling effect on innovation and increase costs. If emissions control regulations are implemented, there is a risk that tech companies move elsewhere,” said Wang.
In an opinion piece published on Singapore’s Centre for International Law’s website, Jon Truby, an expert on AI and technology law and a visiting professor at professor at the research centre, said that regulatory interventions could discourage the use of energy-intensive datasets and prevent the externalisation of environmental costs onto taxpayers.
But DeepSeek does demonstrate that AI can be trained in a more efficient way, and hence the pressure should fall on existing providers to reduce the energy intensity of their models to save costs and reduce climate impact, he said.
Tech firms have not had to respond to regulatory pressure to curb the environmental cost of AI, which includes massive water use as well as electronic waste. Nuclear waste could be another consideration if more companies turn to atom-splitting to power their operations, as Google announced it plans to last year.
Most large firms have set climate goals to address their impact, but the tech sector – which accounts for 2 to 3 per cent of global emissions – has been criticised for underreporting and disguising supply chain, or Scope 3, emissions, the bulk of which comes from energy-guzzling data centres.
Google has lobbied the GHG Protocol, a standard for assessing greenhouse gas emissions, to exclude Scope 3 reporting for technology companies. The tech giant committed to achieve net zero emissions across its value chain by 2030, but its emissions have jumped by 48 per cent since 2019, mainly as a result of integrating AI into its products.
Amazon, which has positioned itself as a climate leader in the tech space, has been found to have excluded large chunks of its Scope 3 emissions from its carbon reporting. The company was removed from the Science Based Targets initiative (SBTi)’s list of companies taking action aligned with global climate goals last year.
Microsoft, which in 2020 set the radically optimistic target of being carbon negative by 2030, has seen its emissions soar by a near 30 per cent since 2020 – largely as a result of increasing AI use.
“Big Tech has never really thought about efficiency when programming AI models – it’s growth at all costs,” one tech sector source who chose to withhold their name told Eco-Business.
Why Asia could win at green AI
The boom in data centre building in Southeast Asia to meet global demand for AI services has called into question the carbon impact of new technology on corporate and national net zero ambitions.
Singapore put a moratorium on data centre building in 2019, because of their carbon cost. The country’s ambassador for climate action Ravi Menon said last month that while the city-state aspires to be a leading centre for AI applications, it also wants to reach net zero by 2050, which is “a challenging task given our limited domestic recourse to renewable energy”.
A data centre under construction in Malaysia, Asia’s fastest growing hub for data centres to meet soaring demand for cloud and AI services. Image: CSSEC
According to the International Energy Agency, the energy consumption of AI data centres is projected to double by 2027.
However, Tatiana Collins of Ingenie said that Asia is well positioned to lead in green AI, because of the region’s focus on cost. “Cost can be an indicator of emissions efficiency if the models are effective architecturally – and that is where Asian tech firms could have an advantage.”
“Look at how Chinese firms have managed the cost of solar panels. It is similar with AI. China’s companies have had to make the technology cheaper to be competitive – and that’s why I believe they will win in greening AI.”
